An early diagnosis method for rack rail cracks using acoustic signal analysis

Abstract

The present invention discloses an early diagnosis method for rack cracks using acoustic signal analysis, and belongs to the technical field of sound processing. The present invention applies a multi-frequency fused sound test signal to the rack and collects sound conduction signals at multiple equally spaced detection points, thereby being able to fully capture the acoustic characteristics of the internal structure of the rack, and effectively overcome the limitation that the traditional method is difficult to identify internal cracks. Secondly, by eliminating the test frequency and extracting the frequency set generated by sound conduction, the characteristic frequency caused by the rack can be effectively separated, thereby improving the accuracy and sensitivity of the detection. Then, by extracting the amplitude energy fluctuation sequence and the spectrum fluctuation sequence, the influence of cracks on the propagation of sound waves can be fully analyzed, and a dual convolution channel fusion network is used for crack assessment to accurately assess the degree of rack cracks.

Description

Technical Field

[0001] The invention relates to the technical field of sound processing, and in particular to an early diagnosis method for rack cracks using acoustic signal analysis. Background Art

[0002] Rack rail is a special rail system widely used in various industrial and transportation fields, such as mountain railways, cable car systems and some special industrial conveying equipment. Its uniqueness lies in the rack on the track, which meshes with the gears on the vehicle or equipment, so as to achieve stable operation on steep slopes or occasions requiring precise positioning. This design not only improves the climbing ability, but also ensures the accuracy and safety of operation. However, due to the long-term huge stress and environmental factors, the rack rail may suffer from cracks and other damage, which will not only affect the smoothness and efficiency of operation, but also may cause serious safety accidents.

[0003] Traditional rack inspection methods rely mainly on regular manual visual inspections, but this method has serious limitations. Rack cracks are remarkably hidden and complex, including not only visible cracks on the surface, but also potential internal microscopic cracks. These internal cracks are usually not visible to the naked eye in the early stages, but are the key factors leading to sudden rack failure. Manual visual inspection methods have difficulty identifying these hidden internal defects, especially in the initial stages of cracks. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method for early diagnosis of rack cracks using acoustic signal analysis to solve the problem that the prior art is unable to identify cracks inside the rack.

[0005] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a method for early diagnosis of rack cracks using acoustic signal analysis, comprising the following steps:

[0006] S1, applying a multi-frequency fused sound test signal to one end of the rack section to be detected, setting multiple detection points at equal intervals on the rack section to be detected, and collecting the sound conduction signal;

[0007] S2. Eliminate the test frequencies from the multiple sound conduction signals respectively to obtain multiple sound conduction generation frequency sets;

[0008] S3, extracting the amplitude energy fluctuation sequence and the spectrum fluctuation sequence according to each frequency in the frequency set generated by multiple sound conduction;

[0009] S4. According to the amplitude energy fluctuation sequence and the spectrum fluctuation sequence, the rack crack assessment value is obtained based on the double convolution channel fusion network.

[0010] Furthermore, the expression of the multi-frequency fused sound test signal in S1 is: , where S(t) is the multi-frequency fusion sound test signal, A is the signal amplitude, sin is the sine function, and f 1 is the first test frequency, f 2 is the second test frequency, f 3 is the third test frequency, f 4 is the fourth test frequency, and t is the time.

[0011] Further, S2 includes the following sub-steps:

[0012] S21, transforming each sound conduction signal from the time domain to the frequency domain to obtain a sound conduction frequency set;

[0013] S22. For each sound conduction frequency set, the test frequency contained in the sound test signal is eliminated to obtain a sound conduction generation frequency set.

[0014] Furthermore, S3 includes the following sub-steps:

[0015] S31, extracting common frequencies from a plurality of sound conduction generation frequency sets, wherein the common frequencies are frequencies present in all sound conduction generation frequency sets;

[0016] S32, calculating the amplitude energy value of each detection point according to the common frequency of the detection point;

[0017] S33, subtracting energy values ​​of adjacent detection points to construct an amplitude energy fluctuation sequence;

[0018] S34, calculating the spectrum balance value of each sound conduction generated frequency set;

[0019] S35, subtracting spectrum equalization values ​​of adjacent detection points to construct a spectrum fluctuation sequence.

[0020] Furthermore, the formula for calculating the amplitude energy value of the detection point in S32 is: , where E i is the amplitude energy value of the i-th detection point, F i,j is the amplitude corresponding to the jth common frequency at the i-th detection point, j and i are positive integers, and R is the number of common frequencies;

[0021] The amplitude energy fluctuation sequence in S33 is: X={r 1 ,…,r i ,…,r N-1}, X is the amplitude energy fluctuation sequence, r 1 is the first energy difference, which is the difference between the energy value of the second detection point and the energy value of the first detection point, r iis the i-th energy difference value, the i-th energy difference value is the difference between the energy value of the i+1-th detection point and the energy value of the i-th detection point, r N-1 is the N-1th energy difference value, the N-1th energy difference value is the difference between the energy value of the N-th detection point and the energy value of the N-1th detection point, and N is the number of detection points.

[0022] Furthermore, the formula for calculating the spectrum balance value of each sound conduction frequency set in S34 is: , where se i is the spectrum equalization value of the i-th detection point, f i,n is the nth frequency in the frequency set generated by sound conduction at the i-th detection point, F i,n is the frequency f i,n The corresponding amplitude, M is the number of frequencies in the frequency set generated by sound conduction at the i-th detection point, i and n are positive integers;

[0023] The spectrum fluctuation sequence in S35 is: Y={c 1 ,…,c i ,…,c N-1}, where Y is the spectrum fluctuation sequence, c 1 is the first spectrum balance difference, which is the difference between the spectrum balance value of the second detection point and the spectrum balance value of the first detection point, c i is the ith spectrum equalization difference, the ith spectrum equalization difference is the difference between the spectrum equalization value of the i+1th detection point and the spectrum equalization value of the ith detection point, c N-1 is the N-1th spectrum equalization difference value, the N-1th spectrum equalization difference value is the difference between the spectrum equalization value of the Nth detection point and the spectrum equalization value of the N-1th detection point, and N is the number of detection points.

[0024] Furthermore, the dual convolution channel fusion network in S4 includes: a first feature extraction convolution channel, a second feature extraction convolution channel, a first maximum pooling layer, a second maximum pooling layer, a first Concat layer, a second Concat layer, a first Tanh layer, a second Tanh layer and a fully connected layer;

[0025] The input end of the first feature extraction convolution channel is used to input the amplitude energy fluctuation sequence, and its output end is respectively connected to the input end of the first maximum pooling layer and the first input end of the first Concat layer; the input end of the second feature extraction convolution channel is used to input the spectrum fluctuation sequence, and its output end is respectively connected to the input end of the second maximum pooling layer and the second input end of the first Concat layer; the first input end of the second Concat layer is connected to the output end of the first maximum pooling layer, its second input end is connected to the output end of the second maximum pooling layer, and its output end is connected to the input end of the second Tanh layer; the input end of the first Tanh layer is connected to the output end of the first Concat layer; the input end of the fully connected layer is respectively connected to the output end of the first Tanh layer and the output end of the second Tanh layer, and its output end serves as the output end of the dual convolution channel fusion network.

[0026] Furthermore, the first feature extraction convolution channel includes a first convolution layer, a first attention adjustment layer, and a second convolution layer connected in sequence, and the second feature extraction convolution channel includes a third convolution layer, a second attention adjustment layer, and a fourth convolution layer connected in sequence.

[0027] Furthermore, the output sequence of the first convolutional layer is , the expression of the first attention adjustment layer is: , where x 1,1 is the first eigenvalue in the output sequence of the first convolutional layer, x 1,k is the kth eigenvalue in the output sequence of the first convolutional layer, x 1,K is the Kth eigenvalue in the output sequence of the first convolutional layer, H 1 is the output sequence of the first attention adjustment layer, θ 1,1 is the first attention value in the first attention adjustment layer, θ 1,k is the kth attention value in the first attention adjustment layer, θ 1,K is the Kth attention value in the first attention adjustment layer. The expression of the attention value in the first attention adjustment layer is: , k is a positive integer, K is the sequence length, is element-wise multiplication;

[0028] The output sequence of the third convolutional layer is , the expression of the second attention adjustment layer is: , where x 3,1 is the first eigenvalue in the output sequence of the third convolutional layer, x 3,k is the kth eigenvalue in the output sequence of the third convolutional layer, x 3,K is the Kth eigenvalue in the output sequence of the third convolutional layer, H 2 is the output sequence of the second attention adjustment layer, θ 2,1is the first attention value in the second attention adjustment layer, θ 2,k is the kth attention value in the second attention adjustment layer, θ 2,K is the Kth attention value in the second attention adjustment layer, and the expression of the attention value in the second attention adjustment layer is: .

[0029] Furthermore, the expression of the first Tanh layer is: , the expression of the second Tanh layer is: , where S 1 is the output of the first Tanh layer, S 2 is the output of the second Tanh layer, Tanh is the hyperbolic tangent function, h 1,m is the mth eigenvalue output by the first Concat layer, ω 1,m h 1,m The weight of b 1,m h 1,m The bias, h 2,m is the mth eigenvalue output by the second Concat layer, ω 2,m h 2,m The weight of b 2,m h 2,m The bias, L 1 is the number of eigenvalues ​​output by the first Concat layer, L 2 is the number of eigenvalues ​​output by the second Concat layer, m is a positive integer;

[0030] The expression of the fully connected layer is: , where y is the rack crack assessment value, ω S1 For S 1 The weight of S2 For S 2 The weight of .

[0031] The beneficial effects of the present invention are as follows: by applying a multi-frequency fused sound test signal to the rack and collecting sound conduction signals at multiple equally spaced detection points, the present invention can fully capture the acoustic characteristics of the internal structure of the rack, effectively overcoming the limitation of traditional methods that are difficult to identify internal cracks. Secondly, by eliminating the test frequency and extracting the frequency set generated by sound conduction, the characteristic frequency caused by the rack can be effectively separated, thereby improving the accuracy and sensitivity of the detection. By extracting the amplitude energy fluctuation sequence and the spectrum fluctuation sequence, the influence of cracks on sound wave propagation can be fully analyzed, and a dual convolution channel fusion network is used for crack assessment to accurately assess the degree of rack cracks. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flow chart of a method for early diagnosis of rack cracks using acoustic signal analysis;

[0033] Figure 2 A schematic diagram of setting multiple detection points on the rack;

[0034] Figure 3 Schematic diagram of the structure of the dual convolution channel fusion network. DETAILED DESCRIPTION

[0035] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0036] like Figure 1 As shown, a method for early diagnosis of rack cracks using acoustic signal analysis includes the following steps:

[0037] S1. Apply a multi-frequency fused sound test signal to one end of the rack section to be tested, set multiple test points at equal intervals on the rack section to be tested, and collect sound conduction signals, such as Figure 2 As shown;

[0038] S2. Eliminate the test frequencies from the multiple sound conduction signals respectively to obtain multiple sound conduction generation frequency sets;

[0039] S3, extracting the amplitude energy fluctuation sequence and the spectrum fluctuation sequence according to each frequency in the frequency set generated by multiple sound conduction;

[0040] S4. According to the amplitude energy fluctuation sequence and the spectrum fluctuation sequence, the rack crack assessment value is obtained based on the double convolution channel fusion network.

[0041] In this embodiment, the expression of the multi-frequency fused sound test signal in S1 is: , where S(t) is the multi-frequency fusion sound test signal, A is the signal amplitude, sin is the sine function, and f 1 is the first test frequency, f 2 is the second test frequency, f 3 is the third test frequency, f 4 is the fourth test frequency, and t is the time.

[0042] In this embodiment, f 1 =1000Hz, f 2 =2000Hz, f 3 =5000Hz, f 4=10000Hz, covering a wide frequency range from 1kHz to 10kHz. This helps detect cracks of different sizes and depths, because cracks with different characteristics have different responses to sound waves of different frequencies.

[0043] In this embodiment, S2 includes the following sub-steps:

[0044] S21, transforming each sound conduction signal from the time domain to the frequency domain to obtain a sound conduction frequency set;

[0045] S22. For each sound conduction frequency set, the test frequency contained in the sound test signal is eliminated to obtain a sound conduction generation frequency set.

[0046] The present invention transforms the sound conduction signal from the time domain to the frequency domain through S21, obtains spectrum information, and constructs a sound conduction frequency set. Each sound conduction frequency set is composed of the frequency of the corresponding sound conduction signal. The test frequency contained in the sound test signal is eliminated to avoid interference from the original signal components, and the characteristic frequency caused by the rack structure is obtained to constitute a sound conduction generation frequency set.

[0047] In this embodiment, S3 includes the following sub-steps:

[0048] S31, extracting common frequencies from a plurality of sound conduction generation frequency sets, wherein the common frequencies are frequencies present in all sound conduction generation frequency sets;

[0049] S32, calculating the amplitude energy value of each detection point according to the common frequency of the detection point;

[0050] S33, subtracting energy values ​​of adjacent detection points to construct an amplitude energy fluctuation sequence;

[0051] S34, calculating the spectrum balance value of each sound conduction generated frequency set;

[0052] S35, subtracting spectrum equalization values ​​of adjacent detection points to construct a spectrum fluctuation sequence.

[0053] By focusing on the common frequency, the present invention can eliminate the amplitude interference of accidental frequencies caused by the measurement position or environmental factors, and facilitate the analysis of the influence of the rack on the amplitude. The spectrum balance value provides the overall characteristics of the spectrum distribution and can reflect the overall deviation of the frequency components.

[0054] In this embodiment, the formula for calculating the amplitude energy value of the detection point in S32 is: , where E i is the amplitude energy value of the i-th detection point, F i,j is the amplitude corresponding to the jth common frequency at the i-th detection point, j and i are positive integers, and R is the number of common frequencies;

[0055] The amplitude energy fluctuation sequence in S33 is: X={r 1 ,…,r i ,…,r N-1}, X is the amplitude energy fluctuation sequence, r 1 is the first energy difference, which is the difference between the energy value of the second detection point and the energy value of the first detection point, r i is the i-th energy difference value, the i-th energy difference value is the difference between the energy value of the i+1-th detection point and the energy value of the i-th detection point, r N-1 is the N-1th energy difference value, the N-1th energy difference value is the difference between the energy value of the N-th detection point and the energy value of the N-1th detection point, and N is the number of detection points.

[0056] The present invention extracts the common frequencies in the frequency sets generated by multiple sound conduction, identifies the frequencies that always exist between various detection points, calculates the amplitude energy values ​​based on the amplitudes at these common frequencies, and constructs an amplitude energy fluctuation sequence to reflect the influence of the rack on the amplitude energy. When the sound wave encounters a crack, part of the energy will be reflected, scattered, or converted into heat energy, resulting in a decrease in the amplitude of the sound wave propagating to subsequent detection points. Cracks may produce local resonances at certain specific frequencies, resulting in abnormal enhancement or attenuation of the amplitude of the sound waves at these frequencies. Therefore, constructing an amplitude energy fluctuation sequence can capture the amplitude mutation characteristics.

[0057] In this embodiment, the formula for calculating the spectrum balance value of each sound conduction frequency set in S34 is: , where se i is the spectrum equalization value of the i-th detection point, f i,n is the nth frequency in the frequency set generated by sound conduction at the i-th detection point, F i,n is the frequency f i,n The corresponding amplitude, M is the number of frequencies in the frequency set generated by sound conduction at the i-th detection point, i and n are positive integers;

[0058] The spectrum fluctuation sequence in S35 is: Y={c 1 ,…,c i ,…,c N-1}, where Y is the spectrum fluctuation sequence, c 1 is the first spectrum balance difference, which is the difference between the spectrum balance value of the second detection point and the spectrum balance value of the first detection point, c i is the ith spectrum equalization difference, the ith spectrum equalization difference is the difference between the spectrum equalization value of the i+1th detection point and the spectrum equalization value of the ith detection point, c N-1is the N-1th spectrum equalization difference, the N-1th spectrum equalization difference is the difference between the spectrum equalization value of the Nth detection point and the spectrum equalization value of the N-1th detection point, and N is the number of monitoring points.

[0059] The present invention can effectively extract the acoustic wave features of each detection point by calculating the spectrum balance value. This feature extraction helps to identify the energy distribution of the acoustic wave at different frequencies, and then reflects the structural state of the rack. By constructing a spectrum fluctuation sequence, the spectrum changes between adjacent detection points can be captured. This local change analysis helps to identify the potential crack location and severity, especially in the detection of small cracks.

[0060] like Figure 3 As shown, the dual convolution channel fusion network in S4 includes: a first feature extraction convolution channel, a second feature extraction convolution channel, a first maximum pooling layer, a second maximum pooling layer, a first Concat layer, a second Concat layer, a first Tanh layer, a second Tanh layer and a fully connected layer;

[0061] The input end of the first feature extraction convolution channel is used to input the amplitude energy fluctuation sequence, and its output end is respectively connected to the input end of the first maximum pooling layer and the first input end of the first Concat layer; the input end of the second feature extraction convolution channel is used to input the spectrum fluctuation sequence, and its output end is respectively connected to the input end of the second maximum pooling layer and the second input end of the first Concat layer; the first input end of the second Concat layer is connected to the output end of the first maximum pooling layer, its second input end is connected to the output end of the second maximum pooling layer, and its output end is connected to the input end of the second Tanh layer; the input end of the first Tanh layer is connected to the output end of the first Concat layer; the input end of the fully connected layer is respectively connected to the output end of the first Tanh layer and the output end of the second Tanh layer, and its output end serves as the output end of the dual convolution channel fusion network.

[0062] In this embodiment, the first feature extraction convolution channel includes a first convolution layer, a first attention adjustment layer, and a second convolution layer connected in sequence, and the second feature extraction convolution channel includes a third convolution layer, a second attention adjustment layer, and a fourth convolution layer connected in sequence.

[0063] In the present invention, both the convolutional layer and the pooling layer are one-dimensional.

[0064] The present invention can capture two different types of signal features at the same time by processing the amplitude energy fluctuation sequence and the spectrum fluctuation sequence respectively. This dual-channel structure fuses the outputs of the first and second feature extraction convolution channels through the first Concat layer, so that the subsequent first Tanh layer can fully integrate these two types of features. Considering the importance of the mutation characteristics of amplitude energy and spectrum to crack diagnosis, the present invention introduces a maximum pooling layer to effectively extract the most significant local change features, so that the second Tanh layer can focus on the maximum value of the two types of features, greatly improving the sensitivity of the dual convolution channel fusion network to mutation signals.

[0065] In this embodiment, the output sequence of the first convolutional layer is , the expression of the first attention adjustment layer is: , where x 1,1 is the first eigenvalue in the output sequence of the first convolutional layer, x 1,k is the kth eigenvalue in the output sequence of the first convolutional layer, x 1,K is the Kth eigenvalue in the output sequence of the first convolutional layer, H 1 is the output sequence of the first attention adjustment layer, θ 1,1 is the first attention value in the first attention adjustment layer, θ 1,k is the kth attention value in the first attention adjustment layer, θ 1,K is the Kth attention value in the first attention adjustment layer. The expression of the attention value in the first attention adjustment layer is: , k is a positive integer, K is the sequence length, is element-wise multiplication;

[0066] The output sequence of the third convolutional layer is , the expression of the second attention adjustment layer is: , where x 3,1 is the first eigenvalue in the output sequence of the third convolutional layer, x 3,k is the kth eigenvalue in the output sequence of the third convolutional layer, x 3,K is the Kth eigenvalue in the output sequence of the third convolutional layer, H 2 is the output sequence of the second attention adjustment layer, θ 2,1 is the first attention value in the second attention adjustment layer, θ 2,k is the kth attention value in the second attention adjustment layer, θ 2,K is the Kth attention value in the second attention adjustment layer, and the expression of the attention value in the second attention adjustment layer is: .

[0067] The present invention takes into account the importance of the mutation characteristics of amplitude energy and spectrum to crack diagnosis. Therefore, the attention value of each eigenvalue output by the convolutional layer is calculated. The attention value reflects the proportion of the eigenvalue in the sequence. By calculating the relative importance of each feature (i.e., the attention value), the features that are more important for diagnosis can be highlighted, and the less important features can be suppressed, thereby further highlighting the eigenvalues ​​corresponding to the mutation.

[0068] In this embodiment, the expression of the first Tanh layer is: , the expression of the second Tanh layer is: , where S 1 is the output of the first Tanh layer, S 2 is the output of the second Tanh layer, Tanh is the hyperbolic tangent function, h 1,m is the mth eigenvalue output by the first Concat layer, ω 1,m h 1,m The weight of b 1,m h 1,m The bias, h 2,m is the mth eigenvalue output by the second Concat layer, ω 2,m h 2,m The weight of b 2,m h 2,m The bias, L 1 is the number of eigenvalues ​​output by the first Concat layer, L 2 is the number of eigenvalues ​​output by the second Concat layer, m is a positive integer;

[0069] The expression of the fully connected layer is: , where y is the rack crack assessment value, ω S1 For S 1 The weight of S2 For S 2 The weight of .

[0070] In this embodiment, the weights and biases in the dual convolution channel fusion network can be trained by the gradient descent method.

[0071] In the present invention, the fully connected layer integrates the outputs of the first Tanh layer and the second Tanh layer to improve the accuracy of diagnosis.

[0072] In this embodiment, when the rack crack evaluation value is greater than a threshold value, it can be determined that cracks exist in the rack section.

[0073] The present invention applies a multi-frequency fused sound test signal to the rack and collects sound conduction signals at multiple equally spaced detection points, thereby being able to fully capture the acoustic characteristics of the internal structure of the rack, effectively overcoming the limitation of traditional methods that are difficult to identify internal cracks. Secondly, by eliminating the test frequency and extracting the frequency set generated by sound conduction, the characteristic frequency caused by the rack can be effectively separated, thereby improving the accuracy and sensitivity of the detection. By extracting the amplitude energy fluctuation sequence and the spectrum fluctuation sequence, the effect of cracks on sound wave propagation can be fully analyzed, and a dual convolution channel fusion network is used for crack assessment to accurately assess the extent of the rack crack.

[0074] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for early diagnosis of rack cracks using acoustic signal analysis, characterized in that: The following steps are involved: S1, applying a multi-frequency fused sound test signal to one end of the rack section to be detected, setting multiple detection points at equal intervals on the rack section to be detected, and collecting the sound conduction signal; S2. Eliminate the test frequencies from the multiple sound conduction signals respectively to obtain multiple sound conduction generation frequency sets; S3, extracting the amplitude energy fluctuation sequence and the spectrum fluctuation sequence according to each frequency in the frequency set generated by multiple sound conduction; S4. According to the amplitude energy fluctuation sequence and the spectrum fluctuation sequence, the rack crack assessment value is obtained based on the double convolution channel fusion network; The S3 comprises the following sub-steps: S31, extracting common frequencies from a plurality of sound conduction generation frequency sets, wherein the common frequencies are frequencies present in all sound conduction generation frequency sets; S32, calculating the amplitude energy value of each detection point according to the common frequency of the detection point; S33, subtracting energy values ​​of adjacent detection points to construct an amplitude energy fluctuation sequence; S34, calculating the spectrum balance value of each sound conduction generated frequency set; S35, subtracting spectrum equalization values ​​of adjacent detection points to construct a spectrum fluctuation sequence; The formula for calculating the spectrum balance value of each sound conduction frequency set in S34 is: , where se i is the spectrum equalization value of the i-th detection point, f i,n is the nth frequency in the frequency set generated by sound conduction at the i-th detection point, F i,n is the frequency f i,n The corresponding amplitude, M is the number of frequencies in the frequency set generated by sound conduction at the i-th detection point, i and n are positive integers; The spectrum fluctuation sequence in S35 is: Y={c1,…,c i ,…,c N-1 }, where Y is the spectrum fluctuation sequence, c1 is the first spectrum balance difference, the first spectrum balance difference is the difference between the spectrum balance value of the second detection point and the spectrum balance value of the first detection point, c i is the ith spectrum equalization difference, the ith spectrum equalization difference is the difference between the spectrum equalization value of the i+1th detection point and the spectrum equalization value of the ith detection point, c N-1 is the N-1th spectrum equalization difference, the N-1th spectrum equalization difference is the difference between the spectrum equalization value of the Nth detection point and the spectrum equalization value of the N-1th detection point, and N is the number of detection points; The dual convolution channel fusion network in S4 includes: a first feature extraction convolution channel, a second feature extraction convolution channel, a first maximum pooling layer, a second maximum pooling layer, a first Concat layer, a second Concat layer, a first Tanh layer, a second Tanh layer and a fully connected layer; The input end of the first feature extraction convolution channel is used to input the amplitude energy fluctuation sequence, and its output end is respectively connected to the input end of the first maximum pooling layer and the first input end of the first Concat layer; the input end of the second feature extraction convolution channel is used to input the spectrum fluctuation sequence, and its output end is respectively connected to the input end of the second maximum pooling layer and the second input end of the first Concat layer; the first input end of the second Concat layer is connected to the output end of the first maximum pooling layer, the second input end is connected to the output end of the second maximum pooling layer, and the output end is connected to the input end of the second Tanh layer; the input end of the first Tanh layer is connected to the output end of the first Concat layer; the input end of the fully connected layer is respectively connected to the output end of the first Tanh layer and the output end of the second Tanh layer, and its output end serves as the output end of the dual convolution channel fusion network.

2. The method for early diagnosis of rack cracks using acoustic signal analysis according to claim 1, characterized in that: The expression of the multi-frequency fused sound test signal in S1 is: , where S(t) is the multi-frequency fused sound test signal, A is the signal amplitude, sin is the sine function, f1 is the first test frequency, f2 is the second test frequency, f3 is the third test frequency, f4 is the fourth test frequency, and t is time.

3. The method for early diagnosis of rack cracks using acoustic signal analysis according to claim 1, characterized in that: The S2 comprises the following sub-steps: S21, transforming each sound conduction signal from the time domain to the frequency domain to obtain a sound conduction frequency set; S22. For each sound conduction frequency set, the test frequency contained in the sound test signal is eliminated to obtain a sound conduction generation frequency set.

4. The method for early diagnosis of rack cracks using acoustic signal analysis according to claim 1, characterized in that: The formula for calculating the amplitude energy value of the detection point in S32 is: , where E i is the amplitude energy value of the i-th detection point, F i,j is the amplitude corresponding to the jth common frequency at the i-th detection point, j and i are positive integers, and R is the number of common frequencies; The amplitude energy fluctuation sequence in S33 is: X={r1,…,r i ,…,r N-1 }, X is the amplitude energy fluctuation sequence, r1 is the first energy difference, the first energy difference is the difference between the energy value of the second detection point and the energy value of the first detection point, r i is the i-th energy difference value, the i-th energy difference value is the difference between the energy value of the i+1-th detection point and the energy value of the i-th detection point, r N-1 is the N-1th energy difference value, the N-1th energy difference value is the difference between the energy value of the N-th detection point and the energy value of the N-1th detection point, and N is the number of detection points.

5. The method for early diagnosis of rack cracks using acoustic signal analysis according to claim 1, characterized in that: The first feature extraction convolution channel includes a first convolution layer, a first attention adjustment layer and a second convolution layer connected in sequence, and the second feature extraction convolution channel includes a third convolution layer, a second attention adjustment layer and a fourth convolution layer connected in sequence.

6. The method for early diagnosis of rack cracks using acoustic signal analysis according to claim 5, characterized in that: The output sequence of the first convolutional layer is , the expression of the first attention adjustment layer is: , where x 1,1 is the first eigenvalue in the output sequence of the first convolutional layer, x 1,k is the kth eigenvalue in the output sequence of the first convolutional layer, x 1,K is the Kth eigenvalue in the output sequence of the first convolutional layer, H1 is the output sequence of the first attention adjustment layer, θ 1,1 is the first attention value in the first attention adjustment layer, θ 1,k is the kth attention value in the first attention adjustment layer, θ 1,K is the Kth attention value in the first attention adjustment layer. The expression of the attention value in the first attention adjustment layer is: , k is a positive integer, K is the sequence length, is element-wise multiplication; The output sequence of the third convolutional layer is , the expression of the second attention adjustment layer is: , where x 3,1 is the first eigenvalue in the output sequence of the third convolutional layer, x 3,k is the kth eigenvalue in the output sequence of the third convolutional layer, x 3,K is the Kth eigenvalue in the output sequence of the third convolutional layer, H2 is the output sequence of the second attention adjustment layer, θ 2,1 is the first attention value in the second attention adjustment layer, θ 2,k is the kth attention value in the second attention adjustment layer, θ 2,K is the Kth attention value in the second attention adjustment layer, and the expression of the attention value in the second attention adjustment layer is: .

7. The method for early diagnosis of rack cracks using acoustic signal analysis according to claim 1, characterized in that: The expression of the first Tanh layer is: , the expression of the second Tanh layer is: , where S1 is the output of the first Tanh layer, S2 is the output of the second Tanh layer, Tanh is the hyperbolic tangent function, and h 1,m is the mth eigenvalue output by the first Concat layer, ω 1,m h 1,m The weight of b 1,m h 1,m The bias, h 2,m is the mth eigenvalue output by the second Concat layer, ω 2,m h 2,m The weight of b 2,m h 2,m , L1 is the number of eigenvalues ​​output by the first Concat layer, L2 is the number of eigenvalues ​​output by the second Concat layer, and m is a positive integer; The expression of the fully connected layer is: , where y is the rack crack assessment value, ω S1 is the weight of S1, ω S2 is the weight of S2.

Images