One-dimensional spectrum recognition method and system based on enhanced feature engineering and hierarchical classifier

Abstract

The present application relates to the field of spectral data processing and automatic identification, in particular to a one-dimensional spectrum identification method and system based on enhanced feature engineering and hierarchical classifier, the method first extracts multi-dimensional enhanced features containing morphological features and Chevron pattern features from one-dimensional spectrum data to construct a feature vector; then a hierarchical classifier containing multi-level binary classification nodes is constructed to simplify the classification logic by using the category hierarchy relationship; then the classification threshold is optimized based on the reachable sample set of each node to improve the discrimination accuracy; finally, the feature vector is input into the hierarchical classifier, the predicted probability is compared with the optimal threshold according to the node order, the classification decision is completed, the spectral details are fully represented by the enhanced features, the hierarchical architecture and adaptive threshold are combined, the problems of insufficient feature extraction, low classification accuracy and misjudgment of similar patterns in traditional methods are effectively solved, the recognition accuracy and reasoning efficiency are considered, and the method is suitable for one-dimensional spectrum automatic identification in quantum computing, spectral analysis and other scenes.

Description

Technical Field

[0001] This invention relates to the field of spectral data processing and automatic identification technology, specifically to a one-dimensional spectral identification method and system based on enhanced feature engineering and hierarchical classifiers. Background Technology

[0002] In fields such as quantum computing, spectral analysis, materials science, and biomedicine, the automatic classification and identification of one-dimensional spectral data is one of the core technologies. With the explosive growth of experimental data, traditional manual interpretation methods can no longer meet the demands for efficient and accurate classification. For example, in superconducting quantum computing, it is necessary to automatically identify qubit excitation spectra, dispersive spectra, etc., to achieve automated calibration and parameter optimization; similarly, in chemistry and biomedicine, the automatic identification of one-dimensional spectra such as infrared spectroscopy, Raman spectroscopy, and fluorescence spectroscopy has a wide range of applications.

[0003] Existing one-dimensional spectral classification techniques have the following main drawbacks: First, feature extraction is insufficient. Traditional methods typically rely on simple statistical features (such as mean, variance, and maximum value) or a few manually designed features, failing to effectively capture the complex morphological features (such as peak symmetry, sharpness, and interpeak distance), frequency domain features (such as power spectrum distribution and frequency centroid), and specific mode features (such as the Chevron mode, i.e., a "V"-shaped peak pair structure) inherent in one-dimensional spectra. This results in a lack of sufficient discriminative information between spectral modes with similar morphology but different categories, such as "single peak," "single peak Chevron," and "single peak + single peak Chevron," leading to misclassification and low classification accuracy.

[0004] Second, the classifier architecture is simplistic. Existing technologies generally employ a single multi-classifier to directly classify all categories. This flat architecture fails to utilize the inherent hierarchical relationships between categories, such as "peakless / peaked," "unimodal / multimodal," and "whether it follows the Chevron pattern." On the one hand, similar categories compete with each other within the same decision space, increasing the risk of misclassification; on the other hand, all samples, regardless of their simplicity or complexity, must undergo a complete classification process, resulting in low computational efficiency.

[0005] Third, the threshold setting lacks adaptability. Most methods use a fixed threshold (such as 0.5) for binary classification decisions. However, in a hierarchical classification architecture, the sample distributions faced by different nodes differ significantly, and a fixed threshold cannot adapt to the actual data distribution of each node, resulting in the overall classification performance not reaching its optimal level.

[0006] Fourth, it is difficult to balance computational efficiency and accuracy. High-precision deep learning models are often computationally complex and difficult to meet the needs of real-time processing; while fast models usually sacrifice accuracy, making it difficult to achieve a good balance between the two.

[0007] Therefore, there is an urgent need for a one-dimensional spectral recognition method and system that can fully extract one-dimensional spectral features, utilize hierarchical relationships of categories, adaptively optimize decision thresholds, and balance accuracy and efficiency. Summary of the Invention

[0008] The purpose of this invention is to provide a one-dimensional spectral recognition method and system based on enhanced feature engineering and hierarchical classifiers to solve the above-mentioned problems.

[0009] This invention is achieved through the following technical solution: The first aspect of this application provides a one-dimensional spectral recognition method based on enhanced feature engineering and hierarchical classifiers, comprising the following steps: Step 1: Extract multi-dimensional enhancement features from the one-dimensional spectral data to be identified to form a feature vector. The feature vector includes at least morphological features for describing peak morphology and Chevron mode features for describing peak pair patterns. Step 2: Construct a hierarchical classifier, which contains multiple binary classification nodes arranged in a hierarchical order of one-dimensional spectral categories, with each node corresponding to a binary classification model; Step 3: For each node in the hierarchical classifier, using only the training sample set that can reach that node, optimize to obtain a classification threshold corresponding to that node that makes the preset evaluation index optimal. Step 4: Input the feature vector into the hierarchical classifier. According to the hierarchical order of the nodes, compare the predicted probability output by the current node with the classification threshold corresponding to the node. Determine the category of the current node based on the comparison result, and decide whether to output the final classification result or move to the next node.

[0010] In one optional embodiment, the morphological features include one or more of the following: peak number, maximum peak height, peak height ratio, peak symmetry score, peak sharpness score, peak width statistics, and interpeak distance statistics; the Chevron pattern features include one or more of the following: peak pair symmetry score, peak pair separation, peak pair height ratio, and Chevron angle.

[0011] In one optional embodiment, the constructed hierarchical classifier is a five-node hierarchical classifier, with the node hierarchy as follows: N0 node: Determine if there is a peak; N1 node: Determines whether it is a single peak, and only executes when N0 node determines that there is a peak; N2 node: Determines whether it is a single-peak Chevron mode, and executes only when N0 node is determined to be peak and N1 node is determined to be non-single-peak; N3 node: Determines whether it is a bimodal node, and only executes when N0 node is a peak, N1 node is not a single peak, and N2 node is not a single peak Chevron. Node N4: Determines whether it is a single-peak + single-peak Chevron mode. It is only executed when node N0 is peaked, node N1 is not single-peaked, node N2 is not single-peaked Chevron, and node N3 is not bimodal.

[0012] In an optional embodiment, step three, optimizing the classification threshold corresponding to each node, specifically involves: for each node, collecting the predicted probability of the binary classification model corresponding to that node on its reachable sample set, searching within a preset classification threshold range with a preset step size, calculating a preset evaluation index for each classification threshold obtained from the search, wherein the preset evaluation index is classification accuracy or F1 score, and determining the classification threshold that makes the preset evaluation index reach its optimal value as the classification threshold corresponding to that node.

[0013] In an optional embodiment, the extracted multi-dimensional enhancement features further include at least one of a basic statistical feature set, a frequency domain feature set, and a signal quality feature set; the basic statistical feature set includes mean, standard deviation, skewness, and kurtosis; the frequency domain feature set includes total power, dominant frequency, frequency centroid, and frequency bandwidth; and the signal quality feature set includes signal-to-noise ratio.

[0014] In one optional embodiment, the binary classification model is automatically trained using an automated machine learning framework, which takes the feature vector as input and outputs binary classification labels based on node hierarchy, with a training time limit of 300-600 seconds.

[0015] A second aspect of this application provides a one-dimensional spectral recognition system based on enhanced feature engineering and a hierarchical classifier, for implementing the method described in any of the above claims, the system comprising: Enhanced feature extraction module: used to extract multi-dimensional enhanced features from one-dimensional spectral data to form feature vectors; Hierarchical classifier module: Used to store and run a hierarchical classifier containing multiple binary classification nodes arranged in hierarchical order of categories; Threshold optimization module: used to optimize and store the corresponding classification threshold for each node in the hierarchical classifier based on its reachable sample set; Serial decision reasoning module: It is used to input the feature vector into the hierarchical classifier, compare the predicted probability output by the current node with its corresponding classification threshold according to the node order, and output the final recognition result.

[0016] In one alternative embodiment, the binary classification model in the hierarchical classifier module includes a CatBoost model or a LightGBM model.

[0017] In an optional embodiment, the system further includes a model update module, which, after obtaining newly labeled training samples, performs a threshold search on the reachable sample set of each node to update the classification threshold of each node, and optionally retrains the binary classification model of each node.

[0018] In one optional embodiment, the multi-dimensional enhancement features extracted by the enhancement feature extraction module include at least a 45-dimensional feature vector, which contains five types of features: basic statistical features, peak morphology features, frequency domain features, Chevron mode features, and signal quality features.

[0019] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) This invention extracts multi-dimensional enhancement features, especially introducing peak morphological features (such as peak symmetry score, peak sharpness score, peak width statistics, and inter-peak distance statistics) and Chevron pattern features (such as peak pair symmetry score, peak pair separation, and Chevron angle), to comprehensively characterize one-dimensional spectral data from multiple dimensions such as time domain, frequency domain, morphology, and specific patterns, effectively solving the problem of insufficient feature extraction in traditional methods. On this basis, a five-node hierarchical classifier is constructed and serial decision-making is performed according to the hierarchy of "no peak → single peak → single peak Chevron → double peak → single peak + single peak Chevron → multi-peak", making full use of the natural hierarchical relationship between categories and avoiding direct competition between similar categories.

[0020] (2) The present invention adopts a hierarchical serial decision architecture, in which most simple samples (such as those without peaks or with single peaks) can be classified and exited in the early nodes without having to go through the calculation of all nodes. At the same time, the binary classification model of each node can adopt fast inference models such as CatBoost or LightGBM, further reducing the processing time of a single sample.

[0021] (3) For each node in the hierarchical classifier, this invention uses only the training sample set (i.e., reachable samples) that can reach the node, and uses classification accuracy or F1 score as the evaluation index to search for the optimal classification threshold specific to the node within a preset threshold range, replacing the traditional fixed threshold. This reachable sample threshold optimization mechanism enables the decision boundary of each node to adapt to the actual sample distribution it faces, effectively overcoming the performance degradation problem caused by sample imbalance or data distribution drift, and improving the system's robustness to edge samples and noise interference.

[0022] (4) This invention uses an automatic machine learning framework to automatically train binary classification models for each node and sets a training time limit of 300 to 600 seconds, eliminating the need for tedious manual parameter tuning and lowering the technical application threshold. In addition, the system also includes a model update module. When newly labeled training samples are obtained, the system can re-search the threshold of the reachable sample set for each node and optionally retrain the model, enabling the system to continuously learn and adapt to changes in data distribution, thus ensuring the recognition accuracy under long-term operation.

[0023] (5) This invention ensures sufficient information through multi-dimensional enhanced feature extraction, guarantees processing efficiency through hierarchical serial decision-making and fast reasoning models, and improves decision accuracy through adaptive threshold optimization. The synergistic effect of these three aspects breaks the dilemma of difficulty in balancing accuracy and efficiency in existing technologies, and provides a high-precision, high-efficiency, and highly adaptive intelligent one-dimensional spectrum recognition solution for fields such as quantum bit spectral analysis and chemical spectrum recognition. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 The flowchart of the one-dimensional spectrum recognition method provided in this application; Figure 2 The overall flowchart of the one-dimensional spectrum recognition method provided in this application; Figure 3 A schematic diagram of the waveform of peakless one-dimensional spectral data; Figure 4 A schematic diagram of the waveform of a single-peak one-dimensional spectral data; Figure 5 A schematic diagram of the waveform of one-dimensional spectral data in unimodal Chevron mode; Figure 6 This is a schematic diagram of the waveform of a bimodal one-dimensional spectral data. Figure 7 A waveform diagram of one-dimensional spectral data in single-peak plus single-peak Chevron mode; Figure 8 A schematic diagram of the waveform of multi-peak one-dimensional spectral data; Figure 9 This is a flowchart of the multi-dimensional enhanced feature extraction process in the embodiments of this application; Figure 10 A flowchart illustrating the hierarchical relationship of the five-node hierarchical classifier provided in this application; Figure 11 This is a flowchart illustrating the optimization of reachable sample thresholds for each node of the hierarchical classifier in an embodiment of this application. Figure 12This is a schematic diagram of a one-dimensional spectral recognition system based on enhanced feature engineering and a hierarchical classifier provided in this application.

[0025] The attached diagram shows the markings and corresponding component names: 10. Enhanced Feature Extraction Module; 20. Hierarchical Classifier Module; 30. Threshold Optimization Module; 40. Serial Decision Reasoning Module. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments and accompanying drawings. The illustrative embodiments and descriptions of this invention are for illustrative purposes only and are not intended to limit the invention. It should be noted that this invention is already in the actual research and development stage.

[0027] Example 1 Traditional automatic classification and recognition techniques for one-dimensional spectral data typically rely on simple statistical features (such as mean, variance, and maximum value) or a few manually designed features, and use a single multi-classifier to directly classify all categories. These methods fail to effectively capture the complex morphological features, frequency domain features, and specific pattern features (such as Chevron patterns) inherent in one-dimensional spectra. This results in a lack of sufficient discriminative information between morphologically similar but different spectral patterns (e.g., single-peak, single-peak Chevron, and single-peak plus single-peak Chevron), easily leading to misclassification. Furthermore, the flat classification architecture fails to utilize the natural hierarchical relationships between categories, and the classification thresholds often use fixed values ​​without optimization for sample distribution at different nodes, making it difficult to meet the efficiency requirements of real-time processing while maintaining high accuracy.

[0028] To address the aforementioned problems, this application provides a one-dimensional spectral recognition method based on enhanced feature engineering and a hierarchical classifier, combined with... Figure 1 The steps and procedures shown are as follows Figure 2 The overall process shown includes the following steps.

[0029] Step S1: Extract multi-dimensional enhancement features from the one-dimensional spectral data to be identified, forming a feature vector. The feature vector includes at least morphological features describing peak shapes and Chevron mode features describing peak pair patterns. Multi-dimensional enhancement features refer to a set of features that comprehensively characterize the physical properties and geometric shape of the one-dimensional spectral data. Morphological features are obtained through peak detection and geometric analysis of the one-dimensional spectral data, used to quantify the number, height, width, symmetry, and sharpness of spectral peaks, distinguishing between peak patterns of different numbers and shapes. Chevron mode features are specifically extracted for common Chevron interference fringe patterns in qubit spectra, obtained by calculating parameters such as peak pair symmetry score, separation, height ratio, and Chevron angle, used to accurately identify spectra with specific interference structures. The feature vector generation process involves concatenating the above features into a fixed-dimensional numerical vector, which serves as input to the subsequent classification model. For example, morphological features may specifically include the number of peaks, maximum peak height, peak height ratio (the ratio of the highest peak to the second highest peak), peak symmetry score (calculated based on the symmetry of data points on the left and right sides of the peak), peak sharpness score (calculated based on the second derivative), and inter-peak distance statistics; Chevron pattern features may specifically include peak pair symmetry score, mean peak pair separation, peak pair height ratio, and Chevron angle. This multi-dimensional enhanced feature extraction process is as follows: Figure 9 As shown, this illustrates the entire process from inputting raw one-dimensional spectral data, through parallel processing branches including basic statistics, peak morphology analysis, frequency domain transformation, and pattern matching, ultimately converging to form a feature vector containing approximately 45 or more dimensions of features. This multi-dimensional feature extraction method significantly improves the ability to represent complex spectral patterns, particularly effectively distinguishing easily confused categories such as single peaks and single-peak Chevrons, providing a sufficient data foundation for subsequent high-precision classification.

[0030] Step S2: Construct a hierarchical classifier. The hierarchical classifier contains multiple binary classification nodes arranged in a hierarchical order according to the one-dimensional spectral categories. Each node corresponds to a binary classification model. The hierarchical classifier refers to a tree-like or chain-like sequential decision architecture, composed of multiple binary classification nodes arranged in a logical hierarchical order according to the one-dimensional spectral categories. Each binary classification node corresponds to an independent binary classification model, which can be a classifier built based on gradient boosting trees or neural networks. The principle behind the predicted probability output by the binary classification model is as follows: The model's original output is not a direct class label, but a real value representing the probability that a sample belongs to the positive class. For logistic regression or neural networks, this value is mapped to a probability between 0 and 1 using the Sigmoid function. For ensemble models such as gradient boosting trees, it is the sum of the predicted scores of all weak learners, which is also calibrated to a probability using transformation functions such as the Sigmoid function. Finally, the model uses this calibrated probability value as the predicted probability, providing a basis for subsequent flexible threshold decisions. The order of the binary classification nodes is determined based on the physical characteristics of the one-dimensional spectral data, typically following a judgment logic from macroscopic to microscopic and from simple to complex. For example, first determine whether there is a peak or not, then determine whether it is single-peaked or multi-peaked, and finally determine the specific pattern type. Each node's function is to divide its input samples into two classes, with one class directly outputting the final classification result, and the other class being passed to the next node for further judgment. For example, a hierarchical classifier can be constructed as a five-node structure: node N0 determines whether there is a peak; node N1 is executed only when N0 determines there is a peak, determining whether it is single-peaked; node N2 determines whether it is a single-peak Chevron pattern in the case of non-single-peaked conditions; node N3 determines whether it is double-peaked; and node N4 determines whether it is a single-peak plus single-peak Chevron pattern. The waveforms of one-dimensional spectral data with no peak, single peak, single-peak Chevron, double peak, single-peak plus single-peak Chevron, and multi-peak are respectively as follows: Figure 3-8 As shown, the hierarchical classifier, based on these morphological differences, progressively classifies signals from N0 to N4. It directly truncates peakless signals at node N0, while guiding complex single-peak plus single-peak Chevron patterns to the deeper node N4 for fine-grained discrimination. This hierarchical architecture utilizes the natural hierarchical relationships between categories, enabling most simple samples to be quickly classified at early nodes, significantly reducing unnecessary computation and improving overall classification efficiency.

[0031] Step S3: For each node in the hierarchical classifier, using only the training sample set that can reach that node, optimize to obtain a classification threshold corresponding to that node that makes the preset evaluation metric optimal. The threshold optimization process is as follows: Figure 11As shown in the diagram. The training sample set that can reach a node refers to the subset of samples that actually flow into the current node after being filtered by the preceding nodes in the logical path of the hierarchical classifier. The optimization process of the classification threshold involves searching within a preset threshold range with a preset step size, calculating the classification accuracy or F1 score of each candidate threshold on the reachable sample set of the current node, and selecting the threshold that optimizes this evaluation metric as the final classification threshold for that node. This step aims to eliminate the bias caused by traditional fixed thresholds, allowing the decision boundary of each node to adapt to its actual sample distribution, thereby solving the performance degradation problem caused by sample imbalance or distribution shift. For example, for node N0, searching using all training samples might yield an optimal threshold of 0.37; while for the deep node N4, searching using only a small number of complex samples filtered by N0, N1, N2, and N3 might yield an optimal threshold of 0.515. This independent optimization mechanism for the reachable sample set ensures that each binary classification model can achieve the best discrimination performance under its specific input distribution, avoids the increase in misclassification rate caused by a uniform threshold, and further improves the system's robustness in recognizing edge samples.

[0032] Step S4: Input the feature vector into the hierarchical classifier. Following the hierarchical order of the nodes, compare the predicted probability output by the current node with the classification threshold corresponding to that node. Based on the comparison result, determine the category of the current node and decide whether to output the final classification result or proceed to the next node. The node hierarchy and execution logic are as follows: Figure 10As shown in the diagram, the sequential decision-making inference process involves passing the feature vector of the sample to be identified through each node of the hierarchical classifier. At each node, the binary classification model outputs a predicted probability representing a certain category. The system compares this probability with the node-specific classification threshold obtained through optimization. If the comparison result meets a preset condition, the category corresponding to the current node is determined to be valid, the final classification result is output directly, and the process terminates. If the condition is not met, it is determined that the sample is more likely to enter the next category, and the sample is passed to the next node to continue the comparison operation. This step is executed by a processor or inference engine that has deployed a hierarchical classifier, and its output is the final one-dimensional spectral category label. For example, when a feature vector of a unimodal Chevron pattern is input, the probability of a peak is first calculated at node N0, which is 0.95, greater than the threshold of 0.37, thus classifying it as peaked and proceeding to node N1. At node N1, the probability of a unimodal pattern is calculated, which is 0.25, less than the threshold of 0.30, thus classifying it as non-unimodal and proceeding to node N2. At node N2, the probability of a unimodal Chevron pattern is calculated, which is 0.85, greater than the threshold of 0.415, thus classifying it as correct. The unimodal Chevron classification result is then output, and the process ends, without proceeding to subsequent nodes N3 and N4. Through this step-by-step comparison and dynamic routing mechanism, efficient and accurate sequential decision-making is achieved, ensuring both precise recognition of complex patterns and rapid response to simple samples, effectively solving the technical challenge of balancing computational efficiency and accuracy.

[0033] This application achieves high-precision automatic identification of one-dimensional spectral data through the synergistic effect of multi-dimensional enhanced feature engineering, hierarchical classifier architecture, reachable sample threshold optimization, and serial decision reasoning. By extracting multi-dimensional enhanced features including morphological features and Chevron pattern features, the ability to represent complex spectral patterns is significantly improved, especially in effectively distinguishing easily confused categories such as single peaks and single-peak Chevrons. Based on this, a hierarchical classifier arranged according to category hierarchy is constructed. Utilizing the natural hierarchical relationship between categories, the complex multi-class problem is decomposed into multiple simple binary classification problems, enabling most samples to be classified at early nodes, greatly improving classification efficiency. Furthermore, for each node, threshold search is performed only based on training samples that can reach that node, replacing the traditional fixed threshold. This makes the decision boundary of each node more adaptable to the actual sample distribution, thereby improving the overall classification accuracy. Finally, through a serial decision mechanism, the predicted probability is compared with the optimized threshold step by step, realizing full automation and intelligence from feature extraction to result output. Significant technical effects have been achieved in application scenarios such as quantum bit spectroscopy.

[0034] Example 2 The method also includes further refined definition and extraction of morphological features and Chevron pattern features from the above embodiments.

[0035] Step 1: Extract multi-dimensional enhancement features from the one-dimensional spectral data to be identified, forming a feature vector. The feature vector includes at least morphological features describing peak morphology and Chevron pattern features describing peak pair patterns. Specifically, morphological features include one or more of the following: peak number, maximum peak height, peak height ratio, peak symmetry score, peak sharpness score, peak width statistics, and inter-peak distance statistics. Peak number refers to the total number of local maxima points identified in the one-dimensional spectral data using the peak-finding algorithm, directly reflecting the complexity of the spectrum. For example, when multiple obvious response signals exist in the spectrum, the peak number is greater than 1. Maximum peak height refers to the maximum signal intensity among all identified peaks, used to characterize the intensity level of the main response signal. Peak height ratio refers to the ratio of the intensity of the highest peak to the intensity of the second highest peak. This parameter is used to distinguish the relationship between primary and secondary peaks. For example, in a single-peak mode, this ratio approaches infinity or a maximum value, while in a bimodal or multi-peak mode, this ratio is a finite value. Peak symmetry score is a quantitative indicator calculated based on the symmetry of the distribution of a predetermined number of data points (e.g., 10 points on each side) on the left and right sides of the peak. It measures the degree of symmetry of a single peak; a higher symmetry score indicates that the peak shape is closer to an ideal Gaussian or Lorentz distribution. Peak sharpness score is a value calculated based on the second derivative or rate of curvature change of the signal within a predetermined window (e.g., 11 data points) around the peak. It reflects the steepness of the peak; a higher sharpness score indicates a narrower and steeper peak shape. Peak width statistics include the mean, standard deviation, or specific quantiles of the peak width, used to describe the broadening of the peak on the frequency or time axis. Inter-peak distance statistics include the mean, standard deviation, minimum, or maximum distance between adjacent peaks, used to describe the spatial distribution density of multiple peaks in the spectrum.

[0036] Chevron mode features specifically include one or more of the following: peak pair symmetry score, peak pair separation, peak pair height ratio, and Chevron angle. These features are specifically designed to detect paired resonance peaks in a one-dimensional spectrum that are distributed in a V-shape or cross shape. The peak pair symmetry score is an indicator that comprehensively evaluates the symmetry of a pair of peaks relative to the center frequency or a specific reference point, used to identify typical Chevron cross structures. Peak pair separation refers to the distance between the pair of peaks constituting the Chevron mode on the horizontal axis; this distance exhibits a specific evolutionary pattern with changes in control parameters. The peak pair height ratio refers to the intensity ratio of the two peaks in the Chevron mode, used to assist in determining the peak's affiliation and mode type. The Chevron angle is the angle formed by connecting the vertices of the pair of peaks to the horizontal axis or reference line; this angle visually reflects the changing trend of energy level repulsion or coupling strength. For example, in the dispersive spectrum of superconducting qubits, when the bit frequency and the resonant cavity frequency cross in opposite directions, a typical Chevron pattern is formed. At this time, by calculating the Chevron angle and peak-pair separation, the single-peak Chevron mode can be accurately distinguished from the ordinary single-peak or double-peak mode.

[0037] By combining the aforementioned morphological features with Chevron pattern features, the structural information of one-dimensional spectral data can be comprehensively characterized, from the microscopic morphology of individual peaks to the macroscopic evolution patterns of peak pairs. Peak symmetry and peak sharpness scores effectively enhance the ability to capture subtle morphological differences in the main peak, solving the problem that traditional methods struggle to distinguish spectral patterns that are morphologically similar but of different categories. Meanwhile, peak pair separation and Chevron angle provide strong discriminative features for specific physical phenomena (such as energy level anti-crossing). These features together constitute a high-dimensional discriminative space, significantly improving the accuracy of subsequent hierarchical classifiers in identifying easily confused categories such as single peaks, single-peak Chevrons, and single-peak plus single-peak Chevrons.

[0038] This application, by refining the definition of morphological and Chevron pattern features, not only enriches the information dimensions of feature vectors but also specifically addresses the technical challenge of lacking fine-grained quantitative descriptions of peak morphology and Chevron patterns in existing technologies. Utilizing peak quantity and inter-peak distance statistics, the system can quickly filter out samples with no peaks or simple peak structures; combining peak symmetry scores and Chevron angles, the system can deeply analyze complex spectral interferograms. This multi-level, multi-angle feature extraction strategy enables the classification model to achieve more robust classification performance without significantly increasing computational complexity, effectively avoiding misjudgments caused by insufficient feature representation, and laying a solid data foundation for subsequent threshold optimization based on reachable samples and sequential decision inference.

[0039] Example 3 This method also includes constructing a specific five-node hierarchical classifier structure, whose node hierarchy and execution logic are as follows: Figure 10 As shown.

[0040] Step A1: Node N0 determines whether there is a peak. As the root node of the hierarchical classifier, the N0 node's role is to perform preliminary screening of the one-dimensional spectral data to be identified, determining whether there are valid signal peaks in the data. The input to this node is a multi-dimensional enhanced feature vector extracted from the one-dimensional spectral data to be identified, and the output is a binary classification result (peak present or peak absent). Specifically, the binary classification model corresponding to the N0 node makes a judgment based on the difference in feature distribution between peak-less and peak-containing features learned from the training samples. For example, when the peak quantity feature in the input feature vector is 0 and the signal quality feature is below a preset noise threshold, the N0 node determines that there is no peak; conversely, if a significant peak feature is detected, it determines that there is a peak. The N0 node is located at the top of the decision tree and is a necessary checkpoint for all subsequent classification paths. If the N0 node determines that there is no peak, it directly outputs the peak-less category and terminates the classification process; if the determination result is that there is a peak, it activates the next level N1 node. By setting the N0 node, invalid data can be effectively filtered out, avoiding unnecessary calculations of noisy data by subsequent complex models, thus improving the overall processing efficiency of the system.

[0041] Step A2: Node N1 determines whether a sample is unimodal, executing only if node N0 determines it is unimodal. The execution of node N1 has strict preconditions; it is triggered only when node N0 outputs a unimodal classification result. The technical purpose of this node is to separate the unimodal category from all unimodal samples. Node N1 uses key indicators such as the number of peaks and peak symmetry score in morphological features to distinguish between a single dominant peak and a multimodal mixed pattern. For example, when the feature vector shows a peak count of 1 and a peak symmetry score higher than a specific threshold, node N1 classifies it as unimodal. Node N1 is directly connected to the unimodal branch of node N0. If node N1 determines it is unimodal (i.e., confirms it as unimodal), it outputs the unimodal classification result and ends the process; if it determines it is not unimodal (i.e., not unimodal), it indicates that the sample belongs to a more complex multimodal or special pattern category, and the process continues to the N2 node. This design ensures that the unimodal category does not directly compete with subsequent Chevron patterns or bimodal categories, reducing the false positive rate.

[0042] Step A3: Node N2 determines whether it is a single-peak Chevron mode, and this step is executed only if node N0 determines it has a peak and node N1 determines it is not a single-peak mode. Node N2 is specifically designed to identify single-peak Chevron modes with specific physical significance, and its execution is limited to node N0 determining it has a peak and node N1 determining it is not a single-peak mode. This node relies heavily on Chevron mode features, such as peak pair symmetry scores and Chevron angles, to capture the unique V-shaped or inverted V-shaped interference fringes in the spectrum. Specifically, even if a sample may macroscopically appear as a single-peak, if it contains substructures that conform to Chevron geometry, node N2 will identify it as a single-peak Chevron. For example, when peak pair separation is detected within a specific range and the Chevron angle conforms to quantum ratio dispersion spectral characteristics, it is determined to be a positive sample. Node N2 is located downstream of the non-single-peak branch of node N1. If node N2 determines "yes," then the single-peak Chevron result is output; if it determines "no," it means the sample is neither a normal single-peak nor a single-peak Chevron pattern, and further verification is needed at node N3. This step effectively solves the technical problem in traditional methods of easily misclassifying a single-peak Chevron as a normal single-peak.

[0043] Step A4: Node N3 determines whether a sample is bimodal, executing only when N0 is peaked, N1 is non-single-peaked, and N2 is a non-single-peak Chevron. After excluding single-peaked and single-peak Chevron patterns, node N3 focuses on identifying standard bimodal structures. Its execution logic relies on the negative results of the first three nodes, ensuring that all samples entering this node are complex multi-peak candidates. Node N3 primarily uses inter-peak distance statistics, peak height ratio, and peak width statistics for discrimination. For example, when the feature vector contains two significant main peaks, and the distance between the two peaks is greater than a preset minimum separation threshold, while the peak height ratio is within a certain range, node N3 determines it to be bimodal. Node N3 follows the negative branch of node N2. If the determination is yes, the bimodal category is output; if the determination is no, it means the sample has a more complex structure than bimodal, and the process proceeds to the final node N4. By independently determining bimodality at a specific level, bimodal samples are prevented from being incorrectly classified as multimodal or other mixed patterns.

[0044] Step A5: Node N4 determines whether it is a single-peak plus single-peak Chevron pattern. This step is executed only when N0 is peaked, N1 is non-single-peak, N2 is non-single-peak Chevron, and N3 is non-bimodal. Node N4, as the last dedicated decision node in the hierarchical classifier, is used to identify the most complex single-peak plus single-peak Chevron mixture pattern. This node has the most stringent execution conditions, requiring samples to pass through N0 through N3 sequentially without being identified as belonging to any of the aforementioned categories. Node N4 comprehensively analyzes basic statistical features, frequency domain features, and higher-order morphological features to distinguish between simple multi-peak patterns and this specific mixture pattern. For example, when the spectrum simultaneously contains an independent strong peak and a weak peak with Chevron characteristics, and its frequency domain centroid and bandwidth characteristics conform to the mixture pattern distribution, node N4 determines it as a positive sample. Node N4 is located at the end of the decision tree. If node N4 determines it is positive, it outputs the single-peak plus single-peak Chevron result; if it determines it is negative, it uniformly classifies all remaining unmatched samples into multi-peak or other undefined categories. This hierarchical design ensures that complex mixing patterns can be accurately captured without being lost in generic multi-peak categories.

[0045] This application achieves deep synergy of technical features by constructing the aforementioned five-node hierarchical classifier. Node N0, as the main entry point, quickly eliminates invalid data, providing a clean set of peaked samples for subsequent nodes. Node N1 uses serial logic to pre-divide simple single-peak patterns, reducing the processing burden on subsequent nodes. Node N2 specifically identifies easily confused Chevron patterns, preventing them from directly competing with single-peak patterns. Nodes N3 and N4, while gradually narrowing the sample range, respectively refine the discrimination of bimodal and mixed patterns. This layered, conditionally triggered node connection relationship allows each binary classification model to focus only on the specific classification task of its current level, significantly reducing the learning difficulty and overfitting risk of individual models. The overall architecture not only clarifies the logical dependencies between categories but also ensures the uniqueness and determinism of the classification path through a strict execution order, thereby significantly improving the accuracy and robustness of one-dimensional spectral recognition, particularly effectively solving the technical problem of distinguishing similar spectral patterns.

[0046] Example 4 The method also includes a specific implementation of the classification threshold optimization process, the optimization process as follows: Figure 11 As shown.

[0047] Step B1: For each node, collect the predicted probabilities of the binary classification model corresponding to that node on its reachable sample set. The reachable sample set refers to the subset of training samples that actually flow into the current node after the one-dimensional spectral data has been judged by the preceding nodes of the hierarchical classifier. Because the hierarchical classifier uses a serial decision mechanism, only samples that meet the judgment conditions of all preceding nodes (e.g., N0 is judged as peaked, N1 as non-single-peaked, etc.) can reach the current node. Therefore, the distribution characteristics of the reachable sample set differ significantly from the full training set, often exhibiting a skewed distribution or specific confidence clustering characteristics. Collecting the predicted probabilities specifically means inputting each sample in the reachable sample set into the pre-trained binary classification model of the current node, obtaining the probability value output by the model that belongs to the positive class, forming a probability distribution sequence specific to that node. For example, for node N2, which determines whether it is a single-peaked Chevron, its reachable samples only contain those spectral lines judged as peaked and non-single-peaked. The probability values ​​output by the model for these specific data constitute the predicted probability set of node N2.

[0048] The generation of the aforementioned predicted probability relies on the probability mapping mechanism within the binary classification model. Specifically, after the binary classification model (such as a gradient boosting tree or neural network) of the current node operates on the input feature vector, its original output is a real value representing the decision tendency, not the final probability. This real value needs to be converted into a value between 0 and 1 through a sigmoid function with equal probability calibration mapping. This value is then used as the predicted probability output representing the likelihood that the sample belongs to the positive class.

[0049] This probabilistic collection method based on reachable samples can accurately reflect the data distribution faced by the current node in the actual reasoning scenario, providing a reliable data foundation for the accurate positioning of subsequent thresholds.

[0050] Step B2: Within a preset threshold range, search with a preset step size to select the threshold that optimizes the classification accuracy or F1 score as the classification threshold for that node. The preset threshold range is a search space defined by the effective range of the binary classification model's output probabilities, typically set to 0.1 to 0.9 to avoid overfitting or underfitting due to extreme probability values. The preset step size is the smallest granularity in the search process, typically set to 0.01 or smaller to ensure search accuracy. The search process is as follows: Within the preset threshold range, iterate through each candidate threshold sequentially with the preset step size, compare the collected predicted probabilities with the current candidate threshold, statistically analyze the classification results at that candidate threshold, and calculate the corresponding classification accuracy or F1 score. The F1 score is the harmonic mean of precision and recall, providing a more objective evaluation of classification performance in cases of class imbalance. This process demonstrates starting from the initial threshold, iteratively calculating the evaluation metrics until the entire search range is traversed, ultimately locking in the threshold corresponding to the maximum value of the evaluation metrics. For example, when optimizing node N0 (determining whether there is a peak), if the highest F1 score is obtained at 0.37, while the F1 score is lower at the traditional fixed threshold of 0.5, then 0.37 is selected as the final classification threshold for node N0. Similarly, for node N1, an optimal threshold of 0.30 may be found. This grid search mechanism ensures that each node can find the best decision boundary that adapts to its specific sample distribution, rather than blindly using a universal threshold of 0.5.

[0051] This application achieves adaptive optimization of the classification threshold through targeted collection of reachable sample sets and grid search based on evaluation metrics, as described in the above steps. Due to the serial structure of hierarchical classifiers, different nodes face varying sample difficulty and distribution patterns, making it impossible to consider the performance requirements of all nodes with a fixed threshold. By limiting the threshold optimization process to the reachable sample set and using accuracy and F1 score as quantitative metrics for refined searching, each binary classification node can achieve a locally optimal discrimination state. This node-level threshold customization not only significantly improves the classification accuracy of individual nodes and eliminates misclassifications caused by improper threshold settings (such as misclassifying a low-confidence unimodal classifier as a non-modal classifier, or misclassifying a complex Chevron pattern as a common multimodal classifier), but also, through a cascading effect, cumulatively improves the final recognition accuracy of the overall hierarchical classification system, effectively solving the suboptimal classification performance problem caused by rigid thresholds in traditional methods.

[0052] Example 5 The extracted multi-dimensional enhanced features are further expanded to construct a more complete feature representation system. The feature extraction process is as follows: Figure 9 As shown.

[0053] Multi-dimensional enhancement features are extracted from the one-dimensional spectral data to be identified, forming a feature vector. This feature vector includes at least morphological features describing peak shapes and Chevron pattern features describing peak pair patterns. In addition to the extracted morphological and Chevron pattern features, the multi-dimensional enhancement features also include at least one of a basic statistical feature set, a frequency domain feature set, and a signal quality feature set to comprehensively characterize the statistical distribution, frequency domain characteristics, and data reliability of the one-dimensional spectral data. Specifically, the basic statistical feature set reflects the overall distribution trend and dispersion of the signal, and includes the mean, standard deviation, skewness, and kurtosis. The mean characterizes the average level of signal strength, the standard deviation describes the amplitude of signal fluctuations around the mean, skewness reflects the asymmetry of signal distribution, and kurtosis measures the steepness or flatness of the distribution curve. For example, for a one-dimensional spectrum with baseline drift, its mean will shift significantly; while for a signal containing anomalous spike noise, its kurtosis value will be much larger than the theoretical value of a normal distribution. By introducing these basic statistical features, the classifier can be provided with fundamental criteria for the overall energy distribution and morphological deviation of the signal, which complements the local peak morphological features.

[0054] Frequency domain feature sets are used to reveal the composition and structure of a signal in the frequency domain, helping to identify periodic or resonant characteristics hidden in time-domain waveforms. These features include total power, dominant frequency, frequency centroid, and frequency bandwidth. These features are typically calculated by performing a Fast Fourier Transform on one-dimensional spectral data to obtain the power spectrum. Total power reflects the sum of the signal's energy across the entire frequency band; the dominant frequency corresponds to the frequency component with the highest energy in the power spectrum, often corresponding to the main resonant mode in the spectrum; the frequency centroid describes the location of the center of gravity of the spectral energy; and the frequency bandwidth characterizes the frequency range occupied by the effective signal. For example, when weak periodic interference or specific qubit coupling modes exist in the one-dimensional spectrum, they may exhibit irregular fluctuations in the time domain, making them difficult to capture directly through morphological features, but they will show obvious dominant frequency peaks or specific bandwidth characteristics in the frequency domain. By fusing frequency domain features, the system can effectively distinguish spectral modes with similar time-domain morphology but different frequency-domain structures.

[0055] Signal quality feature sets, including the signal-to-noise ratio (SNR), are used to evaluate the reliability and signal-to-noise level of acquired data. The SNR is calculated as the ratio of the power of the effective signal component to the power of the background noise, directly reflecting the purity of the data. For example, in low SNR scenarios, even if a suspected peak shape is detected, it may be a false positive caused by random noise. In this case, the SNR feature can be used as a weighting factor or filtering condition to help the classifier reduce the false positive rate. By incorporating the SNR into the feature vector, the hierarchical classifier can perceive the quality of the input data, thereby dynamically adjusting the decision boundary or triggering corresponding robust processing mechanisms when facing strong noise interference or weak signal samples.

[0056] By integrating features from three dimensions—basic statistics, frequency domain analysis, and signal quality assessment—along with the original morphological and Chevron pattern features, a high-dimensional, multi-perspective feature space is constructed. This multi-domain feature fusion strategy not only preserves the local details of the signal but also supplements global statistical properties and frequency domain fingerprints, significantly enhancing the feature vector's adaptability to complex environments (such as baseline drift, high-frequency noise, and weak signals), and providing sufficient and diverse information input for the accurate discrimination of subsequent hierarchical classifiers.

[0057] This application constructs a comprehensive one-dimensional spectral data representation system through the synergistic effect of basic statistical feature sets, frequency domain feature sets, signal quality feature sets, morphological features, and Chevron pattern features. Basic statistical features provide a stable global baseline for classification, preventing overall misclassification due to local noise; frequency domain features uncover periodic patterns that are difficult to detect in the time domain, effectively distinguishing spectral modes with similar morphologies but different physical mechanisms; signal quality features introduce a data reliability dimension, enabling the classification decision process to adapt to noisy environments. Based on this, combined with the serial decision-making mechanism of a hierarchical classifier, the system can utilize these multi-dimensional features for targeted discrimination at different nodes: for example, at node N0, invalid noise data is quickly eliminated using signal-to-noise ratio and total power; at nodes N1 to N4, fine features such as skewness, kurtosis, and dominant frequency are used to distinguish single-peak, double-peak, and complex Chevron patterns. This deep fusion of multiple feature dimensions fundamentally solves the problem of insufficient feature extraction caused by traditional methods relying solely on simple statistics, significantly improving the resolution between similar spectral modes and the overall classification accuracy.

[0058] Example 6 The binary classification model in step S2 is automatically trained using an automated machine learning framework. This framework is a machine learning system capable of automating the entire process, including model selection, feature preprocessing, hyperparameter optimization, and model ensemble. The framework uses the multi-dimensional enhanced feature vectors extracted in step S1 as input data and the binary classification labels obtained based on the node hierarchy as the output target. Specifically, for each binary classification node in the hierarchical classifier, the automated machine learning process is run independently once. During training, the framework automatically traverses multiple candidate algorithms, evaluates the performance of different algorithms on the sample set reachable at the current node, and automatically adjusts the hyperparameter combinations of each algorithm, ultimately outputting the binary classification model or model ensemble scheme with the best predictive performance at that node. This automation avoids the subjectivity and limitations of manual model selection based on experience and manual parameter tuning, ensuring that each node obtains the best classifier adapted to its specific data distribution. For example, for node N0, which determines whether there is a peak, the automated machine learning framework may search for a tree model with strong noise robustness; while for node N2, which distinguishes subtle morphological differences, the framework may automatically combine multiple weak classifiers to form a strong classifier. This significantly lowers the technical threshold for model building and improves the overall generalization ability and classification accuracy of the hierarchical classifier.

[0059] The automated machine learning framework takes feature vectors as input and outputs binary classification labels based on node hierarchy, with a training time limit of 300 to 600 seconds. This training time limit refers to the maximum allowed runtime set when starting the automated machine learning framework, constraining the model search and optimization process. This time range of 300 to 600 seconds is determined based on a balance between model convergence speed and computational resource consumption. Specifically, when the training time is less than 300 seconds, the automated search space may not be fully explored, resulting in a non-globally optimal model configuration, affecting classification accuracy. While a small performance improvement may be achieved when the training time exceeds 600 seconds, the computational cost increases dramatically, and the marginal benefit diminishes, hindering rapid model iteration and engineering deployment. Upon reaching the preset time limit, regardless of whether the search process is completely finished, the framework will forcibly terminate the current task and return the optimal model configuration found so far. For example, when training a model for node N3 (to determine if it is bimodal), the training time is set to 450 seconds. Within this time, the framework will try dozens of different hyperparameter combinations and output the model with the highest F1 score when the time runs out. By setting this explicit time boundary, not only is the controllability and repeatability of the model training process guaranteed, but the method also enables efficient batch training of five-node models with limited computing resources, meeting the dual requirements of industrial applications for development efficiency and real-time performance.

[0060] This application achieves intelligent construction of each binary classification node in a hierarchical classifier by employing an automated machine learning framework combined with strict training time constraints. The automated machine learning framework utilizes the mapping relationship between feature vectors and binary classification labels to adaptively select the optimal model structure for each node, solving the problem that traditional manual parameter tuning struggles to achieve global optimization. The 300-600 second training time limit serves as a crucial constraint, ensuring that the automated process does not get bogged down in infinite search, thus significantly shortening the model development cycle while maintaining model performance. The synergistic effect of these two aspects enables the entire one-dimensional spectral recognition method to possess both high-accuracy classification capabilities and an efficient model construction and maintenance process, effectively overcoming the technical challenge of balancing computational efficiency and accuracy in existing technologies.

[0061] Example 7 This application also provides a one-dimensional spectrum recognition system based on enhanced feature engineering and hierarchical classifiers to implement the aforementioned one-dimensional spectrum recognition method. The system structure is as follows: Figure 12 As shown, it includes: an enhanced feature extraction module 10, a hierarchical classifier module 20, a threshold optimization module 30, and a serial decision reasoning module 40.

[0062] The enhanced feature extraction module 10 is used to extract multi-dimensional enhanced features from one-dimensional spectral data to form a feature vector. This module serves as the system's data entry point and processing foundation, responsible for converting unstructured or semi-structured one-dimensional spectral signals into high-dimensional numerical vectors that can be processed by a computer. In its implementation, this module can extract various types of features from one-dimensional spectral data, including basic statistical features, peak morphological features, frequency domain features, Chevron mode features, and signal quality features. For example, basic statistical features may include mean, standard deviation, skewness, and kurtosis; peak morphological features may include peak number, maximum peak height, and peak symmetry score; and Chevron mode features may include peak pair symmetry score and peak pair separation. These features together constitute a multi-dimensional enhanced feature vector, providing sufficient discriminative information for subsequent classification decisions. This module interacts with the hierarchical classifier module 20 through feature vectors, and its output feature vector directly serves as the input to the hierarchical classifier module 20, determining the upper limit of accuracy in the subsequent classification process.

[0063] The hierarchical classifier module 20 stores and runs a hierarchical classifier containing multiple binary classification nodes arranged in a hierarchical order of categories. As the core decision engine of the system, this module internally maintains a node network arranged in a specific category hierarchy. Each node corresponds to an independent binary classification model used to determine whether the current sample belongs to a specific category or subset of categories. For example, the first node can be used to determine if there is a peak, the second node can determine whether it is a single peak given the presence of a peak, and so on. This hierarchical structure allows the system to utilize the logical dependencies between categories, avoiding the confusion caused by flattened competitive classification of all categories. This module works closely with the serial decision inference module 40, receiving control commands from the inference module, activating the corresponding node models sequentially, and feeding back the node output probabilities to the inference module.

[0064] The threshold optimization module 30 optimizes and stores the corresponding classification threshold for each node in the hierarchical classifier based on its reachable sample set. This module addresses the problem that traditional fixed thresholds cannot adapt to the differences in sample distribution across different nodes. In practice, this module operates based on the reachable sample set; that is, for nodes at deeper levels, it only uses training samples that meet the path conditions in the judgment of preceding nodes for threshold search. For example, for a node determining whether it is a unimodal Chevron, it only uses samples that are judged to have peaks but are not unimodal for optimization. This module can scan within a preset range with a preset step size, select the value that optimizes the classification accuracy or F1 score as the classification threshold for that node, and store this threshold for use by the serial decision inference module 40. This mechanism ensures that the decision criteria for each node are tailored to the actual sample distribution it faces, thereby improving the robustness of the overall system.

[0065] The serial decision-making inference module 40 is used to input feature vectors into the hierarchical classifier. Following the node order, it compares the predicted probability output by the current node with its corresponding classification threshold and outputs the final recognition result. This module acts as the system's process scheduler, coordinating the working sequence of the enhanced feature extraction module 10, the hierarchical classifier module 20, and the threshold optimization module 30. Specifically, after receiving the feature vector, this module first calls the starting node in the hierarchical classifier module 20 to obtain the predicted probability, then reads the classification threshold corresponding to that node from the threshold optimization module 30 and compares the two. Based on the comparison result, this module decides whether to directly output the final classification result (e.g., no peak) or guide the process to the next node for further judgment. This serial decision-making mechanism achieves a short-circuit effect, meaning that most simple samples can be classified in early nodes without going through the calculations of all nodes, thus significantly reducing the average inference time and improving the system's response speed.

[0066] Specifically, the working process and principle of this application system are as follows: After the system starts, the original one-dimensional spectral data first enters the enhanced feature extraction module 10. This module parses the data and calculates a multi-dimensional feature vector including morphological, frequency domain, and pattern features. Subsequently, the feature vector is transmitted to the serial decision inference module 40. According to the preset hierarchical logic, the inference module first activates the root node (such as node N0) in the hierarchical classifier module 20 and obtains the predicted probability of the current sample for that node. At the same time, the inference module retrieves the optimal classification threshold corresponding to the root node from the threshold optimization module 30. The inference module compares the predicted probability with the threshold. If the exit condition is met (such as the probability being lower than the threshold and being judged as having no peak), the no-peak label is directly output and the process ends; if not, the next level child node (such as node N1) is activated according to the hierarchical relationship, and the above process of probability acquisition, threshold comparison, and flow judgment is repeated. This process is carried out sequentially until a certain node outputs a positive classification result or all nodes are traversed and a default category (such as multi-peak) is output. Throughout the process, the threshold optimization module 30 pre-optimizes the threshold of each node independently based on historical training data, ensuring the accuracy of each judgment step.

[0067] As a preferred embodiment, the solution of this application is implemented as follows: Assume the system is applied to an automatic classification scenario of superconducting quantum bit excitation spectra. When the system receives a new quantum bit spectrum data, the enhanced feature extraction module 10 first performs a fast Fourier transform and peak detection on the spectrum, extracting approximately 45-dimensional feature vectors including signal-to-noise ratio, peak sharpness, Chevron angle, etc. The serial decision inference module 40 then takes over the vector, first calling node N0 (peak or no-peak judgment) in the hierarchical classifier module 20. Assuming the peak probability output by node N0 is 0.85, and the threshold stored in the threshold optimization module 30 for node N0 is 0.37, since 0.85 is greater than 0.37, the inference module determines to proceed to the next round. Next, the inference module calls node N1 (single-peak or non-single-peak judgment). If the single-peak probability output by node N1 is 0.25, which is lower than its optimization threshold of 0.30, the inference module determines it to be non-single-peak and continues to node N2. The N2 node is specifically designed to identify single-peak Chevron patterns. If its output probability is higher than a set threshold, the system directly outputs the single-peak Chevron classification result and terminates subsequent calculations; otherwise, it continues to judge downwards. Through this step-by-step screening, the system can accurately identify complex spectral patterns within tens of milliseconds and effectively distinguish between spectral patterns that are similar in appearance but have different physical meanings.

[0068] Through the above technical solutions, this application achieves modular and efficient one-dimensional spectral recognition. By employing the enhanced feature extraction module 10, the system can fully capture the subtle morphological and frequency domain features of the spectrum, solving the misjudgment problem caused by insufficient feature representation in traditional systems. By constructing a hierarchical classifier module 20 and combining it with the serial decision reasoning module 40, the system utilizes hierarchical logic between categories to achieve step-by-step stripping and rapid classification of similar patterns, improving classification efficiency. By setting up an independent threshold optimization module 30, the system can adaptively adjust the decision criteria for sample distribution at different levels, overcoming the defect of poor applicability of fixed thresholds at different nodes. The collaborative work of each module forms a closed-loop intelligent recognition system, ensuring both high accuracy and meeting the engineering requirements of real-time processing.

[0069] Example 8 The binary classification models in the hierarchical classifier module 20 include the CatBoost model or the LightGBM model.

[0070] The CatBoost model refers to a gradient boosting decision tree model built based on a category feature optimization algorithm. In this one-dimensional spectral recognition scheme, the CatBoost model serves as a specific implementation at the hierarchical classifier node, its function being to utilize an ordered boosting mechanism to process multi-dimensional enhancement features extracted from one-dimensional spectral data. This model, in conjunction with other nodes in the hierarchical classifier module 20, receives feature vectors output by the enhancement feature extraction module 10 and performs layer-by-layer judgment on the input data through an internally integrated decision tree sequence. When processing data containing latent category attributes or high-dimensional numerical attributes, such as morphological features and Chevron pattern features, the CatBoost model can automatically handle the combination relationships between features without the need for cumbersome one-hot encoding, thus effectively reducing the risk of overfitting while maintaining high classification accuracy. Its output is the predicted probability corresponding to the current node, which is compared by the serial decision reasoning module 40 with a preset classification threshold to determine the category of the spectral data or to proceed to the next node's judgment process.

[0071] The LightGBM model refers to a decision tree model built based on a lightweight gradient boosting machine algorithm. In this one-dimensional spectral recognition scheme, the LightGBM model serves as an alternative implementation for the hierarchical classifier nodes. Its name derives from its use of a histogram-based decision tree algorithm and leaf growth strategy, aiming to achieve efficient training and inference speeds. This model works in conjunction with the reachable sample set in the hierarchical classifier module 20, optimizing for the specific training sample distribution of each node. By supporting parallel learning and caching optimization, the LightGBM model can quickly process feature vectors of 45 dimensions or more involved in this application. Especially in applications requiring real-time response, its low memory footprint and high throughput significantly shorten the inference time per sample. Its output generates the classification prediction probability for the current node, which directly serves as the input to the serial decision logic, ensuring that the system can complete path determination for various categories, from no peaks to multi-peaks, within milliseconds.

[0072] Specifically, in the technical solution of this application, the binary classification nodes within the hierarchical classifier module 20 do not necessarily adopt the same model architecture. Instead, they can flexibly choose either the CatBoost model or the LightGBM model, or use a combination of these two models in different nodes, depending on the resource constraints and performance requirements of the actual deployment environment. For example, for the N0 node located at the top of the hierarchy and processing the largest sample size (determining whether there is a peak), the LightGBM model, with its faster inference speed, can be used to accelerate the overall process. For the N2 or N3 nodes located deep in the hierarchy and needing to distinguish subtle morphological differences (such as unimodal Chevron versus bimodal), the CatBoost model, with its stronger ability to capture class features and complex nonlinear relationships, can be used to improve discrimination accuracy. Both models can be adapted to the reachable sample threshold optimization step in this application. That is, the predicted probability distribution output by the model can be used by the subsequent threshold search algorithm. By scanning within a specific threshold range, the cutoff point that optimizes the F1 score or accuracy of the node on the reachable sample set can be found, thereby achieving the best match between the model's predictive ability and the decision threshold.

[0073] The specific implementation is as follows: When constructing the five-node hierarchical classifier, for each node from N0 to N4, a CatBoost classifier instance or a LightGBM classifier instance is initialized respectively. The system first loads the feature vector dataset formed after multi-dimensional enhanced feature engineering. This dataset has been divided into corresponding binary labels according to the node hierarchy (for example, the label of node N1 is only unimodal and non-unimodal). If the CatBoost model is selected, its iteration number, learning rate, and depth parameters are configured, and the automatic category feature processing function is enabled. The model is trained using the reachable sample set of node N0 until the model converges. If the LightGBM model is selected, its parameters such as the maximum number of leaves and the minimum number of samples in the leaves are configured, and the histogram algorithm is used to accelerate the split point search process during training. After training, the models of each node are saved to the hierarchical classifier module 20. During the inference phase, when new one-dimensional spectral data is input, the feature vector first enters the CatBoost or LightGBM model corresponding to node N0. The model outputs a predicted probability value between 0 and 1, which is then fed into the comparison logic. If the value is greater than the optimized classification threshold of that node (e.g., 0.37), it is determined to have a peak and is passed to node N1; otherwise, a peakless result is directly output and the process terminates. Subsequent nodes follow this process until the final spectral category label is output.

[0074] Through the above technical solution, this application realizes the flexible application of high-performance gradient boosting tree models under a hierarchical classification architecture. By employing the CatBoost or LightGBM model as the core algorithm for binary classification nodes, leveraging their advantages in handling high-dimensional sparse features and nonlinear relationships, it can effectively distinguish similar spectral patterns (such as single-peak plus single-peak Chevron versus single-peak Chevron) that are difficult to identify using traditional methods. Simultaneously, thanks to the inherent fast inference characteristics of these two models, the problem of excessive computation time for traditional complex models in real-time processing scenarios is solved. This achieves the technical effect of reducing the average prediction time per sample to approximately 0.091 seconds while maintaining high classification accuracy (e.g., above 87.52%), meeting the real-time requirements for automatic spectral recognition in scenarios such as quantum computing calibration.

[0075] Example 9 The model update module is used to re-search the threshold of the reachable sample set of each node after obtaining newly labeled training samples, so as to update the classification threshold of each node, and optionally retrain the binary classification model of each node.

[0076] The model update module refers to a functional unit or software process within the system, configured to monitor data input interfaces or storage areas to detect the inflow of newly labeled training samples. This module operates in standby or periodically activated state during system operation. Once a preset number of newly labeled training samples is detected or a preset time interval condition is met, the update process is triggered. This module interacts with the hierarchical classifier module 20 and the threshold optimization module 30, integrating new samples into the existing training data system to provide a data foundation for continuous model optimization.

[0077] Newly labeled training samples refer to one-dimensional spectral data and their corresponding enhanced feature vectors that have been manually interpreted, confirmed by experts, or processed through a high-confidence automatic labeling process, and have clear category labels. These samples represent the latest data distribution characteristics in the current application scenario and may contain spectral morphology drift caused by equipment aging, environmental changes, or adjustments to experimental conditions. The source of the newly labeled training samples can be set according to the actual situation. For example, it can be newly generated test data in the laboratory, or data collected in real time from the production line and subsequently verified. This application embodiment does not impose any special limitations on this.

[0078] Re-searching the reachable sample set for each node involves the model update module invoking threshold optimization logic to merge newly labeled training samples with historical training samples, forming an updated reachable sample set. For each binary classification node in the hierarchical classifier, only samples that can reach the node in the current hierarchical decision path (i.e., samples judged by all preceding nodes) are selected to form the node's exclusive reachable sample set. Based on this, the threshold optimization process is re-executed, i.e., searching within a preset threshold range with a preset step size, calculating the classification accuracy or F1 score under different thresholds, thereby determining a new classification threshold that optimizes the evaluation metrics. This process ensures that the classification threshold can dynamically adapt to changes in data distribution, avoiding a decline in classification performance due to a fixed threshold.

[0079] Updating the classification threshold for each node means replacing the old threshold stored in the system with the new optimal threshold obtained from the search. The updated threshold will be directly used by the serial decision reasoning module 40 in the next reasoning task, thereby reflecting the latest data features in real time. The threshold can be updated by a full replacement or by a partial update for specific nodes that have experienced significant drift; this embodiment does not impose any special limitations on this.

[0080] Optional retraining of the binary classification model at each node refers to the model update module deciding whether to initiate the model retraining process based on a preset strategy while updating the threshold. If the number of new samples is large enough and the data distribution has changed significantly, simply updating the threshold may not be sufficient to compensate for the model's lack of generalization ability. In this case, an automatic machine learning framework can be invoked to retrain the corresponding binary classification model using the updated reachable sample set. The retraining process can be consistent with the automatic training process described in the above embodiments, including setting training time limits and selecting model types. If the number of new samples is small or the data distribution is relatively stable, the retraining step can be skipped, and only the threshold update can be performed to save computational resources. This optional mechanism balances the relationship between system performance improvement and computational overhead.

[0081] Specifically, the working process and principle of this application are as follows: The model update module first monitors the inflow of new data. When the triggering condition is met, it locks the range of nodes that need to be updated. Next, the module merges the new samples with the original sample library and re-divides the reachable sample set for each node. Subsequently, a threshold scanning algorithm is executed for each node to find a new optimal decision boundary. If it is determined that the model needs to be retrained, the binary classification model training task for each node is started in parallel or serially, and the model parameters are updated after training is completed. Finally, the newly generated threshold and (if any) new model are deployed to the online service to complete a complete iterative update. The entire process realizes the system's self-evolution capability, ensuring recognition accuracy under long-term operation.

[0082] The specific implementation is as follows: Assume the system has been deployed in a quantum bit spectral analysis scenario, and the thresholds for each node have been optimized in the initial state. After one month of operation, due to minor adjustments in the superconducting quantum chip's manufacturing process, the baseline noise characteristics of the acquired spectral signals changed slightly, causing some single-peak signals to be misclassified as peakless. At this time, the operator imported a batch of 500 newly labeled spectral data. After detecting this batch of data, the model update module immediately incorporated it into the training set. For node N0 (determining whether there is a peak), the module extracted all reachable samples (i.e., all samples, since N0 is the first node), rescanned the threshold range of 0.1 to 0.9, and found that the original threshold of 0.37 was no longer optimal, and the new optimal threshold changed to 0.42. At the same time, given the significant change in data distribution, the module decided to trigger retraining, using the AutoGluon framework to incrementally train or fully retrain the N0 node model on the dataset containing the 500 new data and historical data, taking approximately 400 seconds to obtain the new model. After the update was completed, the system automatically switched to the new threshold and the new model, and the classification accuracy of subsequently input spectral data immediately returned to a high level. For nodes N1 to N4, if the new samples also contain the corresponding changes in category distribution, the same update process is performed; if there are no significant changes, the threshold can be fine-tuned or left unchanged.

[0083] Through the above technical solutions, this application achieves the technical effects of endowing the system with continuous learning capabilities and adapting to data distribution drift. By setting up a model update module, the system can respond promptly to the addition of newly labeled samples, thus solving the technical problem of performance degradation of static models due to changes in the data environment after long-term use. By adopting a re-threshold search mechanism based on the reachable sample set, the decision boundaries of each classification node are always kept in an optimal state, thereby maintaining the high accuracy of the overall hierarchical classifier. By providing an optional retraining mechanism, unnecessary waste of computational resources is avoided while improving the model's generalization ability, achieving flexibility and efficiency in system lifecycle management.

[0084] Example 10 The enhanced feature extraction module 10 extracts multi-dimensional enhanced features, including at least 45-dimensional feature vectors. These feature vectors encompass five categories: basic statistical features, peak morphology features, frequency domain features, Chevron mode features, and signal quality features. The feature composition and extraction process is as follows: Figure 9 As shown.

[0085] The 45-dimensional feature vector refers to a high-dimensional data representation composed of multiple feature components of different categories. The number of dimensions can be set according to actual conditions, such as 45, 48, 50, or 52 dimensions, etc., and this application embodiment does not impose any special limitation on this. The function of this feature vector in the overall technical solution is to carry multi-source information of the one-dimensional spectral data, serving as the input basis for subsequent hierarchical classifiers to make discriminative decisions. By constructing a feature space containing at least 45 components, it is possible to ensure that the input data has sufficient information richness to support the differentiation of complex spectral patterns.

[0086] Basic statistical features refer to the fundamental mathematical statistics used to describe the overall distribution of one-dimensional spectral data. In this application, basic statistical features work in conjunction with other features such as peak morphology features and frequency domain features to form a complete feature description system. Specific implementation methods may include one or more of the following: mean, standard deviation, skewness, kurtosis, maximum value, minimum value, range, median, variance, quartiles, and energy. These features are mainly used to characterize the overall strength and distribution of the signal, providing a basic data distribution reference for the classifier.

[0087] Peak morphology features refer to a set of features used to quantitatively describe the geometry, number, and arrangement of peaks in a one-dimensional spectrum. These features are closely linked to Chevron mode features, both jointly undertaking the task of finely characterizing the spectral microstructure. In practical implementation, peak morphology features can include peak number, maximum peak height, peak height ratio, peak symmetry score, peak sharpness score, peak width statistics, and inter-peak distance statistics. For example, the peak symmetry score can be obtained by calculating the symmetry of data points within a preset range on both sides of the peak apex; the peak sharpness score can be calculated based on the second derivative of data near the peak apex. These features enable the system to effectively distinguish between different peak forms such as single peaks, double peaks, and multiple peaks.

[0088] Frequency domain features refer to the characteristics reflecting the frequency component distribution of a signal extracted after performing a frequency domain transformation on one-dimensional spectral data. This technique supplements the limitations of time domain features in the overall scheme, revealing the energy distribution characteristics of the signal in the frequency domain. Its implementation can include performing a Fast Fourier Transform on the original signal to obtain the power spectrum, and then calculating the total power, dominant frequency, frequency centroid, frequency bandwidth, high-frequency energy proportion, or spectral flatness, etc. Combined with basic statistical features, frequency domain features can comprehensively describe signal characteristics from both time and frequency dimensions, improving the model's robustness to noise interference or specific frequency patterns.

[0089] Chevron mode features refer to characteristic parameters specifically used to identify and quantify specific Chevron interference fringe patterns in a one-dimensional spectrum. These features are key to distinguishing easily confused categories such as single-peak Chevron and single-peak-plus-single-peak Chevron patterns. Combined with peak morphology features, they focus on modeling the relative relationships between peak pairs. Specifically, Chevron mode features can include peak pair symmetry scores, peak pair separation, peak pair height ratio, and Chevron angles. For example, peak pair separation can be used to measure the distance between two associated peaks; Chevron angles can be used to characterize the opening of interference fringes. By introducing these features, the system can capture specific physical patterns that traditional statistical features cannot reflect.

[0090] Signal quality features are metrics used to evaluate the reliability and noise level of one-dimensional spectral data. These features play a supporting role in system decision-making; when signal quality is too low, they can be combined with other features to comprehensively assess the reliability of the classification results. Implementation methods can include signal-to-noise ratio (SNR), signal quality scoring, or signal stability indices. For example, the SNR can be obtained by calculating the ratio of peak signal power to background noise power.

[0091] Specifically, the working process and principle of this application are as follows: The enhanced feature extraction module 10 receives the input one-dimensional spectral data and performs the extraction operations of the above five types of features in parallel or serially. First, the basic statistics of the data are calculated, and peaks are identified and morphological parameters are calculated. Then, frequency domain transformation is performed to extract frequency domain parameters. Subsequently, Chevron mode parameters are calculated for specific peak pair structures. Finally, the signal quality is evaluated. All extracted feature components are concatenated into a feature vector with a dimension of at least 45. This feature vector is then transmitted to the hierarchical classifier module 20 as input data for each binary classification node. Since the feature vector covers comprehensive information from macroscopic statistics to microscopic morphology, from the time domain to the frequency domain, and from general features to specific mode features, each node in the hierarchical classifier can make an accurate binary classification judgment based on sufficient information, thereby achieving high-precision spectral recognition.

[0092] As a preferred embodiment, the specific implementation of this application's scheme is as follows: Assuming a quantum bit excitation spectrum data to be identified is input, the enhanced feature extraction module 10 first preprocesses it to remove baseline drift. Then, the module calculates 13 basic statistical features of the spectrum, including mean, standard deviation, skewness, and kurtosis; it uses a peak-finding algorithm to identify the peaks on the spectrum and calculates 15 peak morphological features, including peak number, peak height, peak width, and symmetry; it performs an FFT transform on the spectrum intensity sequence to extract 9 frequency domain features, including total power and dominant frequency; it detects whether Chevron interference fringes exist in the spectrum, and if so, calculates 8 Chevron mode features, including peak pair spacing and angle; finally, it calculates 3 signal quality features, including signal-to-noise ratio. All the above features are concatenated to form a 48-dimensional feature vector. This vector is then input into a pre-trained hierarchical classifier, which sequentially passes through nodes N0 to N4 for judgment, and finally outputs the category label of the spectrum, such as single-peak Chevron. If different feature combination strategies are adopted, such as selecting only the top few features with the highest contribution in each category, a 45-dimensional feature vector that meets the minimum dimensionality requirement can also be formed. This application does not impose any restrictions on this.

[0093] Through the above technical solution, this application ensures the sufficiency and systematic nature of feature engineering by specifying minimum dimensions and category composition. Because it constructs feature vectors with a dimension of at least 45, encompassing five categories of features—basic statistics, peak morphology, frequency domain, Chevron pattern, and signal quality—it provides the classifier with sufficient and multi-dimensional discrimination criteria. Especially when distinguishing complex and similar spectral patterns such as single-peak plus single-peak Chevron patterns, it effectively avoids misjudgments caused by insufficient feature dimensions or missing information, thereby improving the generalization ability and classification accuracy of the overall recognition model. The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A one-dimensional spectral recognition method based on enhanced feature engineering and hierarchical classifiers, characterized in that, Includes the following steps: Step 1: Extract multi-dimensional enhancement features from the one-dimensional spectral data to be identified to form a feature vector. The feature vector includes at least morphological features for describing peak morphology and Chevron mode features for describing peak pair patterns. Step 2: Construct a hierarchical classifier, which contains multiple binary classification nodes arranged in a hierarchical order of one-dimensional spectral categories, with each node corresponding to a binary classification model; Step 3: For each node in the hierarchical classifier, using only the training sample set that can reach that node, optimize to obtain a classification threshold corresponding to that node that makes the preset evaluation index optimal. Step 4: Input the feature vector into the hierarchical classifier. According to the hierarchical order of the nodes, compare the predicted probability output by the current node with the classification threshold corresponding to the node. Determine the category of the current node based on the comparison result, and decide whether to output the final classification result or move to the next node. The morphological features include one or more of the following: peak number, maximum peak height, peak height ratio, peak symmetry score, peak sharpness score, peak width statistics, and interpeak distance statistics; The Chevron mode features include one or more of the following: peak pair symmetry score, peak pair separation, peak pair height ratio, and Chevron angle; The hierarchical classifier constructed in step two is a five-node hierarchical classifier, and its node hierarchy is as follows: N0 node: Determine if there is a peak; N1 node: Determines whether it is a single peak, and only executes when N0 node determines that there is a peak; N2 node: Determines whether it is a single-peak Chevron mode, and executes only when N0 node is determined to be peak and N1 node is determined to be non-single-peak; N3 node: Determines whether it is a bimodal node, and only executes when N0 node is a peak, N1 node is not a single peak, and N2 node is not a single peak Chevron. N4 node: Determines whether it is a single-peak + single-peak Chevron mode. It is only executed when N0 node is peaked, N1 node is not single-peaked, N2 node is not single-peaked Chevron, and N3 node is not bimodal. In step three, the optimization of the classification threshold corresponding to each node is specifically as follows: For each node, the predicted probability of the binary classification model corresponding to the node on its reachable sample set is collected, and a search is performed within the preset classification threshold range with a preset step size. For each classification threshold obtained by the search, a preset evaluation index is calculated. The preset evaluation index is the classification accuracy or F1 score, and the classification threshold that makes the preset evaluation index reach the optimal value is determined as the classification threshold corresponding to the node.

2. The one-dimensional spectral recognition method according to claim 1, characterized in that, The multi-dimensional enhancement features extracted in step one also include at least one of the following: a basic statistical feature set, a frequency domain feature set, and a signal quality feature set; The basic statistical feature set includes mean, standard deviation, skewness, and kurtosis; The frequency domain feature set includes total power, dominant frequency, frequency centroid, and frequency bandwidth; The signal quality feature set includes the signal-to-noise ratio.

3. The one-dimensional spectral recognition method according to claim 1, characterized in that, The binary classification model in step two is automatically trained using an automatic machine learning framework. The automatic machine learning framework takes the feature vector as input and outputs binary classification labels based on node hierarchy. The training time is limited to 300-600 seconds.

4. A one-dimensional spectral recognition system based on enhanced feature engineering and hierarchical classifiers, characterized in that, The system for implementing the one-dimensional spectral identification method according to any one of claims 1 to 3, the system comprising: Enhanced feature extraction module: used to extract multi-dimensional enhanced features from one-dimensional spectral data to form feature vectors; Hierarchical classifier module: Used to store and run a hierarchical classifier containing multiple binary classification nodes arranged in hierarchical order of categories; Threshold optimization module: used to optimize and store the corresponding classification threshold for each node in the hierarchical classifier based on its reachable sample set; Serial decision reasoning module: It is used to input the feature vector into the hierarchical classifier, compare the predicted probability output by the current node with its corresponding classification threshold according to the node order, and output the final recognition result.

5. The one-dimensional spectral recognition system according to claim 4, characterized in that, The binary classification model in the hierarchical classifier module includes either the CatBoost model or the LightGBM model.

6. The one-dimensional spectral recognition system according to claim 4, characterized in that, It also includes a model update module, which, after obtaining newly labeled training samples, performs a threshold search on the reachable sample set of each node to update the classification threshold of each node, and optionally retrains the binary classification model of each node.

7. The one-dimensional spectral recognition system according to claim 4, characterized in that, The enhanced feature extraction module extracts multi-dimensional enhanced features including at least 45-dimensional feature vectors, which contain five types of features: basic statistical features, peak morphology features, frequency domain features, Chevron mode features, and signal quality features.

Images