Multimodal fusion method for LIBS (laser-induced breakdown spectroscopy)

Abstract

The invention discloses an LIBS laser-induced spectrometer multi-modal fusion method, and the method comprises the steps: carrying out the dynamic decomposition of collected multi-modal spectrum data through a data structure fidelity preprocessing module based on tensor dynamic modal decomposition, and obtaining a target dynamic modal; performing fusion analysis on the target dynamic mode by using an attention enhancement depth fusion module based on interactive distillation and dispersion loss to obtain an analysis result; in the analysis process, a closed-loop adaptive scanning and control module is used for intelligent monitoring and dynamic regulation and control until all analysis results reach preset confidence and quality standards; and an interactive decision support module based on operable anti-fact explanation is utilized to convert the analysis result into decision support valuable to the user. In this way, analysis precision loss caused by information dimension reduction and structural damage is avoided.

Description

Technical Field

[0001] This application relates to the field of LIBS laser-induced spectrometer technology, and in particular to a multimodal fusion method for LIBS laser-induced spectrometers. Background Technology

[0002] Spectroscopic analysis, as a core means of detecting the composition and structure of substances, plays an indispensable role in many fields. Currently, there are two main types of spectroscopic analysis techniques: one is molecular spectroscopy, represented by visible-near-infrared-short-wave infrared spectroscopy. It excels at identifying the molecular structure information of substances, especially sensitive to the vibrational absorption of specific chemical bonds (such as hydrogen-containing groups), but it is difficult to accurately resolve the specific composition and content of elements. The other is atomic emission spectroscopy, represented by laser-induced breakdown spectroscopy. It uses high-energy lasers to excite samples and can accurately obtain the elemental types and quantitative information of substances, but it is powerless to obtain structural information at the molecular level.

[0003] To overcome the limitations of single technologies, multimodal spectral fusion technology has emerged, aiming to integrate the information advantages of two technologies. However, current mainstream fusion methods still suffer from significant technical bottlenecks, limiting their performance in complex application scenarios: 1. Disruption of the intrinsic structure of data: Current methods generally "flatten" spectral data from different sources, forcibly converting multidimensional data with spatial or temporal structure (such as spectral cubes obtained from scanning imaging) into one-dimensional vectors before stitching and fusing them. This processing method ignores and destroys the intrinsic correlation of data in dimensions such as spatial distribution and temporal evolution, resulting in the loss of a large amount of key information.

[0004] 2. Decision bias in fusion models: When using deep learning models for fusion analysis, the training process is often a "black box." The internal attention mechanism is easily misled by noise or non-essential spurious correlation features in the data, causing the model's "attention mechanism to fail." Although it may perform well on the training set, its generalization ability and robustness of the analysis results will be significantly reduced when facing new samples or complex environments.

[0005] 3. Static vs. Open-Loop Analytical Processes: Most existing spectral analysis systems follow a fixed, linear "open-loop" workflow: data acquisition, processing, analysis, and result output are completed according to a preset procedure. The entire process is static and lacks the ability to dynamically adjust based on real-time analysis conditions. Unlike human experts, the system cannot proactively perform supplementary detection or adjust analytical strategies when anomalies are detected or analytical results are uncertain.

[0006] 4. Lack of interpretability and decision support in analytical results: While most advanced analytical systems can provide highly accurate analytical conclusions (e.g., determining whether a sample is qualified or not), they cannot explain the specific basis for their judgments. Users are faced with results that are "what" but not "why." More importantly, when the results do not meet expectations, the system cannot provide any actionable or forward-looking improvement suggestions, and cannot answer the crucial question in actual production and research: "How can we achieve the desired results?"

[0007] Existing multimodal spectral fusion technologies, due to their inherent data structure corruption, black-box model decision-making, static analysis processes, and uninterpretable results, cannot simultaneously meet the comprehensive requirements of high precision, high efficiency, high robustness, and strong interpretability when facing analytical tasks with complex compositions, variable states, or those requiring fine-tuning. In particular, they lack an intelligent analysis capability that can simulate expert thinking, achieve proactive exploration and closed-loop self-correction, and ultimately provide users with operable and executable decision support. Summary of the Invention

[0008] The multimodal fusion method for LIBS laser-induced spectrometers provided in this application can avoid the loss of analytical accuracy caused by information dimensionality reduction and structural destruction.

[0009] Firstly, this application provides a multimodal fusion method for LIBS laser-induced spectrometers. The LIBS laser-induced spectrometer multimodal fusion method includes: dynamically decomposing the acquired multimodal spectral data using a data structure fidelity preprocessing module based on tensor dynamic mode decomposition to obtain the target dynamic mode; performing fusion analysis on the target dynamic mode using an attention-enhanced deep fusion module based on interactive distillation and diffusion loss to obtain analysis results; wherein, the analysis process utilizes a closed-loop adaptive scanning and control module for intelligent monitoring and dynamic adjustment until all analysis results reach preset confidence levels and quality standards; and using an interactive decision support module based on operable counterfactual interpretation to transform the analysis results into valuable decision support for the user.

[0010] The data structure fidelity preprocessing module based on tensor dynamic mode decomposition includes: a data tensor quantization construction unit, a time-shifted tensor pair construction unit, a tensor singular value decomposition and dimensionality reduction unit, a dimensionality reduction state transition operator solving unit, and a dynamic mode extraction and reconstruction unit. This module dynamically decomposes the acquired multimodal spectral data to obtain the target dynamic mode. Specifically, it uses the data tensor quantization construction unit to perform a unified structured representation of the acquired multimodal spectral data, obtaining a state snapshot; the time-shifted tensor pair construction unit stitches the state snapshots together in time or depth to obtain the preceding and following state tensors; the tensor singular value decomposition and dimensionality reduction unit performs tensor singular value decomposition on the preceding state tensor to obtain a low-rank subspace; the dimensionality reduction state transition operator solving unit solves from the low-rank subspace and the following state tensor to obtain the dimensionality reduction state transition operator; and the dynamic mode extraction and reconstruction unit performs eigenvalue decomposition on the dimensionality reduction state transition operator to obtain the target dynamic mode.

[0011] The method involves using tensor singular value decomposition and dimensionality reduction units to perform tensor singular value decomposition on the preceding state tensor to obtain a low-rank subspace. This includes: using tensor singular value decomposition and dimensionality reduction units to perform tensor singular value decomposition on the preceding state tensor, decomposing the preceding state tensor into the product of three tensors; retaining the part with the largest energy in the product of the three tensors to obtain a low-rank subspace.

[0012] Among them, the deep fusion module for attention enhancement based on interactive distillation and diffusion loss deploys a deep learning model. The deep fusion module for attention enhancement based on interactive distillation and diffusion loss is used to perform fusion analysis on the target dynamic mode and obtain analysis results, including: using a deep learning model to perform fusion analysis on the target dynamic mode and obtain analysis results.

[0013] The deep learning model is trained as follows: training spectral data is input into the teacher model and the deep learning model respectively to obtain the teacher attention map in the teacher model and the target attention map in the deep learning model; the difference between the teacher attention map and the target attention map is calculated and used as the attention alignment loss; during training, for all samples in a batch, feature vectors output by any two different samples in any feature layer of the deep learning model are extracted to obtain any two feature vectors; the diffusion loss is obtained based on the two feature vectors; and the model parameters in the deep learning model are adjusted based on the attention alignment loss and the diffusion loss.

[0014] The closed-loop adaptive scanning and control module includes a strategy synthesis unit, a process monitoring unit, and an intelligent repair unit. The analysis process utilizes this module for intelligent monitoring and dynamic adjustment until all analysis results reach preset confidence levels and quality standards. This includes: when the analysis task starts, the strategy synthesis unit performs global intelligent planning to generate the optimal initial analysis strategy; executing the initial analysis strategy, the process monitoring unit continuously senses and evaluates various states during the analysis process; when the process monitoring unit detects an abnormal event, the intelligent repair unit performs causal diagnosis of the abnormal event and selects and executes the most appropriate countermeasure from a predefined repair strategy library based on the diagnosis results.

[0015] The process monitoring unit continuously senses and evaluates various states during the analysis process, including: receiving instrument status data, spectral quality data, and model analysis feedback; and detecting abnormal events in the instrument status data, spectral quality data, and model analysis feedback.

[0016] The process involves selecting and executing the most appropriate response from a predefined repair strategy library based on the diagnostic results. This includes: responding to a diagnostic result indicating the discovery of a suspected critical area, selecting and executing a first repair strategy from the predefined repair strategy library; the first repair strategy includes: pausing the original scanning path, immediately generating a high-density, refined scanning subtask around the area, and then returning to the main path after completion; responding to a diagnostic result indicating a decline in regional data quality, selecting and executing a second repair strategy from the predefined repair strategy library; the second repair strategy includes: automatically increasing the integration time of spectral acquisition in the area, or, for laser-induced breakdown spectra, automatically performing a pre-ablation pulse to clean the sample surface before re-measuring; responding to a diagnostic result indicating blurred boundaries, selecting and executing a third repair strategy from the predefined repair strategy library; the third repair strategy includes: initiating a boundary fine-tuning subroutine, automatically performing zigzag or interpolation scans on both sides of the current position to accurately determine the boundary location with higher spatial resolution.

[0017] The interactive decision support module based on actionable counterfactual interpretation includes: a high-dimensional probability density landscape unit, a counterfactual optimization objective unit, a path planning algorithm-based solution unit, and a unit for generating and presenting actionable suggestions. This module transforms the analysis results into valuable decision support for users, including: using the high-dimensional probability density landscape unit to learn from training data to obtain a high-dimensional probability density distribution function, and constructing this function as a data landscape; using the counterfactual optimization objective unit to construct an optimization objective based on the user-input initial point and target area, with the objective being: to find a target path in the data landscape that starts from the initial point and ultimately reaches the target area; using the path planning algorithm-based solution unit to solve the optimization objective and obtain the optimal target path; and using the unit for generating and presenting actionable suggestions to translate the optimal target path into language that users can understand and execute, thus providing valuable decision support.

[0018] The process of translating the optimal target path into a language that users can understand and execute by using the actionable suggestion generation and presentation unit includes: interpreting the optimal target path using the actionable suggestion generation and presentation unit to obtain the interpretation results; generating an interactive report based on the interpretation results; wherein the interactive report includes overall suggestions, step-by-step paths, and sensitivity analysis.

[0019] The beneficial effects of this application are as follows: Unlike existing technologies, the LIBS laser-induced spectrometer multimodal fusion method provided in this application includes: dynamically decomposing the acquired multimodal spectral data using a data structure fidelity preprocessing module based on tensor dynamic mode decomposition to obtain the target dynamic mode; performing fusion analysis on the target dynamic mode using an attention-enhanced deep fusion module based on interactive distillation and diffusion loss to obtain analysis results; wherein, the analysis process utilizes a closed-loop adaptive scanning and control module for intelligent monitoring and dynamic adjustment until all analysis results reach preset confidence levels and quality standards; and transforming the analysis results into valuable decision support for users using an interactive decision support module based on operable counterfactual interpretation. Through the above methods, the intrinsic structure and dynamic evolution information of spectral data in multiple dimensions such as space and time can be fully preserved and utilized throughout the fusion analysis process, avoiding the loss of analytical accuracy caused by information dimensionality reduction and structural destruction. Furthermore, it guides the deep fusion model beyond simple statistical fitting, forming an expert-like "cognitive ability," enabling it to accurately focus on the key characteristics that determine the properties of matter and shield the interference of irrelevant information, thereby achieving more reliable and robust judgments on unknown samples. Furthermore, it breaks away from rigid analytical processes, constructing a closed-loop intelligent system with self-awareness and self-adjustment capabilities. This system can dynamically optimize analytical strategies based on real-time feedback during the analysis process, achieving optimal analytical efficiency and effectiveness. It also upgrades the analytical system from a simple "detection tool" to a "decision-making partner," not only providing "what" and "why," but also offering users specific, quantifiable, and clearly operational "how-to" improvement solutions based on a deep understanding of the underlying mechanisms.

[0020] In other words, this application proposes a complete and systematic intelligent analysis system architecture. The system consists of four core modules: data preprocessing, deep fusion analysis, closed-loop adaptive control, and interactive decision support, forming a complete technical closed loop from data fidelity to intelligent decision-making.

[0021] The system's workflow begins with structured, high-fidelity input of multidimensional spectral data, followed by high-precision fusion analysis using attention-enhanced deep models. The entire analysis process is intelligently monitored and dynamically controlled by a closed-loop adaptive framework, and finally, through an operable counterfactual interpretation module, the analysis results are transformed into valuable decision support for users. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating an embodiment of the LIBS laser-induced spectrometer multimodal fusion method provided in this application; Figure 2 yes Figure 1 A flowchart illustrating an embodiment of step 11; Figure 3 yes Figure 1 A flowchart illustrating an embodiment of step 12; Figure 4 yes Figure 1 A flowchart of an embodiment of step 13; Figure 5 yes Figure 1 A flowchart illustrating an embodiment of step 14; Figure 6 This is a schematic diagram of the structure of an embodiment of the multimodal spectral intelligent analysis system provided in this application; Figure 7 This is a schematic diagram of an embodiment of the computer-readable storage medium provided in this application. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It is understood that the specific embodiments described herein are only for explaining this application and not for limiting it. Furthermore, it should be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all structures. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0024] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0025] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the LIBS laser-induced spectrometer multimodal fusion method provided in this application. The LIBS laser-induced spectrometer multimodal fusion method includes: Step 11: Use the data structure fidelity preprocessing module based on tensor dynamic mode decomposition to dynamically decompose the acquired multimodal spectral data to obtain the target dynamic mode.

[0026] In some embodiments, the data structure fidelity preprocessing module based on tensor dynamic mode decomposition includes: a data tensor quantization construction unit, a time-shifted tensor pair construction unit, a tensor singular value decomposition and dimensionality reduction unit, a dimensionality reduction state transition operator solving unit, and a dynamic mode extraction and reconstruction unit. (See also...) Figure 2 Step 11 can be the following process: Step 21: Use data tensor quantization to construct a unified structured representation of the collected multimodal spectral data to obtain a state snapshot.

[0027] Step 22: Use the constructed time-shift tensor pair to stitch together the state snapshots in time or depth to obtain the preceding state tensor and the following state tensor.

[0028] Step 23: Perform tensor singular value decomposition on the preceding state tensor using tensor singular value decomposition and dimensionality reduction unit to obtain a low-rank subspace.

[0029] In some embodiments, tensor singular value decomposition and dimensionality reduction units are used to perform tensor singular value decomposition on the preceding state tensor, decomposing the preceding state tensor into the product of three tensors; the part with the largest energy in the product of the three tensors is retained to obtain a low-rank subspace.

[0030] Step 24: Use the solution of the dimension reduction state transition operator unit to obtain the dimension reduction state transition operator from the low-rank subspace and the subsequent state tensor.

[0031] Step 25: Use the dynamic mode extraction and reconstruction unit to perform eigenvalue decomposition on the dimensionality reduction state transition operator to obtain the target dynamic mode.

[0032] In one application scenario, the data structure fidelity preprocessing module based on tensor dynamic mode decomposition is described as follows: 1. Module objectives and core ideas.

[0033] The core objective of this module is to address the problem of data structure loss caused by traditional "flattening" methods in the preprocessing stage of multimodal spectral data fusion, thereby providing inputs with higher information fidelity and better feature quality for subsequent deep fusion analysis.

[0034] Traditional data processing methods, such as flattening multidimensional spectral image scan data (including spatial coordinates and spectral bands) into two-dimensional tables or one-dimensional vectors, completely destroy the spatial neighborhood relationships between pixels and their evolution patterns over time or depth sequences. This approach essentially simplifies a structured, dynamic, and complex system into an unstructured, static dataset, leading to irreversible information loss.

[0035] To overcome this bottleneck, this module introduces an advanced data analysis paradigm. Its core idea is to treat the entire multimodal spectral dataset as a complex dynamic system and analyze it directly within the system's high-dimensional original form (i.e., tensor form). Instead of viewing the data as a collection of isolated points, it is considered as a whole, aiming to extract the dominant intrinsic patterns that describe the spatiotemporal evolution of the entire system.

[0036] Specifically, this method originates from techniques used to analyze dynamic systems such as complex fluids and video sequences. This technique can decompose high-dimensional time-series data into a set of spatially fixed but time-varying spatial modes, along with their respective temporal dynamic characteristics.

[0037] However, standard methods are limited to processing two-dimensional matrix data. The key innovation of this module lies in its successful extension and application to the field of multidimensional spectral data, forming a complete process for tensor dynamic mode decomposition. This extension enables this application to perform operations directly in multidimensional tensor space, fundamentally avoiding the structural destruction and information distortion caused by data "flattening".

[0038] 2. Technical implementation process.

[0039] The implementation process of this module is rigorous and systematic, mainly including the following five steps: Step 1: Data Tensorization Construction.

[0040] First, the acquired multimodal spectral data is represented in a unified structure. For example, for a scanning analysis performed in a specific region, this application can construct a four-dimensional data tensor, whose dimensions can be defined as: (image height, image width, spectral band, time / depth layer). The "spectral band" dimension integrates data from visible-near-infrared-short-wave infrared spectroscopy and laser-induced breakdown spectroscopy; the "time / depth layer" dimension can represent different time points of dynamic chemical reactions, or different depths of layer-by-layer ablation on the sample by laser-induced breakdown spectroscopy. This application refers to the complete data at a certain moment or depth as a state snapshot.

[0041] Step 2: Construct time-shifted tensor pairs.

[0042] In order to analyze the dynamic evolution of the system, this application requires the construction of two large tensors with a one-unit offset in time or depth.

[0043] The preceding state tensor is formed by stitching together snapshots of states from the 0th to the (T-1th)th state along a new dimension.

[0044] The post-order state tensor is formed by stitching together the snapshots of the first to the Tth states along a new dimension.

[0045] These two tensors fully describe the state of the system at each time step and the state at the next time step, and are the basis for subsequent dynamic analysis.

[0046] Step 3: Tensor Singular Value Decomposition and Dimensionality Reduction.

[0047] To capture the core structural features from massive datasets and reduce the complexity of subsequent computations, this application performs tensor singular value decomposition on the "preceding state tensor." This is a generalization of matrix singular value decomposition in high-dimensional space, which decomposes the original data tensor into the product of three tensors. By retaining the part with the highest energy (i.e., the largest singular value), this application obtains a low-rank subspace that optimally represents all the major, common spatial structural patterns in the original data.

[0048] Step 4: Solve for the dimensionality reduction state transition operator.

[0049] This step is the core of the entire method. It aims to solve for a linear state transition operator that describes how the system's main spatial patterns evolve from the current time step to the next. Solving this operator directly in high-dimensional space is very difficult; therefore, using the low-rank subspace obtained in the previous step, a dimensionality-reduced state transition operator is solved. This dimensionality-reduced operator accurately describes the evolution of data within the core feature space.

[0050] Step 5: Dynamic modality extraction and reconstruction.

[0051] By performing eigenvalue decomposition on the "dimensionality reduction state transition operator", this application can obtain a set of eigenvalues ​​and corresponding eigenvectors.

[0052] Dynamic modes: Projecting these feature vectors back into the original high-dimensional space yields the dynamic modes of the system. Each mode represents a fundamental, coherent structural pattern in both spatial and spectral dimensions (e.g., the distribution characteristics of a particular mineral phase).

[0053] Modal dynamics: The eigenvalues ​​corresponding to each mode describe the evolution of that mode over time or depth, such as whether it is growing, decaying, or oscillating.

[0054] 3. Mathematical formula expression.

[0055] For clarity, the mathematical relationships of the core steps are expressed as follows: 1. Basic assumptions about the dynamic evolution of the system: State snapshot (t+1) = State transition operator * State snapshot (t).

[0056] 2. Construction of time-shifted tensor pairs: The preceding state tensor = [state snapshot_0, state snapshot_1, ..., state snapshot_T-1].

[0057] The subsequent state tensor = [state snapshot_1, state snapshot_2, ..., state snapshot_T].

[0058] 3. Tensor Singular Value Decomposition (TSVD): The preceding state tensor = left orthogonal tensor * diagonal tensor * (right orthogonal tensor) transpose.

[0059] Note: The multiplication here refers to tensor multiplication, and the transpose refers to tensor transpose.

[0060] 4. Solving for the dimension reduction state transition operator: Dimension reduction transition operator = (left orthogonal tensor) transpose * post-order state tensor * right orthogonal tensor * (diagonal tensor) inverse.

[0061] 5. Extraction of dynamic modes: First, perform eigenvalue decomposition on the dimension reduction transition operator: Dimensionality reduction transition operator = modality factor tensor * eigenvalue diagonal tensor * (modality factor tensor) inverse.

[0062] Then, the complete dynamic mode is reconstructed: Dynamic mode = Left orthogonal tensor * Mode factor tensor.

[0063] 4. Module output and technical advantages.

[0064] The final output of this module is a series of dynamic modes and their respective dynamic characteristic parameters. These outputs offer unparalleled technical advantages compared to traditional preprocessing results: High information fidelity: Analysis is performed directly in tensor space, preserving the spatiotemporal structure of the data and avoiding information loss.

[0065] Excellent feature quality: The extracted dynamic modes are globally optimal features that can represent the core evolutionary laws of the system and have clear physical meaning, far superior to traditional features based on a single pixel or local region.

[0066] High computational efficiency: Dimensionality reduction greatly reduces the computational burden of subsequent complex model fusion analysis.

[0067] In summary, this module, by directly and dynamically decomposing high-dimensional data, not only solves the problem of structure fidelity in the data preprocessing stage, but also provides high-quality features containing rich spatiotemporal evolution laws for subsequent fusion analysis. This forms a solid foundation for the entire intelligent analysis system to achieve high precision and deep insights.

[0068] Step 12: Use the attention-enhanced deep fusion module based on interactive distillation and diffusion loss to perform fusion analysis on the target dynamic modes and obtain the analysis results.

[0069] In some embodiments, a deep learning model is deployed in the attention-enhanced deep fusion module based on interactive distillation and diffusion loss. Step 12 may involve using the deep learning model to perform fusion analysis on the target dynamic modality to obtain the analysis results.

[0070] In some embodiments, see Figure 3 The deep learning model is trained in the following way: Step 31: Input the training spectral data into the teacher model and the deep learning model respectively to obtain the teacher attention map in the teacher model and the target attention map in the deep learning model.

[0071] Step 32: Calculate the difference between the teacher attention map and the target attention map, and use the difference as the attention alignment loss.

[0072] Step 33: During the training process, for all samples in a batch, extract the feature vectors output by any feature layer of any two different samples in the deep learning model to obtain any two feature vectors.

[0073] Step 34: Obtain the diffusion loss based on any two eigenvectors.

[0074] Step 35: Adjust the model parameters in the deep learning model based on the attention alignment loss and diffusion loss.

[0075] In one application scenario, the attention-enhanced deep fusion module based on interactive distillation and diffusion loss is introduced as follows: 1. Module objectives and core ideas.

[0076] The core objective of this module is to address the issues of decision bias and insufficient generalization ability in deep learning models when fusing multimodal spectral data, due to the inherent "black box" nature and limitations of the training paradigm. The aim is to endow the fusion model with an expert-like "cognitive ability," enabling it to accurately focus on the key features that determine the properties of matter and to form a clear and distinguishable internal understanding of different substances.

[0077] Traditional deep learning model training typically focuses only on the accuracy of the final output, without constraining its complex internal decision-making process. This leads to two main problems: 1) Attention Hacking: During training, the model may "sneak" learning pseudo-correlation in the data that is not essential for causation. For example, if a certain ore in the training data is always accompanied by specific background noise, the model may mistakenly treat the background noise as an important feature for identifying the ore. When analyzing in a new environment without this noise, the model's performance will drop sharply. This phenomenon is essentially the model's attention mechanism being "hijacked" or "deceived" by irrelevant information, failing to focus on the truly crucial spectral features.

[0078] 2) Representation Collapse: In the model's internal feature space, the feature representations of different categories of substances may be too similar to distinguish. Especially for substances with similar compositions and subtle differences in spectral characteristics, the internal feature vectors learned by the model will cluster tightly together in space, forming what is known as "representation collapse." This makes it difficult for subsequent classifiers to define clear decision boundaries, leading to classification errors.

[0079] To address the aforementioned issues, this module introduces two advanced deep learning training strategies. The core idea is to go beyond simply supervising the model's "behavioral outcomes" and delve deeper into its "thought processes" to guide and shape them. This application employs a "teacher-student" interaction model to force the model to learn the "thinking style" of an expert; simultaneously, through a special loss function, it forces the model to internally recognize and clearly distinguish between different things.

[0080] 2. Technical implementation process.

[0081] The implementation process of this module consists of two innovative technologies that work together during model training to enhance the model's cognitive abilities.

[0082] Part 1: Attention guidance based on interactive distillation.

[0083] This technology aims to address the problem of "attention deficit" by calibrating the attention mechanisms of a deployed "student model" by mimicking a more powerful "teacher model".

[0084] 1) Constructing or Introducing a Teacher Model: This application first requires a "teacher model." This model can be a larger, more complex, ultra-large-scale model pre-trained on massive amounts of multi-source data, or an expert system that has been validated as extremely reliable in a specific domain. This application assumes that this teacher model has a deeper and more fundamental understanding of spectral data, and that its internal attention allocation pattern is closer to the true causal relationships.

[0085] 2) Extracting the Teacher's Attention Map: When the teacher model processes spectral data, this application can extract the attention map generated by its internal attention layer (e.g., in a Transformer structure). This map is a matrix that quantitatively describes the degree of attention the model pays to different parts of the input spectral data (such as different bands or different spatial locations) during analysis. This application considers this map to be the teacher model's "thought trajectory" or "distribution of attention points."

[0086] 3) Attention Alignment Training: When training the student model of this application, in addition to the conventional main loss function (such as cross-entropy loss) used to match the final output results, this application introduces an attention alignment loss function. The specific process is as follows: The same spectral data is input into both the teacher model and the student model.

[0087] Attention maps of corresponding layers within the teacher and student models are extracted separately.

[0088] Calculate the difference between the two attention maps (e.g., using mean squared error or KL divergence).

[0089] This difference is used as the attention alignment loss and added to the total loss function, and then optimized together through backpropagation.

[0090] In this way, the student model learns not only how to make correct judgments, but also how to "think correctly." It is guided to focus on the feature regions that the teacher model considers important, while ignoring spurious or irrelevant noisy regions, thereby greatly improving the model's generalization ability and robustness.

[0091] Part Two: Feature Space Reconstruction Based on Diffusion Loss.

[0092] This technology aims to solve the problem of "representation collapse" by actively reshaping the feature space inside the model through a special loss function, making it more conducive to classification.

[0093] 1) Limitations of Traditional Contrastive Learning: Traditional contrastive learning methods construct "positive sample pairs" (e.g., different spectral measurements of the same substance) and "negative sample pairs" (spectral data of different substances) in the feature space to "bring" positive sample pairs closer together and "pull" negative sample pairs further apart. This method is effective, but it heavily relies on the construction of high-quality positive and negative sample pairs, which is often very difficult in practical applications.

[0094] 2) Introducing a diffusion loss without positive samples: This module employs a more ingenious and efficient diffusion loss. The design philosophy of this loss function is that within a training batch, the feature representations of any two different samples should be considered as "negative sample pairs" and pushed away from each other. It completely eliminates the need for "positive sample pairs".

[0095] 3) Implementation of the diffusion loss: During training, for all samples within a batch, this application extracts their feature vectors output by a certain feature layer of the model. Then, the diffusion loss function calculates the distance between any two feature vectors and drives the optimization process to maximize this distance. Its effect is to generate a "repulsive force" in the feature space, causing the feature representations of all samples to be dispersed as much as possible and evenly distributed in the feature space.

[0096] In this way, even very similar substances are forced to separate their feature representations, thus solving the "representation collapse" problem. This makes it easier for subsequent classifiers to learn clear decision boundaries, significantly improving classification accuracy, especially when dealing with substances with subtle differences.

[0097] 3. Mathematical formula expression.

[0098] The core loss function of this module is expressed as follows (all parameters are in Chinese): 1). Total Loss Function: Total loss = Main loss + Attention alignment weight * Attention alignment loss + Diffusion loss weight * Diffusion loss.

[0099] 2). Main Loss Function (taking classification task as an example): Main loss = -Σ (true label * log(model predicted probability)).

[0100] 3). Attention Alignment Loss Function (taking mean squared error as an example): Attention alignment loss = Σ ((Teacher attention map - Student attention map)^2) Note: Σ represents the summation of all elements in the graph.

[0101] 4). Dispersion loss function: Diffusion loss = log ( Σ_{i≠j} exp( - distance(feature_i, feature_j) / temperature coefficient) ).

[0102] Feature_i and feature_j are feature vectors of any two different samples within the same batch.

[0103] Distance(...) is a distance metric function, such as the square of the Euclidean distance.

[0104] The temperature coefficient is a hyperparameter used to adjust the strength of the repulsive force.

[0105] Σ_{i≠j} represents summing over all possible sample pairs within the batch.

[0106] 4. Module output and technical advantages.

[0107] After model training is complete, this module outputs a deeply optimized deep learning model with powerful fusion analysis capabilities. Compared to models trained using traditional methods, it possesses the following significant technical advantages: More reliable decision-making logic: By learning the "attention patterns" of experts, the model's decision-making basis is closer to the physical essence of things, and its robustness and generalization ability are significantly enhanced.

[0108] Clearer internal representation: By reshaping the feature space, the model's internal cognition (feature representation) of different substances has clearer boundaries, greatly improving its ability to distinguish similar substances.

[0109] Higher training efficiency: Diffusion loss eliminates the need to construct complex positive and negative sample pairs, simplifying the training process and making it more feasible and efficient in practical applications.

[0110] In summary, this module, through deep intervention and guidance in the model training process, has achieved a paradigm shift from "behavioral imitation" to "mind shaping," and is the core engine for the entire intelligent analysis system to achieve high-precision and high-robustness analysis.

[0111] Step 13: The analysis process utilizes a closed-loop adaptive scanning and control module for intelligent monitoring and dynamic adjustment until all analysis results reach the preset confidence level and quality standards.

[0112] In some embodiments, the closed-loop adaptive scanning and control module includes a strategy synthesis unit, a process monitoring unit, and an intelligent repair unit. (See also...) Figure 4 Step 13 can be the following process: Step 41: When the analysis task starts, the strategy synthesis unit is used to perform global intelligent planning and generate the optimal initial analysis strategy.

[0113] Step 42: Execute the initial analysis strategy and use the process monitoring unit to continuously sense and evaluate various states during the analysis process.

[0114] In some embodiments, the process monitoring unit receives instrument status data, spectral quality data, and model analysis feedback; and performs abnormal event detection on the instrument status data, spectral quality data, and model analysis feedback respectively.

[0115] Step 43: When the process monitoring unit detects an abnormal event, it uses the intelligent repair unit to perform causal diagnosis on the abnormal event, and selects and executes the most appropriate response measures from the predefined repair strategy library based on the diagnosis results.

[0116] In some embodiments, in response to a diagnostic result indicating the discovery of a suspected critical area, a first repair strategy is selected from a predefined repair strategy library and executed; wherein, the first repair strategy includes: pausing the original scanning path, immediately generating a high-density fine-grained scanning subtask around the area, and then returning to the main path after completion.

[0117] In some embodiments, in response to a diagnostic result indicating a decline in regional data quality, a second repair strategy is selected from a predefined repair strategy library and executed; wherein the second repair strategy includes: automatically increasing the integration time of spectral acquisition in the region, or, for laser-induced breakdown spectra, automatically performing a pre-ablation pulse to clean the sample surface before re-measuring.

[0118] In some embodiments, in response to a diagnostic result indicating blurred boundaries, a third repair strategy is selected from a predefined repair strategy library and executed; wherein the third repair strategy includes: initiating a boundary fine-tuning subroutine to automatically perform zigzag or interpolation scans on both sides of the current position to accurately determine the boundary position with higher spatial resolution.

[0119] In one application scenario, the closed-loop adaptive scanning and control module based on "strategy synthesis - process monitoring - intelligent repair" is described below: 1. Module objectives and core ideas.

[0120] The core objective of this module is to completely transform the static, open-loop operating mode of traditional spectral analysis systems, solving their rigid and passive processes. The aim is to upgrade the analysis system from a "tool" that can only passively execute preset programs into an "intelligent agent" capable of autonomous planning, proactive exploration, and self-correction based on real-time feedback.

[0121] Traditional spectral analysis workflows are like following a map in the dark, strictly adhering to pre-set scanning paths and parameters. The fundamental flaw of this approach is that it assumes the analytical task is completely deterministic and predictable, ignoring the ubiquitous unknowns and dynamics of the real world. For example, unexpected anomalies may exist in the sample, some areas may have poor signal quality due to surface contamination, or preliminary analysis results may show high uncertainty. In these situations, the rigid workflow is incapable of handling them, either missing crucial discoveries or drawing erroneous conclusions based on low-quality data.

[0122] To overcome this limitation, this module introduces a closed-loop intelligent control framework derived from the field of advanced autonomous robot control. Its core idea is to treat each spectral analysis task as a dynamic, exploratory "task execution" process, rather than a static "program execution" process. The system no longer completes the entire analysis in a single, one-time event, but rather continuously optimizes its analytical behavior within a persistent "perception-decision-action" closed loop until the task objective is achieved with the highest quality. This closed loop consists of three tightly coupled components: strategy synthesis (Synthesize), process monitoring (Monitor), and intelligent repair (Repair).

[0123] 2. Technical implementation process.

[0124] The implementation process of this module is a dynamic loop, which breaks down complex analysis tasks into a series of iterative optimization steps.

[0125] Part 1: Strategy Synthesize – The Initial Blueprint for Intelligent Planning

[0126] At the start of the analysis task, the system does not blindly begin scanning, but first performs a global intelligent planning to generate an optimal initial analysis strategy.

[0127] 1) Spatiotemporal Optimization Modeling: This application abstracts the entire analysis task into a spatiotemporal optimization scheduling problem. The goal of this problem is to maximize the value of information acquisition under limited time and resource constraints. This draws on mathematical methods used in logistics, manufacturing, and other fields to optimize resource allocation and path planning.

[0128] Objective Function: This application defines a comprehensive objective function that aims to minimize the total analysis time while maximizing the expected information gain. The information gain can be quantified based on prior knowledge of the sample (e.g., known that certain regions exhibit more drastic compositional variations) or specific analytical objectives (e.g., focusing on finding specific elements).

[0129] Constraints: Constraints include physical limitations of the instrument (such as the scanning stage speed and laser ablation rate) and analytical accuracy requirements.

[0130] 2) Generating an Initial Strategy: By solving this optimization model (e.g., using mixed-integer linear programming or heuristic algorithms), the system generates an initial analysis strategy containing a detailed sequence of operations. This strategy is not just a scanning path; it may also include preset parameters such as laser energy and spectral integration time for different regions. This strategy represents the system's best prediction of "how to complete the task most efficiently" given the current information.

[0131] Part Two: Process Monitoring – Real-time Sensing and Analysis of Status.

[0132] Once the system begins executing the initial strategy, the process monitoring phase is activated simultaneously, acting like an alert "central nervous system" to continuously perceive, evaluate, and analyze various states during the process.

[0133] 1) Multimodal data stream monitoring: The monitoring module receives and processes data streams from multiple sources in real time. Instrument status data: real-time position of the scanning stage, energy output stability of the laser, detector temperature of the spectrometer, etc.

[0134] Spectral quality data: Real-time acquired spectral signal-to-noise ratio, signal saturation, background noise level, etc.

[0135] Model analysis feedback: The real-time acquired spectral data is "fed" into the deep fusion model of the second module to obtain the model's preliminary analysis results for the current analysis point, as well as a crucial indicator—model confidence. This confidence quantifies the model's "certainty" in its own judgment.

[0136] 2) Anomaly Detection: The system continuously compares monitored data with preset normal thresholds or expected models. Any significant deviation is identified as an anomaly. For example: In a certain region, the spectral signal-to-noise ratio suddenly and continuously falls below the threshold.

[0137] The model's confidence level for a certain region remains low, indicating that the model is "confused" about that region.

[0138] An unexpected spectral peak not found in the spectral feature library was discovered, suggesting the possible presence of unknown substances.

[0139] Part Three: Smart Repair – Dynamically adapting to change.

[0140] When the monitoring module detects an abnormal event, it immediately triggers the intelligent repair mechanism. This is the most intelligent feature of this module; the system will autonomously and dynamically modify or "repair" the original analysis strategy based on the type and context of the abnormality.

[0141] 1) Causal Diagnosis and Decision Making: The system first performs causal diagnosis on abnormal events. For example, a low signal-to-noise ratio may be due to sample surface contamination, while a low model confidence may indicate the encounter with a component boundary or a complex mixture.

[0142] 2) Dynamic replanning of strategies: Based on the diagnostic results, the system will select and execute the most appropriate countermeasures from a predefined "repair strategy library", which is essentially a dynamic replanning of the subsequent analysis plan.

[0143] Case 1: Discovering a Suspected Key Area. If the model finds a high concentration of target elements in an unexpected location, or if the classification result for a certain area has extremely low confidence (suggesting that the area is complex and contains a lot of information), the system will determine that this area is a "high-value exploration area." The remediation strategy will be: pause the original scanning path, immediately generate a high-density, refined scanning subtask around the area, and then return to the main path after completion.

[0144] Case 2: Degraded Data Quality. If the spectral signal-to-noise ratio of a certain area remains below standard, the system will diagnose it as "poor measurement conditions." The remediation strategy might be to automatically increase the integration time of the spectral acquisition in that area, or, for laser-induced breakdown spectra, automatically perform a pre-ablation pulse to clean the sample surface before re-measuring.

[0145] Case 3: Blurred Boundaries. When searching for the boundary between two substances, if the confidence level of the model fluctuates repeatedly in the boundary region, the system will activate the boundary fine-tuning subroutine, which will automatically perform a zigzag or interpolation scan on both sides of the current position to accurately determine the boundary location with higher spatial resolution.

[0146] This closed loop of "synthesis-monitoring-repair" will run through the entire analysis task, iterating continuously until the analysis results of all regions reach the preset confidence level and quality standards.

[0147] 3. Mathematical formula expression.

[0148] This application can provide a conceptual mathematical expression for this closed-loop control process: 1). Initial strategy synthesis: Initial policy_0 = argmin_{all possible policies} (analysis time (policy) - information value weight * expected information gain (policy)).

[0149] 2). Monitoring function: Abnormal signal(t) = monitoring function(instrument status(t), spectral quality(t), model confidence(t)).

[0150] Note: Repair is triggered when the abnormal signal (t) exceeds the threshold.

[0151] 3). Intelligent Repair (Policy Update): New strategy_{t+1} = Repair function(current strategy_t, abnormal signal(t)) Note: The repair function is a decision system that selects the optimal strategy adjustment plan based on the anomaly type.

[0152] 4). Looping and Iterating: The system executes a portion of the current policy _t at time step t, and simultaneously calculates an anomaly signal (t). If a repair is triggered, a new policy _{t+1} is generated and executed from the next moment; otherwise, the current policy _t continues to be executed.

[0153] 4. Module output and technical advantages.

[0154] This module runs continuously throughout the analysis task, and its final implicit output is a high-quality, high-confidence complete spectral dataset. This dataset represents the "best result" obtained by the system through dynamic optimization and proactive exploration. Its technological advantages are revolutionary: It has achieved a leap from "blind execution" to "intelligent exploration": the system possesses intuition and judgment similar to human experts, and can intelligently focus limited analytical resources on the most valuable areas, greatly improving analytical efficiency and the probability of discovering new phenomena.

[0155] Extremely robust: Faced with real-world uncertainties such as sample contamination, instrument status fluctuations, and unknown components, the system can proactively adapt to and overcome these challenges through closed-loop feedback and self-correction, ensuring the reliability of the final analytical results.

[0156] The analysis process is "traceable" and "understandable": The system records a "decision log" for every strategy adjustment, detailing when, where, why, and what remedial measures were taken. This makes the entire intelligent analysis process no longer a black box, but completely transparent and traceable, providing valuable contextual information for subsequent manual review and in-depth research.

[0157] In summary, this module serves as a bridge connecting high-precision models with the complex real world. It endows the analysis system with unprecedented autonomy, intelligence, and robustness, and is the key to transforming the entire technical solution from an "advanced calculator" into an "intelligent analysis partner."

[0158] Step 14: Transform the analysis results into valuable decision support for users using an interactive decision support module based on actionable counterfactual interpretation.

[0159] In some embodiments, the interactive decision support module based on actionable counterfactual interpretation includes: constructing high-dimensional probability density landscape units, defining counterfactual optimization objective units, a solution unit based on a path planning algorithm, and generating and presenting actionable suggestions. See also Figure 5 Step 14 can be the following process: Step 51: Utilize the construction of high-dimensional probability density landscape units to obtain high-dimensional probability density distribution functions by learning from training data, and construct the high-dimensional probability density distribution functions as data landscapes.

[0160] Step 52: Utilize the defined counterfactual optimization target unit to construct an optimization target based on the user-input initial point and target area. The optimization target is to find a target path in the data landscape that starts from the initial point and eventually reaches the target area.

[0161] Step 53: Use the solution unit based on the path planning algorithm to solve the optimization objective and obtain the optimal objective path.

[0162] Step 54: Use generated and presented actionable suggestion units to translate the optimal target path into a language that users can understand and execute, thereby providing valuable decision support to users.

[0163] In some embodiments, the optimal target path is interpreted using the generating and presenting actionable suggestion unit to obtain the interpretation result; an interactive report is generated based on the interpretation result; wherein the interactive report includes overall suggestions, step-by-step paths, and sensitivity analysis.

[0164] In one application scenario, we will introduce an interactive decision support module based on actionable counterfactual interpretation: 1. Module objectives and core ideas.

[0165] The core objective of this module is to address the "black box" and "inoperability" problems inherent in traditional analytical systems at the decision support level. It aims to transform the system's role from merely a detector providing "what" to a decision partner capable of deeply explaining "why" and offering concrete "how" guidance.

[0166] Traditional analysis systems, no matter how advanced, typically output a classification label (such as "qualified" / "unqualified") or a quantitative value. This result represents a closed, one-way information transmission for the user, and has two fatal flaws: 1) Lack of Explanation: Users cannot know the specific reasons why the system makes this judgment. For example, the system determines that a certain alloy is "unqualified," but the user has no way of knowing which element's content is excessive or whether it is caused by an imbalance in the proportion of multiple elements.

[0167] 2) Lack of operability: When results are unsatisfactory, the system cannot provide any improvement suggestions. The user's most pressing question—"How should I adjust the process parameters to make the product qualified?"—remains unanswered. Users are forced to rely on personal experience for blind trial and error, which is costly and inefficient.

[0168] To overcome this challenge, this module introduces a powerful theory from the field of interpretable artificial intelligence—Actionable Counterfactual Explanations. Its core idea is to construct a "data landscape" simulating all possibilities in the real world, and to transform users' improvement needs into a planning problem of finding the "optimal path" within this landscape. This provides users with actionable improvement solutions that conform to objective laws and have the lowest implementation cost.

[0169] "Counterfactual" here refers to exploring questions like "What if...?" For example, "If I increase the content of element A in the sample by 0.1%, will the final analysis result become acceptable?" This module aims to answer these kinds of questions intelligently and efficiently, and find the optimal solution.

[0170] 2. Technical implementation process.

[0171] This module cleverly transforms the complex decision support problem into an intuitive and computable geometric path planning problem, mainly including the following four steps: Step 1: Construct a high-dimensional probability density landscape.

[0172] This forms the basis for counterfactual explanations. The system first needs to learn a data model that can describe "what is possible" and "what is realistically plausible".

[0173] 1) Learning Data Distribution: This application no longer treats training data as isolated points, but uses them to learn a continuous, smooth, high-dimensional probability density distribution function. This function assigns a probability density value to each point in the feature space (representing a possible sample state). Regions with higher probability density represent those that are more likely to occur in the real world; this application refers to these as "high-density regions" or "data manifolds." Conversely, regions with extremely low probability density represent sample combinations that are almost impossible or nonexistent.

[0174] 2) Constructing a Data Landscape: This probability density distribution function can be conceptually imagined as a high-dimensional geographical landscape. The "mountains" and "plains" composed of a large number of real sample points are high-probability-density areas, while the "canyons" and "abysses" with sparse samples are low-probability-density areas. Any real sample is a point in this landscape.

[0175] Part Two: Defining Counterfactual Optimization Objectives

[0176] When a user initiates a counterfactual query, the system will transform it into a clear optimization target.

[0177] 1) User Input: The user inputs the sample instance that needs to be analyzed and improved (referred to as the initial point in this application), and the desired target state (target area). For example, the initial point is "currently non-compliant alloy sample", and the target area is "all alloys rated as superior".

[0178] 2) Optimization Objective: The system's task is to find a path within the data landscape constructed in the previous step, starting from the "initial point" and ultimately reaching the "target area." This path needs to simultaneously satisfy two conditions: Feasibility: The path must be as close as possible to "mountains" and "plains," meaning it must always be located in high-probability-density areas. This ensures that every change along the path is realistically achievable, avoiding the generation of unrealistic and ineffective suggestions.

[0179] Economy: Among all feasible paths, find the path with the minimum total cost. This "cost" can be the geometric length of the path (representing the minimum change) or a weighted length that assigns different weights to the difficulty of changing different features.

[0180] Step 3: Solving based on path planning algorithms.

[0181] This application transforms this constrained optimization problem into a classic path planning problem and solves it using an efficient algorithm.

[0182] 1) Mathematical representation of the landscape: This application uses the negative logarithm of the probability density function as the "cost map" for path planning. This means that the higher the probability density of a region (mountain range), the lower its negative logarithm value, and the lower the cost of traversing the path; conversely, the lower the probability density of a region (canyon), the higher the cost, and the path planning algorithm will naturally avoid these regions.

[0183] 2) Employing Advanced Optimization Algorithms: This application employs advanced multi-objective optimization algorithms, such as the Non-Dominated Sorting Genetic Algorithm II (NSGA-II), to find the optimal path. These algorithms are well-suited for solving such complex problems because they can: Simultaneously optimize multiple objectives (e.g., both short path and high average probability density along the path).

[0184] By employing genetic and mutation techniques, we can efficiently explore a vast, high-dimensional solution space and avoid getting trapped in local optima.

[0185] Step 4: Generate and present actionable suggestions.

[0186] The output of the path planning algorithm is an optimal path from the initial point to the target region, consisting of a series of intermediate points. The final step of this module is to "translate" this mathematical path into a language that the user can understand and execute.

[0187] 1) Interpretation of the path: Each point on this path represents a tiny evolution of the sample's state. The entire path from the initial point to the end point constitutes a step-by-step, gradual transformation plan.

[0188] 2) Generate an interactive report: The system will present the user with an interactive report, which may include: General recommendation: Clearly indicate which core characteristics (such as element content, molecular peak intensity) need to be changed in what direction and magnitude from the initial state to the target state.

[0189] Step-by-step path: The entire optimization path is displayed in a visual way. Users can drag points on the path to view the changes in the spectral characteristics of the samples and the expected results in real time at different modification stages.

[0190] Sensitivity analysis: Informs users which changes in features have the greatest impact on the results and which have a minor impact, providing a basis for users to weigh costs and benefits in actual operation.

[0191] 3. Mathematical formula expression.

[0192] This application expresses the core concepts of this module mathematically: 1). Probability density landscape (cost function): Cost (sample state) = -log(probability density (sample state)).

[0193] 2). Path cost integral (the objective to be minimized): Total path cost (path) = ∫_{path start point}^{path end point} Cost (points on the path) * d (path length).

[0194] Note: This is a line integral that sums the cost of each point on the path. The goal of path planning is to find a path that minimizes this total cost.

[0195] 3). Formalization of the counterfactual optimization problem: Find the optimal path * = argmin_{all possible paths} (total path cost(path)).

[0196] Constraints: a. Path start point = Initial sample state.

[0197] b. The path endpoint ∈ the target state region.

[0198] c. For any point on the path, its classification prediction result = the expected category.

[0199] d. For any point on the path, its probability of existence is greater than the minimum probability threshold.

[0200] 4. Module output and technical advantages.

[0201] The final output of this module is a dynamic, interactive, and actionable decision support solution. It achieves several unexpected results that traditional technologies cannot match: This achieves a shift from a "black box" to a "white box": users not only know the results, but more importantly, they can deeply understand the form and reasons of the model's decision boundaries through counterfactual exploration, truly achieving "knowing why it is so".

[0202] It has achieved a leap from "diagnosis" to "treatment": the system is no longer just a diagnostic tool for discovering problems, but has been upgraded to a decision advisor that can provide specific, quantitative, and economically optimal "treatment plans", and its application value has undergone a qualitative leap.

[0203] It empowers users with an unprecedented sense of control and insight: through an interactive interface, users can proactively explore the possibilities of different improvement paths and simulate the consequences of different decisions. This transforms the relationship between humans and machines from a one-way "query-response" model into a two-way, collaborative "exploration-discovery" partnership.

[0204] In summary, this module, by transforming abstract machine learning models into an explorable and concrete "data landscape" and introducing powerful path planning tools, successfully converts complex counterfactual reasoning problems into an intuitive solution process. This is key to the entire intelligent analysis system's transformation from an efficient analytical tool into an innovative intelligent decision-making platform. See Figure 6 , Figure 6 This is a schematic diagram of an embodiment of the multimodal spectral intelligent analysis system provided in this application. The multimodal spectral intelligent analysis system 60 includes a processor 61 and a memory 62 coupled to the processor 61; The memory 62 is used to store computer programs, and the processor 61 is used to execute the computer programs to implement the following methods: The system utilizes a data structure fidelity preprocessing module based on tensor dynamic mode decomposition to dynamically decompose the acquired multimodal spectral data, obtaining the target dynamic modes. A deep fusion module based on interactive distillation and diffusion loss is then used to perform fusion analysis on the target dynamic modes, yielding the analysis results. The analysis process employs a closed-loop adaptive scanning and control module for intelligent monitoring and dynamic adjustment until all analysis results reach preset confidence and quality standards. Finally, an interactive decision support module based on actionable counterfactual interpretation transforms the analysis results into valuable decision support for users.

[0205] In some embodiments, the processor 61 is also configured to execute a computer program to implement the methods of any of the above embodiments.

[0206] In some embodiments, the multimodal spectral intelligent analysis system includes a data structure fidelity preprocessing module based on tensor dynamic mode decomposition, an attention-enhanced deep fusion module based on interactive distillation and diffusion loss, a closed-loop adaptive scanning and control module, and an interactive decision support module based on operable counterfactual interpretation. The data structure fidelity preprocessing module based on tensor dynamic mode decomposition, the attention-enhanced deep fusion module based on interactive distillation and diffusion loss, the closed-loop adaptive scanning and control module, and the interactive decision support module based on operable counterfactual interpretation cooperate to implement the method of any of the above embodiments.

[0207] The core mathematical framework of the multimodal spectral intelligent analysis system is as follows: The ultimate goal of this system is not merely to achieve accurate one-time classification of samples, but to act as an intelligent agent, dynamically interacting with the analysis task to find an optimal analysis strategy π. This strategy encompasses the entire process from scan path planning to final decision recommendations, aiming to maximize a comprehensive cumulative reward function J(π). This function is a comprehensive quantification of the efficiency, accuracy, robustness, and decision value of the entire analysis process.

[0208] Formula 0: Overall system optimization objective function.

[0209] The optimal policy * = argmax_{all possible policies π} J(π).

[0210] Where J(π) = E_{trajectory ~ p(trajectory|strategy)} [ Σ_{t=0}^{end of analysis} (γ^t * instant reward(state_t, action_t)) ].

[0211] Parameter description: Optimal Strategy*: The ultimate goal pursued by the system, a complete sequence of actions that maximizes total reward.

[0212] Strategy π: A function that determines the system's next action based on the current state.

[0213] J(π): The expected cumulative total return that the system can obtain under policy π.

[0214] E[...] : Mathematical expectation, representing a weighted average of all possible analytical trajectories.

[0215] Trajectory: A complete analysis process of a system from start to finish, consisting of a series of states and actions.

[0216] p(trajectory|policy): The probability of a specific trajectory occurring under policy π.

[0217] γ: Discount factor (0 < γ ≤ 1), used to balance the importance of immediate returns with future returns.

[0218] Instant reward (...): The single-step reward obtained by the system after performing an action at time t. Its specific composition will be discussed in subsequent modules.

[0219] Module 1 (corresponding to the data structure fidelity preprocessing module based on tensor dynamic mode decomposition): Data structure fidelity preprocessing based on tensor dynamic mode decomposition The core of this module lies in decomposing high-dimensional data streams into more physically interpretable intrinsic modes, the reward of which is providing high-quality features for subsequent analysis.

[0220] Formula 1.1: State evolution equation in tensor space State snapshot_{t+1} = State transition tensor ∘ State snapshot_t + Process noise_t.

[0221] State snapshot_t: A high-dimensional tensor representing all multimodal spectral information at time t.

[0222] State transition tensor: A linear operator (high-dimensional tensor) that describes the inherent dynamic laws of a system.

[0223] ∘: Tensor shrinking operation, representing the effect of an operator on a state tensor.

[0224] Process noise _t: represents random disturbances in the system evolution.

[0225] Formula 1.2: Dominant Subspace Projection Based on Tensor Singular Value Decomposition (TSVD) Let the preceding state tensor = U ∘ S ∘ V^T (singular value decomposition of tensors).

[0226] Projection operator _U = U ∘ U^T.

[0227] Dimensionality reduction state snapshot_t = projection operator_U ∘ state snapshot_t.

[0228] U, S, V: Left orthogonal tensor, diagonal singular value tensor, and right orthogonal tensor obtained through TSVD decomposition.

[0229] Projection operator _U: An operator that projects data onto an optimal low-rank subspace spanned by U.

[0230] Dimensionality-reduced state snapshot_t: A feature tensor that contains more than 99% of the energy of the original data, but with a significantly reduced dimensionality.

[0231] Formula 1.3: Definition of Dynamic Mode (Dimensionality reduction transition operator) ∘ Dynamic mode_k = Eigenvalue_k * Dynamic mode_k.

[0232] Dimensionality reduction transition operator: The core operator that describes the evolution of a system in a low-rank subspace.

[0233] Dynamic mode_k: The k-th feature tensor of the dimension reduction transition operator, representing a fundamental, spatiotemporally coherent spectral structure mode.

[0234] Eigenvalue _k: A complex number whose real part determines the growth or decay rate of mode _k, and whose imaginary part determines its oscillation frequency.

[0235] Instant Rewards for Contributions: Instantaneous return_Module 1 = -log(TSVD truncation error) + log(Σ_k |eigenvalue_k|) This reward is given to preprocessing results that concisely describe complex data with a few powerful dynamic modalities, while penalizing information loss.

[0236] Module 2 (corresponding to the Attention-Enhanced Deep Fusion Module Based on Interactive Distillation and Diffusion Loss): Interactive Diffusion Regularization in Attention-Enhanced Deep Fusion This module enhances the cognitive ability of the model by constraining its internal "thought process." Its core lies in constructing a complex, multi-objective loss function.

[0237] Formula 2.1: Total loss function of the fusion model Total loss = Main loss (prediction, label) + λ_attention * attention alignment loss + λ_diffusivity * diffuse loss.

[0238] • λ_attention, λ_diffuse: hyperparameters used to balance the weights of different loss terms.

[0239] Formula 2.2: Hierarchical Attention Alignment Loss Attention alignment loss = Σ_{l=1}^{number of network layers} w_l * D_KL(Attention graph_{student, l}||Attention graph_{teacher, l}).

[0240] w_l: Layer weights, which allow different alignment strengths to be applied to network layers of different depths.

[0241] D_KL(...||...): KL divergence, an asymmetric measure of the difference between two probability distributions (in this case, the attention weight distribution).

[0242] Attention graph _{..., l}: The attention weight matrix calculated at the l-th layer of the network.

[0243] Formula 2.3: Hinge-Dispersive Loss Diffusion loss = E_{i,j ~ samples within the batch} [max(0, Boundary margin - D_feature(feature_i,feature_j) + α * D_label(label_i, label_j))^2].

[0244] Boundary margin: A preset minimum distance that forces the feature representations of different samples to be at least this far apart.

[0245] D_feature(...): Distance metric in feature space, such as weighted Euclidean distance.

[0246] D_label(...): Distance in the label space. When two samples have different labels, D_label is 0, indicating that they should be pushed apart; when the labels are the same (samples generated by data augmentation), D_label is a large positive number to counteract the repulsive force and allow them to come closer.

[0247] α: A hyperparameter that adjusts the influence of tag information.

[0248] This formula is an improvement on the basic diffusion loss, introducing label information to avoid over-pushing similar samples away, making it more intelligent.

[0249] Instant Rewards for Contributions: Instant reward_Module 2 = log(model confidence) - total loss.

[0250] This reward is given to models that make correct judgments with high confidence, and whose internal "thought processes" conform to expert paradigms and have high feature differentiation.

[0251] Module 3 (corresponding to the closed-loop adaptive scanning and control module): Stochastic Bellman optimal equations for closed-loop adaptive control This module models the analysis process as a Markov Decision Process (MDP) and seeks the optimal analysis strategy.

[0252] Formula 3.1: Bellman Equation for the Optimal Action-Value Function (Q-function) Optimal Q-value * (state, action) = E_{next state ~ p(next state | state, action)} [instant reward (state, action) + γ * max_{next action} (optimal Q-value * (next state, next action)) ].

[0253] Optimal Q-value * (state, action): The expected cumulative reward that can be obtained by performing an action in a given state and then following the optimal policy thereafter.

[0254] p(next state|...): State transition probability, that is, the probability of transitioning to a certain next state after performing an action in the current state.

[0255] Formula 3.2: Definition of Intelligent Repair Strategy Repair action * (state) = argmax_{all possible actions a} (optimal Q value * (state, a) ).

[0256] When an anomaly (a special state) is detected, the "repair action" performed by the system is the one that maximizes its subsequent cumulative rewards.

[0257] Formula 3.3: Quantification of Abnormal Signals in Process Monitoring Abnormal signal = w_signal-to-noise ratio * f(signal-to-noise ratio) + w_confidence * g(1 - model confidence) + w_novelty * h(spectral residual).

[0258] f, g, h: Nonlinear mapping functions that convert physical quantities into risk or value signals.

[0259] Spectral residual: The difference between the current spectrum and the most similar spectrum in the existing knowledge base; the larger the value, the higher the novelty.

[0260] w_...: Weighting coefficient, which reflects the system's sensitivity to different types of anomalies.

[0261] Instant Rewards for Contributions: Instant Reports_Module 3 = - Analysis Time Consumption - w_Abnormality* (Abnormal Signal)^2.

[0262] This reward is a penalty, which penalizes analysis processes that take too long or result in abnormal states, thereby driving the system to learn more efficient and robust strategies.

[0263] Module 4 (corresponding to the interactive decision support module based on actionable counterfactual interpretation): Actionable counterfactual interpretation of geodesic path integral This module transforms the problem of finding the optimal improvement scheme into the problem of finding the shortest geodesic on a Riemannian manifold.

[0264] Formula 4.1: Riemannian metric tensor of the data landscape Metric tensor(x) = (1 / (probability density(x) + ε)) * Euclidean metric tensor + β * H(probability density(x)).

[0265] The metric tensor (x) defines the local spatial geometry at point x in the data landscape (how distances and angles are measured).

[0266] The first term is the conformal transformation term. It scales the standard metric of Euclidean space according to the inverse of the probability density. Where the probability density is high, the space is "shrinked," and the path length becomes shorter; where the probability density is low, the space is "stretched," and the path length increases dramatically.

[0267] ε is a small positive constant to prevent the denominator from being zero.

[0268] H(...): The Hessian matrix (second-order partial derivative matrix) of the probability density function, where β is the weight. This term introduces the curvature of the space, causing the path to tend to move along the "ridge" of the probability density field, rather than simply staying at higher elevations.

[0269] Equation 4.2: Energy functional of an operable counterfactual path (the objective to be minimized) Path energy(path(s)) = ∫_{0}^{1}< path velocity(s), path velocity(s)>_{metric tensor(path(s))} ds.

[0270] Path(s): A parameterized path from the initial point to the target region, where s is the path parameter.

[0271] Path velocity (s): The tangent vector of the path, representing the direction and rate of change.

[0272] <... , ...>_{...} : The inner product defined by the metric tensor at the point path(s).

[0273] This functional calculates the "energy" or weighted length of a path in spacetime curved by probability density. The path with the lowest energy is the optimal operational counterfactual interpretation.

[0274] Instant Rewards for Contributions: Instant reward_Module 4 = 1 / (Path energy + ε).

[0275] This reward is given to the final analysis results that can find low-cost, highly feasible improvement solutions for users.

[0276] The above formulas together constitute a logically rigorous and interconnected complete system, from top-level strategic objectives to bottom-level tactical execution. This application can be understood as a pyramid structure or a chain of command.

[0277] The systematic relationship between the formulas is explained in detail below: At the top of the pyramid: the ultimate goal (Formula 0) Formula 0: Overall Optimization Objective Function of the System Optimal strategy* = argmax J(π).

[0278] This is the overarching principle and supreme commander of the entire system. It defines the sole purpose of the system's existence: to find an optimal strategy π* to maximize a long-term cumulative return called J(π).

[0279] It doesn't care about the details: at this level, the system doesn't care how the "instant reward" is calculated or what the "state" actually is. It's only responsible for setting the ultimate goal.

[0280] It drives everything: all the modules and formulas at the bottom of the pyramid exist to serve this ultimate goal. They are either for more accurately defining and calculating "instant rewards" or for more effectively executing "actions" and perceiving "states," thus making the process of finding argmax J(π) possible.

[0281] The middle layer of the pyramid: Integrating core decision-making and returns. The middle layer of the pyramid serves as a bridge connecting the top-level goals and the bottom-level execution. At its core is Module 3, but it relies on the "reward" component provided by other modules.

[0282] Key join point: Instant response (state_t, action_t).

[0283] In Formula 0, this "instant reward" is an abstract function. In reality, it is a composite function, a weighted sum of the "instant reward contributions" of each module: Instant reward (state_t, action_t) = w_1 * Instant reward_module 1 + w_2 * Instant reward_module 2 + w_3 * Instant reward_module 3 + w_4 * Instant reward_module 4 (*Note: w here is the weight of the returns of each module, used to balance the importance of different sub-goals*).

[0284] Now let's look at the core of the decision-making process: Module 3: Closed-Loop Adaptive Control (Equations 3.1 - 3.3) This is the system's central processing unit (CPU) or decision-making brain.

[0285] Formula 3.1 (Bellman Equation): This is the core of the learning algorithm. It uses the integrated "immediate reward" and expectations for the future to learn and update the optimal Q-value function. This Q-value function is essentially a value assessment table that tells the system how good it is to perform a certain "action" in a certain "state". It directly serves argmax J(π) in Formula 0 and is the specific mathematical tool for finding the optimal policy.

[0286] Formula 3.2 (Intelligent Repair Strategy): This is the executor of the decision. Once the optimal Q-value function becomes sufficiently accurate through learning, this formula defines the policy π itself: in any state, always choose the action with the highest Q-value.

[0287] Formula 3.3 (Abnormal Signal Quantization): This is the sensing system of Module 3, responsible for monitoring the "health" of the process. Its output not only defines "Instant Report_Module 3" (penalty for abnormalities and time consumption), but also provides a richer connotation for the concept of "state". When the abnormal signal is high, the system's "state" enters a region requiring special handling.

[0288] Summary of the middle-level relationships: Module 3 is the decision-maker. It uses the Bellman equation (Formula 3.1) as a learning tool to formulate the optimal action strategy (Formula 3.2) based on the "immediate rewards" gathered from all modules, and ultimately achieves the grand goal of Formula 0.

[0289] The base of the pyramid: state awareness, task execution, and value assessment At the base of the pyramid are various specialized execution and evaluation departments. They are responsible for handling specific tasks and quantifying the quality of task completion into their respective "immediate reward contributions," which are then reported to the decision-making brain in the middle level.

[0290] Module 1: Data Structure Fidelity Preprocessing (Formulas 1.1 - 1.3) This is the system's perception and feature engineering department.

[0291] Relationship: It is responsible for defining and processing the core part of the concept of "state"—spectral data. Formulas 1.1 to 1.3 define how to extract more valuable, lower-dimensional state snapshots and dynamic modes from the original high-dimensional data stream. This high-quality feature is part of the "state" and will be input into Modules 2 and 3. At the same time, it quantifies the quality of preprocessing (less information loss, stronger dominant modes) into an immediate reward—Module 1—which is reported to the decision-making brain.

[0292] Module 2: Attention-Enhanced Deep Fusion (Formulas 2.1 - 2.3) This is the system's cognitive and classification department.

[0293] Relationship: It is responsible for performing core analytical tasks (such as classification and identification). Formulas 2.1 to 2.3 define how to evaluate its "work quality". This evaluation is multi-dimensional, considering not only the accuracy of the results (main loss) but also the quality of the "thinking process" (attention alignment, feature diffusion). This comprehensive evaluation result, along with the model's confidence level, constitutes the immediate reward module 2, which is reported to the decision-making brain. Predictions with high confidence and low loss will receive higher rewards.

[0294] Module 4: Actionable Counterfactual Explanations (Formulas 4.1 - 4.2) This is the system's strategic consulting and solution development department.

[0295] Relationship: It is responsible for providing valuable improvement suggestions to users after the analysis is completed. Formulas 4.1 and 4.2 define how to find a low-cost path from the "current state" to the "ideal state" in a space distorted by data probability density. The reciprocal of the "energy" or "cost" of this path constitutes the immediate return - Module Four. Finding an easily achievable and effective improvement path will bring huge returns to the entire analysis process.

[0296] System Relationship Summary Diagram The entire system can be imagined as the following command and information flow: 1. Top-level objective (Formula 0): Maximize long-term returns J(π).

[0297] 2. Decision-making cycle (dominated by Module 3): a. Perceiving the current state: The system integrates information from various aspects to understand its current state.

[0298] Obtain from Module 1: Processed, high-quality spectral features (dimensionality-reduced state snapshot_t).

[0299] Obtain from Module 2: the prediction results and confidence level of the current model.

[0300] Obtain abnormal signals from Module 3 itself: process monitoring.

[0301] b. Evaluate all available actions: The system uses Equation 3.1 (Bellman equation) and the learned optimal Q-value function to predict the total future reward that may be obtained after performing each action _a.

[0302] c. Execute the optimal action: The system selects the action _a with the highest Q value to execute according to Formula 3.2.

[0303] d. Receive immediate reward: After the action is executed, the system enters a new state _{t+1} and immediately receives an immediate reward. This reward is: The feedback from Module 1 (How well was the data processed?) How well did the model predict the returns in Module Two? The rewards of Module Three (Is the process efficient and robust?) The rewards of Module Four (Was the final recommendation valuable?) e. Learning and updating: The system uses the immediate rewards it receives to update its optimal Q-value function again via Equation 3.1 (Bellman equation), making future decisions more informed.

[0304] 3. Iterative Process: This cycle of "perception-evaluation-action-reward-learning" repeats continuously until the analysis ends. The goal of the entire process is to continuously optimize the optimal Q-value*, ultimately approaching the optimal strategy* that maximizes formula 0.

[0305] Therefore, this is a highly integrated intelligent system framework that is top-down goal-driven, bottom-up information feedback, and centrally controlled by the reinforcement learning core (Module 3). Each formula plays an indispensable and logically clear role within this framework.

[0306] Through this series of interconnected and progressive mathematical formulas, this application not only provides a solid theoretical foundation for each module, but also constructs a unified intelligent decision-making framework aimed at maximizing comprehensive returns, fully demonstrating the creativity and systematic nature of its design.

[0307] Through the deep coupling and collaborative innovation of the above four modules, this solution generates a series of comprehensive technical effects that transcend the simple combination of individual technologies and have non-obvious effects, forming a complete logical closed loop.

[0308] 1. A paradigm shift in analysis: from "static snapshots of matter" to "dynamic process insights" Results: Traditional techniques can only provide a "snapshot" of a sample's composition under a fixed condition. This system, by combining structure-fidelity dynamic data processing with closed-loop adaptive active exploration, can reveal the evolution of material composition along spatial gradients or its dynamic changes during reaction processes. For example, it can not only identify the final products of catalytic reactions but also track the formation, diffusion, and disappearance of key intermediates in real time and depict their spatiotemporal distribution on the catalyst surface. This represents a cognitive leap from answering "what it is" to deeply understanding "how it evolves."

[0309] Unexpectedness: The unexpectedness of this effect lies in the fact that it does not simply speed up the analysis, but fundamentally changes the dimension of the analysis. It is precisely because closed-loop control can make "predictive" proactive detection based on the early dynamic trends revealed by tensor decomposition that the system can accurately capture those fleeting or averaged key process information in traditional uniform scanning. This has revolutionary value for mechanism research and process optimization.

[0310] 2. Enhanced analytical depth: From "statistical data fitting" to "underlying causal insights" Results: Traditional fusion models aim for a mathematical fit to the training data, resulting in fragile and opaque decision-making logic. This system learns the "thinking paradigm" of expert models through "interactive distillation" and strengthens the discriminative power of essential features using "diffuse loss," making its decision-making logic more reliant on the inherent causal relationships between things. Combined with "counterfactual interpretation," the system can clearly reveal the inherent logical chain between "minor changes in input variables" and "qualitative changes in output results."

[0311] Unexpectedness: The unexpected aspect of this effect is that, without sacrificing or even improving analytical accuracy, it significantly enhances the interpretability and credibility of the model, successfully breaking the long-standing inherent contradiction between "performance" and "interpretability" in complex models. The system not only provides a more robust answer but also offers a more insightful explanation.

[0312] 3. The transformation of the system's role: from a "passive execution tool" to a "proactive analytical agent". Results: In operation, this system functions as an intelligent agent with basic "scientific discovery" capabilities. Faced with a vast sample of unknown composition, it can autonomously plan the most efficient preliminary exploration route, much like an experienced geologist. When signs of "mineral veins" are discovered, it can conduct a detailed review and ultimately submit an in-depth analysis report that includes not only a detailed compositional map but also "potential optimization solutions."

[0313] Unexpectedness: The emergence of this "intelligent agent" behavior is an inevitable result of the deep coupling of multiple seemingly unrelated technologies, such as spatiotemporal optimization, closed-loop control, deep learning, and counterfactual reasoning, within the specific scenario of multimodal spectral analysis. It models and automates the analytical intuition, decision-making logic, and optimization strategies of human experts, achieving an exponential improvement in analytical efficiency and the potential for discovering new knowledge.

[0314] 4. Synergistic Emergence of Comprehensive Performance: Breaking through traditional performance constraints and achieving multi-dimensional synchronous optimization. Results: Ultimately, as a whole, this system achieves superior overall performance compared to existing single optimization schemes in all four core dimensions of analysis: accuracy, efficiency, robustness, and decision support capabilities.

[0315] Unexpectedness: Traditional technological improvements often face the dilemma of "pressing down one gourd and another floats up"—for example, increasing accuracy by adding scanning points inevitably sacrifices analysis efficiency. The unexpected aspect of this solution lies in its systematic intelligent design, which achieves a synergistic improvement in multiple performance indicators. For instance, the proactive exploration of closed-loop control not only improves analysis efficiency by focusing on key areas but also enhances accuracy and robustness through meticulous anomaly review. Furthermore, counterfactual interpretation, building upon all of this, imbues the results with unprecedented decision-making value. This multi-dimensional, "positive feedback"-style synchronous growth in performance is the most innovative and non-obvious comprehensive technical effect of this solution.

[0316] See Figure 7 , Figure 7 This is a schematic diagram of an embodiment of the computer-readable storage medium provided in this application. The computer-readable storage medium 70 is used to store a computer program 71, which, when executed by a processor, implements the following method: The system utilizes a data structure fidelity preprocessing module based on tensor dynamic mode decomposition to dynamically decompose the acquired multimodal spectral data, obtaining the target dynamic modes. A deep fusion module based on interactive distillation and diffusion loss is then used to perform fusion analysis on the target dynamic modes, yielding the analysis results. The analysis process employs a closed-loop adaptive scanning and control module for intelligent monitoring and dynamic adjustment until all analysis results reach preset confidence and quality standards. Finally, an interactive decision support module based on actionable counterfactual interpretation transforms the analysis results into valuable decision support for users.

[0317] In some embodiments, when executed by a processor, computer program 71 is also used to implement the methods of any of the above embodiments.

[0318] In the several embodiments provided in this application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0319] If the integrated units in the other embodiments described above are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0320] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A multimodal fusion method for a LIBS laser-induced spectrometer, characterized in that, The LIBS laser-induced spectrometer multimodal fusion method includes: The acquired multimodal spectral data is dynamically decomposed using a data structure fidelity preprocessing module based on tensor dynamic mode decomposition to obtain the target dynamic mode; The target dynamic mode is fused and analyzed using an attention-enhanced deep fusion module based on interactive distillation and diffusion loss to obtain analysis results. The analysis process utilizes a closed-loop adaptive scanning and control module for intelligent monitoring and dynamic adjustment until all analysis results reach the preset confidence level and quality standards. The analysis results are transformed into valuable decision support for users using an interactive decision support module based on actionable counterfactual interpretation.

2. The LIBS laser-induced spectrometer multimodal fusion method according to claim 1, characterized in that, The data structure fidelity preprocessing module based on tensor dynamic mode decomposition includes: a data tensor quantization construction unit, a time-shifted tensor pair construction unit, a tensor singular value decomposition and dimensionality reduction unit, a dimensionality reduction state transition operator solving unit, and a dynamic mode extraction and reconstruction unit; the process of using the data structure fidelity preprocessing module based on tensor dynamic mode decomposition to dynamically decompose the acquired multimodal spectral data to obtain the target dynamic mode includes: The data tensor construction unit is used to perform a unified structured representation of the collected multimodal spectral data to obtain a state snapshot; The state snapshot is spliced ​​in time or depth using the construction time-shift tensor pair unit to obtain the preceding state tensor and the following state tensor. The tensor singular value decomposition and dimensionality reduction unit is used to perform tensor singular value decomposition on the preceding state tensor to obtain a low-rank subspace. The dimensionality reduction state transition operator is obtained from the low-rank subspace and the subsequent state tensor using the aforementioned solution unit for the dimensionality reduction state transition operator; The target dynamic mode is obtained by performing eigenvalue decomposition on the dimensionality reduction state transition operator using the dynamic mode extraction and reconstruction unit.

3. The LIBS laser-induced spectrometer multimodal fusion method according to claim 2, characterized in that, The step of performing tensor singular value decomposition on the preceding state tensor using the tensor singular value decomposition and dimensionality reduction unit to obtain a low-rank subspace includes: The tensor singular value decomposition and dimensionality reduction unit is used to perform tensor singular value decomposition on the preceding state tensor, decomposing the preceding state tensor into the product of three tensors. The low-rank subspace is obtained by retaining the part with the largest energy in the product of the three tensors.

4. The LIBS laser-induced spectrometer multimodal fusion method according to claim 1, characterized in that, The attention-enhanced deep fusion module based on interactive distillation and diffusion loss deploys a deep learning model. The module is used to perform fusion analysis on the target dynamic modality, yielding analysis results including: The deep learning model is used to perform fusion analysis on the target dynamic modality to obtain the analysis results.

5. The LIBS laser-induced spectrometer multimodal fusion method according to claim 1, characterized in that, The deep learning model is trained in the following way: The training spectral data is input into the teacher model and the deep learning model respectively to obtain the teacher attention map in the teacher model and the target attention map in the deep learning model; Calculate the difference between the teacher attention map and the target attention map, and use the difference as the attention alignment loss; And during the training process, for all samples in a batch, extract the feature vectors output by any two different samples in any feature layer of the deep learning model to obtain any two feature vectors. The diffusion loss is obtained based on any two eigenvectors. The model parameters in the deep learning model are adjusted based on the attention alignment loss and the diffusion loss.

6. The LIBS laser-induced spectrometer multimodal fusion method according to claim 1, characterized in that, The closed-loop adaptive scanning and control module includes a strategy synthesis unit, a process monitoring unit, and an intelligent repair unit. The analysis process utilizes the closed-loop adaptive scanning and control module for intelligent monitoring and dynamic adjustment until all analysis results reach preset confidence levels and quality standards, including: When the analysis task is started, the strategy synthesis unit is used to perform global intelligent planning and generate the optimal initial analysis strategy. The initial analysis strategy is executed, and the process monitoring unit continuously senses and evaluates various states during the analysis process. When the process monitoring unit detects an abnormal event, it uses the intelligent repair unit to perform causal diagnosis on the abnormal event, and selects and executes the most appropriate countermeasure from the predefined repair strategy library based on the diagnosis results.

7. The LIBS laser-induced spectrometer multimodal fusion method according to claim 6, characterized in that, The process monitoring unit continuously senses, evaluates, and analyzes various states during the process, including: The process monitoring unit receives instrument status data, spectral quality data, and model analysis feedback. Abnormal events are detected for the instrument status data, the spectral quality data, and the model analysis feedback, respectively.

8. The LIBS laser-induced spectrometer multimodal fusion method according to claim 6, characterized in that, The step of selecting and executing the most appropriate response from a predefined remediation strategy library based on the diagnostic results includes: In response to the diagnostic result indicating the discovery of a suspected critical area, a first repair strategy is selected from a predefined repair strategy library and executed; wherein, the first repair strategy includes: pausing the original scanning path, immediately generating a high-density fine-grained scanning subtask around the area, and returning to the main path after completion; In response to the diagnostic result indicating a decline in regional data quality, a second repair strategy is selected from a predefined repair strategy library and executed; wherein, the second repair strategy includes: automatically increasing the integration time of spectral acquisition in the region, or for laser-induced breakdown spectroscopy, automatically performing a pre-ablation pulse to clean the sample surface, and then re-measuring; In response to the diagnostic result indicating blurred boundaries, a third repair strategy is selected from a predefined repair strategy library and executed; wherein, the third repair strategy includes: starting a boundary fine positioning subroutine, automatically performing zigzag or interpolation scans on both sides of the current position to accurately determine the boundary position with higher spatial resolution.

9. The LIBS laser-induced spectrometer multimodal fusion method according to claim 1, characterized in that, The interactive decision support module based on actionable counterfactual interpretation includes: constructing high-dimensional probability density landscape units, defining counterfactual optimization objective units, a solution unit based on path planning algorithms, and generating and presenting actionable suggestions; the step of using the interactive decision support module based on actionable counterfactual interpretation to transform the analysis results into decision support valuable to users includes: The high-dimensional probability density landscape unit is constructed by learning training data to obtain a high-dimensional probability density distribution function, and the high-dimensional probability density distribution function is constructed into a data landscape. The defined counterfactual optimization target unit constructs an optimization target based on the user-input initial point and target area. The optimization target is to find a target path in the data landscape that starts from the initial point and eventually reaches the target area. The optimal target path is obtained by solving the optimization objective using the solution unit based on the path planning algorithm. The unit that generates and presents actionable suggestions translates the optimal target path into a language that the user can understand and execute, thereby providing valuable decision support to the user.

10. The LIBS laser-induced spectrometer multimodal fusion method according to claim 9, characterized in that, The step of translating the optimal target path into a language that the user can understand and execute using the unit that generates and presents actionable suggestions includes: The optimal target path is interpreted using the unit that generates and presents actionable suggestions, and the interpretation result is obtained. An interactive report is generated based on the interpretation results; wherein the interactive report includes overall recommendations, step-by-step paths, and sensitivity analysis.

Images