Intelligent palladium-copper material purity analysis method and system based on cooperative detection

Abstract

The invention provides a palladium-copper material purity intelligent analysis method and system based on collaborative detection, and relates to the technical field of collaborative detection.The method comprises the steps that spectral data and physical and chemical parameters are obtained, features are extracted and fused through a deep convolutional neural network, and dimensionality reduction is conducted through a Monte Carlo method; and inputting the dimension reduction features into a double-path parallel verification network for beam search prediction and greedy search verification, and determining purity or triggering early warning according to a prediction difference value. According to the method, the accuracy and reliability of purity analysis of the palladium-copper material are improved, and the misjudgment rate is reduced.

Description

Technical Field

[0001] This invention relates to collaborative detection technology, and more particularly to a method and system for intelligent analysis of the purity of palladium-copper materials based on collaborative detection. Background Technology

[0002] Palladium-copper (PCC) is an important functional material widely used in high-tech fields such as electronics, energy, and catalysts. In practical applications, the purity of PCC directly affects its physicochemical properties and service life; therefore, accurate detection and analysis of PCC purity is of significant practical value. Traditional PCC purity detection mainly relies on chemical analysis and physical measurement methods, such as atomic absorption spectrometry, X-ray fluorescence spectrometry, and inductively coupled plasma mass spectrometry. With the development of artificial intelligence technology, combining deep learning with materials detection technology has provided a new research direction for PCC purity analysis.

[0003] Existing purity testing technologies for palladium-copper materials have the following shortcomings: First, traditional single-method testing is insufficient to comprehensively reflect the complex compositional characteristics of palladium-copper materials. Relying solely on spectral analysis or physicochemical parameter analysis often leads to incomplete results, failing to consider the multi-dimensional characteristics of the material and thus affecting the accuracy of purity assessment. Second, existing analytical models typically employ simple linear or shallow neural network structures, which have limited ability to extract weak features and nonlinear relationships in palladium-copper spectral data. This is particularly problematic in detecting low-concentration impurities, easily leading to missed detections or misjudgments, reducing the reliability of purity analysis. Furthermore, the lack of an effective verification mechanism is also a significant deficiency in current technology. Single prediction paths are susceptible to data noise and model uncertainty, failing to achieve automatic identification and early warning of abnormal data, resulting in potential safety hazards during quality control.

[0004] As the application of palladium-copper materials in high-precision fields continues to expand, the market has placed higher demands on the accuracy, reliability, and adaptability of purity testing technologies. There is an urgent need to develop intelligent analysis methods based on multi-source data fusion and deep learning to improve the accuracy and reliability of palladium-copper material purity testing. Summary of the Invention

[0005] This invention provides a method and system for intelligent analysis of the purity of palladium-copper materials based on collaborative detection, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a method for intelligent purity analysis of palladium-copper materials based on collaborative detection, comprising: Acquire the spectral data and physicochemical parameters of the palladium-copper material to be tested, establish a spectral curve based on the spectral data, calculate the basic characteristic values ​​of the palladium-copper material based on the physicochemical parameters, and use the spectral curve and the basic characteristic values ​​as the data to be analyzed. A deep convolutional neural network is used to perform feature extraction. Multi-layer convolution operations are performed on the data to be analyzed to extract spectral feature vectors. The spectral feature vectors are then fused with the basic feature values ​​through adaptive weights to obtain a multi-dimensional feature tensor. The multi-dimensional feature tensor is then dimensionality-reduced based on the Monte Carlo iterative sampling method to obtain dimensionality-reduced feature data. The dimensionality-reduced feature data is input into a pre-trained dual-path parallel verification network. The purity of the dimensionality-reduced feature data is predicted by a beam search path to obtain a beam search prediction value. The purity of the dimensionality-reduced feature data is verified by a greedy search path to obtain a greedy search prediction value. When the difference between the beam search prediction value and the greedy search prediction value is less than a preset prediction threshold, the weighted average of the two prediction values ​​is calculated based on preset weights as the final purity analysis result of the palladium-copper material. When the difference is greater than or equal to the preset prediction threshold, anomaly detection is triggered and a warning message is output.

[0007] A spectral curve is established based on the spectral data, and the fundamental characteristic values ​​of the palladium-copper material are calculated based on the physicochemical parameters. The spectral curve and the fundamental characteristic values ​​are used as data to be analyzed. The spectral data is divided into a training dataset and a validation dataset. Wavelet transform is performed on the training dataset to obtain multi-scale spectral features. Frequency domain information in the multi-scale spectral features is extracted by Fourier transform. The frequency domain information is combined with the original spectral data to construct a spectral curve. Based on the physicochemical parameters, a feature mapping relationship is constructed. The feature mapping relationship is optimized by a nonlinear optimization algorithm to obtain basic feature values ​​that reflect the physicochemical properties of palladium-copper materials. The basic feature values ​​are cross-validated with the validation dataset. When the validation error is less than a preset error threshold, the spectral curve and the basic feature values ​​are combined to form the data to be analyzed.

[0008] The spectral feature vector and the basic feature value are fused using adaptive weights to obtain a multidimensional feature tensor. The multidimensional feature tensor is then dimensionality-reduced using a Monte Carlo iterative sampling method to obtain dimensionality-reduced feature data, including: Calculate the correlation coefficient between the spectral feature vector and the basic feature value, construct an adaptive weight calculation function based on the correlation coefficient, dynamically adjust the fusion weight between the spectral feature vector and the basic feature value through the adaptive weight calculation function, and perform tensor operations on the fusion weight to obtain a multidimensional feature tensor. The multidimensional feature tensor is repeatedly randomly sampled using the Monte Carlo iterative sampling method. Based on the sampling results, the feature importance score is calculated. The feature dimension with the highest feature contribution is selected according to the feature importance score, and the feature dimensions are combined to generate dimensionality-reduced feature data.

[0009] The multidimensional feature tensor is repeatedly randomly sampled using the Monte Carlo iterative sampling method. Based on the sampling results, the feature importance score is calculated, including: The Monte Carlo method is used to repeatedly random sample the multidimensional feature tensor. The sampling probability distribution is dynamically adjusted in each round of sampling. When the variance of the sampling results of two adjacent rounds is less than the preset variance threshold, the sampling interval is increased and the sampling operation is repeated until the sampling results converge or the preset maximum number of samplings is reached. The converged sampling results are subjected to feature decomposition, and the contribution of each feature dimension in different sampling intervals is calculated. The contribution is combined with the entropy value of the sampling probability distribution, and the feature importance score is obtained by nonlinear weighting calculation.

[0010] The dimensionality-reduced feature data is input into a pre-trained dual-path parallel validation network. The purity of the dimensionality-reduced feature data is predicted using a beam search path to obtain a beam search prediction value. The purity of the dimensionality-reduced feature data is verified using a greedy search path to obtain a greedy search prediction value, including: Calculate the statistical and distribution characteristics of the dimensionality-reduced feature data, and construct a feature scoring standard based on the statistical and distribution characteristics; perform dual-path parallel computation on the dimensionality-reduced feature data, wherein the beam search path constructs a hierarchical search space by setting the search width and search depth, retains candidate results that meet the feature scoring standard at each search level, calculates the state transition matrix based on the historical search path, predicts the search direction of the next layer based on the state transition matrix, and uses the purity calculation result corresponding to the optimal search path as the beam search prediction value; In the greedy search path, the dimensionality-reduced feature data is validated hierarchically based on the feature scoring criteria. At each validation layer, a feature combination scheme is determined through local optimal selection, and the feature combination scheme is input into the validation function to calculate the purity index. An adaptive step size strategy is used to dynamically adjust the validation direction until the purity index converges to obtain the greedy search prediction value.

[0011] The dimensionality-reduced feature data is subjected to dual-path parallel computation, wherein the beam search path constructs a hierarchical search space by setting the search width and search depth, and retains candidate results that meet the feature scoring criteria at each search level, including: Calculate the dimensionality distribution characteristics of the dimensionality-reduced feature data, and divide the dimensionality-reduced feature data into a bundle search channel and a parallel search channel; in the bundle search channel, determine the initial value of the search width and the initial value of the depth according to the dimensionality distribution characteristics, and construct a multi-layer search space structure based on the initial value of the search width and the initial value of the depth; after each layer of search is completed, record the feature combination method of the search results of that layer to form a search history. Perform feature correlation analysis on the search history, calculate the score value of each candidate feature result based on the result of the feature correlation analysis, and filter and retain candidate results whose score values ​​exceed a preset score threshold; A feature state transition matrix is ​​constructed based on the search history. The feature transition probability is calculated using the feature state transition matrix. The search direction of the next search layer is determined based on the feature transition probability. The candidate results passed to each layer according to the transfer direction are sorted in descending order according to the score value. The feature combination ranked first is selected as the optimal feature combination of that layer. When the search reaches the last layer, the optimal feature combination of the last layer is used as the output result of the bundle search channel.

[0012] A second aspect of the present invention provides an intelligent analysis system for the purity of palladium-copper materials based on collaborative detection, comprising: The first unit is used to acquire the spectral data and physicochemical parameters of the palladium-copper material to be tested, establish a spectral curve based on the spectral data, calculate the basic characteristic values ​​of the palladium-copper material based on the physicochemical parameters, and use the spectral curve and the basic characteristic values ​​as data to be analyzed. The second unit is used to perform feature extraction using a deep convolutional neural network. It performs multi-layer convolution operations on the data to be analyzed to extract spectral feature vectors, and fuses the spectral feature vectors with the basic feature values ​​through adaptive weights to obtain a multi-dimensional feature tensor. It then performs dimensionality reduction processing on the multi-dimensional feature tensor based on the Monte Carlo iterative sampling method to obtain dimensionality-reduced feature data. The third unit is used to input the dimensionality-reduced feature data into a pre-trained dual-path parallel verification network, perform purity prediction on the dimensionality-reduced feature data through a beam search path to obtain a beam search prediction value, and perform purity verification on the dimensionality-reduced feature data through a greedy search path to obtain a greedy search prediction value. When the difference between the beam search prediction value and the greedy search prediction value is less than a preset prediction threshold, the weighted average of the two prediction values ​​is calculated based on preset weights as the final purity analysis result of the palladium-copper material. When the difference is greater than or equal to the preset prediction threshold, anomaly detection is triggered and a warning message is output.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] The beneficial effects of this application are as follows: By acquiring the spectral data and physicochemical parameters of the palladium-copper material to be tested, and combining them with deep convolutional neural networks to extract features, a comprehensive analysis of the spectral and physicochemical properties of the palladium-copper material was achieved. This significantly improved the accuracy and reliability of purity analysis and overcame the limitations of traditional analytical methods with a single data source.

[0016] The adaptive weight fusion and Monte Carlo iterative sampling methods are used to process features, which effectively solves the problems of noise interference and data redundancy in high-dimensional data analysis, improves the robustness of the algorithm, and enables the system to adapt to the analysis needs of palladium-copper material samples from different batches and sources.

[0017] A dual-path parallel verification network structure was designed, which uses two different strategies, beam search and greedy search, to predict and verify purity. An anomaly detection mechanism was also set up to automatically identify abnormal samples and issue warnings, which greatly reduces the false judgment rate and improves the reliability of the analysis process and its industrial application value. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the intelligent purity analysis method for palladium-copper materials based on collaborative detection, as described in an embodiment of the present invention. Figure 2 This is a flowchart of the beam search channel feature analysis and decision optimization process according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0021] Figure 1 This is a flowchart illustrating the intelligent purity analysis method for palladium-copper materials based on collaborative detection, as described in an embodiment of the present invention. Figure 1 As shown, the method includes: Acquire the spectral data and physicochemical parameters of the palladium-copper material to be tested, establish a spectral curve based on the spectral data, calculate the basic characteristic values ​​of the palladium-copper material based on the physicochemical parameters, and use the spectral curve and the basic characteristic values ​​as the data to be analyzed. A deep convolutional neural network is used to perform feature extraction. Multi-layer convolution operations are performed on the data to be analyzed to extract spectral feature vectors. The spectral feature vectors are then fused with the basic feature values ​​through adaptive weights to obtain a multi-dimensional feature tensor. The multi-dimensional feature tensor is then dimensionality-reduced based on the Monte Carlo iterative sampling method to obtain dimensionality-reduced feature data. The dimensionality-reduced feature data is input into a pre-trained dual-path parallel verification network. The purity of the dimensionality-reduced feature data is predicted by a beam search path to obtain a beam search prediction value. The purity of the dimensionality-reduced feature data is verified by a greedy search path to obtain a greedy search prediction value. When the difference between the beam search prediction value and the greedy search prediction value is less than a preset prediction threshold, the weighted average of the two prediction values ​​is calculated based on preset weights as the final purity analysis result of the palladium-copper material. When the difference is greater than or equal to the preset prediction threshold, anomaly detection is triggered and a warning message is output.

[0022] In one optional implementation, a spectral curve is established based on the spectral data, and the fundamental characteristic values ​​of the palladium-copper material are calculated based on the physicochemical parameters. The spectral curve and the fundamental characteristic values ​​are used as data to be analyzed, including: The spectral data is divided into a training dataset and a validation dataset. Wavelet transform is performed on the training dataset to obtain multi-scale spectral features. Frequency domain information in the multi-scale spectral features is extracted by Fourier transform. The frequency domain information is combined with the original spectral data to construct a spectral curve. Based on the physicochemical parameters, a feature mapping relationship is constructed. The feature mapping relationship is optimized by a nonlinear optimization algorithm to obtain basic feature values ​​that reflect the physicochemical properties of palladium-copper materials. The basic feature values ​​are cross-validated with the validation dataset. When the validation error is less than a preset error threshold, the spectral curve and the basic feature values ​​are combined to form the data to be analyzed.

[0023] The partitioning of spectral data employs a combination of random sampling and stratified sampling to construct the training and validation datasets. The partitioning process considers the wavelength and intensity distribution characteristics of the spectral data, ensuring statistical consistency between the training and validation sets. Stratified sampling divides the spectral data into several levels according to intensity range, with samples from each level allocated proportionally to both the training and validation sets. The training dataset typically comprises 70% to 80% of the total data, while the validation dataset accounts for 20% to 30%. The partitioning process also considers the time-series nature of the spectral data, preventing data from adjacent time points from appearing simultaneously in both the training and validation sets, thus preventing data leakage.

[0024] The wavelet transform process decomposes the training dataset into multiple scales by selecting appropriate wavelet basis functions. The selection of wavelet basis functions considers the characteristics of the spectral signal: the Daubechies wavelet is chosen for spectral signals with sharp peaks, while the Morlet wavelet is chosen for smooth, continuous spectral signals. The wavelet transform is implemented by convolving the original spectral signal with wavelet basis functions at different scales, generating wavelet coefficients at different scales and time points. The multi-scale decomposition process is recursive; each decomposition divides the signal into approximate and detail components. The approximate components reflect the low-frequency characteristics of the signal, while the detail components reflect its high-frequency characteristics.

[0025] The extraction of multi-scale spectral features is achieved by analyzing wavelet coefficients at different scale levels. Feature extraction includes the calculation of energy features, peak features, and distribution features. Energy features are obtained by calculating the sum of squares of wavelet coefficients at each scale level, reflecting the signal intensity at that scale. Peak features are obtained by identifying local maxima and minima of the wavelet coefficients, reflecting the abrupt changes in the signal at that scale. Distribution features are obtained by calculating the statistical moments of the wavelet coefficients, including mean, variance, skewness, and kurtosis, reflecting the signal's distribution characteristics at that scale.

[0026] The Fast Fourier Transform (FFT) algorithm transforms multi-scale spectral features from the time domain to the frequency domain. The transformation process uses wavelet coefficients at each scale level as input signals, and obtains the frequency domain representation through complex number operations. The frequency domain representation includes amplitude and phase spectra; the amplitude spectrum reflects the intensity of each frequency component, and the phase spectrum reflects the phase relationship between the frequency components. The FFT improves computational efficiency by decomposing the transformation process into several smaller-scale transformations using a divide-and-conquer strategy. The transformation process also incorporates window functions, such as the Hanning window or Blackman window, to reduce spectral leakage.

[0027] The extraction of frequency domain information is achieved by analyzing the frequency characteristics of the Fourier transform results. Frequency domain information includes the dominant frequency component, band energy distribution, and frequency peak positions. The dominant frequency component is obtained by identifying the frequency corresponding to the maximum value in the amplitude spectrum, reflecting the main frequency characteristics of the spectral signal. The band energy distribution is obtained by dividing the frequency range into several sub-bands and calculating the energy proportion within each sub-band. The frequency peak positions are determined by detecting local peaks in the amplitude spectrum; these peak positions correspond to the characteristic frequencies of the spectral signal. Frequency domain information also includes frequency stability analysis, which assesses the frequency stability of the signal by comparing the frequency domain characteristics of different time windows.

[0028] The spectral curve is constructed by fusing frequency domain information with the original spectral data. The fusion process employs a weighted average method, linearly combining the original spectral data and frequency domain information according to preset weights. Weighting is based on information quality assessment, which is evaluated using the signal-to-noise ratio (SNR) and feature significance. The SNR is calculated as the ratio of signal power to noise power; information with a higher ratio receives greater weight. Feature significance is assessed through statistical significance testing; frequency domain features with higher significance receive greater weight. The spectral curve also undergoes smoothing filtering to remove high-frequency noise, using either moving average or Gaussian filtering methods.

[0029] The construction of the feature mapping relationship is based on the correlation analysis between physicochemical parameters and the properties of palladium-copper materials. The mapping relationship includes two types: linear and nonlinear. Linear mapping establishes a direct relationship between parameters and properties through linear regression. Nonlinear mapping establishes complex mapping relationships through multinomial regression, radial basis functions, or neural network methods. The construction of the mapping relationship considers the interaction effects of physicochemical parameters, capturing the synergistic effects between parameters through cross terms and higher-order terms. The parameters of the mapping function are determined using the least squares method or maximum likelihood estimation method to ensure that the mapping relationship has good fitting accuracy.

[0030] The application of nonlinear optimization algorithms involves iteratively searching for the optimal mapping parameters. These algorithms include gradient descent, Newton's method, and genetic algorithms. Gradient descent calculates the gradient of the objective function with respect to the parameters and updates the parameter values ​​along the gradient descent direction. Newton's method calculates the second derivative of the objective function and utilizes curvature information to accelerate the convergence process. Genetic algorithms simulate biological evolution, searching for the optimal parameter combination through selection, crossover, and mutation operations. The optimization process also includes constraint handling, using Lagrange multiplier methods or penalty function methods to address parameter constraints.

[0031] The calculation of fundamental eigenvalues ​​transforms physicochemical parameters into eigenvalues ​​reflecting the properties of palladium-copper materials through optimized mapping relationships. Eigenvalues ​​are categorized into physical and chemical eigenvalues. Physical eigenvalues ​​reflect physical properties such as density, hardness, electrical conductivity, and thermal conductivity. Chemical eigenvalues ​​reflect chemical properties such as chemical composition, bond strength, reactivity, and stability. The eigenvalue calculation process considers dimensional uniformity of the parameters, eliminating the influence of different parameter dimensions through dimensionless processing. Dimensionless methods include standardization and normalization. Standardization is achieved by subtracting the mean and dividing by the standard deviation, while normalization is achieved by dividing by the maximum value or mapping to a specific interval.

[0032] Cross-validation evaluates the reliability of feature values ​​by comparing them with the validation dataset. The validation process employs k-fold cross-validation, dividing the validation dataset into k subsets. In each iteration, k-1 subsets are used for training, and the remaining subset is used for testing. Validation error is calculated by measuring the difference between predicted and true values. Measures of difference include mean squared error, mean absolute error, and correlation coefficient. Mean squared error is calculated by averaging the squared differences between predicted and true values, and mean absolute error is calculated by averaging the absolute values ​​of these differences. The correlation coefficient is calculated by assessing the linear correlation between predicted and true values.

[0033] The error threshold setting process is based on prediction accuracy requirements and practical application needs. Threshold setting considers the accuracy of the measuring instrument, the variation range of material properties, and the fault tolerance of the application scenario. For applications requiring high accuracy, the error threshold is set to 1 to 2 times the measurement accuracy. For applications requiring general accuracy, the error threshold is set to 3 to 5 times the measurement accuracy. Threshold setting also considers statistical significance, determining a reasonable threshold range through confidence interval analysis. When the verification error is less than the preset threshold, the basic characteristic value is considered to have sufficient reliability.

[0034] The data integration process involves combining spectral curves and fundamental characteristic values ​​in a specific format. This format employs a multi-dimensional array structure, with spectral curve data as the first dimension and fundamental characteristic value data as the second. Data integration considers data synchronization, ensuring that the spectral curves and fundamental characteristic values ​​correspond to the same material samples and measurement conditions. The integration process also includes data quality checks, ensuring data quality through outlier detection and consistency testing. Outlier detection uses statistical methods to identify data points deviating from the normal range, while consistency testing uses cross-validation to check the internal consistency of the data. The combined data possesses complete characteristic information, providing a reliable data foundation for subsequent analysis and processing.

[0035] In one optional implementation, the spectral feature vector and the basic feature values ​​are fused using adaptive weights to obtain a multidimensional feature tensor. The multidimensional feature tensor is then dimensionality-reduced using a Monte Carlo iterative sampling method to obtain dimensionality-reduced feature data, including: Calculate the correlation coefficient between the spectral feature vector and the basic feature value, construct an adaptive weight calculation function based on the correlation coefficient, dynamically adjust the fusion weight between the spectral feature vector and the basic feature value through the adaptive weight calculation function, and perform tensor operations on the fusion weight to obtain a multidimensional feature tensor. The multidimensional feature tensor is repeatedly randomly sampled using the Monte Carlo iterative sampling method. Based on the sampling results, the feature importance score is calculated. The feature dimension with the highest feature contribution is selected according to the feature importance score, and the feature dimensions are combined to generate dimensionality-reduced feature data.

[0036] The correlation coefficient between spectral eigenvectors and fundamental eigenvalues ​​is calculated by analyzing the correlation strength between each element in the spectral eigenvector and each eigenvalue in the set of fundamental eigenvalues. The correlation calculation employs Pearson correlation analysis, obtaining the correlation coefficient by dividing the covariance of the spectral eigenvector element and the fundamental eigenvalue by the product of their standard deviations. The covariance calculation involves multiplying the difference between the spectral eigenvector element and its mean by the difference between the fundamental eigenvalue and its mean, summing the results over all data points, and then dividing by the number of data points minus one. The standard deviation is calculated by taking the square root of the sum of the squares of the differences between each value and the mean, divided by the number of data points minus one. The correlation coefficient ranges from -1 to +1, with positive values ​​indicating positive correlation and negative values ​​indicating negative correlation. The absolute value reflects the strength of the correlation.

[0037] The adaptive weight calculation function is constructed based on the distribution characteristics of the correlation coefficient and the weight adjustment strategy. The weight function is constructed using a piecewise linear interpolation method, dividing the correlation coefficient interval into several sub-intervals, each corresponding to a different weight adjustment rule. When the absolute value of the correlation coefficient is greater than 0.8, a high-weight allocation strategy is adopted, with the weight value set as an adjustment factor equal to the absolute value of the correlation coefficient multiplied by 1.2. When the absolute value of the correlation coefficient is between 0.5 and 0.8, a medium-weight allocation strategy is adopted, with the weight value set as an adjustment factor equal to the absolute value of the correlation coefficient multiplied by 1.0. When the absolute value of the correlation coefficient is less than 0.5, a low-weight allocation strategy is adopted, with the weight value set as an adjustment factor equal to the absolute value of the correlation coefficient multiplied by 0.8.

[0038] The dynamic weight adjustment process is achieved by monitoring changes in the feature fusion effect. The adjustment process establishes a weight update mechanism, adjusting weight allocation based on changes in the information entropy of the fused features. Information entropy is calculated by dividing feature values ​​into several probability intervals after sorting them by numerical value, and then calculating the negative weighted sum of the probabilities of each interval. When the information entropy of the fused feature is significantly higher than that before fusion, the fusion weight of the corresponding feature is increased. When the information entropy change is not significant or decreases, the fusion weight of the corresponding feature is decreased. The magnitude of the weight adjustment is determined based on the degree of information entropy change; features with larger changes receive larger weight adjustments.

[0039] The tensor operation process for fusion weights achieves feature data fusion through multidimensional array operations. Tensor operations include element-wise multiplication of the weight matrix and eigenvectors, and tensor summation. Element-wise multiplication multiplies each element of the spectral feature vector with its corresponding fusion weight, and each element of the basic eigenvalue is multiplied by its corresponding fusion weight. Tensor summation sums the weighted spectral feature vector with the weighted basic eigenvalues ​​according to their respective dimensions. The summation process considers dimension matching; when the dimensions of the spectral feature vector and the basic eigenvalues ​​are inconsistent, dimension alignment is achieved through dimension expansion or reduction.

[0040] The construction of multidimensional feature tensors is achieved by reorganizing the fused feature data into a high-dimensional array structure. Tensor construction employs a hierarchical organization, arranging the fused features according to source type, importance level, and correlation strength. The first layer contains the fusion results from spectral features, the second layer contains the fusion results from basic features, and the third layer contains the interactive fusion results from cross-type features. Within each layer, features are sorted according to their importance scores, with higher-importance features appearing first. The tensor's dimensional design considers the computational efficiency of subsequent processing, typically employing a three-dimensional or four-dimensional structure to balance information preservation and computational complexity.

[0041] The Monte Carlo iterative sampling method generates sampling locations and parameters using a random number generator. Random sampling locations are determined by generating random numbers that conform to a uniform distribution, with the range of values ​​corresponding to the dimensionality range of the multidimensional feature tensor. Sampling parameters include the sampling window size and sampling step size. The sampling window size determines the number of features covered in each sampling iteration, and the sampling step size determines the distance between adjacent sampling locations. The iterative sampling process is repeated a specified number of times, generating new random sampling locations and parameters in each iteration, and extracting a subset of features from the corresponding locations for analysis.

[0042] The execution process of repeated random sampling involves the design of diverse sampling strategies, including three types: completely random sampling, hierarchical random sampling, and systematic random sampling. Completely random sampling selects sampling locations in a completely random manner, ensuring unbiased sampling. Hierarchical random sampling divides the multidimensional feature tensor into several levels according to feature type, and performs random sampling within each level, ensuring that all types of features are fully sampled. Systematic random sampling selects sampling locations in the tensor at fixed intervals, ensuring a uniform distribution of samples. Each sampling strategy performs multiple rounds of sampling, the number of rounds being determined based on the tensor size and accuracy requirements.

[0043] The calculation of feature importance scores is based on statistical analysis of the sampling results. The importance score calculation considers the frequency of feature occurrence across different samples, the variance of feature values, and the strength of the association between the feature and the target variable. Frequency of occurrence is obtained by dividing the number of times the feature is selected in all sampling results by the total number of samples; features with high frequency are considered to have higher importance. Eigenvalue variance is obtained by calculating the dispersion of feature values ​​across different samples; features with moderate variance are considered to have good stability and discriminative ability. Association strength is obtained by calculating the mutual information between the feature and the preset target variable; features with high mutual information are considered to have strong predictive ability.

[0044] The evaluation of feature contribution is achieved by comprehensively considering multiple dimensions of importance scores. A multi-dimensional evaluation index system is established for contribution assessment, including independent contribution index, collaborative contribution index, and stability index. The independent contribution index measures the performance of the feature when used alone, with performance evaluated by prediction accuracy or classification precision. The collaborative contribution index measures the performance improvement when the feature is combined with other features; a greater improvement indicates a higher collaborative contribution. The stability index measures the performance variation coefficient of the feature across different data subsets; features with smaller variation coefficients exhibit better stability.

[0045] The feature selection process filters high-contribution features by setting a contribution threshold. This threshold is based on the distribution characteristics of all feature contributions, typically selecting features ranking in the top 20% to 30% as high-contribution features. Threshold setting also considers computational resource constraints and the complexity requirements of subsequent processing; when computational resources are limited, the threshold is appropriately increased to reduce the number of selected features. The feature selection process also includes redundant feature removal. Highly correlated feature pairs are identified by calculating the correlation between features, retaining features with higher contributions and removing redundant features.

[0046] The feature combination process determines the optimal feature combination scheme through optimization algorithms. The combination process considers the complementarity and synergistic effects between features, and searches for the optimal feature combination using genetic algorithms or particle swarm optimization algorithms. Genetic algorithms encode feature selection schemes as chromosomes and evolve the feature combination schemes through selection, crossover, and mutation operations. The fitness function is based on a comprehensive performance evaluation of the feature combination, including prediction accuracy, computational efficiency, and generalization ability. After multiple generations of evolution, the feature combination with the highest fitness is selected as the final scheme.

[0047] The generation process of dimensionality-reduced feature data is achieved by reorganizing the selected feature dimensions according to their importance. The generation process establishes a feature index mapping relationship, recording the correspondence between the original feature dimensions and the dimensionality-reduced feature dimensions. The dimensionality-reduced feature data preserves the numerical and distribution characteristics of the original features, and normalization ensures that the numerical ranges of different feature dimensions are consistent. The normalization method uses min-max normalization, mapping the value of each feature dimension to the interval between zero and one. The output format of the dimensionality-reduced feature data is designed as a standard numerical matrix structure, facilitating subsequent processing and analysis by algorithms.

[0048] In one optional implementation, the multidimensional feature tensor is repeatedly randomly sampled using a Monte Carlo iterative sampling method, and the feature importance score is calculated based on the sampling results, including: The Monte Carlo method is used to repeatedly random sample the multidimensional feature tensor. The sampling probability distribution is dynamically adjusted in each round of sampling. When the variance of the sampling results of two adjacent rounds is less than the preset variance threshold, the sampling interval is increased and the sampling operation is repeated until the sampling results converge or the preset maximum number of samplings is reached. The converged sampling results are subjected to feature decomposition, and the contribution of each feature dimension in different sampling intervals is calculated. The contribution is combined with the entropy value of the sampling probability distribution, and the feature importance score is obtained by nonlinear weighting calculation.

[0049] To obtain the multidimensional feature tensor data to be analyzed, in practical applications, this data comes from fields such as image processing, speech recognition, or user behavior analysis. For example, in an e-commerce recommendation system, a tensor dataset containing multidimensional features such as user age, gender, shopping history, and browsing duration can be constructed. Suppose a four-dimensional feature tensor is constructed, with dimensions of user features (100 dimensions), product features (50 dimensions), time features (24 dimensions), and environmental features (10 dimensions).

[0050] Initialize the relevant parameters for Monte Carlo iterative sampling, including the initial sampling interval, initial sampling probability distribution, preset variance threshold, and maximum number of sampling iterations. For example, the initial sampling interval can be set to [0.2, 0.8] for each dimension, i.e., extreme cases are not considered; the initial sampling probability distribution can be set to a uniform distribution; the preset variance threshold can be set to 0.001, indicating that when the variance of the sampling results of two adjacent rounds is less than this value, the sampling is considered to be converging; and the maximum number of sampling iterations can be set to 1000 rounds to prevent infinite loops caused by failure to converge.

[0051] The Monte Carlo iterative sampling process is executed. In each round of sampling, samples are drawn from the feature tensor according to the current sampling probability distribution. For example, in the first round of sampling, 60% of the features can be randomly selected from the user feature dimension, 50% from the product feature dimension, 70% from the time feature dimension, and 40% from the environmental feature dimension, forming a subset of features. This subset of features is then used to perform the target task (such as predicting user click-through rate), and performance metrics (such as accuracy) are recorded.

[0052] The sampling probability distribution is dynamically adjusted, updating the probability distribution for the next round of sampling based on the results of the previous round. In practice, a Bayesian update method can be used to adjust the sampling probability. For example, if removing the user age feature significantly degrades model performance, the probability of retaining that feature is increased in the next round of sampling; conversely, if removing a feature doesn't significantly affect model performance, the probability of that feature being sampled in the next round is decreased.

[0053] The variance of sampling results between two adjacent rounds is calculated. In this embodiment, sampling results refer to model performance metrics obtained using different feature subsets. For example, the variance of model accuracy over 10 consecutive rounds of sampling can be calculated. Assuming the accuracies obtained from rounds 11 to 20 are 0.824, 0.819, 0.825, 0.820, 0.823, 0.821, 0.822, 0.824, 0.823, and 0.821, respectively, the calculated variance is 0.0000356.

[0054] The system checks if the variance is less than a preset threshold. If the variance is less than the preset threshold (e.g., 0.001), it indicates that the sampling results within the current sampling interval have stabilized, and the sampling interval needs to be expanded. For example, the sampling interval can be expanded from [0.2, 0.8] to [0.1, 0.9] to consider more extreme feature combinations. If the variance is greater than the preset threshold, sampling continues within the current sampling interval until the results converge or the maximum number of samplings is reached.

[0055] Feature decomposition is performed on the converged sampling results to analyze the performance of each feature across different sampling intervals. Specifically, the performance difference between models when each feature is selected and when it is not can be statistically analyzed. For example, for the user age feature, the average difference in model accuracy between including and excluding this feature can be calculated. Assuming that in 500 rounds of sampling, the average accuracy is 0.835 for the 300 rounds including the user age feature and 0.792 for the 200 rounds excluding it, then the performance contribution difference for this feature is 0.043.

[0056] Calculate the contribution of each feature dimension across different sampling intervals, and for each sampling interval, calculate the relative importance of the feature. In practice, the sampling interval can be divided into multiple sub-intervals, such as [0.1, 0.3], [0.3, 0.5], [0.5, 0.7], and [0.7, 0.9], and the contribution of the feature in each sub-interval is calculated separately. For example, if the contribution of the user age feature in these four sub-intervals is 0.025, 0.038, 0.051, and 0.060 respectively, it indicates that this feature contributes more when the sampling ratio is higher (the feature is more complete).

[0057] The entropy value of the sampling probability distribution is calculated. The entropy value reflects the uncertainty in the sampling process and can be used to adjust the feature importance score. For example, if the sampling probability of the user gender feature is consistently around 0.9, it indicates that the feature is highly valued, and its sampling probability distribution has a low entropy value; while if the sampling probability of the product price feature fluctuates between 0.4 and 0.6, it indicates that the importance of the feature is uncertain, and its entropy value is high.

[0058] By combining contribution and entropy values, a feature importance score is calculated using a non-linear weighted average. In practice, the inverse relationship between contribution and entropy can be adopted, meaning features with high contribution and low entropy receive higher importance scores. For example, the average contribution of the user age feature is 0.0435, and the entropy of its sampling probability distribution is 0.15, so its importance score can be calculated as 0.0435 / (0.15+0.1)×10=1.74; while the average contribution of the user gender feature is 0.0210, and the entropy is 0.32, so its importance score is 0.0210 / (0.32+0.1)×10=0.50, indicating that the user age feature is more important than the gender feature.

[0059] The above steps complete the entire process of importance scoring for multidimensional feature tensors based on the Monte Carlo iterative sampling method. This method can effectively identify feature dimensions that have a key impact on the target task, providing a scientific basis for feature selection and model optimization.

[0060] In one optional implementation, the dimensionality-reduced feature data is input into a pre-trained dual-path parallel verification network. A bundle search prediction value is obtained by performing purity prediction on the dimensionality-reduced feature data using a bundle search path, and a greedy search prediction value is obtained by performing purity verification on the dimensionality-reduced feature data using a greedy search path. Calculate the statistical and distribution characteristics of the dimensionality-reduced feature data, and construct a feature scoring standard based on the statistical and distribution characteristics; perform dual-path parallel computation on the dimensionality-reduced feature data, wherein the beam search path constructs a hierarchical search space by setting the search width and search depth, retains candidate results that meet the feature scoring standard at each search level, calculates the state transition matrix based on the historical search path, predicts the search direction of the next layer based on the state transition matrix, and uses the purity calculation result corresponding to the optimal search path as the beam search prediction value; In the greedy search path, the dimensionality-reduced feature data is validated hierarchically based on the feature scoring criteria. At each validation layer, a feature combination scheme is determined through local optimal selection, and the feature combination scheme is input into the validation function to calculate the purity index. An adaptive step size strategy is used to dynamically adjust the validation direction until the purity index converges to obtain the greedy search prediction value.

[0061] After receiving the dimensionality-reduced feature data, the dual-path parallel verification network immediately performs statistical feature analysis on the input data. The statistical feature analysis process includes calculating basic statistics such as the mean, variance, skewness, and kurtosis of the feature vectors. The mean is calculated by summing all feature values ​​and dividing by the number of feature dimensions, reflecting the central tendency of the feature data. The variance is calculated by dividing the sum of the squares of the differences between each feature value and the mean by the number of feature dimensions minus one, measuring the dispersion of the feature data. Skewness is calculated by summing the cubes of the differences between the feature values ​​and the mean, reflecting the asymmetry of the data distribution. Kurtosis is calculated by summing the fourth power of the differences between the feature values ​​and the mean, reflecting the sharpness of the data distribution.

[0062] The distribution characteristic analysis process is achieved by constructing the probability density function and cumulative distribution function of the feature data. The probability density function is obtained by dividing the feature value space into several intervals and counting the frequency of feature values ​​in each interval. The cumulative distribution function is obtained by calculating the cumulative probability value less than or equal to a specific feature value. Distribution characteristics also include quantile analysis of feature values, which calculates the 25th, 50th, and 75th quantiles after arranging the feature values ​​in ascending order, reflecting the distribution interval characteristics of the data.

[0063] The feature scoring criteria are constructed based on a comprehensive consideration of statistical and distributional characteristics. The scoring criteria include feature importance weight calculation, which determines the weight value by analyzing the contribution of each feature dimension to the target purity prediction. The weight calculation employs the information gain method, determining the importance of a feature by calculating the improvement in purity prediction accuracy after adding a specific feature. Feature correlation analysis identifies redundant features and adjusts weight allocation by calculating the Pearson correlation coefficient between features. Feature stability evaluation ensures that the selected features have good generalization ability by analyzing the coefficient of variation of feature values ​​across different data subsets.

[0064] The hierarchical search space construction process for the bundle search path includes setting parameters for search width and search depth. The search width determines the number of candidate results retained at each search level, typically set to 3 to 10. The search depth determines the number of levels in the search process and is adjusted based on feature dimensions and computational resource constraints. At each search level, the highest-scoring candidate results are retained by evaluating the scores of candidate feature combinations. The score calculation for candidate results involves inputting feature combinations into a scoring function, comprehensively considering feature importance, relevance, and stability to obtain a final score.

[0065] The state transition matrix calculation process is based on statistical analysis of historical search paths. Historical search path records include feature selection decisions for each search, corresponding scoring results, and the final purity prediction accuracy. The state transition matrix is ​​obtained by analyzing the probability distribution of transitions from the current feature combination state to the next feature combination state. The transition probability calculation considers the success rate of similar feature combinations in historical paths; transition paths with higher success rates receive higher transition probabilities. The state transition matrix also includes an exploration factor to balance the use of historical experience with the exploration of new paths.

[0066] The prediction of the next-level search direction for the bundle search path is achieved through matching analysis between the state transition matrix and the current candidate results. The prediction process takes each candidate result at the current level as a state input and calculates the probability of reaching each state at the next level using the state transition matrix. The probability calculation results guide the selection of the next-level search direction, prioritizing the exploration of feature combinations with high transition probabilities. The selection of the search direction also considers feature diversity constraints to avoid the search process becoming overly focused on similar feature combinations.

[0067] The generation of bundle search predictions is achieved through purity calculation of the optimal search path. The optimal search path is determined by comparing the comprehensive scores of all complete search paths. The comprehensive score considers the cumulative scores of candidate results at each level along the path and the path length penalty. The purity calculation process inputs the feature combination corresponding to the optimal path into the purity prediction model. The prediction model learns the mapping relationship between the feature combination and the purity index through a neural network structure. The output of the purity prediction model is processed by an activation function to obtain a purity probability value between 0 and 1.

[0068] The layered, progressive validation process of the greedy search path starts with an initial feature set and filters and combines features layer by layer. Each validation layer determines the best feature combination for the current layer through a local optimum selection strategy. The local optimum selection process evaluates the comprehensive score of the currently available features combined with the selected features, and selects the feature with the highest score to add to the current combination. The generation of feature combination schemes considers the interaction effects between features, and adjusts the combination strategy by calculating the synergistic effect strength between feature pairs.

[0069] The calculation process for the purity index of the validation function includes evaluation of the effectiveness of feature combinations and evaluation of the accuracy of purity prediction. Effectiveness evaluation is achieved by calculating the information entropy and Gini impurity index of the feature combinations. Information entropy reflects the information content of the feature combinations, and Gini impurity reflects the classification purity of the feature combinations. The accuracy of purity prediction is evaluated by inputting the feature combinations into the validation dataset and calculating the mean squared error and correlation coefficient between the predicted results and the true purity values.

[0070] The adaptive step-size strategy's dynamic adjustment mechanism determines the adjustment direction based on the changing trend of the purity index during the verification process. When the purity index continues to improve, the step size remains at its current value or is moderately increased to accelerate convergence. When the purity index fluctuates or deteriorates, the step size is decreased to improve search accuracy. Step-size adjustment also considers gradient information, determining the adjustment magnitude by calculating the sensitivity of the purity index to feature selection.

[0071] The convergence of the greedy search prediction is determined by the change in the purity index across multiple consecutive validation layers. The algorithm is considered convergent when the improvement in the purity index across three consecutive validation layers is less than a preset threshold. This convergence threshold is set considering computational accuracy requirements and resource constraints, and is typically between 0.001 and 0.01. The converged greedy search prediction is obtained by calculating the purity of the current optimal feature combination; this result serves as the final output of the greedy search path.

[0072] In one optional implementation, dual-path parallel computation is performed on the dimensionality-reduced feature data, wherein the beam search path constructs a hierarchical search space by setting the search width and search depth, and at each search level, candidate results that meet the feature scoring criteria are retained, including: Calculate the dimensionality distribution characteristics of the dimensionality-reduced feature data, and divide the dimensionality-reduced feature data into a bundle search channel and a parallel search channel; in the bundle search channel, determine the initial value of the search width and the initial value of the depth according to the dimensionality distribution characteristics, and construct a multi-layer search space structure based on the initial value of the search width and the initial value of the depth; after each layer of search is completed, record the feature combination method of the search results of that layer to form a search history. Perform feature correlation analysis on the search history, calculate the score value of each candidate feature result based on the result of the feature correlation analysis, and filter and retain candidate results whose score values ​​exceed a preset score threshold; A feature state transition matrix is ​​constructed based on the search history. The feature transition probability is calculated using the feature state transition matrix. The search direction of the next search layer is determined based on the feature transition probability. The candidate results passed to each layer according to the transfer direction are sorted in descending order according to the score value. The feature combination ranked first is selected as the optimal feature combination of that layer. When the search reaches the last layer, the optimal feature combination of the last layer is used as the output result of the bundle search channel.

[0073] like Figure 2 As shown, the method includes: After the dimensionality-reduced feature data enters the dual-path parallel computing module, dimensionality distribution feature analysis is immediately performed. Dimensionality distribution feature analysis is achieved by calculating the numerical distribution of the feature vectors across each dimension, including the numerical range, numerical density, and numerical clustering degree for each dimension. The numerical range is obtained by calculating the difference between the maximum and minimum values ​​for each dimension, reflecting the data span of that dimension. The numerical density is obtained by dividing the numerical interval of each dimension into several sub-intervals and counting the frequency of values ​​occurring in each sub-interval, reflecting the degree of concentration of values ​​in that dimension. The degree of numerical clustering is obtained by calculating the clustering coefficient of the values ​​within a dimension; a high clustering coefficient indicates that the data has significant clustering characteristics in that dimension.

[0074] Based on the dimensionality distribution characteristic analysis, the dimensionality-reduced feature data is divided into bundle search channels and parallel search channels according to dimensional characteristics. This division is achieved by setting a dimensional separation threshold. When the clustering coefficient of a dimension exceeds the separation threshold and the numerical density distribution exhibits multimodal characteristics, that dimension and its related dimensions are assigned to the bundle search channel. The separation threshold is set based on the overall distribution characteristics of the feature data, typically as the average clustering coefficient plus a standard deviation. Dimensions with unimodal distribution characteristics or low clustering coefficients are assigned to the parallel search channel. Channel allocation also considers the correlation between dimensions. The correlation strength between dimensions is determined by calculating the mutual information between dimensions, and dimensions with high correlation strength are assigned to the same channel.

[0075] The initial value of the search width for a beam search channel is determined based on the number of dimensions allocated to that channel and the complexity of the dimension distribution. The initial search width is obtained by multiplying the number of dimensions within the channel by a complexity adjustment factor, which is determined according to the strength of the multimodal characteristics of the dimension distribution. When the dimension distribution exhibits strong multimodal characteristics, a larger complexity adjustment factor is used, typically between 1.5 and 2.0. When the multimodal characteristics of the dimension distribution are weak, a smaller complexity adjustment factor is used, typically between 1.0 and 1.3. The initial search width is also constrained by computational resource limitations; when computational resources are limited, the initial search width is restricted to a reasonable range.

[0076] The initial search depth is determined based on the hierarchical complexity of the feature data and the expected search accuracy requirements. Hierarchical complexity is obtained by analyzing the depth of dependencies between features, which reflects the number of feature stacking levels required for effective prediction results. The initial search depth is typically set to 1.2 to 1.5 times the dependency depth to ensure the search process covers all important feature combination paths. The search depth is also constrained by convergence requirements; an excessively deep search will lead to excessively high computational complexity and fail to converge within a reasonable timeframe.

[0077] The construction of the multi-layered search space structure is achieved by expanding the search nodes layer by layer. Each layer contains several search nodes, and each search node represents a feature combination method. The first layer of the search space contains all individual features as initial search nodes. The second layer of the search space is formed by combining the features of the first layer with other features pairwise, taking into account the compatibility constraints between features. Feature compatibility is determined by calculating the covariance and conditional entropy between features; feature combinations with excessively high covariance or excessively low conditional entropy are excluded. Subsequent layers of the search space are formed by adding new features based on the feature combinations of the previous layer, following the principle of the effectiveness of feature combinations.

[0078] The search history is generated by recording detailed information about each layer of search results. This information includes the feature combination method for that layer, the weight allocation of each feature in the combination, the combined score, and the prediction accuracy metric. The feature combination method is represented by a sequence of feature indexes, and the weight allocation is represented by a normalized weight vector. The score includes a comprehensive score for the combination and individual scores for each evaluation dimension. The prediction accuracy metric is obtained by applying the feature combination to a validation dataset and includes prediction error and correlation coefficient.

[0079] The feature correlation analysis process is based on the performance of feature combinations in search history. Correlation analysis calculates the synergistic effect strength of each feature with other features, obtained by comparing the predictive performance difference between using features alone and using them in combination. When the predictive performance of a feature combination is significantly better than the sum of the performance of individual features, a positive synergistic effect is considered to exist between the features. Correlation analysis also includes identifying competition effects between features; when combinations of certain features lead to a decrease in predictive performance, a competition effect is considered to exist between the features.

[0080] The scoring process for candidate features comprehensively considers the feature's independent contribution, synergistic effect strength, and stability. Independent contribution is measured by the predictive accuracy when used alone; features with high contribution receive higher base scores. Synergistic effect strength is measured by the performance improvement when combined with other features; features with strong synergistic effects receive synergistic bonuses. Stability is measured by the performance variation of the feature across different data subsets; features with good stability receive stability bonuses.

[0081] The scoring threshold is set based on the distribution characteristics of candidate feature scores. The threshold is determined by calculating the quantiles of all candidate feature scores, typically set to the 75th or 80th quantile. Threshold setting also considers search width constraints; when the number of candidate features is too large, the scoring threshold is increased to reduce the number of retained candidate results. When the number of candidate features is too small, the scoring threshold is appropriately decreased to ensure the effectiveness of the search process.

[0082] The construction of the feature state transition matrix is ​​based on state transition statistics from the search history. The matrix records the frequency and success rate of transitions from one feature combination state to another. Transition frequency is obtained by counting the number of times the same state transition occurs in historical searches. Transition success rate is calculated by determining the probability of an increase in the feature combination score after the transition. The state transition matrix also includes transition cost information, reflecting the computational resources and time costs required for state transitions.

[0083] The calculation of feature transition probabilities is based on statistical information in the state transition matrix and the current search state. The transition probability is obtained by multiplying the transition frequency by the success rate and then dividing by the transition cost to obtain a normalized probability value. The probability calculation also considers the balance between exploration and exploitation, introducing an exploration factor to encourage the search process to explore transition paths that are less frequently encountered in historical records. The exploration factor is dynamically adjusted as the search progresses; a larger exploration factor in the early stages of the search encourages extensive exploration, while a smaller exploration factor in the later stages focuses on exploiting discovered high-quality paths.

[0084] The process of determining the search direction for the next search layer is achieved by ranking and selecting transition probabilities. All candidate feature combinations from the current layer are used as starting states, and the transition probability from each starting state to each target state in the next layer is calculated. The target states with the highest transition probabilities are selected as the search directions for the next layer. The number of search directions selected is limited by the search width; when the transition probabilities are relatively even, more search directions are selected to maintain search diversity.

[0085] The descending sorting process for candidate results is based on a comparison of score values. This sorting considers multiple dimensions of the score values. When the overall scores are the same, the sorting order is determined by comparing secondary indicators such as prediction accuracy, stability, and computational efficiency. The sorting result guides the selection of the optimal feature combination; the feature combination ranked first is determined as the optimal feature combination for that layer. Information about the optimal feature combination is recorded in the search history, providing a reference for subsequent layers. When the search reaches the last layer, the optimal feature combination of that last layer becomes the final output of the entire bundle search channel.

[0086] The method further includes: During sample pretreatment, the palladium-copper-containing material to be analyzed is thoroughly digested using an acidic solution system to ensure complete dissolution of palladium and copper elements, which exist in ionic form in the solution. After digestion, the solution is transferred to a volumetric flask for final volume adjustment. The volume is adjusted according to the estimated concentration of palladium and copper in the sample, typically to 100 mL or 250 mL. After final volume adjustment, the solution is shaken well and allowed to stand to ensure a uniform concentration distribution.

[0087] Accurately transfer a specific volume of the test solution from the diluted sample solution into an Erlenmeyer flask. The volume selected depends on the palladium-copper content range in the sample; for samples with higher palladium content, a smaller volume can be transferred, while for samples with lower content, a larger volume is required to ensure measurement accuracy. The transfer operation should be performed using a precision pipette or pipette to guarantee the accuracy of the volume measurement.

[0088] Add disodium ethylenediaminetetraacetate (EDTA) standard titration solution to the conical flask. Disodium EDTA acts as a chelating agent, forming stable chelates with palladium and copper ions. The volume of disodium EDTA solution added should be sufficient to completely complex all palladium and copper ions in the sample solution, with a suitable excess, controlled within approximately 10 mL. The concentration of the disodium EDTA standard solution is typically prepared at 0.05 mol / L, and the accuracy of the concentration is calibrated using standardized titration.

[0089] The pH of the solution was adjusted using a 10% sodium hydroxide solution, added dropwise to the test solution using a dropper. During the adjustment process, the solution was continuously stirred, and the pH was monitored using a pH meter. The pH was precisely adjusted to a range of 6.0 to 6.5, which is the optimal pH range for the stable existence of the palladium-copper complex. After pH adjustment, 3 drops of 0.5% dimethylphenol orange indicator were added to the solution.

[0090] The first zinc chloride standard titration is performed. The concentration of the zinc chloride standard titration solution is prepared according to the palladium-copper content in the sample, and the commonly used concentration is 0.02 mol / L to 0.05 mol / L. During the titration, the solution is continuously stirred, and the color change is observed. The titration endpoint is reached when the solution color changes from yellow to red. At this point, the titration is stopped, and the volume of zinc chloride standard titration solution consumed is recorded. This step determines the amount of excess disodium ethylenediaminetetraacetate that is replaced by zinc chloride after binding with the indicator.

[0091] Immediately after the first titration, 15 mL of an ammonia-ammonium chloride buffer solution with a pH of approximately 10 was added to the solution. The buffer solution alters the acid-base environment of the solution, causing the complex of palladium ions and disodium ethylenediaminetetraacetate to dissociate under the new pH conditions. The ammonia-ammonium chloride buffer solution is prepared by mixing ammonia water and ammonium chloride in a specific ratio to ensure that the buffer solution has a stable pH value and good buffering capacity.

[0092] Immediately after adding the buffer solution, a second zinc chloride standard titration was performed. Due to the change in pH, the disodium ethylenediaminetetraacetate (EDTA) that was previously bound to palladium ions was released again. This released disodium EDTA then bound to the indicator again, requiring additional zinc chloride for titration. The color change of the solution was observed during the titration process. The endpoint of the second titration was reached when the solution color changed from yellow to red. The volume of zinc chloride standard titration solution consumed in the second titration was recorded.

[0093] The palladium content is calculated based on the volume of zinc chloride consumed in the second titration, which directly corresponds to the number of moles of palladium ions in the sample. The palladium content is calculated by multiplying the volume of the consumed zinc chloride standard titration solution by its molar concentration, then by the atomic mass of palladium (106.4), dividing by the sample mass, multiplying by the dilution factor, and finally multiplying by 100%.

[0094] Calculating the copper content requires considering both the total amount of disodium ethylenediaminetetraacetate (EDTA) added and the total consumption of the two zinc chloride titrations. Multiply the volume of the EDTA standard titration solution by its molar concentration, then subtract the sum of the volumes consumed in the two zinc chloride titrations multiplied by the zinc chloride molar concentration to obtain the number of moles of EDTA bound to copper ions. Multiply this number by the atomic mass of copper (63.55), divide by the sample mass, multiply by the dilution factor, and finally multiply by 100% to obtain the copper mass percentage.

[0095] In actual measurements, samples containing ammonium ions require special treatment, as ammonium ions can interfere with the formation of palladium-copper complexes and the determination of titration endpoints. The treatment method involves evaporating the sample solution to near dryness under heating conditions to remove the ammonium ions, and then redissolving the sample for subsequent measurement operations.

[0096] When the sample contains heavy metal ions such as iron and aluminum, these ions will form competitive complexes with disodium ethylenediaminetetraacetate, affecting the accuracy of palladium-copper determination. To eliminate this interference, 3.0 g of ammonium fluoride is added to the sample as a masking agent before titration. Ammonium fluoride can form stable fluoride complexes with heavy metal ions such as iron and aluminum, thereby eliminating their interference with palladium-copper determination.

[0097] Key control points of the entire assay include precise pH adjustment, accurate determination of the titration endpoint, and precise calibration of various standard solution concentrations. The stirring speed during titration must be kept moderate; excessively rapid stirring will affect the observation of indicator color changes, while excessively slow stirring will lead to incomplete reactions. All operations should be performed at room temperature to avoid the influence of temperature changes on the complexation reaction equilibrium. This method is applicable to the analysis of various palladium-copper containing materials, including alloys, composite materials, and catalyst supports, covering a range of needs from trace analysis to macro analysis.

[0098] A second aspect of the present invention provides an intelligent analysis system for the purity of palladium-copper materials based on collaborative detection, comprising: The first unit is used to acquire the spectral data and physicochemical parameters of the palladium-copper material to be tested, establish a spectral curve based on the spectral data, calculate the basic characteristic values ​​of the palladium-copper material based on the physicochemical parameters, and use the spectral curve and the basic characteristic values ​​as data to be analyzed. The second unit is used to perform feature extraction using a deep convolutional neural network. It performs multi-layer convolution operations on the data to be analyzed to extract spectral feature vectors, and fuses the spectral feature vectors with the basic feature values ​​through adaptive weights to obtain a multi-dimensional feature tensor. It then performs dimensionality reduction processing on the multi-dimensional feature tensor based on the Monte Carlo iterative sampling method to obtain dimensionality-reduced feature data. The third unit is used to input the dimensionality-reduced feature data into a pre-trained dual-path parallel verification network, perform purity prediction on the dimensionality-reduced feature data through a beam search path to obtain a beam search prediction value, and perform purity verification on the dimensionality-reduced feature data through a greedy search path to obtain a greedy search prediction value. When the difference between the beam search prediction value and the greedy search prediction value is less than a preset prediction threshold, the weighted average of the two prediction values ​​is calculated based on preset weights as the final purity analysis result of the palladium-copper material. When the difference is greater than or equal to the preset prediction threshold, anomaly detection is triggered and a warning message is output.

[0099] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0100] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0101] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart method for purity analysis of palladium-copper materials based on collaborative detection, characterized in that, include: Acquire the spectral data and physicochemical parameters of the palladium-copper material to be tested, establish a spectral curve based on the spectral data, calculate the basic characteristic values ​​of the palladium-copper material based on the physicochemical parameters, and use the spectral curve and the basic characteristic values ​​as the data to be analyzed. A deep convolutional neural network is used to perform feature extraction. Multi-layer convolution operations are performed on the data to be analyzed to extract spectral feature vectors. The spectral feature vectors are then fused with the basic feature values ​​through adaptive weights to obtain a multi-dimensional feature tensor. The multi-dimensional feature tensor is then dimensionality-reduced based on the Monte Carlo iterative sampling method to obtain dimensionality-reduced feature data. The dimensionality-reduced feature data is input into a pre-trained dual-path parallel verification network. The purity of the dimensionality-reduced feature data is predicted by a beam search path to obtain a beam search prediction value. The purity of the dimensionality-reduced feature data is verified by a greedy search path to obtain a greedy search prediction value. When the difference between the beam search prediction value and the greedy search prediction value is less than a preset prediction threshold, the weighted average of the two prediction values ​​is calculated based on preset weights as the final purity analysis result of the palladium-copper material. When the difference is greater than or equal to the preset prediction threshold, anomaly detection is triggered and a warning message is output.

2. The method according to claim 1, characterized in that, A spectral curve is established based on the spectral data, and the fundamental characteristic values ​​of the palladium-copper material are calculated based on the physicochemical parameters. The spectral curve and the fundamental characteristic values ​​are used as data to be analyzed. The spectral data is divided into a training dataset and a validation dataset. Wavelet transform is performed on the training dataset to obtain multi-scale spectral features. Frequency domain information in the multi-scale spectral features is extracted by Fourier transform. The frequency domain information is combined with the original spectral data to construct a spectral curve. Based on the physicochemical parameters, a feature mapping relationship is constructed. The feature mapping relationship is optimized by a nonlinear optimization algorithm to obtain basic feature values ​​that reflect the physicochemical properties of palladium-copper materials. The basic feature values ​​are cross-validated with the validation dataset. When the validation error is less than a preset error threshold, the spectral curve and the basic feature values ​​are combined to form the data to be analyzed.

3. The method according to claim 1, characterized in that, The spectral feature vector and the basic feature value are fused using adaptive weights to obtain a multidimensional feature tensor. The multidimensional feature tensor is then dimensionality-reduced using a Monte Carlo iterative sampling method to obtain dimensionality-reduced feature data, including: Calculate the correlation coefficient between the spectral feature vector and the basic feature value, construct an adaptive weight calculation function based on the correlation coefficient, dynamically adjust the fusion weight between the spectral feature vector and the basic feature value through the adaptive weight calculation function, and perform tensor operations on the fusion weight to obtain a multidimensional feature tensor. The multidimensional feature tensor is repeatedly randomly sampled using the Monte Carlo iterative sampling method. Based on the sampling results, the feature importance score is calculated. The feature dimension with the highest feature contribution is selected according to the feature importance score, and the feature dimensions are combined to generate dimensionality-reduced feature data.

4. The method according to claim 3, characterized in that, The multidimensional feature tensor is repeatedly randomly sampled using the Monte Carlo iterative sampling method. Based on the sampling results, the feature importance score is calculated, including: The Monte Carlo method is used to repeatedly random sample the multidimensional feature tensor. The sampling probability distribution is dynamically adjusted in each round of sampling. When the variance of the sampling results of two adjacent rounds is less than the preset variance threshold, the sampling interval is increased and the sampling operation is repeated until the sampling results converge or the preset maximum number of samplings is reached. The converged sampling results are subjected to feature decomposition, and the contribution of each feature dimension in different sampling intervals is calculated. The contribution is combined with the entropy value of the sampling probability distribution, and the feature importance score is obtained by nonlinear weighting calculation.

5. The method according to claim 1, characterized in that, The dimensionality-reduced feature data is input into a pre-trained dual-path parallel validation network. The purity of the dimensionality-reduced feature data is predicted using a beam search path to obtain a beam search prediction value. The purity of the dimensionality-reduced feature data is verified using a greedy search path to obtain a greedy search prediction value, including: Calculate the statistical and distribution characteristics of the dimensionality-reduced feature data, and construct a feature scoring standard based on the statistical and distribution characteristics; perform dual-path parallel computation on the dimensionality-reduced feature data, wherein the beam search path constructs a hierarchical search space by setting the search width and search depth, retains candidate results that meet the feature scoring standard at each search level, calculates the state transition matrix based on the historical search path, predicts the search direction of the next layer based on the state transition matrix, and uses the purity calculation result corresponding to the optimal search path as the beam search prediction value; In the greedy search path, the dimensionality-reduced feature data is validated hierarchically based on the feature scoring criteria. At each validation layer, a feature combination scheme is determined through local optimal selection, and the feature combination scheme is input into the validation function to calculate the purity index. An adaptive step size strategy is used to dynamically adjust the validation direction until the purity index converges to obtain the greedy search prediction value.

6. The method according to claim 5, characterized in that, The dimensionality-reduced feature data is subjected to dual-path parallel computation, wherein the beam search path constructs a hierarchical search space by setting the search width and search depth, and retains candidate results that meet the feature scoring criteria at each search level, including: Calculate the dimensionality distribution characteristics of the dimensionality-reduced feature data, and divide the dimensionality-reduced feature data into a bundle search channel and a parallel search channel; in the bundle search channel, determine the initial value of the search width and the initial value of the depth according to the dimensionality distribution characteristics, and construct a multi-layer search space structure based on the initial value of the search width and the initial value of the depth; after each layer of search is completed, record the feature combination method of the search results of that layer to form a search history. Perform feature correlation analysis on the search history, calculate the score value of each candidate feature result based on the result of the feature correlation analysis, and filter and retain candidate results whose score values ​​exceed a preset score threshold; A feature state transition matrix is ​​constructed based on the search history. The feature transition probability is calculated using the feature state transition matrix. The search direction of the next search layer is determined based on the feature transition probability. The candidate results passed to each layer according to the transfer direction are sorted in descending order according to the score value. The feature combination ranked first is selected as the optimal feature combination of that layer. When the search reaches the last layer, the optimal feature combination of the last layer is used as the output result of the bundle search channel.

7. A palladium-copper material purity intelligent analysis system based on collaborative detection, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to acquire the spectral data and physicochemical parameters of the palladium-copper material to be tested, establish a spectral curve based on the spectral data, calculate the basic characteristic values ​​of the palladium-copper material based on the physicochemical parameters, and use the spectral curve and the basic characteristic values ​​as data to be analyzed. The second unit is used to perform feature extraction using a deep convolutional neural network. It performs multi-layer convolution operations on the data to be analyzed to extract spectral feature vectors, and fuses the spectral feature vectors with the basic feature values ​​through adaptive weights to obtain a multi-dimensional feature tensor. It then performs dimensionality reduction processing on the multi-dimensional feature tensor based on the Monte Carlo iterative sampling method to obtain dimensionality-reduced feature data. The third unit is used to input the dimensionality-reduced feature data into a pre-trained dual-path parallel verification network, perform purity prediction on the dimensionality-reduced feature data through a beam search path to obtain a beam search prediction value, and perform purity verification on the dimensionality-reduced feature data through a greedy search path to obtain a greedy search prediction value. When the difference between the beam search prediction value and the greedy search prediction value is less than a preset prediction threshold, the weighted average of the two prediction values ​​is calculated based on preset weights as the final purity analysis result of the palladium-copper material. When the difference is greater than or equal to the preset prediction threshold, anomaly detection is triggered and a warning message is output.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images