Multi-channel passive intermodulation amplitude prediction method and device based on echo state network

Abstract

The application belongs to the technical field of passive intermodulation amplitude prediction, and discloses a multi-channel passive intermodulation amplitude prediction method and device based on an echo state network. In the process of transmitting a data sequence, a training data sequence is inserted into the data sequence according to a preset proportion, and the preset proportion is the length ratio of the training data sequence to the data sequence. An echo state network is initialized, and the echo state network includes an input layer, a reservoir layer, an enhancement layer and an output layer. The echo state network is trained, the training data sequence input by each signal transmission channel at time t and the passive intermodulation amplitude at time t-1 are split into real part and imaginary part data, which are input into the input nodes of the echo state network, the passive intermodulation amplitude at time t is split into real part and imaginary part data as the output data of the network, and the network is trained to obtain the output weight; and the passive intermodulation amplitude is predicted. The method has the advantages of low computational complexity, small resource cost and stable performance.

Description

Technical Field

[0001] This invention belongs to the field of passive intermodulation interference amplitude prediction technology, and particularly relates to a multi-channel passive intermodulation amplitude prediction method and device based on echo state network. Background Technology

[0002] Passive intermodulation (PIM) refers to a family of products generated by the nonlinear effects of two or more frequency components through passive devices, whose frequencies are linear combinations of the fundamental frequency. When these linear combination products fall within the passband of a communication receiver, they constitute interference. PIM is widespread in communication systems and is difficult to completely eliminate; it was first discovered in the 1970s. With increasing communication demands, high power and high sensitivity have become the development direction. The antenna sizes required for communication are getting larger, forcing the adoption of shared transmit and receive technologies due to platform limitations. In shared transmit and receive communication systems with low duplexer isolation, passive intermodulation products are more likely to enter the receiving frequency band and cause interference, and cannot be filtered out by frequency domain filtering. Taking a mobile communication repeater as an example, passive intermodulation products of high-power downlink transmitted signals can fall into the uplink receiving frequency band, affecting the communication system.

[0003] Existing adaptive passive intermodulation prediction methods based on digital signal processing suffer from high complexity and slow convergence speed, and cannot well adapt to the time-varying characteristics of passive intermodulation signals. Echo State Networks (ESNs), as a type of neural network model, possess high prediction performance and low computational cost, achieving good results in time series prediction.

[0004] Therefore, it is urgent to optimize the echo state network and apply it to the field of multi-channel passive intermodulation amplitude prediction in order to solve the problems of high computational complexity, slow convergence speed and system instability of existing multi-channel passive intermodulation amplitude prediction methods. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for predicting the amplitude of passive intermodulation interference in a multi-channel system based on an echo state network, which can improve convergence performance, reduce computational complexity, and enable real-time prediction of the amplitude of passive intermodulation interference caused by multi-channel signals.

[0006] To achieve the above objectives, the present invention adopts the following technical solution.

[0007] In a first aspect, the present invention provides a method for predicting the amplitude of multi-channel passive intermodulation based on an echo state network, comprising the following steps: During the transmission of the transmitted data sequence, multiple signal transmission channels insert training data sequences into the transmitted data sequence according to a preset ratio, where the preset ratio is the ratio of the length of the training data sequence to the length of the transmitted data sequence. Initialize the echo state network, which includes an input layer, a reservoir layer, an enhancement layer, and an output layer; initialize the number of input nodes in the input layer, the number of output nodes in the output layer, the number of reservoir nodes in the reservoir layer, and the number of enhancement layer nodes in the enhancement layer; initialize the reservoir update rate; initialize the shrinkage coefficient; initialize the spectral radius; initialize the length ratio of the idle data sequence to the training data sequence; initialize the regularization coefficient and the activation function; The echo state network is trained by splitting the training data sequence input to each signal transmission channel at time t and the passive intermodulation amplitude at time t-1 into real and imaginary data, and inputting them into the input node of the echo state network. The passive intermodulation amplitude at time t is split into real and imaginary data as the output data of the network to train the network and obtain the output weights. Passive intermodulation amplitude prediction involves inputting the real and imaginary parts of all transmitted data sequences into the trained echo state network to obtain the real and imaginary parts of the passive intermodulation amplitude prediction value, and based on this, obtaining the passive intermodulation amplitude prediction value.

[0008] As one possible implementation, the preset ratio is 1:N, 3≤N≤5, that is, a training data sequence is inserted between the N adjacent sets of transmitted data sequences.

[0009] As one possible implementation, after inserting a training data sequence into the transmitted data sequence according to a preset ratio during the transmission of the transmitted data sequence, the multi-channel passive intermodulation amplitude preset method further includes: The data sequence, including the transmitted data sequence and the training data sequence, is normalized.

[0010] As one possible implementation, the number of signal transmission channels is M, in which case the number of input nodes is 2M+2 and the number of output nodes is two; the number of reservoir nodes is determined using principal component analysis; the enhancement layer enhances the characteristics of the reservoir through multiple contracting neurons, and the reservoir characteristic is a good nonlinear fitting ability between the time series of input and output; the enhancement layer and the reservoir are connected by random weights and biases.

[0011] As one possible implementation, training the echo state network includes the following steps: Determine the training data sequence input to the multiple signal transmission channels at time t, and the passive intermodulation amplitude output at time t-1; All training data sequences and the real and imaginary parts of the passive intermodulation amplitude are separated as inputs, and the real and imaginary parts of the passive intermodulation amplitude are used as outputs; the internal state of the reservoir and the internal state of the enhancement layer before the first iteration are determined; the connection weights from the input layer to the reservoir, between reservoirs, and from the reservoir to the enhancement layer are set. Randomly generate matrices of connection weights from the input layer to the reservoir, matrices of connection weights between reservoirs, and matrices of connection weights from the reservoir to the enhancement layer; The matrix of connection weights between reservoirs is scaled based on the spectral radius; Configure a state update model for the reservoir and enhancement layer to record the reservoir state during the time period from the end of the idle data sequence to the end of the training data sequence. In the first iteration, principal component analysis is performed on the reservoir state to determine the number of reservoir nodes; The output weights are determined based on the reservoir state, regularization coefficient, number of input nodes, number of reservoir nodes, number of enhancement layer nodes, and output matrix.

[0012] As one possible implementation, in the first iteration, principal component analysis is performed on the reservoir state to determine the number of reservoir nodes included in the reservoir, including: The eigenvalues ​​of the covariance matrix of the reservoir neuron portion in the reservoir state are determined, and the variance contribution rate of the eigenvalues ​​is determined according to the order of the eigenvalues. Based on this, the number of principal components whose cumulative variance contribution rate exceeds a preset threshold is determined. The number of principal components that exceed the preset threshold is the number of reservoir nodes.

[0013] As one possible implementation, the preset threshold is greater than or equal to 90%.

[0014] As one possible implementation, the real and imaginary parts of all transmitted data sequences are input into the trained echo state network to obtain the real and imaginary parts of the passive intermodulation amplitude prediction value. Based on this, the passive intermodulation amplitude prediction value is obtained, including: The reservoir initialization state of the trained echo state network is the state obtained in the last iteration, and the output weights are the output weights determined in the training echo state network steps. The real and imaginary parts of all the data sequences to be sent are input into the trained echo state network to obtain the output data sequence; By performing inverse normalization on the output data sequence, the predicted value of the passive intermodulation amplitude can be obtained.

[0015] Compared with the prior art, the present invention has the following advantages: 1. Due to the low computational cost and applicability of echo state networks in time series prediction, their application in multi-channel passive intermodulation amplitude prediction has advantages such as low computational complexity, low resource cost, and ease of hardware implementation compared to existing adaptive passive intermodulation prediction methods based on digital signal processing.

[0016] 2. When initializing the echo state network, specifically during the first iteration of the reservoir, the number of reservoir nodes is determined using principal component analysis. Based on this, the size of the reservoir neurons can be optimized, reducing computational complexity.

[0017] 3. Improve the echo state network by adding an enhancement layer. This not only has the advantages of stable performance, strong system adaptability, and good real-time performance, but also has the advantage of low system complexity compared with the existing adaptive passive intermodulation prediction method based on digital signal processing.

[0018] 4. The multi-channel passive intermodulation amplitude prediction method can also accurately predict the amplitude of passive intermodulation interference.

[0019] In a second aspect, the present invention also provides a multi-channel passive intermodulation amplitude prediction device based on an echo state network, the multi-channel passive intermodulation amplitude prediction device being used to execute the multi-channel passive intermodulation amplitude prediction method provided in the first aspect.

[0020] Compared with the prior art, the beneficial effects of the multi-channel passive intermodulation amplitude prediction device based on echo state network provided by the present invention are the same as the beneficial effects of the multi-channel passive intermodulation amplitude prediction method based on echo state network provided by the first aspect and / or any implementation of the first aspect, and will not be elaborated here. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This invention provides a method for predicting the amplitude of multi-channel passive intermodulation based on echo state networks. Figure 2 This is a structural diagram of the echo state network; Figure 3 The training flowchart for the echo state network; Figure 4 The simulation results of the multi-channel passive intermodulation amplitude prediction method based on echo state network are shown in the figure. Figure 5 The figure shows the simulation results of the time domain comparison before and after PIM interference cancellation, based on the prediction results of this invention. Detailed Implementation

[0022] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0023] It should be noted that when a component is referred to as being "fixed to" or "set on" another component, it can be directly on or indirectly on that other component. When a component is referred to as being "connected to" another component, it can be directly connected to or indirectly connected to that other component.

[0024] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. "Several" means one or more, unless otherwise explicitly specified.

[0025] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0026] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0027] In practical applications, passive intermodulation (PIM) occurs when multiple signal transmission channels transmit data sequences. This invention aims to predict the amplitude of this PIM using an improved echo state network. Specifically, the multi-channel PIM amplitude prediction method based on the echo state network employs a multi-tasks echo state network (MT-ESN). This network uses principal component analysis (PCA) to determine the size of the echo state network reservoir and adds several enhancement layers with shrinking neurons after the reservoir to enhance the reservoir's feature extraction capabilities. Finally, a dual-timeslot mechanism, utilizing the training data sequence time slot and the transmission data sequence time slot, is used to predict the amplitude of the multi-channel PIM interference signal in real time. During the training sequence time slot, the output weight parameters of the prediction network are trained and stored. During the data transmission interval, the amplitude of the PIM signal is predicted using the known transmitted data and the weight parameters of the prediction network.

[0028] See Figure 1 In a first aspect, embodiments of the present invention provide a method for predicting the amplitude of multi-channel passive intermodulation based on an echo state network, comprising the following steps: S10. During the transmission of the transmitted data sequence, multiple signal transmission channels insert training data sequences into the transmitted data sequence according to a preset ratio, where the preset ratio is the length ratio of the training data sequence to the transmitted data sequence.

[0029] As an example, the preset ratio is 1:N, where 3 ≤ N ≤ 5, meaning that a training data sequence is inserted between N adjacent sets of transmitted data sequences. For instance, when N = 3, the preset ratio is 1:3, in which case a training data sequence is inserted between 3 adjacent sets of transmitted data sequences.

[0030] As one possible implementation, after inserting training data sequences into the transmitted data sequence according to a preset ratio during the transmission of the transmitted data sequence, the multi-channel passive intermodulation amplitude preset method also includes: Normalize the data sequence, which includes the transmitted data sequence and the training data sequence.

[0031] As an example, the normalization process is performed in the following way: in, For standardized intervals, and These are the minimum and maximum values ​​in the data. For data before normalization, This is the normalized data.

[0032] As an example, the training data sequence uses a pseudo-random sequence. Commonly used pseudo-random sequences include m-sequences, M-sequences, and Gold sequences. This invention can use an m-sequence as the training sequence. The full name of an m-sequence is the longest linear feedback shift register, which is a binary pseudo-random sequence with the longest period generated by a shift register with linear feedback. It is the most typical sequence among all pseudo-random sequences, possessing properties such as balance and autocorrelation. Different m-sequences are used for the training sequences of different transmission channels.

[0033] S11. Initialize the echo state network, which includes an input layer, a reservoir layer, an enhancement layer, and an output layer (see [link]). Figure 2 The initialization steps are as follows: Initialize the number of input nodes in the input layer, the number of output nodes in the output layer, the number of reservoir nodes in the reservoir layer, and the number of enhancement layer nodes in the enhancement layer; initialize the reservoir update rate; initialize the shrinkage coefficient; initialize the spectral radius; initialize the length ratio of the idle data sequence to the training data sequence; initialize the regularization coefficient and the activation function. The idle data sequence refers to the data input into the network without recording the reservoir state.

[0034] As an example, the number of signal transmission channels is M. In this case, the number of input nodes is 2M+2, and the number of output nodes is two. Principal component analysis is used to determine the number of reservoir nodes. The enhancement layer enhances the reservoir characteristics through multiple contractile neurons, and the reservoir characteristics are characterized by good nonlinear fitting ability between the input and output time series. The enhancement layer and the reservoir are connected by random weights and biases.

[0035] For example, when there are two signal transmission channels, there are six input nodes and two output nodes. Principal component analysis (PCA) is used to determine the number of reservoir nodes. If the input signal of the training task changes, PCA is used again to redetermine the number of reservoir nodes. The enhancement layer enhances the reservoir characteristics through multiple contracting neurons, and the reservoir characteristics are good nonlinear fitting ability between the input and output time series. The enhancement layer and the reservoir are connected by random weights and biases.

[0036] Further explanation is needed regarding the principle of the enhancement layer. It consists of several neurons, which are connected to the previous storage layer via fully connected connections. When the features extracted from the storage layer are input to the enhancement layer, the enhancement layer further amplifies these features. After the enhancement layer outputs the amplified features, the input channel state information, the extracted features, and the amplified extracted features are input together to the output layer. The output weights are then estimated using a least-squares method.

[0037] In other words, the enhancement layer is a layer structure containing a small number of neurons added after the original complete network, which has the effect of network contraction.

[0038] S12. Train the echo state network by splitting the training data sequence input to each signal transmission channel at time t and the passive intermodulation amplitude at time t-1 into real and imaginary data, and inputting them into the input node of the echo state network. The passive intermodulation amplitude at time t is split into real and imaginary data as the output data of the network to train the network and obtain the output weights.

[0039] As one possible implementation, training the echo state network may specifically include the following steps: S120. Determine the training data sequence input to multiple signal transmission channels at time t, and the passive intermodulation amplitude output at time t-1; S121. Split all training data sequences and the real and imaginary parts of the passive intermodulation amplitude as inputs, and use the real and imaginary parts of the passive intermodulation amplitude as outputs; determine the internal state of the reservoir and the internal state of the enhancement layer before the first iteration; set the connection weights from the input layer to the reservoir, between reservoirs, and from the reservoir to the enhancement layer. S122. Randomly generate matrices of connection weights from the input layer to the reservoir, matrices of connection weights between reservoirs, and matrices of connection weights from the reservoir to the enhancement layer; S123. Scaling the matrix of connection weights between reservoirs based on spectral radius; S124. Configure state update functions for the reservoir and enhancement layer to record the reservoir state during the time period from the end of the idling data sequence to the end of the training data sequence.

[0040] As an example, the reservoir state update function is as follows: in, For reservoir renewal rate; For the update time; This represents the internal state of the reservoir at time t-1. The connection weights from the input layer to the reservoir; To send a data sequence; The connection weights between reservoirs.

[0041] The state update function of the enhancement layer is as follows: in, The shrinkage coefficient of the reinforcing layer; The connection weights from the reservoir to the reinforcement layer; Let t represent the internal state of the reservoir.

[0042] S125. In the first iteration, principal component analysis is performed on the reservoir state to determine the number of reservoir nodes. The eigenvalues ​​of the covariance matrix of the reservoir neurons in the reservoir state are determined, and the variance contribution rate of the eigenvalues ​​is determined according to their order. Based on this, the number of principal components whose cumulative variance contribution rate exceeds a preset threshold is determined. The number of principal components exceeding the preset threshold is the number of reservoir nodes. The preset threshold is greater than or equal to 90%.

[0043] S126. Determine the output weights based on reservoir state, regularization coefficient, number of input nodes, number of reservoir nodes and number of enhancement layer nodes, and output matrix.

[0044] S13. Passive intermodulation amplitude prediction: Input the real and imaginary parts of all transmitted data sequences into the trained echo state network to obtain the real and imaginary parts of the passive intermodulation amplitude prediction value. Based on this, the passive intermodulation amplitude prediction value is obtained.

[0045] As an example, it may include the following steps: S130. The reservoir initialization state of the trained echo state network is the state obtained in the last iteration, and the output weights are the output weights determined in the training echo state network steps. S131. Input the real and imaginary parts of all the data sequences to be sent into the trained echo state network to obtain the output data sequence; S132. By performing inverse normalization on the output data sequence, the predicted value of the passive intermodulation amplitude can be obtained.

[0046] Compared with the prior art, the present invention has the following advantages: 1. Due to the low computational cost and applicability of echo state networks in time series prediction, their application in multi-channel passive intermodulation amplitude prediction has advantages such as low computational complexity, low resource cost, and ease of hardware implementation compared to existing adaptive passive intermodulation prediction methods based on digital signal processing.

[0047] 2. When initializing the echo state network, specifically during the first iteration of the reservoir, the number of reservoir nodes is determined using principal component analysis. Based on this, the size of the reservoir neurons can be optimized, reducing computational complexity.

[0048] 3. Improve the echo state network by adding an enhancement layer. This not only has the advantages of stable performance, strong system adaptability, and good real-time performance, but also has the advantage of low system complexity compared with the existing adaptive passive intermodulation prediction method based on digital signal processing.

[0049] 4. The multi-channel passive intermodulation amplitude prediction method can also accurately predict the amplitude of passive intermodulation interference.

[0050] To facilitate understanding of the multi-channel passive intermodulation amplitude prediction method based on echo state network provided in the embodiments of the present invention, a specific communication system is used as an example for illustration below. It should be understood that the following examples are for explanation only and are not intended to be limiting.

[0051] To verify the feasibility of the method, we assume that the communication system transmits a dual-channel signal with the following specific parameters: (1) Modulation method: QPSK; (2) Uplink carrier frequency: 13MHz; (3) Downlink carrier frequency: 125MHz, 155MHz; (4) Oversampling rate: 16; (5) Symbol rate: 0.7 Msps (6) Roll-off factor: 0.35; (7) Passive intermodulation generation model: fractional order model, which can be expressed as: .

[0052] in, The input excitation signal, This is the output PIM signal. For each coefficient, the order is... It is a positive real number. Here... The values ​​are 1 and 0.6. The values ​​are 1.6 and 8.9.

[0053] S20. Insert a training data sequence with a length ratio of 1:3 to the transmitted data sequence. Then, normalize the data sequence containing both the training and transmitted data sequences. The specific method of normalization is as described above.

[0054] S21. Initialize the echo state network. Set the number of input nodes for the echo state network. The output number of nodes is 6. The number of reservoir nodes is 2. 2000; number of enhancement layer nodes 200; reservoir renewal rate The shrinkage coefficient is 0.9. It is 0.7; spectral radius The ratio is 0.2; the ratio of the length of the idle sequence to the length of the training sequence. : The ratio is 1:4; regularization coefficient 1e-8; Activation function selection function.

[0055] S22. Training the echo-state network. Let the dual-channel training sequence and its generated passive intermodulation interference sequence be respectively... , and ,in To determine the training sequence length, it is split into real and imaginary parts. , , , , and As input, let it be... , and For output, set to The internal state of the reservoir is as follows: ,set up The internal state of the enhancement layer is as follows: ,set up . , as well as This represents the connection weights input to the reservoir, between reservoirs, and from the reservoir to the enhancement layer. For output weights.

[0056] Randomly generate matrices, , , .

[0057] Scaling matrix, .in, denoted as spectral radius.

[0058] At time t, the reservoir state update equation is: ,in This refers to the reservoir renewal rate.

[0059] The state update equation for the enhancement layer is: Record number To the The reservoir state of this step is denoted as... .

[0060] During the first iteration Perform PCA processing to find Middle reservoir neuron part The eigenvalues ​​of the covariance matrix are obtained, and the variance contribution rate of the eigenvalues ​​is calculated in order of eigenvalues. The number of principal components with a cumulative variance contribution rate of more than 90% is calculated.

[0061] Among them, the cumulative contribution rate of variance, i.e. ; The expression is: .

[0062] More than 90% of the principal components have a number of k.

[0063] reservoir size After setting it to k, learning is conducted via the network.

[0064] The output weights are: ,in is the regularization coefficient.

[0065] S23. Passive Intermodulation Amplitude Prediction The reservoir initialization state of the echo state network is the state obtained in the last iteration of the previous step, and the output weights are the values ​​calculated in the previous step. A data sequence will be sent. Substituting these values ​​into the echo state prediction network yields the network output value. Then, by performing inverse normalization on the obtained output sequence, the predicted passive intermodulation amplitude sequence can be obtained. The real parts of the predicted result and the actual result are as follows: Figure 4 As shown in the figure, the simulation results demonstrate that the echo state network has high prediction accuracy. The horizontal axis represents time, the vertical axis represents amplitude, the cross-shaped line represents the real part of the actual passive intermodulation amplitude, and the circular line represents the real part of the predicted passive intermodulation amplitude. The figure shows that the predicted value almost coincides with the actual value, indicating that the echo state network has high prediction accuracy for passive intermodulation amplitude. Using the multi-channel passive intermodulation amplitude prediction method based on the echo state network provided in this invention, a two-channel PIM signal with an environmental signal-to-weight ratio of 10 dB was suppressed, resulting in the time-domain comparison before and after PIM interference shown in Figure 5. It can be seen that after interference cancellation using the prediction method provided in this invention, the recovered signal almost coincides with the signal without PIM interference, proving that this method can effectively eliminate PIM interference signals.

[0066] In a second aspect, the present invention also provides a multi-channel passive intermodulation amplitude prediction device based on an echo state network, the multi-channel passive intermodulation amplitude prediction device being used to execute the multi-channel passive intermodulation amplitude prediction method provided in the first aspect.

[0067] Compared with the prior art, the beneficial effects of the multi-channel passive intermodulation amplitude prediction device based on echo state network provided by the present invention are the same as the beneficial effects of the multi-channel passive intermodulation amplitude prediction method based on echo state network provided by the first aspect and / or any implementation of the first aspect, and will not be elaborated here.

[0068] In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0069] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the amplitude of multi-channel passive intermodulation based on echo state networks, characterized in that, Includes the following steps: During the transmission of the transmitted data sequence, multiple signal transmission channels insert training data sequences into the transmitted data sequence according to a preset ratio, where the preset ratio is the ratio of the length of the training data sequence to the length of the transmitted data sequence. Initialize the echo state network, which includes an input layer, a reservoir layer, an enhancement layer, and an output layer; initialize the number of input nodes in the input layer, the number of output nodes in the output layer, the number of reservoir nodes in the reservoir layer, and the number of enhancement layer nodes in the enhancement layer; initialize the reservoir update rate; initialize the shrinkage coefficient; initialize the spectral radius; initialize the length ratio of the idle data sequence to the training data sequence; initialize the regularization coefficient and the activation function; The echo state network is trained by splitting the training data sequence input to each signal transmission channel at time t and the passive intermodulation amplitude at time t-1 into real and imaginary data, and inputting them into the input node of the echo state network. The passive intermodulation amplitude at time t is split into real and imaginary data as the output data of the network to train the network and obtain the output weights. Passive intermodulation amplitude prediction involves inputting the real and imaginary parts of all transmitted data sequences into the trained echo state network to obtain the real and imaginary parts of the passive intermodulation amplitude prediction value. Based on this, the passive intermodulation amplitude prediction value is obtained. Training the echo-state network includes the following steps: Determine the training data sequence input to the multiple signal transmission channels at time t, and the passive intermodulation amplitude output at time t-1; All training data sequences and the real and imaginary parts of the passive intermodulation amplitude are separated as inputs, and the real and imaginary parts of the passive intermodulation amplitude are used as outputs; the internal state of the reservoir and the internal state of the enhancement layer before the first iteration are determined. Set the connection weights from the input layer to the reservoir, between reservoirs, and from the reservoir to the enhancement layer; Randomly generate matrices of connection weights from the input layer to the reservoir, matrices of connection weights between reservoirs, and matrices of connection weights from the reservoir to the enhancement layer; The matrix of connection weights between reservoirs is scaled based on the spectral radius; Configure a state update model for the reservoir and enhancement layer to record the reservoir state during the time period from the end of the idle data sequence to the end of the training data sequence. In the first iteration, principal component analysis is performed on the reservoir state to determine the number of reservoir nodes; The output weights are determined based on the reservoir state, regularization coefficient, number of input nodes, number of reservoir nodes, number of enhancement layer nodes, and output matrix. In the first iteration, principal component analysis is performed on the reservoir state to determine the number of reservoir nodes included in the reservoir, including: The eigenvalues ​​of the covariance matrix of the reservoir neuron portion in the reservoir state are determined, and the variance contribution rate of the eigenvalues ​​is determined according to the order of the eigenvalues. Based on this, the number of principal components whose cumulative variance contribution rate exceeds a preset threshold is determined. The number of principal components that exceed the preset threshold is the number of reservoir nodes.

2. The multi-channel passive intermodulation amplitude prediction method based on echo state networks according to claim 1, characterized in that, The preset ratio is 1:N, 3≤N≤5, that is, a training data sequence is inserted between the N adjacent sets of transmitted data sequences.

3. The multi-channel passive intermodulation amplitude prediction method based on echo state networks according to claim 1, characterized in that, During the transmission of the transmitted data sequence, after inserting a training data sequence into the transmitted data sequence according to a preset ratio, the multi-channel passive intermodulation amplitude preset method further includes: The data sequence, including the transmitted data sequence and the training data sequence, is normalized.

4. The multi-channel passive intermodulation amplitude prediction method based on echo state networks according to claim 1, characterized in that, The number of signal transmission channels is M. At this time, the number of input nodes is 2M+2, and the number of output nodes is two. The number of reservoir nodes is determined by principal component analysis. The enhancement layer enhances the characteristics of the reservoir through multiple contractile neurons. The reservoir characteristics have good nonlinear fitting ability between the time series of input and output. The enhancement layer and the reservoir are connected by random weights and biases.

5. The multi-channel passive intermodulation amplitude prediction method based on echo state networks according to claim 1, characterized in that, The preset threshold is greater than or equal to 90%.

6. The multi-channel passive intermodulation amplitude prediction method based on echo state networks according to claim 3, characterized in that, The real and imaginary parts of all data sequences to be transmitted are input into the trained echo state network to obtain the real and imaginary parts of the passive intermodulation amplitude prediction value. Based on this, the passive intermodulation amplitude prediction value is obtained, including: The reservoir initialization state of the trained echo state network is the state obtained in the last iteration, and the output weights are the output weights determined in the training echo state network steps. The real and imaginary parts of all the data sequences to be sent are input into the trained echo state network to obtain the output data sequence; By performing inverse normalization on the output data sequence, the predicted value of the passive intermodulation amplitude can be obtained.

7. The multi-channel passive intermodulation amplitude prediction method based on echo state networks according to claim 1, characterized in that, The reservoir state update function is as follows: in, For reservoir renewal rate; For the update time; This represents the internal state of the reservoir at time t-1. The connection weights from the input layer to the reservoir; To send a data sequence; The connection weights between reservoirs; The state update function of the enhancement layer is as follows: in, The shrinkage coefficient of the reinforcing layer; The connection weights from the reservoir to the reinforcement layer; Let t represent the internal state of the reservoir.

Images