An AI-based method for evaluating the bit error rate of backscattering systems in the environment.

Abstract

This invention, belonging to the field of data processing methods, describes an artificial intelligence-based method for evaluating the bit error rate (BER) of environmental backscattering systems. The method includes the following steps: S1, data acquisition and labeling; S2, data augmentation using the SMOTE algorithm based on quantum state evolution; S3, feature extraction model training using a neural network algorithm optimized by light and shadow tracking; S4, feature dimensionality reduction using an improved autoencoder neural network; S5, classifier model training using a random forest algorithm based on steady-state redundancy removal; and S6, BER classification and evaluation of environmental backscattering systems. This invention not only increases sample diversity but also maintains high information fidelity, thereby improving the model's training effect and final prediction accuracy; it avoids getting trapped in local optima, enhancing the model's adaptability to new data; it reduces the risk of overfitting and improves the model's predictive ability for unknown data.

Description

Technical Field

[0001] This invention belongs to the field of data processing methods, specifically relating to an artificial intelligence-based method for evaluating the bit error rate of environmental backscattering systems. Background Technology

[0002] Environmental backscattering systems are a communication technology that utilizes existing scattering objects in the environment to propagate wireless signals. They can achieve effective communication connections in complex environments or special scenarios. However, in practical applications, this system is susceptible to various factors, such as signal attenuation, multipath effects, and environmental changes. These factors can all lead to bit errors during communication, affecting its stability and reliability.

[0003] Traditional methods for assessing bit error rate (BER) in communications often rely on simplified channel models and theoretical analysis. These methods may fail to accurately capture and assess the actual occurrence of BER in complex and ever-changing real-world communication environments. Furthermore, traditional methods suffer from time-consuming and labor-intensive data acquisition, labeling, and preprocessing processes, insufficient training samples, and poor model generalization ability, all of which limit the accuracy and practicality of BER assessment.

[0004] In response to this, Chinese invention patent application CN202210567236.8 proposes a performance evaluation device for a telemetry, tracking, and command (TT&C) communication unit, comprising a signal generation module, a signal processing module, and a test evaluation module, all connected to the TT&C communication unit. The signal generation module generates channel simulation parameters based on imported orbit data, generates uplink radio frequency (RF) signals based on the channel simulation parameters, and sends the uplink RF signals to the TT&C communication unit. The signal processing module receives downlink RF signals from the TT&C communication unit, parses the downlink RF signals, and obtains downlink test data. The test evaluation module calculates the performance parameters of the TT&C communication unit in its on-orbit state based on the downlink RF signals or downlink test data, and evaluates the orbit determination accuracy of the TT&C communication unit. This application's device integrates TT&C testing and channel simulation functions, improving testing efficiency, facilitating operation, and providing different test evaluation scenarios according to testing needs. This application also discloses a corresponding method for evaluating the performance of a TT&C communication unit. Chinese invention patent application CN202311029523.4 proposes a method for optimizing the performance of stochastic resonance communication signal processing based on bit error rate. At the signal receiver, a stochastic resonance system is used to denoise the received signal containing random noise, and the result is compared with the received signal before and after denoising. The bit error rate of the received signal before and after processing is calculated. The effectiveness of stochastic resonance processing under different initial signal-to-noise ratios is evaluated, and further optimization steps are set to comprehensively select the optimal result from the two received signals before and after stochastic resonance processing as the final received signal. This invention utilizes a bistable stochastic resonance system to process a QPSK signal mixed with Gaussian white noise, evaluates the improvement in bit error rate after demodulation of the transmitted signal at the signal receiver, and corrects for potential side effects of stochastic resonance. Chinese invention patent application number CN202010819436.9 proposes an evaluation method for the operational indicators of a quantum communication network based on the entropy method, including the following steps: Step 1: Obtain the operational data of each QKD network node at each acquisition interval t from day s-k+1 to day s, as the statistical basis data for day s, where k is the number of days in the evaluation period; Step 2: Calculate the anomaly index of each QKD network node based on the statistical basis data of day s; Step 3: Normalize each anomaly index of each QKD network node calculated in Step 2; Step 4: Calculate the anomaly weight value of each anomaly index of each QKD network node after normalization; Step 5: Calculate the entropy value and difference factor corresponding to each anomaly index; Step 6: Calculate the final evaluation weight value of each anomaly index based on the difference factor; Step 7: Calculate the comprehensive evaluation value of each QKD network node, and use the magnitude of the comprehensive evaluation value as the measure of the priority of anomaly handling for that QKD network node. The method of this invention overcomes the limitations of traditional methods that only evaluate abnormal states at a certain moment in the operation of quantum networks, and is more able to reflect the ever-changing operating states of quantum networks.

[0005] Research has revealed the following issues that require further resolution regarding the aforementioned technologies: 1. Existing data generation techniques may lack refinement in processing high-dimensional data, failing to effectively control sample diversity and authenticity, thus limiting model training effectiveness and generalization ability. 2. Existing technologies may not effectively capture dynamically changing data features and are prone to getting trapped in local optima during training, limiting the model's performance in practical applications. 3. When performing data dimensionality reduction, traditional methods may fail to effectively maintain the structural integrity of the original data, leading to reduced accuracy in model reconstruction or data interpretation. 4. Existing models may exhibit overfitting when faced with complex or unseen data, lacking sufficient generalization ability and limiting their application effectiveness in dynamic environments. Summary of the Invention

[0006] The purpose of this invention is to propose an AI-based method for evaluating the bit error rate of environmental backscattering systems. This method not only increases the diversity of samples but also maintains high information fidelity, thereby improving the training effect and final prediction accuracy of the model; it avoids getting trapped in local optima, enhances the model's adaptability to new data, reduces the risk of overfitting, and improves the model's ability to predict unknown data.

[0007] The technical solution adopted in this invention is:

[0008] An AI-based method for evaluating the bit error rate of backscattered environmental communication systems includes the following steps:

[0009] S1. Data Acquisition and Labeling

[0010] The data collected by this invention comes from the communication nodes in the environmental backscattering system. Communication error events are monitored and recorded in real time through a dedicated hardware interface. The collected data is stored in a structured vector data format, with each data point including multiple attribute fields.

[0011] The collected data is manually labeled, and the labeling categories include: no bit error, single bit error, multi-bit error, parity error, synchronization failure error, and other bit error.

[0012] S2, Data Expansion

[0013] In this invention, the acquisition, annotation, and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor model generalization ability and affect model accuracy. This invention employs the SMOTE (Synthetic Minority Over-sampling Technique) algorithm based on quantum state evolution to generate samples, thereby achieving data augmentation. While traditional SMOTE algorithms balance class distribution by interpolating new sample points between minority class samples, this invention utilizes quantum state evolution theory and the principles of quantum state superposition and interference to allow for more precise control of the sample generation process in high-dimensional data space, optimizing sample diversity and information fidelity during data augmentation.

[0014] The process of sample generation based on the SMOTE algorithm of quantum state evolution is as follows:

[0015] S201. Separate minority class samples from the labeled training set. Based on quantum state evolution theory, calculate the quantum state representation of each minority class sample point, b. i Mapped to a quantum state ψ i , represented as:

[0016]

[0017] In the formula, |k> represents the ground state of the qubit, and α ik It is the complex probability magnitude, specifically for sample b. i The probability amplitude in the ground state |k> satisfies the normalization condition. K is the dimension of the quantum state space;

[0018] S202. Select sample pairs that are close to each other and perform quantum state superposition operations to simulate new sample points using the principle of quantum state interference; where, sample pairs (b i ,b j Perform a superposition operation on quantum states to generate a new quantum state ψ. ij , represented as:

[0019]

[0020] In the formula, ψ i and ψ j Sample b i and b j The quantum state representation; |||| is the normalization operator, specifically the L1 norm normalization, ||ψ i +ψ j || represents the new state ψ generated by the superposition of quantum states. ijStandardized processing and standardized operation ensure ψ ij It remains a valid quantum state; E() is the quantum entanglement operation function;

[0021] The quantum entanglement operation function E() for sample b i and b j quantum state ψ i and ψ j The calculation method using entanglement operations is expressed as follows:

[0022]

[0023] In the formula, Tensor products representing quantum states;

[0024] ||ψ i +ψ j The calculation method for || is expressed as follows:

[0025]

[0026] In the formula, |α ik +α jk | 2 The calculation is the squared magnitude of the probability amplitude of the new superposition state in the ground state |k>;

[0027] S203. The newly generated sample undergoes a quantum measurement process to ensure its validity and rationality in the classical data space, while maintaining the diversity and complexity of the data. The quantum measurement process is used to convert the quantum state ψ... ij Convert back to classic sample b new The method is expressed as:

[0028] b new =measure(ψ) new )

[0029]

[0030] In the formula, measure(ψ) is the quantum measurement function, and ψ is the input to the quantum measurement function; β k Through quantum state ψ ij The probability obtained from the measurement is specifically β. k In the quantum state ψ ij After measurement, the probability of the ground state |k> occurring determines the values ​​of the new sample in each dimension; ψ new The quantum state of the heart sample;

[0031] ψ new The calculation method is expressed as follows:

[0032]

[0033] In the formula, <ψ ij |ψ ij > represents ψ ij The inner product; G() is the dynamic adaptive quantum gate operation function;

[0034] S204, Newly generated sample b new After a verification and adjustment process to ensure its adaptability and diversity in data distribution, it is represented as:

[0035]

[0036] In the formula, λ vf To adjust the factor, controlling how closely the new sample resembles the original sample pair; Median operation representing sample features; Indicates sample b i and b j The eigenmedian;

[0037] The calculation method is expressed as follows:

[0038]

[0039] In the formula, The calculation is for sample b. i and b j The arithmetic mean of the two characteristics represents the midpoint of their features;

[0040] S3, Feature Extraction Model Training

[0041] The expanded data is input into the feature extraction model for training. This invention adopts a neural network structure optimized based on light and shadow tracking, which uses the pattern of light and shadow changes to capture and extract dynamic features in the input data, and accurately extracts communication error features. This invention adopts a loop restart strategy, which allows the model to reset network parameters according to preset conditions or performance thresholds during training, thereby avoiding getting stuck in local optima and enhancing the model's adaptability to new data.

[0042] The training process of the neural network algorithm optimized based on light and shadow tracking is as follows:

[0043] S301. Set the initial architecture of the neural network, including an input layer, two hidden layers, and an output layer; during model initialization, the network parameters θ include weights w and biases b, and the parameter initialization is expressed as follows:

[0044]

[0045] In the formula, l represents the network level, i and j represent the neuron indices of that layer, and σ 2It is the variance of the initial weights. The weights of the l-th layer are... For the bias of the l-th layer, This indicates a range with a mean of 0 and a variance of σ. 2 Sampling from a normal distribution;

[0046] S302. Using the augmented training data, calculate the output of each layer through forward propagation, and then adjust the network parameters through backpropagation to minimize the output error. The network output y and the target y are then compared. * The error between them is calculated using a loss function:

[0047]

[0048] Furthermore, during the parameter update process, the method of backpropagating to update the weights is represented as follows:

[0049]

[0050] In the formula, and The weights and biases before the update. and Here are the updated weights and biases, and y is the network output. * For the real goal, and These are the partial derivatives of the loss function with respect to the weights and biases of the l-th layer, respectively, and I(W,b) is the update impact factor.

[0051] S303. At the end of each training cycle, evaluate the model's performance. If the model's performance does not reach the predetermined improvement standard or training progress stagnates, adjust the learning rate or reinitialize the network parameters according to the cycle restart strategy. The cycle restart strategy is implemented by dynamically adjusting the learning rate η, based on the changes in the loss function L during training, and is thus expressed as:

[0052] η new =η×λ vg t / τ

[0053] In the formula, λ vg η is the decay factor, t is the current training cycle, τ is the decay time step, and η is the decay factor. new The updated learning rate;

[0054] Attenuation factor λ vg The calculation method is expressed as follows:

[0055] λ vg =e -α

[0056] In the formula, α is a parameter based on the growth rate of the model performance evaluation index, and its calculation method is expressed as:

[0057]

[0058] In the formula, ΔL represents the change in the loss function over two consecutive training periods, and L is the loss function value for the current period.

[0059] S304. Repeat steps S301-S303 until the preset stopping iteration condition is met, which means that the model training is complete.

[0060] S4: Feature Dimensionality Reduction Model Training

[0061] The feature vectors obtained after feature extraction are input into a feature dimensionality reduction model for training. This invention employs an improved autoencoder neural network for feature dimensionality reduction. The autoencoder neural network consists of an encoder and a decoder. The encoder is responsible for converting the input data into low-dimensional feature representations, while the decoder attempts to reconstruct the original data from these low-dimensional representations. Addressing the problem of traditional autoencoders' poor performance when handling complex data structures, this invention adopts a manifold learning strategy. The autoencoder can better understand and compress the inherent structure of the data during the encoding stage, rather than simply reducing data dimensionality, which helps to reconstruct more accurate original data during the decoding stage.

[0062] The training process for the improved autoencoder neural network is as follows:

[0063] S401. Initialize the parameters of the autoencoder neural network using a Gaussian distribution strategy, where each parameter p... θ The initialization method is expressed as follows:

[0064]

[0065] In the formula, p θ For network parameters; This indicates that the mean is 0 and the variance is σ. 2 The normal distribution; σ is the initial standard deviation, which controls the degree of dispersion of the initial parameters;

[0066] S402. The input data is propagated forward through the encoder. A manifold learning method is used to compress features while preserving the local structure of the data. Specifically, the input data x is converted into a low-dimensional representation z, as follows:

[0067] z = Sig(W) e x+b e )

[0068] In the formula, x represents the input data; W e b is the weight matrix of the encoder;e Here is the encoder's bias vector; Sig() is the Sigmoid activation function.

[0069] The Sigmoid function is calculated as follows:

[0070]

[0071] In the formula, v is the input to the activation function;

[0072] S403. The compressed features are reconstructed by the decoder in an attempt to restore their original data form. During this process, the mean squared error loss function is used to constrain the reconstruction process. The reconstruction operation is represented as follows:

[0073]

[0074] In the formula, To reconstruct the data; z represents the encoded low-dimensional feature; W d b is the weight matrix of the decoder; d This is the bias vector for the decoder;

[0075] S404. After each round of training, a uniformization strategy is used to adjust the weights of each feature dimension. The weights W and bias b are adjusted to uniformly distribute the feature information, as follows:

[0076]

[0077] In the formula, ||W|| F is the Frobenius norm of W, used to normalize the weights; ||b||1 is the 1-norm of b, used to normalize the biases;

[0078] Frobenius norm ||W|| F The calculation method is expressed as follows:

[0079]

[0080] In the formula, W ij It is the element in the i-th row and j-th column of matrix W;

[0081] The 1-norm ||b||1 is expressed as follows:

[0082]

[0083] In the formula, b i It is the i-th element in vector b;

[0084] S405. Calculate the gradient based on the loss function and update the network parameters using gradient descent, as follows:

[0085]

[0086] In the formula, θ represents the network parameters, including W e ,b e W d ,b d η bv It is the learning rate; It is the gradient of the loss function L with respect to θ, where L is the mean squared error.

[0087] gradient of loss function The calculation method is expressed as follows:

[0088]

[0089] S406. Repeat steps S401-S405 until the preset stopping iteration condition is met, which means that the model training is complete.

[0090] S5: Classifier Model Training

[0091] The dimensionality-reduced data is input into the classifier for training the neutral classifier model. This invention employs a random forest algorithm based on steady-state redundancy removal. This algorithm comprises multiple decision trees, each independently constructed during training. The final classification result is determined through a voting mechanism during classification decisions. To enhance the model's generalization ability and reduce the risk of overfitting, this invention dynamically removes trees that provide repetitive information by evaluating the similarity between decision trees, retaining decision trees with high diversity.

[0092] The training process of the random forest algorithm based on steady-state redundancy removal is as follows:

[0093] S501. Assume a random forest consists of N independent decision trees, each tree having a value of T. i During initialization, samples and features are randomly selected. The initialization process is represented as follows:

[0094] T i =initialize(Data,F qi ,S qi ), i = 1, 2, ..., N

[0095] In the formula, initialize() is the initialization function; T i Let F be the i-th tree; Data is the training dataset; F qi S is a subset of features randomly selected from the feature set; qi This refers to a subset of samples randomly selected from the sample.

[0096] S502. For each tree, train using the topologically transformed data. The topological transformation is achieved through principal component analysis to make the data suitable for decision tree splitting, as shown below:

[0097]

[0098] In the formula, X qi Selected sample data; PCA() is the PCA function; X′ qi The data after PCA transformation;

[0099] S503. After the tree is constructed, calculate the redundancy between the trees; if a high redundancy relationship is found, remove the tree with the highest redundancy and only retain the tree that provides independent information. Let Sim(T) = ... i ,T j The function ) represents the similarity calculation function between two trees, and the calculation method is expressed as follows:

[0100]

[0101] In the formula, X is the dataset used for evaluation. This is an indicator function; if the two trees classify the same data point x, the function value is 1, otherwise it is 0.

[0102] S504. Each tree is trained independently, and each tree is trained using the optimal split point to ensure that the decision boundary can be learned from the training data. The selection of the split point is based on maximizing the gain, expressed as:

[0103]

[0104] In the formula, InfoGain(s) represents the gain at the split point s;

[0105] The gain InfoGain(s) is calculated as follows:

[0106]

[0107] In the formula, H(D) is the entropy of the dataset D, and D left (s) and D right (s) represent the left and right subsets of the dataset after the split point s;

[0108] The entropy H(D) is calculated as follows:

[0109]

[0110] In the formula, p k It represents the probability of category k in dataset D;

[0111] S505. The outputs of all decision trees are merged through a voting mechanism. Each tree provides a prediction result for an input sample. The final result is the category with the most votes, represented as:

[0112]

[0113] In the formula, c i C is the classification result of the i-th tree; C is the final classification result, where c represents the possible categories. The indicator function is defined if the output c of the i-th tree... i If the value equals category c, the function value is 1; otherwise, it is 0.

[0114] S506. During model training, the number and depth of decision trees, as well as the parameters of topology transformation, are dynamically adjusted based on feedback from validation data to optimize model performance. The adjustment methods for the number of trees N and the number of PCA components k are expressed as follows:

[0115] N new =N+ΔN

[0116] k new =k+Δk

[0117] In the formula, ΔN and Δk are the number of adjustments made based on the model performance, respectively;

[0118] S6: Classification and Assessment of Bit Error Rate in Environmental Backscattering Systems

[0119] The newly collected environmental backscattering system data is processed using pre-trained feature extraction models, data dimensionality reduction models, and classifier models to evaluate the communication bit error rate.

[0120] Further, in step S201, the complex probability amplitude α ik Based on sample b i The eigenvalues ​​are calculated and expressed as:

[0121]

[0122] In the formula, μ k The center position of the k-th quantum ground state is represented, specifically the characteristic center of the quantum ground state |k>; γ is a hyperparameter controlling the scale; ||b i -μ k || 2 Indicates sample b i With ground state center μ k The square of the Euclidean distance between them.

[0123] Furthermore, in step S203, the dynamic adaptive quantum gate G() depends on the quantum operation of the data characteristics, and the dynamic adaptive quantum gate G() acts on a single quantum state ψ.i or quantum state pair ψ i ,ψ j Above, to adjust the quantum characteristics, it is expressed as:

[0124] G(ψ i )=U(θ i )ψ i

[0125]

[0126] In the formula, U(θ i It depends on the parameter θ. i The unitary transform, where H is the corresponding Hamiltonian operator, θ i These parameters are automatically learned from data features and are solved through the following optimization problem:

[0127]

[0128] In the formula, L() is the cross-entropy loss function, f() represents the quantum measurement function with parameter θ, and y i It is sample b i The target value;

[0129] The entire process is repeated until the predetermined amount of data or sample distribution balance is reached.

[0130] Furthermore, in step S302, when calculating the updated influence factor I(W,b), the importance weight of each parameter needs to be dynamically evaluated and adjusted in each iteration to further identify the most critical parameters for the current training stage and assign higher priority to these parameters, thereby improving the overall efficiency and effectiveness of network training. The updated influence factor I is defined as:

[0131] I(W,b)=tanh(ν ps ·C(W,b)+μ ps ·S(W,b))

[0132] In the formula, tanh() is the hyperbolic tangent function, and ν ps and μ ps These are the coefficients for adjusting and updating the impact factor, where C(W,b) is the parameter contribution and S(W,b) is the parameter sensitivity.

[0133] The parameter contribution C(W,b) is calculated as follows:

[0134]

[0135] In the formula, ⊙ denotes element-wise multiplication. This represents the partial derivative of the current network weights with respect to the loss function. Let W represent the partial derivative of the current network bias with respect to the loss function, W represent the current network weights, and b represent the current network bias.

[0136] The parameter sensitivity S(W,b) is calculated as follows:

[0137] S(W, b) = exp(-γ ps ||C(W,b)||)

[0138] In the formula, exp() represents the exponential function, and γ ps is the sensitivity adjustment coefficient, and |||| is the L1 norm normalization operation.

[0139] Further, in step S402, the weight matrix W e The update method is represented as:

[0140]

[0141] In the formula, The updated weight matrix, The weight matrix before the update is η. ry L is the learning rate, and L is the loss function; for W e partial derivatives Represented as:

[0142]

[0143] The beneficial effects of this invention are:

[0144] 1. New samples are generated using the SMOTE algorithm based on quantum state evolution. By using the superposition and interference principles of quantum states, the generation of samples is controlled in the high-dimensional data space to increase the diversity and information fidelity of the data.

[0145] 2. By using a neural network structure optimized based on light and shadow tracking, dynamic features in the input data can be captured, and a loop restart strategy can be used to avoid the model getting stuck in local optima.

[0146] 3. Using an improved autoencoder neural network, the data is compressed during the encoding stage through a manifold learning strategy, and the original data is reconstructed during the decoding stage to maintain the inherent integrity of the data structure.

[0147] 4. A random forest algorithm based on steady-state redundancy removal is used to enhance the model's generalization ability and reduce the risk of overfitting by evaluating the similarity between decision trees and removing trees with duplicate information.

[0148] Based on the above innovations, the resulting technical effects are as follows:

[0149] 1. The SMOTE algorithm based on quantum state evolution is used to generate new samples. By using the superposition and interference principle of quantum states, the generation of samples is controlled in the high-dimensional data space, which not only increases the diversity of samples, but also maintains a high degree of information fidelity, thereby improving the training effect of the model and the final prediction accuracy.

[0150] 2. By adopting a neural network structure optimized based on light and shadow tracking, dynamic features in the input data can be effectively captured. Combined with the loop restart strategy, the network parameters can be reset during the training process according to the performance threshold or preset conditions, thereby avoiding getting stuck in local optima and enhancing the model's adaptability to new data.

[0151] 3. The improved autoencoder neural network compresses data through a manifold learning strategy during the encoding stage and attempts to reconstruct the original data during the decoding stage. This helps to maintain the inherent integrity of the data structure while reducing data dimensionality, which is beneficial for the model to understand and process complex data more accurately.

[0152] 4. Enhance the generalization ability of the model by using a random forest algorithm based on steady-state redundancy removal, evaluate the similarity between decision trees, remove trees that provide duplicate information, and retain only decision trees with high diversity, thereby reducing the risk of overfitting and improving the model's ability to predict unknown data. Attached Figure Description

[0153] Figure 1 This is a flowchart of the evaluation method of the present invention;

[0154] Figure 2 This is a flowchart of the feature extraction model training process;

[0155] Figure 3 This is a schematic diagram of the training and application process of the present invention. Detailed Implementation

[0156] like Figure 1 As shown, the method for evaluating the bit error rate of an environment backscattering system based on artificial intelligence includes the following steps:

[0157] S1. Data Acquisition and Labeling

[0158] The data collected in this invention comes from the communication nodes in the environmental backscattering system. Communication error events are monitored and recorded in real time through a specific hardware interface. The collected data is stored in a structured vector data format, with each data point including multiple attribute fields.

[0159] In this embodiment, the attributes of the data include:

[0160] ax: Signal strength, measured during communication;

[0161] ay: Signal-to-noise ratio, the ratio of signal strength to background noise;

[0162] az: Frequency offset, the difference between the communication frequency and the standard frequency;

[0163] aa: timestamp, the specific time the data was collected;

[0164] ab: Ambient temperature, the external temperature at which data was collected;

[0165] ac: humidity, the relative humidity of the environment;

[0166] ad: Location information, the geographical coordinates of the communication device;

[0167] ae: Device type, the model of the device involved in the communication;

[0168] af: Operating system, the version of the operating system running on the device;

[0169] ag: Transmission rate, the speed at which data is transmitted.

[0170] It should be emphasized that this embodiment is only to illustrate one data format and type of the present invention. In practical applications, the data usually has more than 10 attributes, and the number of data attributes may reach dozens or even hundreds.

[0171] The collected data is manually labeled, and the labeling categories include: no bit error, single bit error, multi-bit error, parity error, synchronization failure error, and other bit error.

[0172] S2, Data Expansion

[0173] In this invention, the acquisition, annotation, and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor model generalization ability and affect model accuracy. This invention employs the SMOTE (Synthetic Minority Over-sampling Technique) algorithm based on quantum state evolution to generate samples, thereby achieving data augmentation. While traditional SMOTE algorithms balance class distribution by interpolating new sample points between minority class samples, this invention utilizes quantum state evolution theory and the principles of quantum state superposition and interference to allow for more precise control of the sample generation process in high-dimensional data space, optimizing sample diversity and information fidelity during data augmentation.

[0174] The process of sample generation based on the SMOTE algorithm of quantum state evolution is as follows:

[0175] S201. Separate minority class samples from the labeled training set. Based on quantum state evolution theory, calculate the quantum state representation of each minority class sample point, b. i Mapped to a quantum state ψ i , represented as:

[0176]

[0177] In the formula, |k> represents the ground state of the qubit, and α ik It is the complex probability magnitude, specifically for sample b. i The probability amplitude in the ground state |k> satisfies the normalization condition. K is the dimension of the quantum state space;

[0178] Complex probability amplitude α ik Based on sample b i The eigenvalues ​​are calculated and expressed as:

[0179]

[0180] In the formula, μ k The center position of the k-th quantum ground state is represented, specifically the characteristic center of the quantum ground state |k>; γ is a hyperparameter controlling the scale; ||b i -μ k || 2 Indicates sample b i With ground state center μ k The square of the Euclidean distance between them; preferably, γ re Set it to 0.3.

[0181] S202. Select sample pairs that are close to each other and perform quantum state superposition operations to simulate new sample points using the principle of quantum state interference; where, sample pairs (b i ,b j Perform a superposition operation on quantum states to generate a new quantum state ψ. ij , represented as:

[0182]

[0183] In the formula, ψ i and ψ j Sample b i and b j The quantum state representation; |||| is the normalization operator, specifically the L1 norm normalization, ||ψ i +ψ j || represents the new state ψ generated by the superposition of quantum states. ij Standardized processing and standardized operation ensure ψ ijIt remains a valid quantum state; E() is the quantum entanglement operation function;

[0184] The quantum entanglement operation function E() for sample b i and b j quantum state ψ i and ψ j The calculation method using entanglement operations is expressed as follows:

[0185]

[0186] In the formula, Tensor products representing quantum states;

[0187] ||ψ i +ψ j The calculation method for || can be expressed as:

[0188]

[0189] In the formula, |α ik +α jk | 2 It calculates the squared magnitude of the probability amplitude of the new superposition state in the ground state |k>.

[0190] S203. The newly generated sample undergoes a quantum measurement process to ensure its validity and rationality in the classical data space, while maintaining the diversity and complexity of the data. The quantum measurement process is used to convert the quantum state ψ... ij Convert back to classic sample b new The method is expressed as:

[0191] b new =measure(ψ) new )

[0192]

[0193] In the formula, measure(ψ) is the quantum measurement function, and ψ is the input to the quantum measurement function; β k Through quantum state ψ ij The probability obtained from the measurement is specifically β. k In the quantum state ψ ij After measurement, the probability of the ground state |k> occurring determines the values ​​of the new sample in each dimension; ψ new The quantum state of the heart sample;

[0194] ψ new The calculation method is expressed as follows:

[0195]

[0196] In the formula, <ψij |ψ ij > represents ψ ij The inner product; G() is the dynamic adaptive quantum gate operation function;

[0197] The dynamic adaptive quantum gate G() is a quantum operation dependent on data characteristics, which acts on a single quantum state ψ. i or quantum state pair ψ i ,ψ j Above, to adjust its quantum characteristics, it is expressed as:

[0198] G(ψ i )=U(θ i )ψ i

[0199]

[0200] In the formula, U(θ i It depends on the parameter θ. i The unitary transform, where H is the corresponding Hamiltonian operator, θ i These parameters are automatically learned from data features and are solved through the following optimization problem:

[0201]

[0202] In the formula, L() is the cross-entropy loss function, f() represents the quantum measurement function with parameter θ, and y i It is sample b i The target value;

[0203] The entire process is repeated until the predetermined amount of data or sample distribution balance is achieved.

[0204] S204, Newly generated sample b new After a verification and adjustment process to ensure its adaptability and diversity in data distribution, it is represented as:

[0205]

[0206] In the formula, λ vf To adjust the factor, controlling how closely the new sample resembles the original sample pair; Median operation representing sample features; Indicates sample b i and b j The eigenvalue median; preferably, λ vf Set it to 0.5.

[0207] The calculation method is expressed as follows:

[0208]

[0209] In the formula, The calculation is for sample b. i and b j The arithmetic mean of the two characteristics represents the midpoint of their features;

[0210] S3, Feature Extraction Model Training

[0211] The expanded data is input into the feature extraction model for training. This invention adopts a neural network structure optimized based on light and shadow tracking, which uses the pattern of light and shadow changes to capture and extract dynamic features in the input data, and accurately extracts communication error features. This invention adopts a loop restart strategy, which allows the model to reset network parameters according to preset conditions or performance thresholds during training, thereby avoiding getting stuck in local optima and enhancing the model's adaptability to new data.

[0212] The training process of the neural network algorithm based on motion tracking optimization is as follows: Figure 2 As shown:

[0213] S301. Set the initial architecture of the neural network, including an input layer, two hidden layers, and an output layer; during model initialization, the network parameters θ include weights w and biases b, and the parameter initialization is expressed as follows:

[0214]

[0215] In the formula, l represents the network level, i and j represent the neuron indices of that layer, and σ 2 It is the variance of the initial weights. The weights of the l-th layer are... For the bias of the l-th layer, This indicates a range with a mean of 0 and a variance of σ. 2 Sampling from a normal distribution;

[0216] S302. Using the augmented training data, calculate the output of each layer through forward propagation, and then adjust the network parameters through backpropagation to minimize the output error. The network output y and the target y are then compared. * The error between them is calculated using a loss function:

[0217]

[0218] Furthermore, during the parameter update process, the method of backpropagating to update the weights is represented as follows:

[0219]

[0220] In the formula, and The weights and biases before the update. and Here are the updated weights and biases, and y is the network output. * For the real goal, and These are the partial derivatives of the loss function with respect to the weights and biases of the l-th layer, respectively, and I(W,b) is the update impact factor.

[0221] When calculating the update impact factor I(W,b), the importance weight of each parameter needs to be dynamically evaluated and adjusted in each iteration. This further identifies the parameters most critical to the current training stage and assigns them higher priority, thereby improving the overall efficiency and effectiveness of network training. The update impact factor I is defined as:

[0222] I(W,b)=tanh(ν ps ·C(W,b)+μ ps ·S(W,b))

[0223] In the formula, tanh() is the hyperbolic tangent function, and ν ps and μ ps These are the coefficients for adjusting and updating the impact factor, where C(W,b) is the parameter contribution and S(W,b) is the parameter sensitivity; preferably, ν and μ are set to 0.3 and 0.7, respectively.

[0224] The parameter contribution C(W,b) is calculated as follows:

[0225]

[0226] In the formula, ⊙ denotes element-wise multiplication. This represents the partial derivative of the current network weights with respect to the loss function. Let W represent the partial derivative of the current network bias with respect to the loss function, W represent the current network weights, and b represent the current network bias.

[0227] The parameter sensitivity S(W,b) is calculated as follows:

[0228] S(W, b) = exp(-γ ps ||C(W,b)||)

[0229] In the formula, exp() represents the exponential function, and γ ps γ is the sensitivity adjustment coefficient, |||| is the L1 norm normalization operation; preferably, γ ps Set it to 0.5.

[0230] S303. At the end of each training cycle, evaluate the model's performance. If the model's performance does not reach the predetermined improvement standard or training progress stagnates, adjust the learning rate or reinitialize the network parameters according to the cycle restart strategy. The cycle restart strategy is implemented by dynamically adjusting the learning rate η, based on the changes in the loss function L during training, and is thus expressed as:

[0231] η new =η×λ vg t / τ

[0232] In the formula, λ vg η is the decay factor, t is the current training cycle, τ is the decay time step, and η is the decay factor. new The updated learning rate;

[0233] Attenuation factor λ vg The calculation method is expressed as follows:

[0234] λ vg =e -α

[0235] In the formula, α is a parameter based on the growth rate of the model performance evaluation index, and its calculation method is expressed as:

[0236]

[0237] In the formula, ΔL represents the change in the loss function over two consecutive training periods, and L is the loss function value for the current period.

[0238] S304. Repeat steps S301-S303 until the preset stopping iteration condition is met, which indicates that the model training is complete. In this embodiment, the preset stopping iteration condition is to reach the preset maximum number of iterations, which is set to 1000 times.

[0239] S4: Feature Dimensionality Reduction Model Training

[0240] The feature vectors obtained after feature extraction are input into a feature dimensionality reduction model for training. This invention employs an improved autoencoder neural network for feature dimensionality reduction. The autoencoder neural network consists of an encoder and a decoder. The encoder is responsible for converting the input data into low-dimensional feature representations, while the decoder attempts to reconstruct the original data from these low-dimensional representations. Addressing the problem of traditional autoencoders' poor performance when handling complex data structures, this invention adopts a manifold learning strategy. The autoencoder can better understand and compress the inherent structure of the data during the encoding stage, rather than simply reducing data dimensionality, which helps to reconstruct more accurate original data during the decoding stage.

[0241] The training process for the improved autoencoder neural network is as follows:

[0242] S401. Initialize the parameters of the autoencoder neural network using a Gaussian distribution strategy, where each parameter p... θ The initialization method is expressed as follows:

[0243]

[0244] In the formula, p θ For network parameters; This indicates that the mean is 0 and the variance is σ. 2 The normal distribution; σ is the initial standard deviation, which controls the degree of dispersion of the initial parameters;

[0245] S402. The input data is propagated forward through the encoder. A manifold learning method is used to compress features while preserving the local structure of the data. Specifically, the input data x is converted into a low-dimensional representation z, as follows:

[0246] z = Sig(W) e x+b e )

[0247] In the formula, x represents the input data; W e b is the weight matrix of the encoder; e Here is the encoder's bias vector; Sig() is the Sigmoid activation function.

[0248] The Sigmoid function is calculated as follows:

[0249]

[0250] In the formula, v is the input to the activation function;

[0251] Weight matrix W e The update method is represented as:

[0252]

[0253] In the formula, The updated weight matrix, The weight matrix before the update is η. ry L is the learning rate, and L is the loss function; for W e partial derivatives Represented as:

[0254]

[0255] Preferably, η ry Set to 0.005.

[0256] S403. The compressed features are reconstructed by the decoder in an attempt to restore their original data form. During this process, the mean squared error loss function is used to constrain the reconstruction process. The reconstruction operation is represented as follows:

[0257]

[0258] In the formula, To reconstruct the data; z represents the encoded low-dimensional feature; W d b is the weight matrix of the decoder; d This is the bias vector for the decoder;

[0259] S404. After each round of training, a uniformization strategy is used to adjust the weights of each feature dimension. The weights W and bias b are adjusted to uniformly distribute the feature information, as shown below:

[0260]

[0261] In the formula, ||W|| F is the Frobenius norm of W, used to normalize the weights; ||b||1 is the 1-norm of b, used to normalize the biases;

[0262] Frobenius norm ||W|| F The calculation method is expressed as follows:

[0263]

[0264] In the formula, W ij It is the element in the i-th row and j-th column of matrix W;

[0265] The 1-norm ||b||1 is expressed as follows:

[0266]

[0267] In the formula, b i It is the i-th element in vector b;

[0268] S405. Calculate the gradient based on the loss function and update the network parameters using gradient descent, as follows:

[0269]

[0270] In the formula, θ represents the network parameters, including W e ,b e W d ,b d η bv It is the learning rate; It is the gradient of the loss function L with respect to θ, where L is the mean squared error. η bv Set to 0.004.

[0271] gradient of loss function The calculation method is expressed as follows:

[0272]

[0273] S406. Repeat steps S401-S405 until the preset stopping iteration condition is met, which indicates that the model training is complete. In this embodiment, the preset stopping iteration condition is to reach the preset maximum number of iterations, which is set to 1000 times.

[0274] S5: Classifier Model Training

[0275] The dimensionality-reduced data is input into the classifier for training the neutral classifier model. This invention employs a random forest algorithm based on steady-state redundancy removal. This algorithm comprises multiple decision trees, each independently constructed during training. The final classification result is determined through a voting mechanism during classification decisions. To enhance the model's generalization ability and reduce the risk of overfitting, this invention dynamically removes trees that provide repetitive information by evaluating the similarity between decision trees, retaining decision trees with high diversity.

[0276] The training process of the random forest algorithm based on steady-state redundancy removal is as follows:

[0277] S501. Assume the random forest consists of N independent decision trees, each tree having a value of T. i During initialization, samples and features are randomly selected. The initialization process is represented as follows:

[0278] T i =initialize(Data,F qi ,S qi ), i = 1, 2, ..., N

[0279] In the formula, initialize() is the initialization function; T i Let F be the i-th tree; Data is the training dataset; F qi S is a subset of features randomly selected from the feature set; qi The sample subset is randomly selected from the sample; preferably, the number of trees N in the random forest is set to 100, and the number of features F used per tree is... qi and sample size S qi Each represents 50% of the total.

[0280] S502. For each tree, train using the topologically transformed data. The topological transformation is achieved through principal component analysis to make the data suitable for decision tree splitting, as shown below:

[0281] X′ qi =PCA(X) qi ),X qi ∈S qi

[0282] In the formula, X qi Selected sample data; PCA() is the PCA function; X′ qi The data after PCA transformation;

[0283] S503. After the tree is constructed, calculate the redundancy between the trees; if a high redundancy relationship is found, remove the tree with the highest redundancy and only retain the tree that provides independent information. Let Sim(T) = ... i ,T j The function ) represents the similarity calculation function between two trees, and the calculation method is expressed as follows:

[0284]

[0285] In the formula, X is the dataset used for evaluation. The function is an indicator function. If two trees have the same classification result for the same data point x, the function value is 1, otherwise it is 0. Preferably, the similarity threshold is set to 0.85. Trees with a value higher than this are considered redundant and removed from the model.

[0286] S504. Each tree is trained independently, and each tree is trained using the optimal split point to ensure that the decision boundary can be learned from the training data. The selection of the split point is based on maximizing the gain, expressed as:

[0287]

[0288] In the formula, InfoGain(s) represents the gain at the split point s;

[0289] The gain InfoGain(s) is calculated as follows:

[0290]

[0291] In the formula, H(D) is the entropy of the dataset D, and D left (s) and D right (s) represent the left and right subsets of the dataset after the split point s;

[0292] The entropy H(D) is calculated as follows:

[0293]

[0294] In the formula, p k It represents the probability of category k in dataset D;

[0295] S505. The outputs of all decision trees are merged through a voting mechanism. Each tree provides a prediction result for an input sample. The final result is the category with the most votes, represented as:

[0296]

[0297] In the formula, c i C is the classification result of the i-th tree; C is the final classification result, where c represents the possible categories. The indicator function is defined if the output c of the i-th tree... i If the value equals category c, the function value is 1; otherwise, it is 0.

[0298] S506. During model training, the number and depth of decision trees, as well as the parameters of topology transformation, are dynamically adjusted based on feedback from validation data to optimize model performance. The adjustment methods for the number of trees N and the number of PCA components k are expressed as follows:

[0299] N new =N+ΔN

[0300] k new =k+Δk

[0301] In the formula, ΔN and Δk are the number of adjustments made based on the model performance, respectively;

[0302] S6: Classification and Assessment of Bit Error Rate in Environmental Backscattering Systems

[0303] The newly collected environmental backscattering system data is processed using pre-trained feature extraction models, data dimensionality reduction models, and classifier models to evaluate the communication bit error rate.

[0304] To facilitate understanding of the training and application process of this invention, the following is adopted: Figure 3 In this embodiment, a new set of environmental data is collected. First, preliminary feature extraction is performed using a feature extraction model. Then, a data dimensionality reduction model is used to reduce the computational resources required for processing. Finally, a classifier model is used to predict the bit error rate. In this embodiment, the classification categories include: no bit error, single-bit bit error, multi-bit bit error, parity check bit error, synchronization failure bit error, and other bit errors.

Claims

1. A method for evaluating the bit error rate of communication in an environment backscattering system based on artificial intelligence, characterized in that, Includes the following steps: S1. Data Acquisition and Labeling The collected data comes from the communication nodes in the environmental backscattering system, and communication error events are monitored and recorded in real time through a dedicated hardware interface. The collected data is stored in a structured vector data format, with each data point including multiple attribute fields; The collected data is manually labeled, and the labeling categories include: no bit error, single bit error, multi-bit error, parity error, and synchronization failure error. S2, Data Expansion The SMOTE algorithm based on quantum state evolution is used for sample generation, thereby achieving data augmentation. The process of sample generation using the SMOTE algorithm based on quantum state evolution is as follows: S201. Separate minority class samples from the labeled training set. Based on quantum state evolution theory, calculate the quantum state representation of each minority class sample, b. i Mapped to a quantum state ψ i , represented as: In the formula, |k> represents the ground state of the qubit, and α ik It is the complex probability magnitude, specifically for sample b. i The probability amplitude in the ground state |k> satisfies the normalization condition. K is the dimension of the quantum state space; k is the index of the dimension of the quantum state space; S202. Select sample pairs that are close to each other and perform quantum state superposition operations to simulate new samples using the principle of quantum state interference; where, sample pairs (b i ,b j Perform a superposition operation on quantum states to generate a new quantum state ψ. ij , is represented as: In the formula, ψ i and ψ j Sample b i and b j The quantum state representation; |||| is the normalization operator, specifically the L1 norm normalization, ||ψ i +ψ j || represents the new state ψ generated by the superposition of quantum states. ij Standardized processing and standardized operation ensure ψ ij It remains a valid quantum state; E() is the quantum entanglement operation function; The quantum entanglement operation function E() for sample b i and b j quantum state ψ i and ψ j The calculation method using entanglement operations is expressed as follows: In the formula, Tensor products representing quantum states; ||ψ i +ψ j The calculation method for || is expressed as follows: In the formula, |α ik +α jk | 2 The calculation is the square of the magnitude of the complex probability amplitude of the new superposition state in the ground state |k>; S203. The newly generated sample undergoes a quantum measurement process to ensure its validity and rationality in the classical data space, while maintaining the diversity and complexity of the data. The quantum measurement process is used to convert the quantum state ψ... ij Convert back to classic sample b new The method is expressed as: b new =measure(ψ new ) In the formula, measure(ψ) is the quantum measurement function, and ψ is the input to the quantum measurement function; β k Through quantum state ψ ij The probability obtained from the measurement is specifically β. k In the quantum state ψ ij After measurement, the probability of the ground state |k> occurring determines the values ​​of the new sample in each dimension; ψ new The quantum state of the new sample; ψ new The calculation method is expressed as follows: In the formula, <ψ ij |ψ ij > represents ψ ij The inner product; G() is the dynamic adaptive quantum gate operation function; S204, Newly generated sample b new After a verification and adjustment process to ensure its adaptability and diversity in data distribution, it is represented as: b new =λ vf ·b new +(1-l vf )·(b i ⊕b j ) In the formula, λ vf To adjust the factor, controlling how closely the new sample approximates the original sample pair; ⊕ represents the median operation of the sample features; b i ⊕b j Indicates sample b i and b j The eigenmedian; b i ⊕b j The calculation method is expressed as follows: In the formula, The calculation is for sample b. i and b j The arithmetic mean of the two values ​​represents the midpoint of their characteristics; b i For the i-th generated sample, b j Generate the j-th sample; S3, Feature Extraction Model Training The expanded data is input into the feature extraction model for training. A neural network structure based on light and shadow tracking optimization is adopted to capture and extract dynamic features in the input data by utilizing the pattern of light and shadow changes, and to accurately extract communication error features. The training process of the neural network algorithm optimized based on light and shadow tracking is as follows: S301. Set the initial architecture of the neural network, including an input layer, two hidden layers, and an output layer; during model initialization, the network parameters θ include weights w and biases b, and the parameter initialization is expressed as follows: In the formula, l represents the network layer, i and j are both positive integers, representing the neuron index of that layer, and σ 2 It is the variance of the initial weights. The weights of the l-th layer are... For the bias of the l-th layer, This indicates that the mean is 0 and the variance is σ. 2 The normal distribution; S302. Using the augmented training data, calculate the output of each layer through forward propagation, and then adjust the network parameters through backpropagation to minimize the output error. The network output y and the target y are then compared. * The error between them is calculated using a loss function: Furthermore, during the parameter update process, the method of backpropagating to update the weights is represented as follows: In the formula, and The weights and biases before the update. and Here are the updated weights and biases, and y is the network output. * For the real goal, and These are the partial derivatives of the loss function with respect to the weights and biases of the l-th layer, respectively; I(W,b) is the update influence factor; and η is the learning rate of the neural network. S303. At the end of each training cycle, evaluate the model's performance. If the model's performance does not reach the predetermined improvement standard or training progress stagnates, adjust the learning rate or reinitialize the network parameters according to the cycle restart strategy. The cycle restart strategy is implemented by dynamically adjusting the learning rate η, based on the changes in the loss function L during training, and is thus expressed as: or new =η×λ vg t / τ In the formula, λ vg η is the decay factor, t is the current training cycle, τ is the decay time step, and η is the decay factor. new The learning rate for the updated neural network; Attenuation factor λ vg The calculation method is expressed as follows: l vg =e -α In the formula, α is a parameter based on the growth rate of the model performance evaluation index, and its calculation method is expressed as: In the formula, ΔL represents the change in the loss function over two consecutive training periods, and L is the loss function value for the current period. S304. Repeat steps S301-S303 until the preset stopping iteration condition is met, which means that the model training is complete. S4: Feature Dimensionality Reduction Model Training The feature vectors obtained after feature extraction are input into the feature dimensionality reduction model for training. An improved autoencoder neural network is used for feature dimensionality reduction. The training process of the improved autoencoder neural network is as follows: S401. Initialize the parameters of the autoencoder neural network using a Gaussian distribution strategy, where each parameter p... θ The initialization method is expressed as follows: In the formula, p θ For network parameters; This indicates that the mean is 0 and the variance is σ. 2 The normal distribution; σ is the initial standard deviation, which controls the degree of dispersion of the initial parameters; S402. The input data is propagated forward through the encoder. A manifold learning method is used to compress features while preserving the local structure of the data. Specifically, the input data x is transformed into low-dimensional features z, represented as: z=Sig(W e x+b e ) In the formula, x represents the input data; W e b is the weight matrix of the encoder; e Here is the encoder's bias vector; Sig() is the Sigmoid activation function. The Sigmoid function is calculated as follows: In the formula, v is the input to the activation function; S403. The compressed features are reconstructed by the decoder in an attempt to restore their original data form. During this process, the mean squared error loss function is used to constrain the reconstruction process. The reconstruction operation is represented as follows: In the formula, To reconstruct the data; z represents the encoded low-dimensional feature; W d b is the weight matrix of the decoder; d This is the bias vector for the decoder; S404. After each round of training, a uniformization strategy is used to adjust the weights of each feature dimension. The weights W and bias b are adjusted to uniformly distribute the feature information, as follows: In the formula, ||W|| F is the Frobenius norm of W, used to normalize the weights; ||b||1 is the 1-norm of b, used to normalize the biases; Frobenius norm ||W|| F The calculation method is expressed as follows: In the formula, W ij It is the element in the i-th row and j-th column of matrix W; The 1-norm ||b||1 is expressed as follows: In the formula, b i It is the i-th element in bias b; S405. Calculate the gradient based on the loss function and update the network parameters using gradient descent, as follows: In the formula, θ represents the network parameters, including W e ,b e W d ,b d η bv It is the learning rate of the autoencoder; It is the gradient of the loss function L with respect to θ, where L is the mean squared error. Where n represents the number of samples input to the autoencoder in the current batch; gradient of loss function The calculation method is expressed as follows: S406. Repeat steps S401-S405 until the preset stopping iteration condition is met, which means that the model training is complete. S5: Classifier Model Training The dimensionality-reduced data is input into the classifier for training the classifier model. A random forest algorithm based on steady-state redundancy removal is used, and the training process is as follows: S501. Assume a random forest consists of N independent decision trees, each tree having a value of T. i During initialization, samples and features are randomly selected. The initialization process is represented as follows: T i =initialize(Data,F qi ,S qi ),i=1,2,…,N In the formula, initialize() is the initialization function; T i Let F be the i-th tree; Data is the training dataset; F qi S is a subset of features randomly selected from the feature set; qi This refers to a subset of samples randomly selected from the sample. S502. For each tree, train using the topologically transformed data. The topological transformation is achieved through principal component analysis to make the data suitable for decision tree splitting, as shown below: X′ qi =PCA(X qi ),X qi ∈S qi In the formula, X qi Selected sample data; PCA() is the PCA function; X′ qi The data after PCA transformation; S503. After the tree is constructed, calculate the redundancy between the trees; if a high redundancy relationship is found, remove the tree with the highest redundancy and only retain the tree that provides independent information. Let Sim(T) = ... i ,T j The function ) represents the similarity calculation function between two trees, and the calculation method is expressed as follows: In the formula, X is the dataset used for evaluation. This is an indicator function; if the two trees classify the same data point x, the function value is 1, otherwise it is 0. S504. Each tree is trained using the optimal split point to ensure that the decision boundary can be learned from the training data. The selection of the split point is based on maximizing the gain, expressed as: In the formula, InfoGain(s) represents the gain at the split point s; The gain InfoGain(s) is calculated as follows: In the formula, H(D) is the entropy of the dataset D, and D left (s) and D right (s) represent the left and right subsets of the dataset after the split point s; The entropy H(D) is calculated as follows: In the formula, It is category k c The probability k in dataset D c Indicates a data category index; S505. The outputs of all decision trees are merged through a voting mechanism. Each tree provides a prediction result for an input sample. The final result is the category with the most votes, represented as: In the formula, c i C is the classification result of the i-th tree; C is the final classification result. The indicator function is defined if the output c of the i-th tree... i equal to category k c If the value is 1, the function value is 1; otherwise, it is 0. S506. During model training, dynamically adjust the number and depth of decision trees, as well as the parameters of topology transformation, based on feedback from validation data to optimize model performance; the number of trees N and the number of PCA components k. d The adjustment method is expressed as follows: N new =N+ΔN In the formula, ΔN and Δk d N represents the number of trees adjusted based on model performance and the number of PCA components, respectively. new The adjusted number of trees, k is the number of components in the adjusted PCA. d The number of components in the PCA; S6: Classification and Assessment of Bit Error Rate in Environmental Backscattering Systems The newly collected environmental backscattering system data is processed using pre-trained feature extraction models, data dimensionality reduction models, and classifier models to evaluate the communication bit error rate.

2. The method for evaluating the bit error rate of an environment backscattering system based on artificial intelligence as described in claim 1, characterized in that, In step S201, the complex probability amplitude α ik Based on sample b i The eigenvalues ​​are calculated and expressed as: In the formula, μ k This represents the center position of the k-th quantum ground state, specifically the characteristic center of the quantum ground state |k>. γ re It is a hyperparameter that controls the scale; ||b i -μ k || 2 Indicates sample b i With the ground state center μ k The square of the Euclidean distance between them.

3. The method for evaluating the bit error rate of an environment backscattering system based on artificial intelligence as described in claim 1, characterized in that, In step S203, the dynamic adaptive quantum gate G() is a quantum operation dependent on the data characteristics, and the dynamic adaptive quantum gate G() acts on a single quantum state ψ. i or quantum state pair ψ i ,ψ j Above, to adjust the quantum characteristics, it is expressed as: G(ψ i )=U(θ i )ψ i In the formula, U(θ i It depends on the parameter θ. i The unitary transform, where H is the corresponding Hamiltonian operator, θ i These parameters are automatically learned from data features and are solved through the following optimization problem: In the formula, L() is the cross-entropy loss function, f() represents the quantum measurement function with parameter θ, and y i It is sample b i Target value; The entire process is repeated until the predetermined amount of data or sample distribution balance is reached.

4. The method for evaluating the bit error rate of an environment backscattering system based on artificial intelligence as described in claim 1, characterized in that, In step S302, when calculating the updated impact factor I(W,b), the importance weight of each parameter needs to be dynamically evaluated and adjusted in each iteration. This further identifies the parameters most critical to the current training stage and assigns them higher priority, thereby improving the overall efficiency and effectiveness of network training. The updated impact factor I is defined as: I(W,b)=tanh(ν ps ·C(W,b)+μ ps ·S(W,b)) In the formula, tanh() is the hyperbolic tangent function, and ν ps and μ ps These are the coefficients for adjusting and updating the impact factor, where C(W,b) is the parameter contribution and S(W,b) is the parameter sensitivity. The parameter contribution C(W,b) is calculated as follows: In the formula, ⊙ denotes element-wise multiplication. This represents the partial derivative of the current network weights with respect to the loss function. Let W represent the partial derivative of the current network bias with respect to the loss function, W represent the current network weights, and b represent the current network bias. The parameter sensitivity S(W,b) is calculated as follows: S(W,b)=exp(-γ ps ||C(W,b)||) In the formula, exp() represents the exponential function, and γ ps is the sensitivity adjustment coefficient, and |||| is the L1 norm normalization operation.

5. The method for evaluating the bit error rate of an environment backscattering system based on artificial intelligence as described in claim 1, characterized in that, In step S402, the weight matrix W e The update method is represented as: In the formula, The updated weight matrix, The weight matrix before the update is η. bv L is the learning rate of the autoencoder, and L is the loss function; for W e partial derivatives Represented as:

Images