A method for monitoring the health status of a rotary cutting machine

Abstract

The present invention discloses a method for monitoring the health status of a rotary cutter, which belongs to the technical field of equipment detection. The vibration signal of the rotary cutter motor is collected and processed in three different division methods to generate three division signals, each of which has a different window length and overlap. Then, a short-time Fourier transform is performed on each division signal to obtain a corresponding complex spectrum matrix, and its phase change value is calculated. Subsequently, abnormal data pairs are identified in the complex spectrum matrix and outliers are calculated to form three outlier sequences. At the same time, the number of corresponding abnormal data pairs is counted to construct three weight sequences. Finally, the three outlier sequences are processed by a three-channel recursive estimation model, and the influence of the weight sequence is combined to finally obtain the health status value of the rotary cutter motor. This method improves the accuracy and reliability of monitoring through the application of multidimensional signal processing and intelligent models, and provides an effective basis for equipment maintenance.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment detection, and in particular to a method for monitoring the health status of a rotary cutter. Background Art

[0002] In the field of health monitoring and fault diagnosis of industrial equipment, the real-time monitoring of the motor health status of the rotary cutter as an important processing tool is particularly critical. However, the existing vibration signal analysis methods usually only use a single window division method for data processing. This single processing method has certain limitations. First, different window lengths and overlaps will affect the spectrum analysis results of the short-time Fourier transform. Shorter windows can obtain higher time resolution, but lower frequency resolution; while longer windows can provide better frequency resolution, but the time resolution is reduced. A single window setting is difficult to take into account the analysis requirements of both time and frequency dimensions. Secondly, relying solely on a single window division method cannot fully explore the multi-scale features in the vibration signal. Different vibration features may be hidden in different time-frequency distributions, and a single processing method is difficult to fully reflect the actual health status of the equipment. In addition, the current short-time Fourier transform (STFT) processing method often lacks comprehensive consideration of phase changes in spectrum analysis. Many traditional methods only focus on the changes in frequency amplitude and ignore the important features contained in phase information, which reduces the accuracy of anomaly detection and makes it difficult to effectively identify subtle fault signs. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, a method for monitoring the health status of a rotary cutter provided by the present invention solves the problem of low accuracy of abnormality detection in the prior art.

[0004] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a method for monitoring the health status of a rotary cutter, comprising the following steps:

[0005] S1, collecting the vibration signal of the rotary cutting machine motor, dividing the vibration signal using three different division methods, and obtaining three divided signals, wherein each division method has a different window length and overlap amount;

[0006] S2, performing short-time Fourier transform on each divided signal to obtain three spectrum complex matrices;

[0007] S3. Calculate the phase change value in each spectrum complex matrix;

[0008] S4. In each spectrum complex matrix, find the abnormal data pairs according to the phase change value and calculate the outlier value;

[0009] S5, using the outliers of the same spectrum complex matrix at each moment as elements to construct an outlier sequence, and obtaining three outlier sequences;

[0010] S6, using the number of abnormal data pairs corresponding to the same spectrum complex matrix at each moment as an element to construct a weight sequence, and obtaining three weight sequences;

[0011] S7. A three-channel recursive estimation model is used to process three outlier sequences, and the health status value of the rotary cutting machine motor is obtained based on the attention applied by the three weight sequences.

[0012] Furthermore, the first division method in S1 is specifically: dividing the vibration signal into multiple overlapping time windows, each window length is 100ms, and the window overlap is 50ms;

[0013] The second division method is as follows: the vibration signal is divided into multiple overlapping time windows, each window is 20ms long, and the window overlap is 15ms;

[0014] The third division method is specifically as follows: the vibration signal is divided into multiple overlapping time windows, each window is 50 ms long, and the window overlap is 37 ms.

[0015] Furthermore, the complex spectrum matrix in S2 is: , where X i is the complex matrix of the ith spectrum, x i,(1,1) is the complex value of the first frequency point and the first time window in the i-th division method, x i,(1,2) is the complex value of the first frequency point and the second time window in the i-th division method, x i,(1,N) is the complex value of the first frequency point and the Nth time window in the i-th division method, x i,(2,1) is the complex value of the second frequency point and the first time window in the i-th division method, x i,(2,2) is the complex value of the second frequency point and the second time window in the i-th division method, x i,(2,N) is the complex value of the second frequency point and the Nth time window in the i-th division method, x i,(M,1) is the complex value of the Mth frequency point and the first time window in the i-th division method, x i,(M,2) is the complex value of the Mth frequency point and the second time window in the i-th division method, x i,(M,N) is the complex value of the Mth frequency point and the Nth time window in the i-th division method, the value range of i is a positive integer from 1 to 3, M is the number of frequency points, and N is the number of time windows.

[0016] Furthermore, S3 includes the following sub-steps:

[0017] S31. In each spectrum complex matrix, in the same time window, calculate the phase difference between two adjacent frequency points. , where Δθ(m,n) is the phase difference of the mth frequency point in the nth time window, θ(m+1,n) is the phase value between the m+1th frequency point and the nth time window, θ(m,n) is the phase value between the mth frequency point and the nth time window, m is the number of the frequency point, and n is the number of the time window;

[0018] S32, expand the phase difference to obtain the phase change value: , where Δθ open (m,n) is the phase change value at the mth frequency point in the nth time window, and round is the rounding function.

[0019] Furthermore, the S4 comprises the following sub-steps:

[0020] S41, when the phase change value is greater than the phase change threshold, there is a phase mutation;

[0021] S42, in each spectrum complex matrix, calculating the amplitude difference between two frequency points where there is a phase mutation;

[0022] S43, when the amplitude difference is greater than the amplitude difference threshold, taking the corresponding phase change value and the amplitude difference as a pair of abnormal data;

[0023] S44. In each spectral complex matrix, calculate the outlier value according to each pair of abnormal data.

[0024] Furthermore, the formula for calculating the outlier in S44 is: , where γ is the outlier, α1 is the first weighting coefficient, α2 is the second weighting coefficient, e is a natural constant, Δθ ab,j is the phase change value in the jth abnormal data pair, ΔA ab,j is the amplitude difference in the jth abnormal data pair, | | is the absolute value operation, j is a positive integer, and L is the number of abnormal data pairs.

[0025] Further, the three-channel recursive estimation model in S7 includes: a first feature fusion channel, a second feature fusion channel, a third feature fusion channel, a first recursive unit, a second recursive unit, a third recursive unit and a weighting unit;

[0026] The first feature fusion channel is used to fuse the corresponding features of the first outlier sequence and the first weight sequence to obtain a first fused feature sequence; the first recursive unit is used to process the first fused feature sequence to obtain a first health status component;

[0027] The second feature fusion channel is used to fuse the corresponding features of the second outlier sequence and the second weight sequence to obtain a second fused feature sequence; the second recursive unit is used to process the second fused feature sequence to obtain a second health status component;

[0028] The third feature fusion channel is used to fuse the corresponding features of the third outlier sequence and the third weight sequence to obtain a third fused feature sequence; the third recursive unit is used to process the third fused feature sequence to obtain a third health status component;

[0029] The weighting unit is used to weight the first health status component, the second health status component and the third health status component to obtain the health status value of the rotary cutter motor.

[0030] Furthermore, the feature fusion channel is used to extract features from the outlier sequence to form an outlier feature sequence, extract features from the weight sequence to form a weight feature sequence, and multiply the outlier feature sequence and the weight feature sequence element by element to obtain a fused feature sequence.

[0031] Furthermore, the expressions of the recursive units are all: , where h i,t is the health status component output by the i-th recursive unit at the t-th time, S i,t,k is the kth fusion feature in the i-th fusion feature sequence at the t-th time, ω i,k For S i,t,k The learnable weight term, b i,k For S i,t,k The learnable bias term, K is the number of fusion features in the i-th fusion feature sequence, k is the number of fusion features, and h i,t-τ is the health status component output by the i-th recursive unit at the t-τth moment, ω i,t-τ h i,t-τ The learnable weight term, b i,t-τ h i,t-τ is a learnable bias term, t is the current moment, τ is the historical moment, T is the length of the historical moment, and the value range of i is a positive integer from 1 to 3.

[0032] The beneficial effects of the present invention are:

[0033] 1. The present invention divides the vibration signal by using three different division methods, and can obtain signal characteristics under different window lengths and overlap amounts. This diverse processing method enables the monitoring scheme to capture the spectral characteristics of the vibration signal more comprehensively, and improves the accuracy of abnormality detection.

[0034] 2. Perform short-time Fourier transform on each divided signal to obtain three spectral complex matrices. Each spectral complex matrix reflects the spectral data under each division method. Through the spectral data of the three division methods, abnormal information can be captured more comprehensively.

[0035] 3. The present invention finds abnormal data pairs through phase change values, calculates outliers, and reflects the abnormal conditions of each spectrum complex matrix.

[0036] 4. The outliers of the same spectral complex matrix at each moment of the present invention constitute an outlier sequence, which reflects the working condition of the rotary cutter motor over a period of time, and constructs a weight sequence according to the number of abnormal data pairs to quantify the degree of abnormality. The three-channel recursive estimation model is used to process the outlier sequence, so that the health status value of the rotary cutter motor not only depends on the current signal characteristics, but also is based on historical data, thereby improving the prediction accuracy of the health status value of the rotary cutter motor. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of a method for monitoring the health status of a rotary cutting machine;

[0038] Figure 2 Schematic diagram of the structure of the three-channel recursive estimation model. DETAILED DESCRIPTION

[0039] 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.

[0040] like Figure 1 As shown, a method for monitoring the health status of a rotary cutting machine comprises the following steps:

[0041] S1, collecting the vibration signal of the rotary cutting machine motor, dividing the vibration signal using three different division methods, and obtaining three divided signals, wherein each division method has a different window length and overlap amount;

[0042] S2, performing short-time Fourier transform on each divided signal to obtain three spectrum complex matrices;

[0043] S3. Calculate the phase change value in each spectrum complex matrix;

[0044] S4. In each spectrum complex matrix, find the abnormal data pairs according to the phase change value and calculate the outlier value;

[0045] S5, using the outliers of the same spectrum complex matrix at each moment as elements to construct an outlier sequence, and obtaining three outlier sequences;

[0046] S6, using the number of abnormal data pairs corresponding to the same spectrum complex matrix at each moment as an element to construct a weight sequence, and obtaining three weight sequences;

[0047] S7. A three-channel recursive estimation model is used to process three outlier sequences, and the health status value of the rotary cutting machine motor is obtained based on the attention applied by the three weight sequences.

[0048] In this embodiment, the first division method in S1 is specifically: dividing the vibration signal into multiple overlapping time windows, each window length is 100ms, the window overlap is 50ms, where ms is milliseconds;

[0049] The second division method is as follows: the vibration signal is divided into multiple overlapping time windows, each window is 20ms long, and the window overlap is 15ms;

[0050] The third division method is specifically as follows: the vibration signal is divided into multiple overlapping time windows, each window is 50 ms long, and the window overlap is 37 ms.

[0051] In the present invention, a time window is a vibration signal, which is equivalent to dividing the vibration signal into multiple signal segments with overlapping signal values. Taking the first division method as an example, the length of each signal segment is 100ms, and the overlap between signal segments is 50ms, that is, the signal values ​​between adjacent signal segments are the same for a time period of 50ms.

[0052] The first division method uses a 100ms window length and a 50ms overlap, which can provide a higher frequency resolution and is conducive to capturing low-frequency vibration characteristics. The second division method uses a 20ms window length and a 15ms overlap, which can achieve a higher time resolution and is conducive to identifying transient vibration events. The third division method uses a 50ms window length and a 37ms overlap, which strikes a balance between time and frequency resolution and can take into account the overall characteristics of the vibration signal.

[0053] Apply a window function to each time window. The window function can be a rectangular window, a Hamming window, a Hanning window, etc. Perform a short-time Fourier transform on the signal in each window to obtain a frequency domain representation. The complex values ​​obtained by STFT can form a matrix, where the rows represent the frequencies and the columns represent the time windows. Assuming that the signal is divided into N time windows and M frequency points, the complex matrix of the spectrum in S2 is: , where X i is the complex matrix of the ith spectrum, x i,(1,1)is the complex value of the first frequency point and the first time window in the i-th division method, x i,(1,2) is the complex value of the first frequency point and the second time window in the i-th division method, x i,(1,N) is the complex value of the first frequency point and the Nth time window in the i-th division method, x i,(2,1) is the complex value of the second frequency point and the first time window in the i-th division method, x i,(2,2) is the complex value of the second frequency point and the second time window in the i-th division method, x i,(2,N) is the complex value of the second frequency point and the Nth time window in the i-th division method, x i,(M,1) is the complex value of the Mth frequency point and the first time window in the i-th division method, x i,(M,2) is the complex value of the Mth frequency point and the second time window in the i-th division method, x i,(M,N) is the complex value of the Mth frequency point and the Nth time window in the i-th division method, the value range of i is a positive integer from 1 to 3, M is the number of frequency points, and N is the number of time windows.

[0054] The expression of short-time Fourier transform is: , where X t,f is a complex time-frequency value, is the vibration signal, is the window function, e is a natural constant, f is the frequency, is the integral variable, d is the differential symbol, and j is the imaginary unit.

[0055] Each complex value in the spectral complex matrix contains a real part a and an imaginary part b, with an amplitude equal to , the phase value is equal to , tan is the tangent function.

[0056] In this embodiment, S3 includes the following sub-steps:

[0057] S31. In each spectrum complex matrix, in the same time window, calculate the phase difference between two adjacent frequency points. , where Δθ(m,n) is the phase difference of the mth frequency point in the nth time window, θ(m+1,n) is the phase value between the m+1th frequency point and the nth time window, θ(m,n) is the phase value between the mth frequency point and the nth time window, m is the number of the frequency point, and n is the number of the time window;

[0058] S32, expand the phase difference to obtain the phase change value: , where Δθ open (m,n) is the phase change value at the mth frequency point in the nth time window, and round is the rounding function.

[0059] After obtaining the phase difference, the present invention obtains the phase difference. A jump in the range may result in a complete cycle of phase, so the phase difference needs to be "unfolded".

[0060] In this embodiment, S4 includes the following sub-steps:

[0061] S41, when the phase change value is greater than the phase change threshold, there is a phase mutation;

[0062] S42, in each spectrum complex matrix, calculating the amplitude difference between two frequency points where there is a phase mutation;

[0063] S43, when the amplitude difference is greater than the amplitude difference threshold, taking the corresponding phase change value and the amplitude difference as a pair of abnormal data;

[0064] S44. In each spectral complex matrix, calculate the outlier value according to each pair of abnormal data.

[0065] For example, in Δθ open (m,n) is greater than the phase change threshold, and there is a phase mutation between position (m+1,n) and position (m,n), then find the amplitude difference between position (m+1,n) and position (m,n) in the corresponding spectrum complex matrix.

[0066] In this embodiment, the phase change threshold and the amplitude difference threshold are specifically set based on experiments or experience.

[0067] In this embodiment, the phase change threshold is set to 0.5 rad or 30 degrees. When the phase difference exceeds 0.5 rad or 30 degrees, it can be judged as a phase mutation. If the amplitude is abnormal (such as an increase or decrease in amplitude) while the phase changes significantly, it is more likely to indicate a fault.

[0068] When the present invention captures a sudden phase change, the amplitude of the corresponding spectrum information is observed, and the abnormality is determined through the sudden changes in amplitude and phase.

[0069] In this embodiment, the formula for calculating the outlier in S44 is: , where γ is the outlier, α1 is the first weighting coefficient, α2 is the second weighting coefficient, e is a natural constant, Δθ ab,j is the phase change value in the jth abnormal data pair, ΔA ab,j is the amplitude difference in the jth abnormal data pair, | | is the absolute value operation, j is a positive integer, and L is the number of abnormal data pairs.

[0070] The present invention sets an exponential function to enhance the phase change value and the amplitude difference, calculates outliers according to all abnormal data pairs in a frequency spectrum complex matrix, and reflects the abnormal situation of the frequency spectrum complex matrix.

[0071] In this embodiment, the first weighting coefficient α1 and the second weighting coefficient α2 are used to adjust the ratio, α1+α2=1, 0<α1<1, 0<α2<1, for example, α1=0.5, α2=0.5, and the amplitude and phase abnormalities are considered equally.

[0072] like Figure 2 As shown, the three-channel recursive estimation model in S7 includes: a first feature fusion channel, a second feature fusion channel, a third feature fusion channel, a first recursive unit, a second recursive unit, a third recursive unit and a weighting unit;

[0073] The first feature fusion channel is used to fuse the corresponding features of the first outlier sequence and the first weight sequence to obtain a first fused feature sequence; the first recursive unit is used to process the first fused feature sequence to obtain a first health status component;

[0074] The second feature fusion channel is used to fuse the corresponding features of the second outlier sequence and the second weight sequence to obtain a second fused feature sequence; the second recursive unit is used to process the second fused feature sequence to obtain a second health status component;

[0075] The third feature fusion channel is used to fuse the corresponding features of the third outlier sequence and the third weight sequence to obtain a third fused feature sequence; the third recursive unit is used to process the third fused feature sequence to obtain a third health status component;

[0076] The weighting unit is used to weight the first health status component, the second health status component and the third health status component to obtain the health status value of the rotary cutter motor.

[0077] In this embodiment, the feature fusion channel is used to extract features from the outlier sequence to form an outlier feature sequence, extract features from the weight sequence to form a weight feature sequence, and multiply the outlier feature sequence and the weight feature sequence element by element to obtain a fused feature sequence.

[0078] The larger the element in the weight sequence of the present invention, the greater the abnormality of the motor. Therefore, the corresponding features of the outlier sequence and the weight sequence are fused through each feature fusion channel, so that the weight sequence can be used to pay attention to the outlier sequence and improve the prediction accuracy. The present invention then considers the data of historical moments through a recursive unit to simultaneously consider the current state and the historical state, thereby further improving the prediction accuracy.

[0079] In this embodiment, the feature fusion channel includes: an outlier feature extraction module, a weight feature extraction module and a multiplier; the outlier feature extraction module and the weight feature extraction module both use a fully connected layer, and the expression of the outlier feature extraction module is: , where G r is the rth feature output by the outlier feature extraction module, g r is the rth element in the outlier sequence, ω g,r g r The learnable weight term, b g,r g r The learnable bias term of , r is a positive integer.

[0080] The expression of the weight feature extraction module is: , where Z r is the rth feature output by the weight feature extraction module, z r is the rth element in the weight sequence, ω z,r For z r The learnable weight term, b z,r For z r The features output by the outlier feature extraction module are used to form an outlier feature sequence, and the features output by the weight feature extraction module are used to form a weight feature sequence. In the multiplier, the outlier feature sequence and the weight feature sequence are multiplied bit by bit.

[0081] like Figure 2 As shown, M1, M2 and M3 are all multipliers.

[0082] In this embodiment, the expressions of the recursive units are all: , where h i,t is the health status component output by the i-th recursive unit at the t-th time, S i,t,k is the kth fusion feature in the i-th fusion feature sequence at the t-th time, ω i,k For S i,t,k The learnable weight term, b i,k For S i,t,k The learnable bias term, K is the number of fusion features in the i-th fusion feature sequence, k is the number of fusion features, and h i,t-τ is the health status component output by the i-th recursive unit at the t-τth moment, ω i,t-τ h i,t-τ The learnable weight term, b i,t-τ h i,t-τ is a learnable bias term, t is the current moment, τ is the historical moment, T is the length of the historical moment, and the value range of i is a positive integer from 1 to 3.

[0083] The present invention introduces historical state information h i,t-τ,The model can use past data for prediction and analysis, taking into account the continuity of fault occurrence and avoiding accidental faults.

[0084] In this embodiment, the learnable weight term and the learnable bias term in the three-channel recursive estimation model are calibrated using an existing gradient descent method or a genetic algorithm.

[0085] In this embodiment, the greater the health status value of the peeling machine motor is, the worse the health status of the peeling machine motor is.

[0086] In the present invention, after collecting the signal, the signal is divided into multiple segments, and three division methods are used in each segment, so each segment of the signal will generate three types of spectral complex matrices, so there are multiple types of spectral complex matrices. The present invention calculates the outliers of each spectral complex matrix generated at each moment, and each outlier belonging to the same spectral complex matrix constitutes an outlier sequence, so three outlier sequences are obtained, and the number of abnormal data pairs corresponding to the same spectral complex matrix at each moment is counted, and the number at each moment is used as an element to form a weight sequence. Since there are three types of spectral complex matrices, three weight sequences are obtained.

[0087] The present invention divides the vibration signal by using three different division methods, and can obtain signal characteristics under different window lengths and overlap amounts. This diverse processing method enables the monitoring scheme to capture the spectral characteristics of the vibration signal more comprehensively, thereby improving the accuracy of abnormality detection.

[0088] By performing short-time Fourier transform on each divided signal, three spectral complex matrices are obtained. Each spectral complex matrix reflects the spectral data under each division method. Through the spectral data of the three division methods, abnormal information can be captured more comprehensively.

[0089] The present invention finds abnormal data pairs through phase change values, calculates outliers, and reflects the abnormal situation of each frequency spectrum complex matrix.

[0090] The outlier values ​​of the same spectral complex matrix at each moment of the present invention constitute an outlier sequence, which reflects the working condition of the rotary cutter motor within a period of time, and constructs a weight sequence according to the number of abnormal data pairs to quantify the degree of abnormality. The three-channel recursive estimation model is used to process the outlier sequence, so that the health status value of the rotary cutter motor not only depends on the current signal characteristics, but also is based on historical data, thereby improving the prediction accuracy of the health status value of the rotary cutter motor.

[0091] 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 monitoring the health status of a rotary cutting machine, characterized in that: The following steps are involved: S1, collecting the vibration signal of the rotary cutting machine motor, dividing the vibration signal using three different division methods, and obtaining three divided signals, wherein each division method has a different window length and overlap amount; The first division method is specifically as follows: the vibration signal is divided into multiple overlapping time windows, each window is 100ms long, and the window overlap is 50ms; The second division method is as follows: the vibration signal is divided into multiple overlapping time windows, each window is 20ms long, and the window overlap is 15ms; The third division method is as follows: the vibration signal is divided into multiple overlapping time windows, each window is 50ms long, and the window overlap is 37ms; S2, performing short-time Fourier transform on each divided signal to obtain three spectrum complex matrices; S3. In each spectrum complex matrix, calculate the phase difference between two adjacent frequency points in the same time window. , where Δθ(m,n) is the phase difference of the mth frequency point in the nth time window, θ(m+1,n) is the phase value between the m+1th frequency point and the nth time window, θ(m,n) is the phase value between the mth frequency point and the nth time window, m is the number of the frequency point, and n is the number of the time window; Expand the phase difference to get the phase change value: , where Δθ open (m,n) is the phase change value of the mth frequency point in the nth time window, and round is the rounding function; S4. In each spectrum complex matrix, find the abnormal data pairs according to the phase change value and calculate the outlier value; S5, using the outliers of the same spectrum complex matrix at each moment as elements to construct an outlier sequence, and obtaining three outlier sequences; S6, using the number of abnormal data pairs corresponding to the same spectrum complex matrix at each moment as an element to construct a weight sequence, and obtaining three weight sequences; S7, using a three-channel recursive estimation model to process three outlier sequences, and obtaining the health status value of the rotary cutting machine motor based on the attention applied by the three weight sequences; The three-channel recursive estimation model includes: a first feature fusion channel, a second feature fusion channel, a third feature fusion channel, a first recursive unit, a second recursive unit, a third recursive unit and a weighting unit; The first feature fusion channel is used to fuse the corresponding features of the first outlier sequence and the first weight sequence to obtain a first fused feature sequence; the first recursive unit is used to process the first fused feature sequence to obtain a first health status component; The second feature fusion channel is used to fuse the corresponding features of the second outlier sequence and the second weight sequence to obtain a second fused feature sequence; the second recursive unit is used to process the second fused feature sequence to obtain a second health status component; The third feature fusion channel is used to fuse the corresponding features of the third outlier sequence and the third weight sequence to obtain a third fused feature sequence; the third recursive unit is used to process the third fused feature sequence to obtain a third health status component; The weighting unit is used to weight the first health status component, the second health status component and the third health status component to obtain the health status value of the rotary cutter motor.

2. The method for monitoring the health status of a peeling machine according to claim 1, characterized in that: The complex spectrum matrix in S2 is: , where X i is the complex matrix of the ith spectrum, x i,(1,1) is the complex value of the first frequency point and the first time window in the i-th division method, x i,(1,2) is the complex value of the first frequency point and the second time window in the i-th division method, x i,(1,N) is the complex value of the first frequency point and the Nth time window in the i-th division method, x i,(2,1) is the complex value of the second frequency point and the first time window in the i-th division method, x i,(2,2) is the complex value of the second frequency point and the second time window in the i-th division method, x i,(2,N) is the complex value of the second frequency point and the Nth time window in the i-th division method, x i,(M,1) is the complex value of the Mth frequency point and the first time window in the i-th division method, x i,(M,2) is the complex value of the Mth frequency point and the second time window in the i-th division method, x i,(M,N) is the complex value of the Mth frequency point and the Nth time window in the i-th division method, the value range of i is a positive integer from 1 to 3, M is the number of frequency points, and N is the number of time windows.

3. The method for monitoring the health status of a peeling machine according to claim 1, characterized in that: The S4 comprises the following sub-steps: S41, when the phase change value is greater than the phase change threshold, there is a phase mutation; S42, in each spectrum complex matrix, calculating the amplitude difference between two frequency points where there is a phase mutation; S43, when the amplitude difference is greater than the amplitude difference threshold, taking the corresponding phase change value and the amplitude difference as a pair of abnormal data; S44. In each spectral complex matrix, calculate the outlier value according to each pair of abnormal data.

4. The method for monitoring the health status of a peeling machine according to claim 3, characterized in that: The formula for calculating the outlier value in S44 is: , where γ is the outlier, α1 is the first weighting coefficient, α2 is the second weighting coefficient, e is a natural constant, Δθ ab,j is the phase change value in the jth abnormal data pair, ΔA ab,j is the amplitude difference in the jth abnormal data pair, | | is the absolute value operation, j is a positive integer, and L is the number of abnormal data pairs.

5. The method for monitoring the health status of a peeling machine according to claim 1, characterized in that: The feature fusion channel is used to extract features from the outlier sequence to form an outlier feature sequence, extract features from the weight sequence to form a weight feature sequence, and multiply the outlier feature sequence and the weight feature sequence element by element to obtain a fused feature sequence.

6. The method for monitoring the health status of a peeling machine according to claim 1, characterized in that: The expressions of the recursive units are: , where h i,t is the health status component output by the i-th recursive unit at the t-th time, S i,t,k is the kth fusion feature in the i-th fusion feature sequence at the t-th time, ω i,k For S i,t,k The learnable weight term, b i,k For S i,t,k The learnable bias term, K is the number of fusion features in the i-th fusion feature sequence, k is the number of fusion features, and h i,t-τ is the health status component output by the i-th recursive unit at the t-τth moment, ω i,t-τ h i,t-τ The learnable weight term, b i,t-τ h i,t-τ is a learnable bias term, t is the current moment, τ is the historical moment, T is the length of the historical moment, and the value range of i is a positive integer from 1 to 3.

Images