A method for canceling passive intermodulation in a device

Abstract

The present application belongs to the technical field of wireless communication interference suppression, and particularly relates to a FDD MIMO external passive intermodulation digital cancellation method based on a geometric condition sparse tensor train network. The present application directly introduces array position, finite-length dipole complex vector field, beamforming and candidate source polarization into spatial branches through a near-field vector field geometric encoder, so that the network first satisfies the external PIM near-field propagation and polarization projection relationship, and then learns the residual caused by array mutual coupling, actual unit directional diagram and positioning deviation with fewer parameters. The present application automatically identifies the external PIM source truly activated by the current beam in the candidate source set through sparse source gating, adapts to the problem that the number of effective sources is unknown and changes with the beam, and can prune invalid branches during online reasoning, thereby reducing parameter waste, local minimum risk and computational complexity.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication signal processing and radio frequency interference suppression technology, specifically relating to a method for canceling state condition timing networks of passive intermodulation within a device. Background Technology

[0002] In frequency division duplex (FDD) radio equipment, the downlink transmit and uplink receive links are typically located close to each other or share parts of the RF path through devices such as duplexers, filters, RF switches, coaxial connectors, printed circuit board transition structures, antenna feed points, screws, and spring contact structures. Weak contacts, oxide layers, surface contamination, ferromagnetic materials, and micro-protrusion plastic deformation in these devices can create non-ideal nonlinear nodes, generating passive intermodulation (PIM) components between multiple downlink carriers. When PIM components fall into the uplink receive band, they raise the receive noise floor, reduce receive sensitivity, and cause varying degrees of interference to multiple antenna ports. Compared to passive intermodulation in the external environment, the propagation distance of internal passive intermodulation is shorter, but its excitation reference point, equivalent transmission path, and nonlinear state are more complex. The actual excitation experienced by different internal sources may be closer to the baseband signal before digital predistortion, the baseband signal after digital predistortion, or the power amplifier output coupling signal, respectively. If a single reference point is used for modeling, the amplitude and phase mismatch between the reference signal and the actual excitation will directly result in an unavoidable model error.

[0003] Duplexers and filters exhibit frequency-dependent group delays. After multi-band, multi-carrier signals pass through different transmit branches, common ports, and receive branches, the delay relationships corresponding to each intermodulation combination term are not the same. Ordinary memory polynomials typically enumerate only a limited number of principal delays, making it difficult to simultaneously compensate for cross-band group delays and multi-port path differences with low complexity. While the Volterra model can express richer delay combinations, the number of its basis functions increases rapidly with the nonlinear order, the number of frequency bands, and the memory depth, making it unsuitable for deployment on digital signal processors, field-programmable gate arrays, or application-specific integrated circuits. The static mapping of passive intermodulation within a device typically exhibits approximately odd-symmetric properties; that is, when the complex excitation is inverted, the passive intermodulation output is also inverted. General feedforward neural networks do not explicitly utilize this physical prior, easily introducing redundant parameters and exhibiting reduced generalization when the number of device samples is limited and the operating state coverage is insufficient. Simultaneously, contact structures exhibit memory effects; the current contact resistance and nonlinear state depend not only on the current power or temperature but also on previous thermal cycles, mechanical stress, and switching operation history. Temperature rise, temperature drop, power scanning, and RF switching can alter the equivalent contact state and RF path. Abrupt changes in timing and location can also occur at connector, spring, or screw contacts. While conventional long short-term memory (LSTM) networks can utilize historical states, their hidden states may lag in response to abrupt changes; training only under stable operating conditions can easily lead to large transient prediction residuals near temperature cycles and switching events. In practical multi-band, multi-antenna devices, multiple internal sources such as duplexers, RF switches, and feed connection structures may coexist, and different receiver ports may observe passive intermodulation from the same physical source via different linear paths. Using a single holistic model makes it difficult to clearly distinguish source-related reference positions, group delays, and state changes; building a complete model independently for each port results in an excessive number of parameters and fails to utilize the static nonlinearities that can be shared between similar devices.

[0004] Furthermore, online cancellation needs to avoid misjudging the desired uplink signal as passive intermodulation and eliminating it. Updating the model unprotected when uplink traffic exists, or using full-gain cancellation when the model prediction has low correlation with the received signal, may cause amplitude and phase distortion of the desired signal. Therefore, engineered cancellation of passive intermodulation within devices requires not only a high-precision timing model but also constrained online updates, gain control, and model backoff mechanisms. For these reasons, a state-conditional timing network cancellation method for passive intermodulation within devices is needed, ensuring that the network structure corresponds to the technical challenges of multi-reference excitation, filter group delay, odd-symmetric static nonlinearity, contact memory, random mutations, and multi-source superposition, while providing uplink protection capabilities while maintaining deployable complexity. Summary of the Invention

[0005] The purpose of this invention is to provide a state-condition timing network cancellation method for passive intermodulation within a device, in order to solve problems in the prior art such as inconsistent positions of reference signals from different internal sources, mutual coupling between duplexer group delay and device nonlinearity, failure of general networks to utilize odd-symmetric priors, difficulty in uniformly modeling contact memory and abrupt changes in operating state, excessive number of parameters when multiple internal sources are superimposed, and the possibility of online cancellation accidentally damaging the desired uplink signal.

[0006] This invention constructs a state-conditional symmetric nonlinear compensation time-series network, which includes a multi-reference selection and group delay decoupling module, an odd-symmetric static nonlinear module, a state-conditional nonlinear compensation long short-term memory module, a causal dilated convolutional mutation branch, and a state-conditional sparse hybrid expert routing module. The method involves training the state-conditional symmetric nonlinear compensation time-series network and using the trained network to predict and cancel passive intermodulation within devices generated by at least one of duplexers, RF switches, and feed connection structures. The method includes the following steps.

[0007] S1. Acquire multi-reference complex baseband signal Receive complex baseband signals and auxiliary state vector The multi-reference complex baseband signal originates from two or more of the following processing locations: a location before digital predistortion processing and before or after peak-to-average power ratio (CFR) reduction processing; a location after digital predistortion processing and before digital-to-analog conversion; the transmit coupler; and the forward power monitoring branch. The auxiliary state vector includes one or more of the following: temperature, temperature change, forward power, RF switch state, band coding, receiver port coding, and calibration state, expressed as:

[0008]

[0009] in, Indicates the index of the sampling point or sampling block. Indicates temperature. Indicates the amount of temperature change. Indicates forward power. Indicates the status of the radio frequency switch. Indicates frequency band coding. Indicates the receive port encoding. Indicates calibration status. This indicates transpose. In the above expressions, Indicates the first Multi-reference complex baseband signal, Indicates the reference signal index. Indicates the number of reference signals. This indicates that the signal is a complex baseband signal. This represents the auxiliary state vector.

[0010] S2. In the multi-reference selection and group delay decoupling module, for each candidate internal passive intermodulation expert, the multi-reference selection weight is calculated according to the auxiliary state vector, the complex baseband signals at different reference positions are weighted by state conditions, and the filter group delay and port linear path are decoupled from static nonlinearity by the device-sensing complex finite impulse response filter.

[0011]

[0012]

[0013]

[0014]

[0015]

[0016] in, This indicates the candidate's internal passive inter-invocation expert index. Indicates the number of candidate experts; Indicates the reference signal index. Indicates the number of reference signals; This indicates the weighting of multiple reference selections; Indicates a reference summation index; Indicates the first A weighted combination of signals from candidate experts; This indicates the group delay decoupling characteristics; Indicates the finite impulse response filter length; Indicates the filter tap index; Represents the fundamental complex filtering matrix; This represents a low-rank state modulation complex filter matrix; Indicates the number of low-rank state modulation terms; Represents the state modulation coefficient; , , and These are learnable parameters; Represents the Sigmoid function; This represents an exponential function. Furthermore, This represents the state-conditional complex filtering matrix determined by the candidate expert index, the filter tap index, and the auxiliary state vector. This represents the index of the low-rank state modulation term.

[0017] S3. Input the group delay decoupling feature into the odd-symmetric static nonlinear module. This module uses basis functions with paired shared weights to ensure that the static mapping satisfies the odd mapping constraint that the output is inverted when the input is inverted, and allows multiple ports of the same physical source or the same device type to share basis function parameters.

[0018]

[0019]

[0020]

[0021]

[0022] in, This represents a real-number input formed by in-phase components, quadrature components, or amplitude coupling features. Represents a linear rectified function; Indicates the first Breakpoints in the basis functions; Indicates the number of basis functions; Indicates the basis function weights; Represents a real odd-symmetric mapping; and They represent complex numbers respectively The real and imaginary parts; Represents the imaginary unit; Represents an odd-symmetric mapping for complex numbers; Indicates the first The static nonlinear characteristics of each candidate expert are analyzed. Based on the group delay decoupling characteristics, an odd-order generalized memory polynomial basis function is constructed, and the higher-order intermodulation characteristics are explicitly input into the state conditional nonlinear compensation of the long short-term memory module's compensation branch. Furthermore... Represents the pairwise linear rectified basis functions. Indicates the base function index.

[0023]

[0024]

[0025]

[0026] in, A complex channel component representing the group delay decoupling characteristic; Indicates the order of nonlinearity; Represents sets of odd order; This indicates a delay in main memory; Indicates envelope delay; Indicates the depth of primary memory; Indicates the maximum envelope delay; Represents the basis functions of the generalized memoization polynomial; This represents vectorized concatenation; Represents the basis function vector of the generalized memory polynomial; and These are learnable parameters; Represents a nonlinear compensation vector; Represents the modulus of a complex number; This represents the hyperbolic tangent function.

[0027] S4. The group delay decoupling feature and the static nonlinear feature are concatenated to form a timing input. The hybrid gate, compensation gate and output gate are modulated according to the auxiliary state vector, and the unit state is updated using the nonlinear compensation vector.

[0028]

[0029]

[0030]

[0031]

[0032]

[0033]

[0034]

[0035] in, Represents the characteristics of time-series input; Indicates a hybrid gate; Indicates the candidate cell status; Indicates a compensation gate; Indicates the cell state; Indicates the output gate; Indicates a hidden state; Represents a vector that is all one; This represents element-wise multiplication; , , , And the corresponding biases are learnable parameters. The hybrid gate and its complementary term control the historical state and the current candidate state, respectively, while the compensation gate is used to explicitly inject higher-order features of the generalized memory polynomial. Furthermore, and These represent the cell state and hidden state of the previous sampling point or sampling block, respectively; vectors connected by semicolons or commas within square brackets represent concatenation.

[0036] S5. Construct a short-window mutation input by using the change in the auxiliary state vector, the change in the group delay decoupling feature, and the cancellation residual signal from the previous time step. Obtain the mutation correction vector through causal hole depth separable convolution and fuse it with the hidden state and static nonlinear features.

[0037]

[0038]

[0039]

[0040]

[0041] in, Indicates the input of the mutation branch; Indicates the length of the causal observation window; This indicates a separable convolution operation with causal dilation depth. Indicates the number of convolutional layers; Indicates the dilation rate of each convolutional layer; Represents the mutation correction vector; , and This represents a learnable output mapping matrix; Indicates the first Device-internal passive intermodulation prediction by several candidate experts. Furthermore, and These represent the changes in the auxiliary state vector and the changes in the group delay decoupling features of the corresponding candidate experts, respectively. This represents the cancellation residual signal from the previous sampling point or sampling block; This represents a causal short window sequence consisting of the current sampling point or sampling block and its historical inputs.

[0042] S6. In the state-condition sparse hybrid expert routing module, sparse routing weights are generated based on the auxiliary state vector and known device paths. A portion of the duplexer experts, RF switch experts, power supply connection structure experts, and backup unknown source experts are activated and weighted and fused.

[0043]

[0044]

[0045]

[0046] in, Indicates a hard path mask; and These are learnable routing parameters; Indicates the first The routing weights of each candidate expert; This indicates the total prediction of passive intermodulation within the device; This indicates the cancellation of gain; This represents the received residual signal after cancellation. At least one hard path mask is one at any given time to ensure the routing denominator is non-zero. In the summation of the denominator in the routing weight normalization formula, This represents the candidate internal passive inter-invocation expert summation index.

[0047] S7. Construct a training loss function based on the device’s internal passive intermodulation supervision target and / or the target uplink band residual power, and perform joint or phased optimization on multi-reference selection, group delay decoupling, odd symmetric static nonlinearity, state condition timing memory, mutation correction and sparse routing; after deployment, update the selected parameters only in the time slots that meet the uplink protection conditions.

[0048] Furthermore, the internal passive intermodulation monitoring target of the device can be obtained through uplink silent sampling, known calibration sequence, desired uplink signal stripping, external instrument measurement, or a combination of the above methods; when a complete monitoring target cannot be obtained, the residual power in the target uplink frequency band, the correlation between the predicted and received signals, and state consistency can be used as weak monitoring constraints.

[0049] Furthermore, the candidate internal passive intermodulation experts correspond one-to-one or many-to-one with the physical source types. Multiple receiver ports belonging to the same duplexer, the same RF switch, or the same feed connection structure can share odd-symmetric static nonlinear parameters and employ different complex finite impulse response filtering parameters and output mapping parameters.

[0050] Furthermore, the causal-hole convolutional mutation branch employs one or more depth-separable convolutions with progressively increasing dilation rates; the mutation correction gain is reduced when the temperature change, RF switch state change, or residual derivative does not exceed the corresponding threshold, and increased when the corresponding threshold is exceeded.

[0051] Furthermore, the state-conditional sparse hybrid expert routing module employs low-entropy constraints and Top-... Choose one or more of the following: sparse constraints, expert group constraints, or hard path masks, to avoid all experts being active simultaneously in all working states; among them... This indicates the number of candidate experts retained, and .

[0052] Furthermore, the protected online adaptive method only updates one or more of the low-rank state modulation complex filter parameters, routing parameters, output mapping parameters, and mutation branch end layer parameters, and freezes the main parameters of the odd-symmetric static nonlinear module and the state-conditional nonlinear compensation long short-term memory module.

[0053] The beneficial effects of this invention are as follows:

[0054] 1) Based on the physical location of the duplexer, RF switch and feed connection structure, the present invention dynamically selects the reference before digital predistortion, after digital predistortion or the transmit coupler reference, which can reduce the amplitude and phase mismatch between a single reference point and the real internal source excitation.

[0055] 2) This invention utilizes a device-sensing complex finite impulse response filter to independently process the stable group delay, and uses low-rank state modulation to express the path changes caused by temperature and switching state, thus making the division of labor between linear group delay and static nonlinearity clear;

[0056] 3) This invention explicitly writes the odd mapping physical prior into the basis function structure and allows multiple ports of the same type of device and the same physical source to share basis function parameters, which helps to reduce redundant parameters and improve the generalization ability under small sample working conditions.

[0057] 4) This invention explicitly injects the higher-order features of the generalized memory polynomial into the state-conditional nonlinear compensation long short-term memory module, enabling the network to simultaneously utilize higher-order intermodulation priors, short-term contact memory, and auxiliary information such as temperature, power, and switching state.

[0058] 5) This invention sets up a causal-dilated convolutional mutation branch to perform short-window correction for rapid changes near temperature rise, temperature drop, power scan, and RF switch switching, which helps to reduce the response lag of the recurrent network at the moment of mutation.

[0059] 6) This invention employs state-conditional sparse hybrid expert routing, which enables source-related modeling when multiple internal sources coexist, and reduces invalid expert computation through hard path masks and sparse constraints;

[0060] 7) This invention employs a hierarchical training method that combines static device pre-training, time-series memory training, mutation training, device-level fine-tuning, and protected online adaptive training, enabling the transfer and control of the number of online trainable parameters between different device samples.

[0061] 8) This invention controls the cancellation gain and online updates based on predicted correlation, expected uplink signal detection, maximum cancellation amplitude, phase change rate, and model backoff conditions, which helps to avoid false cancellation of expected uplink signals. Attached Figure Description

[0062] Figure 1A schematic diagram illustrating the physical path for passive intermodulation generated by the internal duplexer, RF switch, and power supply connection structure of the multi-band frequency division duplex wireless device provided in this embodiment of the invention;

[0063] Figure 2 This is a schematic diagram of the overall structure of a state-conditional symmetric nonlinear compensation time-series network provided in an embodiment of the present invention;

[0064] Figure 3 A schematic diagram of the multi-reference selection, device-sensing complex finite impulse response filtering, and odd-symmetric static nonlinear module structure provided in the embodiments of the present invention;

[0065] Figure 4 A schematic diagram of the state condition nonlinear compensation long short-term memory module provided in an embodiment of the present invention;

[0066] Figure 5 This is a schematic diagram of the structure of the causal dilated convolutional mutation branch and state condition sparse hybrid expert routing module provided in an embodiment of the present invention. Detailed Implementation

[0067] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments:

[0068] In this specification, "internal passive intermodulation" refers to passive intermodulation components generated within duplexers, filters, RF switches, connectors, feed line transition structures, antenna feed points, screws, springs, or other passive contact structures of a wireless device, and entering the digital baseband via the receiving branch. "Candidate internal passive intermodulation expert" refers to a predictive subnetwork configured for a specific physical source type, a particular physical source, or a set of similar physical sources sharing parameters.

[0069] The complex vectors described in this specification can be implemented by concatenating real numbers of in-phase and quadrature components; complex matrix multiplication can be implemented by corresponding real block matrix operations. The network can run on a central processing unit, graphics processing unit, digital signal processor, field-programmable gate array, application-specific integrated circuit, or a combination of the above hardware.

[0070] Example:

[0071] like Figure 1 As shown, this embodiment takes a dual-band 4T4R frequency division duplex remote radio unit as an example. The first frequency band can be n25, and the second frequency band can be n66; each frequency band includes two or three discontinuous component carriers, and the component carrier bandwidth can be 5MHz or 20MHz. The two frequency bands enter the duplexer branch after passing through peak-to-average power ratio reduction, digital predistortion, digital-to-analog conversion, up-conversion, and power amplifier, respectively, and may form internal nonlinear nodes at shared or adjacent RF switches, feed connection structures, and antenna ports.

[0072] The RF switch in this embodiment is used for calibration path, redundant path, or multi-antenna port selection, and can have four or more discrete states, allowing for repeated switching. The feed connection structure may include a coaxial connector, a printed circuit board or microstrip transition, an antenna feed point, screws, and spring contacts. The device input power can range from 31 dBm to 49 dBm, and the ambient temperature can range from -40 degrees Celsius to 85 degrees Celsius, including temperature rise, isothermal, and temperature drop phases. These values ​​are merely examples and are not intended to limit the invention.

[0073] During data acquisition, multi-reference complex baseband signals are acquired synchronously for each frequency band and transmit chain. The first reference can come from before digital predistortion; the second reference can come from after digital predistortion and before digital-to-analog conversion; and the third reference, when hardware allows, comes from the transmit coupler after the power amplifier. For devices that cannot provide the transmit coupler complex envelope, the third reference can be reconstructed using forward power, carrier configuration, and digital baseband signals, or a degraded mode containing only the first two types of references can be used.

[0074] The receiver synchronously acquires the received complex baseband signal after passing through the duplexer's receiving branch, low-noise amplifier, down-conversion, and analog-to-digital conversion. During training, acquisition occurs during periods of uplink silence, when the calibration sequence is known, or when the desired uplink signal can be stripped, ensuring that the device's internal passive intermodulation monitoring target is obtainable. The general received signal model is as follows:

[0075]

[0076] in, express The received complex baseband vector of a receiver chain; This indicates a desired upward signal; This indicates the actual number of passive intermodulation sources within the device; Indicates the first Passive intermodulation components generated by an actual source; This represents received noise and unmodeled interference. Actual number of sources. It can be less than or equal to the number of candidate experts. In the above formula, Represents the complex field.

[0077] The auxiliary state vector is constructed using the aforementioned formula. Temperature can be obtained from a temperature sensor near the chassis, duplexer, or connector; temperature change can be obtained by differential sampling of adjacent sampling blocks or differential sampling after low-pass filtering; forward power can be the power of each frequency band or the combined power; RF switch state, frequency band coding, receiver port coding, and calibration state can use one-hot coding, embedded coding, or normalized continuous coding.

[0078] To cover contact memory and random mutations, the acquisition matrix includes single-band and dual-band transmission, multiple component carrier combinations, full-power scanning, temperature rise, isothermal, temperature drop, repeated RF switch switching, lifetime cycling, multiple duplexer samples, multiple RF switch samples, multiple feed connection samples, and all transmit and receive port combinations. The passive intermodulation order is mainly third-order, but it can simultaneously cover fifth-order and seventh-order components.

[0079] In one possible implementation, such as Figures 2 to 5 As shown, the candidate internal passive intermodulation experts are divided into duplexer experts, RF switch experts, feed connection structure experts, and backup unknown source experts. Each candidate expert has independent or partially shared multi-reference selection parameters, complex finite impulse response filtering parameters, odd-symmetric static nonlinear parameters, state condition timing parameters, abrupt branching parameters, and output mapping parameters.

[0080] S101 performs synchronization, segmentation, and normalization of the input. Each reference signal and the received signal are time-aligned according to a common sampling clock. When the sampling rate of the coupler reference is lower than the digital baseband sampling rate, interpolation, hold, or statistical analysis by sampling block can be performed. Complex signals can be normalized according to the training set mean and standard deviation, root mean square amplitude, or rated power. The auxiliary state vector can be linearly normalized according to the physical range.

[0081] For the Road reference signal, denoted as ,in This represents the dimension of the reference features. Reference features can be concatenated by frequency band, transmit chain, and carrier, or the dimension can be compressed first using port-dependent linear mapping. For reference types that are not collected, an availability mask can be set, and the corresponding references can be excluded during the normalization of the multi-reference selection weights.

[0082] S102, multiple reference selections for execution status conditions. For example... Figure 3 As shown, for different physical sources, the required reference location varies with the RF switch state, frequency band combination, and forward power. The weighting of multiple reference selection for each candidate expert is calculated using the following formula:

[0083]

[0084] In the formula, For the first The candidate experts on the first The route score for each available reference is calculated; the denominator is summed over all available references. When a reference availability mask exists, it can be multiplied before the exponent term, or the route score for unavailable references can be set to a finite approximation corresponding to negative infinity.

[0085]

[0086] Weighted combination signal This allows the same candidate expert to smoothly switch between before, after, and between the transmit coupler reference. For devices with a known fixed reference location, the corresponding weight can be fixed to one, and other weights can be fixed to zero.

[0087] S103, Device-Sensitive Complex Finite Impulse Response Filter. The basic complex filter matrix is ​​used to represent the stable group delay formed by duplexers, filters, RF traces, and port connections; the low-rank state-modulated complex filter matrix is ​​used to represent smaller path variations caused by temperature, RF switching states, and calibration states.

[0088]

[0089]

[0090]

[0091] In one implementation, Obtained during the static device pre-training phase and frozen during normal device operation; The parameters for generating the state modulation coefficients are updated during device-level fine-tuning or online adaptation. The number of low-rank state modulation terms... It can be less than the number of complete operating states, thus avoiding the need to store a complete set of filters for each temperature range and each RF switch state.

[0092] Finite impulse response filter length The parameters are determined based on the group delay, digital downsampling rate, and maximum path difference of the duplexer transmit / receive filters. For example, when processing by sample block, The values ​​can range from 9 to 17; at higher sampling rates, longer filters or polyphase filter structures can also be used. The above ranges are for illustrative purposes only.

[0093] S104 performs an odd-symmetric static nonlinear mapping. For each in-phase, orthogonal, or amplitude-coupled real characteristic, a pairwise linear rectified basis function with a shared breakpoint is constructed:

[0094]

[0095]

[0096] The aforementioned paired basis functions share breakpoints and weights at positive and negative inputs, enabling the direct incorporation of odd-symmetric priors into the network. For complex inputs, in-phase / orthogonal component mapping can be employed.

[0097]

[0098]

[0099] In another implementation, a phase-preserving amplitude-type odd mapping can be used. ,in Represents linear coefficients. This represents a piecewise linear magnitude basis function. Since the magnitude function depends only on... This mapping also satisfies .

[0100] When passive intermodulation acknowledgments from multiple receiver ports originate from the same duplexer, the same RF switch, or the same feed connection structure, the basis function breakpoints and basis function weights can be shared, retaining only the linear input mapping, complex finite impulse response filtering, and output mapping for different ports. For cases with significant differences in device samples, small-dimensional sample embeddings or low-rank increments can be superimposed on the shared weights.

[0101] S105, construct the generalized memory polynomial basis functions and perform state-conditional nonlinear compensation for long short-term memory recursion. For example... Figure 4 As shown, the generalized memory polynomial basis functions extract odd-order principal terms and envelope cross-delay terms from the group delay decoupling characteristics:

[0102]

[0103]

[0104] Odd-order sets Can be taken , or Main memory depth The value can be 3 to 6; maximum envelope delay The value can be between 0 and 2. For multi-channel group time delay decoupling features, basis functions can be constructed independently for each channel and then concatenated, or cross basis functions can be constructed only for the channel pairs corresponding to the target intermodulation combination.

[0105]

[0106] Nonlinear compensation vector Instead of being directly used as a complete output, the unit state is injected through a compensation gate, thereby avoiding parameter inflation caused by fully expanding all generalized memory polynomial terms to the output layer.

[0107]

[0108]

[0109]

[0110]

[0111]

[0112]

[0113]

[0114] Mixed gate Its complementary terms Separately control the state of historical cells and the state of current candidate cells to reduce gradient conflicts that may occur between the independent forget gate and the input gate; compensation gate The injection intensity of the generalized memory polynomial compensation vector is controlled; the auxiliary state vector simultaneously enters the mixing gate, compensation gate, and output gate, so that the history hold ratio, current update ratio, and output exposure ratio vary with temperature, forward power, and RF switch state.

[0115] When auxiliary state information is missing, it can be set as the default state vector, or the state-related weight terms can be set to zero, causing the state-conditional nonlinear compensation long short-term memory module to degenerate into a nonlinear compensation long short-term memory module without external state modulation.

[0116] S106, perform mutation correction. For example... Figure 5 As shown, the temperature rise, temperature drop, and repeated contact of the connection structure may change rapidly at unpredictable times. Therefore, this embodiment constructs the following differential component and abrupt input:

[0117]

[0118]

[0119]

[0120] Causal-dilation-separable convolutions use only current and historical inputs, excluding future samples, making them suitable for online inference. (Number of convolutional layers) It can be three layers, with dilation rates of 1, 2, and 4 respectively; other dilation rate combinations can also be used depending on the sampling block length. Depthwise separable convolution first performs temporal convolution on a channel-by-channel basis, and then fuses the channels through pointwise convolution to reduce the number of parameters and multiply-accumulate operations.

[0121]

[0122] In a steady state, the threshold gate or learning gate can be used to reduce [the risk / damping]. The gain is increased when the temperature change, RF switch state change, group delay decoupling characteristic derivative, or historical residual derivative exceeds the threshold. This structure enables the cyclic state to handle short-term memory and the dilated convolution branch to handle fast mutations, avoiding a single timing module handling both time scales simultaneously.

[0123] S107, Execute state-conditional sparse hybrid expert routing. The routing module calculates the activation probability of each candidate expert based on the auxiliary state vector and sets the hard path mask according to the RF switch state and port topology:

[0124]

[0125] Hard path mask It can be directly generated from the device topology table. For example, when a certain RF switch state disconnects a specific branch, the candidate expert mask corresponding to that branch is set to zero; the mask of the backup unknown source expert can always be kept at one to ensure that at any given time there is at least one available expert. Routing weights can also perform Top-... Truncate, keeping only the largest Each weight is assigned a weight, and the retained weights are renormalized, where... .

[0126]

[0127]

[0128] The digital cancellation module can be executed after receiving digital down-conversion, before channel estimation, or at other processing points that will not disrupt synchronization. Cancellation gain. It can be updated by sampling block; it approaches one when the model confidence is high and there is no expected uplink signal, and decreases to zero or an intermediate value when the model confidence is low or an expected uplink signal is detected.

[0129] After training, the model can be evaluated across different frequency band combinations, power ranges, temperature stages, RF switch states, device samples, and port combinations. The normalized mean square error can be expressed as:

[0130]

[0131] in, The numerator represents the normalized mean square error, expressed in decibels; the numerator represents the error energy between the internal passive intermodulation monitoring target and the overall prediction; and the denominator represents the monitoring target energy. In the above formula, This represents the logarithmic function to the base 10. Furthermore, Indicates the number of evaluation samples or sampling blocks. This indicates the target of passive intermodulation monitoring within the device. This represents the L2 norm.

[0132] The target uplink band passive intermodulation cancellation ratio can be expressed as:

[0133]

[0134] in, Indicates the passive intermodulation cancellation ratio; Represents the set of discrete frequency points in the target uplink frequency band; and These represent the received signals before and after cancellation at the [number]th [time]. Frequency domain values ​​at each frequency point.

[0135] For the risk of exceeding limits during temperature cycling, the following formula can be used:

[0136]

[0137] in, This indicates the probability that the passive intermodulation power exceeds a threshold. This represents a random variable representing the power of passive intermodulation. Indicates exceeding the threshold; This represents the probability density function of passive intermodulation power within the corresponding temperature range. In the above formula, This represents the integral infinitesimal element of the passive intermodulation power random variable.

[0138] Evaluation metrics can also include the improvement in receiver sensitivity, recovery time after RF switch switching, peak residual in the abrupt change region, migration error between different device samples, number of parameters, number of floating-point operations per second, and inference delay. Model selection is primarily based on the target uplink band residual power and the worst-case normalized mean square error of the receiver port, rather than solely on the average error across the entire bandwidth.

[0139] To verify the correspondence between each module and the technical pain points, ablation experiments can be performed: removing the device-sensing complex finite impulse response filter to evaluate the group delay decoupling effect; removing the odd-symmetric static nonlinear module to evaluate the physical prior and parameter sharing effects; removing the generalized memory polynomial compensation vector to evaluate the higher-order feature effects; replacing the state-conditional nonlinear compensation long short-term memory module with a common long short-term memory network to evaluate the state-conditional and compensation gate effects; removing the causal dilated convolutional mutation branch to evaluate the mutation response; and removing the auxiliary state vector or sparse routing to evaluate the cross-working state and multi-source modeling capabilities.

[0140] Although the present invention has been described in conjunction with specific embodiments, those skilled in the art can make modifications, substitutions, combinations, or variations to the above embodiments without departing from the spirit and scope of the invention. This specification and accompanying drawings are for illustrative purposes only, and the scope of protection of the present invention should be determined by the claims.

Claims

1. A method for canceling state-conditional timing networks of passive intermodulation within a device, characterized in that, A state-conditional symmetric nonlinear compensation timing network is constructed, comprising a multi-reference selection and group delay decoupling module, an odd-symmetric static nonlinear module, a state-conditional nonlinear compensation long short-term memory module, a causal dilated convolutional mutation branch, and a state-conditional sparse hybrid expert routing module. The state-conditional symmetric nonlinear compensation timing network is trained, and the trained network is used to predict and cancel passive intermodulation within devices caused by at least one of duplexers, RF switches, and feed connection structures. Specifically, the following steps are included: S1, Obtain Ludo reference complex baseband signal Receive complex baseband signals and auxiliary state vector ,in, Indicates the index of the sampling point or sampling block. Indicates the reference signal index. Indicates the number of reference signals; S2. In the multi-reference selection and group delay decoupling module, for Each candidate internal passive intermodulation expert, based on the auxiliary state vector Calculate the multi-reference selection weights, perform weighted combination of the multi-reference complex baseband signals, and use a device-sensing complex finite impulse response filter to perform linear group delay decoupling on the weighted combination signals to obtain the first... Group delay decoupling characteristics of candidate internal passive intermodulation experts ,in, This indicates the number of experts who are passively interleaved within the candidate pool; S3. Decouple the group delay feature Input an odd-symmetric static nonlinear module that satisfies the odd mapping constraint, and obtain the static nonlinear characteristics. And based on the group delay decoupling characteristics, a generalized memory polynomial basis function vector is constructed. ; S4. Decouple the group delay feature The static nonlinear characteristics The generalized memory polynomial basis function vector and the auxiliary state vector Input state condition nonlinear compensation long short-term memory module, recursively obtain the first Unit status of each candidate internal passive inter-modulation expert and hidden state ; S5. Based on the changes in the auxiliary state vector and the group delay decoupling feature, as well as the cancellation residual signal from the previous time step, a mutation correction vector is generated using the causal dilated convolution mutation branch. And by fusing the hidden state, the static nonlinear feature, and the mutation correction vector, the first... Predictions from individual candidate internal passive intermodulation experts ; S6. In the state-conditional sparse hybrid expert routing module, calculate the state-conditional sparse hybrid expert weights based on the auxiliary state vector. The predictions of the passive inter-modulation experts within each candidate are weighted and synthesized, and then canceled out in the digital receiver chain: ， ， in, This indicates the total prediction of passive intermodulation within the device. and , This indicates the canceling gain controlled by the protection mechanism. This represents the received residual signal after cancellation; S7. Construct a training loss function based on the passive intermodulation supervision target and / or the residual power in the uplink frequency band of the target, and optimize the model parameters of the state-conditional symmetric nonlinear compensation time-series network; deploy the trained network on the digital receiving and processing link of the multi-band frequency division duplex radio equipment, and update the selected model parameters online within the time slot that meets the uplink protection conditions.

2. The method for canceling state-condition timing networks of passive intermodulation within a device according to claim 1, characterized in that, The multi-reference complex baseband signal in S1 includes two or more of the following: a complex baseband signal located before digital predistortion processing and before or after peak-to-average power ratio reduction processing; a complex baseband signal located after digital predistortion processing and before digital-to-analog conversion; a complex envelope signal acquired by the transmit coupler; and a complex envelope signal reconstructed from forward power monitoring data. The auxiliary state vector is constructed as follows: ， in, Indicates temperature. Indicates the amount of temperature change. Indicates forward power. Indicates the status of the radio frequency switch. Indicates frequency band coding. Indicates the receive port encoding. Indicates calibration status. This indicates transpose.

3. A method for canceling state-condition timing networks of passive intermodulation within a device according to claim 1 or 2, characterized in that, The multi-reference selection weights, weighted combined signals, and device-sensing complex finite impulse response filtering of the multi-reference selection and group delay decoupling module in S2 are calculated as follows: ， ， ， ， ， in, Indicates the first The candidate internal passive inter-adjustment expert on the first The selection weight of the path reference signal; Indicates the index for summing the reference signal; , , and These are learnable parameters; Indicates the first A weighted combination of signals from candidate internal passive intermodulation experts; Indicates the finite impulse response filter length; Indicates the filter tap index; Represents the state condition complex filtering matrix; Represents the fundamental complex filtering matrix; Indicates the first A low-rank state-modulated complex filter matrix; Indicates the number of low-rank state modulation terms; Represents the state modulation coefficient; Represents the Sigmoid function; This represents an exponential function.

4. The method for canceling state-condition timing networks of passive intermodulation within a device according to claim 3, characterized in that, The odd-symmetric static nonlinear module in S3 provides real characteristics for in-phase components, quadrature components, and / or amplitude coupling. Construct symmetric basis functions and obtain the static nonlinear characteristics as follows: ， ， ， ， in, Indicates the first Symmetric basis functions Represents the linear rectified function. Indicates the first A basis function breakpoint, Indicates the number of basis functions. Indicates the first Basis function weights of each candidate internal passive inter-modulation expert. Represents a real odd symmetric mapping. and They represent complex numbers respectively The real and imaginary parts, Represents the imaginary unit. Indicates the first Complex odd-symmetric mapping of candidate internal passive intermodulation experts; for multiple receiver ports belonging to the same device type or the same physical source, the basis function breakpoints are shared. and the basis function weights And retain the port-related linear input mapping and output mapping respectively; the generalized memory polynomial basis function vector and the nonlinear compensation vector for the long short-term memory module used for state-condition nonlinear compensation are calculated as follows: ， ， ， in, This represents a complex channel component representing the group delay decoupling characteristic; Represents the nonlinear order. It represents the set of orders consisting of at least one odd-numbered order among the third, fifth, and seventh orders; This indicates a delay in main memory; Indicates envelope delay; Indicates the depth of primary memory; Indicates the maximum envelope delay; Represents the basis functions of the generalized memoization polynomial; This represents vectorized concatenation; and These are learnable parameters; Indicates the first Nonlinear compensation vectors for candidate internal passive intermodulation experts; Represents the modulus of a complex number; This represents the hyperbolic tangent function.

5. The method for canceling state-condition timing networks of passive intermodulation within a device according to claim 4, characterized in that, The state-conditional nonlinear compensation long short-term memory module in S4 first constructs the input features and calculates the hybrid gate, candidate cell state, compensation gate, cell state, output gate, and hidden state as follows: ， ， ， ， ， ， ， in, This represents the input feature formed by concatenating the group delay decoupling feature and the static nonlinear feature; A gate that represents a mixture of historical states and current candidate states; Indicates the candidate cell status; Represents a nonlinear compensation gate; This represents the nonlinear compensation vector as described in claim 4; Indicates the output gate; Represents a vector that is all one; This represents element-wise multiplication; , , , And the corresponding bias is a learnable parameter; vectors connected by semicolons or commas within square brackets represent concatenation.

6. A method for canceling state-condition timing networks of passive intermodulation within a device according to claim 5, characterized in that, The predictions of the causal dilated convolutional mutation branch and the candidate internal passive intermodulation expert in S5 are calculated as follows: ， ， ， ， in, and These represent the changes in the auxiliary state vector and the changes in the group delay decoupling characteristics, respectively. This represents the cancellation residual signal from the previous time step; Indicates the input of the mutation branch; Indicates the length of the causal observation window for the mutated branch; Indicates the first The causal dilation depth of each candidate internal passive intermodulation expert can be separable convolution operation; Indicates the number of convolutional layers; Indicates the dilation rate of each convolutional layer; , and The learnable output mapping matrix is ​​represented; the causal dilated convolutional mutation branch increases the mutation correction gain when the temperature change, the RF switch state change, or the residual derivative exceeds the corresponding threshold.

7. The method for canceling state-condition timing networks of passive intermodulation within a device according to claim 6, characterized in that, The state-condition sparse hybrid expert weights in S6 are calculated as follows: ， in, This represents a hard path mask determined by known device paths, RF switch states, and port connection relationships, and at least one such mask exists at any given time. ; This represents the index for summing passive inter-invocation experts within the candidate pool; and These are learnable routing parameters; Represents the routing weight vector; apply low-entropy constraints and Top-to-Top constraints to the routing weight vector. Select constraints and / or expert group sparse constraints so that each working state activates only a subset of the candidate internal passive inter-invocation experts, where... This indicates the number of candidate experts retained, and .

8. The method for canceling state-condition timing networks of passive intermodulation within a device according to claim 7, characterized in that, The training loss function in S7 includes one or more of the following: temporal complex loss, target uplink frequency band spectrum loss, abrupt weighting loss, route sparsity and smoothing loss, cell state smoothing loss, and parameter regularization term, calculated as follows: ， ， ， ， ， ， ， ， ， in, , , , and These represent time-domain complex loss, target uplink frequency band spectrum loss, abrupt weighting loss, route sparsity and smoothing loss, and cell state smoothing loss, respectively. Indicates the total training loss; This refers to the internal passive intermodulation monitoring target of the device obtained through uplink silent sampling, known calibration sequence, desired uplink signal stripping, or external measurement. Indicates the length of the training sequence; Represents the Discrete Fourier Transform; Represents the target uplink frequency band set; Indicates the first Frequency domain mask for the uplink frequency band of each target; Indicates frequency band weight; , and Indicates the mutation weighting coefficient; Indicates characteristic functions; ; Represents routing entropy; Represents a positive number to prevent abnormal values ​​in logarithms or divisions; Indicates a stable operating status indicator; Indicates the route smoothing coefficient; This represents all parameters to be optimized; , , , , and Non-negative loss weights; Indicates the target uplink frequency band index. Indicates the prediction error of complex numbers. Indicates the weight of the mutated sample. and Let them represent the 2-norm and the Frobenius norm, respectively. It represents the natural logarithm.

9. The method for canceling state-condition timing networks of passive intermodulation within a device according to claim 8, characterized in that, The training and online update sequentially include static device pre-training, timing memory training, mutation branch training, device-level fine-tuning, and protected online adaptation. The device-level fine-tuning and protected online adaptation only update one or more of the low-rank state modulation complex filter parameters, routing parameters, output mapping parameters, and mutation branch final layer parameters, and employ at least one of the following protection measures: freezing the online update upon detection of a desired uplink signal; adjusting the cancellation gain based on the correlation between the device's internal passive intermodulation total prediction and the received complex baseband signal. Limit cancellation amplitude, phase change rate, and parameter update step size; fall back to the most recently stable model when online loss increases continuously; set a freeze window or realign window after RF switch state switching.

Images