Classification method for irregular time series

Abstract

The invention discloses a method for classifying irregular time sequences, which belongs to the technical field of artificial intelligence, and comprises the following steps: obtaining irregular time sequences to be classified and mask matrixes corresponding to the irregular time sequences; inputting a to-be-classified irregular time sequence and a mask matrix into the trained frequency-guided adaptive kernel model to obtain an adaptive kernel selection feature OX and an adaptive kernel selection feature OM; processing the adaptive kernel selection feature OX through a self-attention module to obtain a data mode feature; processing the self-adaptive kernel selection feature OM through a self-attention module and a gating structure in sequence to obtain a missing mode feature; performing channel dimension splicing on the data mode features and the missing mode features, and then performing processing through a linear layer and a Sigmoid function layer in sequence to obtain a weight a and a weight 1-a; inputting the weighted fusion feature HX * a + HM * (1-a) into a prototype classifier to obtain a classification result of the irregular time sequence X to be classified; according to the invention, the classification precision and robustness can be improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more particularly to a classification method for irregular time series. Background Art

[0002] Due to factors such as asynchronous sampling by ship sensors, sampling frequency variations, communication interruptions, equipment jitter, sea state interference, and time synchronization errors, collected data often presents irregular time series with a high proportion of missing values. Existing methods for determining sea state levels based on irregular time series output by ship sensors suffer from low classification accuracy and poor robustness.

[0003] Therefore, how to provide a classification method for irregular time series that can improve classification accuracy and robustness is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, an object of the present invention is to provide a classification method for irregular time series.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A classification method for irregular time series includes the following steps:

[0007] Obtain the irregular time series X to be classified and its corresponding mask matrix M; wherein the irregular time series is a time series with missing values;

[0008] The irregular time series X to be classified is input into the trained frequency-guided adaptive kernel model to obtain the adaptive kernel selection feature O X ; Wherein, the trained frequency-guided adaptive kernel model is used to reconstruct and adaptively select the irregular time series X to be classified;

[0009] The mask matrix M is input into the trained frequency-guided adaptive kernel model to obtain the adaptive kernel selection feature O M ; Wherein, the trained frequency-guided adaptive kernel model is used to reconstruct the mask matrix M and perform adaptive kernel selection;

[0010] The adaptive kernel selection feature O X After processing by the self-attention module, the data pattern feature H is obtained X ;

[0011] The adaptive kernel selection feature O M After being processed by the self-attention module and the gate structure, the missing pattern feature H is obtained. M ;

[0012] The data pattern feature H X and the missing pattern characteristic H M After channel dimension splicing, the data pattern feature H is obtained by sequentially processing through the linear layer and the Sigmoid function layer. X The weight a and the missing pattern feature H M The weight of 1-a;

[0013] The weighted fusion feature H X *a+H M *(1-a) is input to the prototype classifier to obtain the classification result of the irregular time series X to be classified.

[0014] Preferably, the mask matrix M is obtained based on the following steps:

[0015] If X c,l If missing, then M c,l =1;

[0016] If X c,l If not missing, then M c,l =0;

[0017] Among them, X c,l M represents the element in the cth row and lth column of the irregular time series X to be classified; c,l Represents the element in the cth row and lth column in the mask matrix M; c=1,2,...,C; l=1,2,...,L; C represents the number of rows of the irregular time series X to be classified; L represents the number of columns of the irregular time series X to be classified.

[0018] Preferably, the irregular time series X to be classified is input into the trained frequency-guided adaptive kernel model to obtain the adaptive kernel selection feature O X ; wherein the trained frequency-guided adaptive kernel model is used to reconstruct and adaptively select the irregular time series X to be classified, specifically comprising the following steps:

[0019] Get the low-frequency component S of the irregular time series X to be classified d and high frequency component S h ;

[0020] Perform discrete Fourier transform on the irregular time series X to be classified to obtain the frequency component X F [u];X F [u] represents the u-th frequency component; u=1, 2, ..., U; U represents the total number of frequency components obtained by performing discrete Fourier transform on the irregular time series X to be classified;

[0021] Based on the frequency component X FThe amplitude of [u] is used to obtain the low-frequency static weight φ d and high-frequency static weight φ h ;

[0022] Based on the low-frequency component S d Calculate the low-frequency adaptive weight λ d Based on the high frequency component S h Calculate high frequency adaptive weight λ h ;

[0023] Based on the low-frequency component S d , the high frequency component S h , the low-frequency static weight φ d , the high frequency static weight φ h , the low-frequency adaptive weight λ d and the high frequency adaptive weight λ h Reconstruct the irregular time series X to be classified to obtain a reconstructed feature X0;

[0024] The reconstructed feature X0 is processed by B convolutional layers in sequence to obtain the feature X b ; where b=1,2,...,B; X b Represents the features of the output of the b-th convolutional layer; B convolutional layers have different convolution kernels and expansion rates;

[0025] The feature X b Input to the 1*1 convolution layer for interaction between channels to obtain features ; where b=1,2,...,B; Represents the weight matrix of the 1*1 convolutional layer;

[0026] The features Perform channel dimension splicing to obtain features ; Where R represents the real number space;

[0027] Feature-based Calculate the average descriptor ω avg and the maximum descriptor ω max ;

[0028] Based on the average descriptor ω avg , the maximum descriptor calculation ω max , Sigmod function calculates the normalized selection factor ;

[0029] Based on the normalization selection factor , the reconstructed feature X0 and the feature Calculate the adaptive kernel selection feature O X .

[0030] Preferably, the low-frequency component S is obtained based on the following formula: d and the high frequency component S h :

[0031] ;

[0032] ;

[0033] in, represents the lth column of data in the irregular time series X to be classified; L represents the number of columns in the irregular time series X to be classified.

[0034] Preferably, the frequency component X is obtained based on the following formula: F [u];

[0035] ;

[0036] Where, e represents the base of natural logarithm; j represents the imaginary unit; π is the ratio of circumference to circumference;

[0037] The low-frequency static weight φ is obtained based on the following formula d and the high frequency static weight φ h ;

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] in, Indicates that when satisfied hour, The value is 1, otherwise it is 0; Indicates taking The real part of Indicates taking The imaginary part of represents the average amplitude;

[0043] The low-frequency adaptive weight λ is obtained based on the following formula d and the high frequency adaptive weight λ h :

[0044] ;

[0045] ;

[0046] Among them, Wd 、b d 、W h 、b h represents a learnable parameter.

[0047] Preferably, the reconstructed feature X0 is obtained based on the following formula:

[0048] .

[0049] Preferably, the feature X is obtained based on the following formula b :

[0050] ;

[0051] in, Represents feature X b The lth column element in ; Represents feature X b-1 The first + d b i column elements; k b represents the kernel size of the b-th convolutional layer; d b represents the expansion rate of the b-th convolutional layer; Represents the weight matrix of the b-th convolutional layer with kernel size i;

[0052] The average descriptor ω is obtained based on the following formula avg and the maximum descriptor ω max ;

[0053] ;

[0054] ;

[0055] in, Representation characteristics The j-th column element in ;

[0056] The normalized selection factor is obtained based on the following formula: :

[0057] ;

[0058] ;

[0059] in, Represents the selection factor The lth column element in [ω avg ;ω max ] represents the average descriptor ω avg and the maximum descriptor ω max Features obtained by channel dimension splicing; [ω avg ;ωmax ] l+i Indicates [ω avg ;ω max ] in the l+ith column; Represents the normalized selection factor The lth column element in .

[0060] Preferably, the adaptive kernel selection feature is obtained based on the following formula: X :

[0061] .

[0062] Preferably, the method of reconstructing the mask matrix M and adaptively selecting the kernel is consistent with the method of reconstructing the irregular time series X to be classified and adaptively selecting the kernel.

[0063] Preferably, the network parameters of the frequency-guided adaptive kernel model are updated by minimizing the loss function L; wherein the network parameters include the weight matrix , learnable parameters W d , learnable parameter b d , learnable parameters W h , learnable parameter b h and the weight matrix ;

[0064] The expression of the loss function L is:

[0065] ;

[0066] Among them, f n represents the weighted fusion feature of the nth training sample input to the prototype classifier; n=1,2,...,N; N represents the number of training samples; C k represents the cluster center of the k-th class training sample, that is, the center point of the weighted fusion features of all k-th class training samples input to the prototype classifier; C j represents the cluster center of the j-th training sample, that is, the center point of the weighted fusion features of all j-th training samples input to the prototype classifier; k=1,2,...,K; j=1,2,...,K; K represents the number of categories of training samples; represents the parameter controlling similarity smoothing; when f n Belongs to C k When y nk =1, otherwise y nk =0; represents the regularization coefficient.

[0067] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a classification method for irregular time series, which can improve classification accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0069] Figure 1 The present invention provides a flow chart of a classification method for irregular time series. DETAILED DESCRIPTION

[0070] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0071] like Figure 1 As shown, the embodiment of the present invention discloses a classification method for irregular time series, comprising the following steps:

[0072] Obtain the irregular time series X to be classified and its corresponding mask matrix M; wherein the irregular time series is a time series with missing values;

[0073] In one embodiment, the irregular time series X to be classified is an irregular time series output by a ship sensor;

[0074] Ship sensors include sensors for measuring roll rate, roll angle, sway speed, pitch rate, pitch angle, sway speed, heading rate, heading angle, and heave speed;

[0075] Due to factors such as asynchronous sampling of ship sensors, sampling frequency differences, communication interruptions, equipment jitter, sea condition interference, and time synchronization errors, the time series output by ship sensors are usually irregular time series (i.e., accompanied by a high proportion of missing values).

[0076] The purpose of the present invention is to determine the sea condition level (i.e., the classification result) based on the irregular time series output by the ship's sensors;

[0077] There are five levels of sea state: Sea State Level 1, Sea State Level 2, Sea State Level 3, Sea State Level 4 and Sea State Level 5:

[0078] Sea state level 1 means very calm sea surface with wave height between 0 and 0.5 meters, which has almost no impact on navigation;

[0079] Sea state level 2 means light waves with wave heights of 0.5 to 1.25 meters;

[0080] Sea state level 3 indicates moderate waves, with wave heights ranging from 1.25 to 2.5 meters, and the ship begins to sway noticeably;

[0081] Sea state level 4 indicates larger waves with a wave height ranging from 2.5 to 4 meters. The sea conditions have a greater impact on ship control.

[0082] Sea state level 5 indicates severe sea conditions with wave heights reaching 4 to 6 meters, and the safety of ship navigation is significantly reduced.

[0083] In one embodiment, the mask matrix M is obtained based on the following steps:

[0084] If X c,l If missing, then M c,l =1;

[0085] If X c,l If not missing, then M c,l =0;

[0086] Among them, X c,l M represents the element in the cth row and lth column of the irregular time series X to be classified; c,l Represents the element in the cth row and lth column in the mask matrix M; c=1,2,...,C; l=1,2,...,L; C represents the number of rows of the irregular time series X to be classified; L represents the number of columns of the irregular time series X to be classified.

[0087] The irregular time series X to be classified is input into the trained frequency-guided adaptive kernel model to obtain the adaptive kernel selection feature O X ; Wherein, the trained frequency-guided adaptive kernel model is used to reconstruct and adaptively select the irregular time series X to be classified;

[0088] In one embodiment, the irregular time series X to be classified is input into the trained frequency-guided adaptive kernel model to obtain the adaptive kernel selection feature O X ; wherein the trained frequency-guided adaptive kernel model is used to reconstruct and adaptively select the irregular time series X to be classified, specifically comprising the following steps:

[0089] Get the low-frequency component S of the irregular time series X to be classified d and high frequency component S h ;

[0090] In one embodiment, the low-frequency component S is obtained based on the following formula: d and the high frequency component S h :

[0091] ;

[0092] ;

[0093] in, represents the lth column of data in the irregular time series X to be classified; L represents the number of columns in the irregular time series X to be classified.

[0094] Perform discrete Fourier transform on the irregular time series X to be classified to obtain the frequency component X F [u];X F [u] represents the u-th frequency component; u=1, 2, ..., U; U represents the total number of frequency components obtained by performing discrete Fourier transform on the irregular time series X to be classified;

[0095] In one embodiment, the frequency component X is obtained based on the following formula: F [u];

[0096] ;

[0097] Where, e represents the base of natural logarithm; j represents the imaginary unit; π is the ratio of circumference to circumference;

[0098] Based on the frequency component X F The amplitude of [u] is used to obtain the low-frequency static weight φ d and high-frequency static weight φ h ;

[0099] In one embodiment, the low-frequency static weight φ is obtained based on the following formula: d and the high frequency static weight φ h ;

[0100] ;

[0101] ;

[0102] ;

[0103] ;

[0104] in, Indicates that when satisfied hour, The value is 1, otherwise it is 0; Indicates taking The real part of Indicates taking The imaginary part of represents the average amplitude;

[0105] Based on the low-frequency component S d Calculate the low-frequency adaptive weight λ d Based on the high frequency component S h Calculate high frequency adaptive weight λ h ;

[0106] In one embodiment, the low-frequency adaptive weight λ is obtained based on the following formula: d and the high frequency adaptive weight λ h :

[0107] ;

[0108] ;

[0109] Among them, W d 、b d 、W h 、b h represents a learnable parameter.

[0110] Based on the low-frequency component S d , the high frequency component S h , the low-frequency static weight φ d , the high frequency static weight φ h , the low-frequency adaptive weight λ d and the high frequency adaptive weight λ h Reconstruct the irregular time series X to be classified to obtain a reconstructed feature X0;

[0111] In one embodiment, the reconstructed feature X0 is obtained based on the following formula:

[0112] .

[0113] The reconstructed feature X0 is processed by B convolutional layers in sequence to obtain the feature X b ; where b=1,2,...,B; X b Represents the features of the output of the b-th convolutional layer; B convolutional layers have different convolution kernels and expansion rates;

[0114] In one embodiment, the feature X is obtained based on the following formula:b :

[0115] ;

[0116] in, Represents feature X b The lth column element in ; Represents feature X b-1 The first + d b i column elements; k b represents the kernel size of the b-th convolutional layer; d b represents the expansion rate of the b-th convolutional layer; Represents the weight matrix of the b-th convolutional layer with kernel size i;

[0117] The feature X b Input to the 1*1 convolution layer for interaction between channels to obtain features ; where b=1,2,...,B; Represents the weight matrix of the 1*1 convolutional layer;

[0118] The features Perform channel dimension splicing to obtain features ; Where R represents the real number space;

[0119] Feature-based Calculate the average descriptor ω avg and the maximum descriptor ω max ;

[0120] In one embodiment, the average descriptor ω is obtained based on the following formula: avg and the maximum descriptor ω max ;

[0121] ;

[0122] ;

[0123] in, Representation characteristics The j-th column element in ;

[0124] Based on the average descriptor ω avg , the maximum descriptor calculation ω max , Sigmod function calculates the normalized selection factor ;

[0125] In one embodiment, the normalized selection factor is obtained based on the following formula: :

[0126] ;

[0127] ;

[0128] in, Represents the selection factor The lth column element in [ω avg ;ω max ] represents the average descriptor ω avg and the maximum descriptor ω max Features obtained by channel dimension splicing; [ω avg ;ω max ] l+i Indicates [ω avg ;ω max ] in the l+ith column; Represents the normalized selection factor The lth column element in .

[0129] Based on the normalization selection factor , the reconstructed feature X0 and the feature Calculate the adaptive kernel selection feature O X .

[0130] In one embodiment, the adaptive kernel selection feature is obtained based on the following formula: X :

[0131] .

[0132] The mask matrix M is input into the trained frequency-guided adaptive kernel model to obtain the adaptive kernel selection feature O M ; Wherein, the trained frequency-guided adaptive kernel model is used to reconstruct the mask matrix M and perform adaptive kernel selection;

[0133] In one embodiment, the method of reconstructing the mask matrix M and adaptively selecting the kernel is consistent with the method of reconstructing the irregular time series X to be classified and adaptively selecting the kernel.

[0134] That is, the adaptive kernel selection feature O is obtained M Method and method for obtaining adaptive kernel selection feature O X The method is consistent.

[0135] The adaptive kernel selection feature O X After processing by the self-attention module, the data pattern feature H is obtained X ;

[0136] The adaptive kernel selection feature O M After being processed by the self-attention module and the gate structure, the missing pattern feature H is obtained.M ;

[0137] The data pattern feature H X and the missing pattern characteristic H M After channel dimension splicing, the data pattern feature H is obtained by sequentially processing through the linear layer and the Sigmoid function layer. X The weight a and the missing pattern feature H M The weight of 1-a;

[0138] The weighted fusion feature H X *a+H M *(1-a) is input to the prototype classifier to obtain the classification result of the irregular time series X to be classified.

[0139] It can be understood that if the irregular time series X to be classified is an irregular time series output by a ship sensor, the classification result is the sea state level.

[0140] In one embodiment, the network parameters of the frequency-guided adaptive kernel model are updated by minimizing the loss function L; wherein the network parameters include the weight matrix , learnable parameters W d , learnable parameter b d , learnable parameters W h , learnable parameter b h and the weight matrix ;

[0141] The expression of the loss function L is:

[0142] ;

[0143] Among them, f n represents the weighted fusion feature of the nth training sample input to the prototype classifier; n=1,2,...,N; N represents the number of training samples; C k represents the cluster center of the k-th class training sample, that is, the center point of the weighted fusion features of all k-th class training samples input to the prototype classifier; C j represents the cluster center of the j-th training sample, that is, the center point of the weighted fusion features of all j-th training samples input to the prototype classifier; k=1,2,...,K; j=1,2,...,K; K represents the number of categories of training samples; represents the parameter controlling similarity smoothing; when f n Belongs to C k When y nk =1, otherwise y nk =0; represents the regularization coefficient.

[0144] It can be understood that a training sample includes an irregular time series and its corresponding mask matrix and category label.

[0145] The method of obtaining the weighted fusion feature of the training sample input to the prototype classifier and the weighted fusion feature H X *a+H M *(1-a) is obtained in the same way.

[0146] It can be understood that: if the irregular time series X to be classified is an irregular time series output by a ship sensor, then a training sample includes an irregular time series output by a ship sensor and its corresponding mask matrix and category label (i.e., sea state level).

[0147] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0148] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A classification method for irregular time series, characterized in that: The following steps are involved: Obtain the irregular time series X to be classified and its corresponding mask matrix M; wherein the irregular time series is a time series with missing values; The irregular time series X to be classified is input into the trained frequency-guided adaptive kernel model to obtain the adaptive kernel selection feature O X ; Wherein, the trained frequency-guided adaptive kernel model is used to reconstruct and adaptively select the irregular time series X to be classified; The mask matrix M is input into the trained frequency-guided adaptive kernel model to obtain the adaptive kernel selection feature O M ; Wherein, the trained frequency-guided adaptive kernel model is used to reconstruct the mask matrix M and perform adaptive kernel selection; The adaptive kernel selection feature O X After processing by the self-attention module, the data pattern feature H is obtained X ; The adaptive kernel selection feature O M After being processed by the self-attention module and the gate structure, the missing pattern feature H is obtained. M ; The data pattern feature H X and the missing pattern characteristic H M After channel dimension splicing, the data pattern feature H is obtained by sequentially processing through the linear layer and the Sigmoid function layer. X The weight a and the missing pattern feature H M The weight of 1-a; The weighted fusion feature H X *a+H M *(1-a) is input to the prototype classifier to obtain the classification result of the irregular time series X to be classified.

2. The classification method of irregular time series according to claim 1, characterized in that: The mask matrix M is obtained based on the following steps: If X c,l If missing, then M c,l =1; If X c,l If not missing, then M c,l =0; Among them, X c,l M represents the element in the cth row and lth column of the irregular time series X to be classified; c,l Represents the element in the cth row and lth column in the mask matrix M; c=1,2,...,C; l=1,2,...,L; C represents the number of rows of the irregular time series X to be classified; L represents the number of columns of the irregular time series X to be classified.

3. The method for classifying irregular time series according to claim 1, characterized in that: The irregular time series X to be classified is input into the trained frequency-guided adaptive kernel model to obtain the adaptive kernel selection feature O X ; wherein the trained frequency-guided adaptive kernel model is used to reconstruct and adaptively select the irregular time series X to be classified, specifically comprising the following steps: Get the low-frequency component S of the irregular time series X to be classified d and high frequency component S h ; Perform discrete Fourier transform on the irregular time series X to be classified to obtain the frequency component X F [u];X F [u] represents the u-th frequency component; u=1, 2, ..., U; U represents the total number of frequency components obtained by performing discrete Fourier transform on the irregular time series X to be classified; Based on the frequency component X F The amplitude of [u] is used to obtain the low-frequency static weight φ d and high-frequency static weight φ h ; Based on the low-frequency component S d Calculate the low-frequency adaptive weight λ d Based on the high frequency component S h Calculate high frequency adaptive weight λ h ; Based on the low-frequency component S d , the high frequency component S h , the low-frequency static weight φ d , the high frequency static weight φ h , the low-frequency adaptive weight λ d and the high frequency adaptive weight λ h Reconstruct the irregular time series X to be classified to obtain a reconstructed feature X0; The reconstructed feature X0 is processed by B convolutional layers in sequence to obtain the feature X b ; where b=1,2,...,B; X b Represents the features of the output of the b-th convolutional layer; B convolutional layers have different convolution kernels and expansion rates; The feature X b Input to the 1*1 convolution layer for interaction between channels to obtain features ; where b=1,2,...,B; Represents the weight matrix of the 1*1 convolutional layer; The features Perform channel dimension splicing to obtain features ; Where R represents the real number space; Feature-based Calculate the average descriptor ω avg and the maximum descriptor ω max ; Based on the average descriptor ω avg , the maximum descriptor calculation ω max , Sigmod function calculates the normalized selection factor ; Based on the normalization selection factor , the reconstructed feature X0 and the feature Calculate the adaptive kernel selection feature O X .

4. The method for classifying irregular time series according to claim 3, characterized in that: The low-frequency component S is obtained based on the following formula d and the high frequency component S h : ； ； in, represents the lth column of data in the irregular time series X to be classified; L represents the number of columns in the irregular time series X to be classified.

5. The method for classifying irregular time series according to claim 4, characterized in that: The frequency component X is obtained based on the following formula F [u]; ； Where, e represents the base of natural logarithm; j represents the imaginary unit; π is the ratio of circumference to circumference; The low-frequency static weight φ is obtained based on the following formula d and the high frequency static weight φ h ; ； ； ； ； in, Indicates that when satisfied hour, The value is 1, otherwise it is 0; Indicates taking The real part of Indicates taking The imaginary part of represents the average amplitude; The low-frequency adaptive weight λ is obtained based on the following formula d and the high frequency adaptive weight λ h : ； ； Among them, W d 、b d 、W h 、b h represents a learnable parameter.

6. The method for classifying irregular time series according to claim 5, characterized in that: The reconstructed feature X0 is obtained based on the following formula: 。 7. The method for classifying irregular time series according to claim 6, characterized in that: The feature X is obtained based on the following formula b : ； in, Represents feature X b The lth column element in ; Represents feature X b-1 The first + d b i column elements; k b represents the kernel size of the b-th convolutional layer; d b represents the expansion rate of the b-th convolutional layer; Represents the weight matrix of the b-th convolutional layer with kernel size i; The average descriptor ω is obtained based on the following formula avg and the maximum descriptor ω max ; ； ； in, Representation characteristics The j-th column element in ; The normalized selection factor is obtained based on the following formula: : ； ； in, Represents the selection factor The lth column element in [ω avg ;ω max ] represents the average descriptor ω avg and the maximum descriptor ω max Features obtained by channel dimension splicing; [ω avg ;ω max ] l+i Indicates [ω avg ;ω max ] in the l+ith column; Represents the normalized selection factor The lth column element in .

8. The method for classifying irregular time series according to claim 7, characterized in that: The adaptive kernel selection feature is obtained based on the following formula: X : 。 9. The method for classifying irregular time series according to claim 8, characterized in that: The method of reconstructing the mask matrix M and adaptively selecting the kernel is consistent with the method of reconstructing the irregular time series X to be classified and adaptively selecting the kernel.

10. The method for classifying irregular time series according to claim 9, characterized in that: Minimize the loss function L to update the network parameters of the frequency-guided adaptive kernel model; wherein the network parameters include the weight matrix , learnable parameters W d , learnable parameter b d , learnable parameters W h , learnable parameter b h and the weight matrix ; The expression of the loss function L is: ； Among them, f n represents the weighted fusion feature of the nth training sample input to the prototype classifier; n=1,2,...,N; N represents the number of training samples; C k represents the cluster center of the k-th class training sample, that is, the center point of the weighted fusion features of all k-th class training samples input to the prototype classifier; C j represents the cluster center of the j-th training sample, that is, the center point of the weighted fusion features of all j-th training samples input to the prototype classifier; k=1,2,...,K; j=1,2,...,K; K represents the number of categories of training samples; represents the parameter controlling similarity smoothing; when f n Belongs to C k When y nk =1, otherwise y nk =0; represents the regularization coefficient.

Images