Method and apparatus for recognizing abnormality of collagen detection index

Abstract

The present application relates to a kind of collagen detection index abnormality identification method and equipment, belong to artificial intelligence technical field.It includes the following steps: original collagen detection index data is collected, forms original collagen characteristic matrix;Original collagen detection index data is adaptively preprocessed, and normalized characteristic matrix is obtained;Normalized characteristic matrix is mapped to high-dimensional multiscale feature space, and collagen damage degree perception matrix is obtained;Collagen damage degree perception matrix and normalized characteristic matrix are input into pre-trained collagen detection index abnormality identification model, and abnormality identification result is obtained.The present application can effectively improve the robustness, accuracy and explainability of collagen detection index abnormality identification.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method and device for identifying abnormalities in collagen detection indicators. Background Technology

[0002] Collagen, as a major structural protein of the extracellular matrix, plays a crucial role in maintaining tissue integrity and biomechanical function. Accurate assessment of collagen's structural integrity and metabolic state is a key task in biomedical research, quality control of tissue-engineered products, and clinical pathological diagnosis. Collagen abnormalities are often closely related to various pathological processes, such as abnormal cross-linking associated with tissue fibrosis, degradation imbalances in wound healing disorders, collagen network disruption caused by osteoarthritis, and glycosylation damage induced by metabolic diseases such as diabetes. Therefore, accurately identifying abnormalities in collagen detection indicators from biological samples has significant scientific and practical value for disease mechanism research, drug efficacy evaluation, and quality control and release of regenerative medicine products.

[0003] Existing technologies have the following drawbacks: When normalizing collagen detection indicators, current methods typically use the mean and standard deviation. However, collagen samples often contain extreme outliers with severe degradation or structural damage. These outliers severely skew the mean and standard deviation, distorting the distribution of normal samples and resulting in poor robustness during preprocessing. Conventional multi-kernel learning methods usually use globally fixed kernel weights for all samples or learn a globally optimal kernel combination. They cannot dynamically adjust the response weights of kernel functions at different scales based on the feature content of each sample, making it difficult to simultaneously address slight and severe deviations. Sensitivity to drastic changes; conventional anomaly detection models mostly only output binary classification probabilities and cannot provide confidence information on the prediction results. When the sample is on the boundary between normal and abnormal or the combination of detection indicators deviates from the training distribution, the model cannot inform the user of the reliability of the judgment result through uncertainty prompts; conventional machine learning methods rely only on statistical correlation for anomaly classification, ignoring the inherent biophysical relationships between collagen indicators, such as the ratio of hydroxyproline content to total amount, the correlation between cross-linking degree and thermal stability, etc., which may cause the model to learn surface patterns that violate the rules of the domain, and limit its generalization ability and interpretability. Summary of the Invention

[0004] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, the present invention provides a method for identifying abnormalities in collagen detection indicators, comprising the following steps: S1. Collect raw collagen detection index data to form a raw collagen feature matrix; S2. Adaptive preprocessing is performed on the original collagen detection index data to obtain a normalized feature matrix with stable distribution and reduced correlation. S3. Map the normalized feature matrix to a high-dimensional multi-scale feature space to obtain the collagen damage degree perception matrix, thereby transforming the linearly inseparable pattern of the original collagen detection index into a linearly separable pattern in a high-dimensional space. S4. Input the collagen damage perception matrix and the normalized feature matrix into the pre-trained collagen detection index anomaly recognition model to obtain the anomaly recognition results.

[0005] Furthermore, step S1 specifically includes: The original collagen detection index data include hydroxyproline content, total collagen, soluble collagen ratio, cross-linking degree, degradation product fragments, thermal denaturation temperature, enzymatic resistance, fluorescence intensity, degree of glycosylation, oxidative damage markers, metal ion content, and moisture content. After completing the detection of all indicators, the 12 test results of each sample are summarized to form an original collagen characteristic matrix. The labeling categories are: normal samples and abnormal samples.

[0006] Furthermore, step S2 specifically includes: Feature matrix of raw collagen Each column is statistically analyzed, and the median of all sample values ​​in each column is calculated to obtain a 12-dimensional robust center vector. For each detection index, first calculate the absolute deviation of each sample value of each detection index from its median, then calculate the median of these absolute deviations to obtain a 12-dimensional robust scaling vector. Using robust center vectors and robust scale vectors to analyze the original collagen feature matrix By centering and scaling column by column, detection indicators of different dimensions fall into a more comparable numerical range, resulting in a robustly scaled feature matrix. This approach effectively unifies the fluctuation scale of different detection indicators, reducing the impact of dimensional differences on subsequent models. Based on the robust scaling feature matrix in the training samples, the covariance matrix among the 12 detection indicators is calculated. A regularization term is introduced to construct a stable matrix of index correlation. Eigenvalue decomposition or singular value decomposition is performed on the stability matrix of index correlation to obtain the collagen index decoupling transformation matrix, thereby weakening the linear correlation between various detection indicators and making the features more statistically independent. The robustly scaled feature matrix is ​​then right-multiplied by the collagen index decoupling transformation matrix to obtain the normalized feature matrix. Each column corresponds to the collagen detection index characteristics after robust normalization and decorrelation processing, and has the characteristics of uniform dimension, weak influence of outlier samples and reduced correlation between features.

[0007] Furthermore, step S3 specifically includes: S31. Perform clustering on all samples in the normalized feature matrix to obtain... There are 1 cluster center, and each cluster center corresponds to a type of collagen detection mode; S32. Take any sample from the normalized feature matrix and input it into the attention network. Evaluate the importance of each kernel function to the current sample based on the sample's feature content, and obtain the raw scores of all kernel function branches. Normalize the raw scores of all kernel function branches and convert them into weight allocation results to obtain the attention weight set. Define Indicates the first The attention weight, i.e. the th attention weight The attention weights of each kernel function for the current sample; S33. Set bandwidth parameters for each kernel function branch, and apply them to the samples. Calculate the distance between each sample and each cluster center, and then... The Gaussian kernel function corresponding to the distance to the cluster centers is used to obtain the first... Collagen abnormality pattern matching degree of each kernel function branch ; will the first Collagen abnormality pattern matching degree of each kernel function branch With the attention weights Multiply to get the first Collagen damage perception value ; for all Each kernel function branch repeats the above calculation to obtain the sample. Collagen damage perception vector ; for all The above process was repeated for each sample to obtain the collagen damage perception matrix. .

[0008] Furthermore, in step S4, the abnormal identification model for collagen detection indicators includes a collagen-specific feature interaction layer: The collagen damage perception matrix The first set of linear mapping layers is input to obtain multi-scale projection features; the normalized feature matrix Input the second set of linear mapping layers to obtain normalized index projection features; then input the collagen damage degree perception matrix. and normalized characteristic matrix Two additional linear mapping layers are input to obtain two sets of hidden representations for interactive computation. Element-wise multiplication of the two sets of hidden representations at the same positions yields the interactive features. The multi-scale projection features, normalized index projection features, and interactive features are then element-wise summed, and finally input into a nonlinear activation function to obtain the index coupling enhancement feature matrix. .

[0009] Furthermore, the anomaly identification model for the collagen detection index also includes an anomaly probability and uncertainty estimation layer based on evidence-based deep learning: Enhance the feature matrix from the coupling of indicators The index coupling enhancement feature vector is extracted from any sample. Couple the indicators to enhance the feature vector Input the classification output layer and calculate the output responses for the normal and abnormal categories respectively; convert the output responses for the normal and abnormal categories into evidence values, defining the evidence value for the normal category as [value]. The value of abnormal evidence is ; Calculate the total evidence for each category based on the two types of evidence values , This leads to the probability of anomalies. , Calculate the prediction uncertainty based on the total amount of evidence. , .

[0010] Furthermore, a multi-objective joint optimization function is employed during training, introducing an evidence regularization term based on the classification loss: Based on the evidence value for each sample, the predicted category is determined by comparing the anomaly probability with 0.5. If the predicted category is inconsistent with the true label of the sample, it indicates that the model has made an incorrect judgment on the sample, and the total evidence is adjusted accordingly. Apply a penalty, with the penalty value set to... or The total loss is multiplied by a weighting factor and then added; if the prediction is correct, no penalty is imposed.

[0011] Furthermore, a domain knowledge constraint term is introduced on top of the classification loss: Domain knowledge constraints are based on the original collagen feature matrix. The system constructs real detection indicators to determine whether the current sample significantly violates known physical or biochemical relationships in the field of collagen, based on the original collagen feature matrix. The following three types of relationships were selected as the basis for constraint: the ratio of hydroxyproline content to total collagen, the product of cross-linking degree and soluble collagen ratio, and the linear relationship between thermal denaturation temperature and cross-linking degree.

[0012] Secondly, the present invention provides an anomaly identification device for collagen detection indicators, which performs the above-described anomaly identification method for collagen detection indicators, including: Data acquisition module: used to collect raw collagen detection index data and form a raw collagen feature matrix; Adaptive preprocessing module: used to adaptively preprocess the raw collagen detection index data to obtain a normalized feature matrix; High-dimensional space perception mapping module: used to map the normalized feature matrix to a high-dimensional multi-scale feature space to obtain the collagen damage degree perception matrix; Anomaly Detection Module: This module is used to input the collagen damage perception matrix and normalized feature matrix into a pre-trained anomaly detection model for collagen detection indicators to obtain anomaly detection results.

[0013] Thirdly, the present invention provides a computer-readable storage medium storing a program that can be executed by a processor to implement the above-described method for identifying abnormalities in collagen detection indicators.

[0014] The advantages of this invention are: This invention employs a robust statistical method based on the median and median absolute deviation to center and scale the original detection indicators, replacing the traditional mean and standard deviation. This effectively resists the interference of extreme outliers in collagen samples on data preprocessing, ensuring that the distribution of normal samples is not skewed by abnormal samples. It utilizes an attention-guided adaptive multi-kernel perceptual mapping, extracting typical abnormal configurations through clustering as kernel function centers, and dynamically assigning weights of different bandwidth kernel functions to each sample using an attention network. This allows the model to be sensitive to both slight indicator shifts and drastic multi-indicator co-variation. A feature interaction layer is constructed to normalize... The feature matrix and damage perception matrix are deeply fused. Through multiple sets of linear mappings and element-wise multiplication operations, the nonlinear coupling relationship between collagen indicators is explicitly modeled to uncover joint anomaly patterns such as imbalance in the ratio of hydroxyproline to total collagen and inverse changes in cross-linking degree and soluble collagen. An evidence-based deep learning framework is used to simultaneously output the anomaly probability and prediction uncertainty. Collagen domain knowledge, such as the ratio of hydroxyproline to total collagen, the product of cross-linking degree and soluble collagen, and the linear relationship between thermal denaturation temperature and cross-linking degree, are encoded as constraint terms in the loss function, so that the model maintains consistency with biophysical laws while learning classification. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0016] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 This is a comparison chart of the robust centers and conventional centers of each indicator in this invention; Figure 3 This is a comparison chart of the robustness scales and conventional scales of the various indicators of this invention; Figure 4 This is a typical abnormal collagen conformation diagram of the present invention; Figure 5 This is a visualization heatmap of the collagen damage perception matrix of the present invention; Figure 6 This is a diagram showing the recognition results of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 In this embodiment, as Figure 1 As shown, this invention provides a method for identifying abnormalities in collagen detection indicators, the specific steps of which include: S1. Collection of collagen detection indicators Raw collagen detection data are collected from biological samples to be tested. These samples can be derived from animal tissues, human tissues, or in vitro cultured cells and their secreted collagen products. During collection, the freshness and integrity of the samples must be ensured to avoid protein degradation or distortion of indicators due to improper sample storage. For each individual sample, standard biochemical detection procedures are used to determine its key collagen-related indicators. In one embodiment, these specifically include hydroxyproline content, total collagen, soluble collagen ratio, cross-linking degree, degradation product fragments, thermal denaturation temperature, enzymatic resistance, fluorescence intensity, degree of glycosylation, oxidative damage markers, metal ion content, and moisture content, as shown in Table 1. Table 1. Names and Units of Collagen Detection Indicators Each indicator must be measured strictly in accordance with the corresponding industry standard operating procedures. For example, the hydroxyproline content is determined by spectrophotometry, degradation product fragments are detected by enzyme-linked immunosorbent assay (ELISA), and thermal denaturation temperature is measured by differential scanning calorimetry.

[0019] After completing the testing of all indicators, the 12 test results of each sample are summarized to form an original collagen feature matrix. Each row of the matrix represents an independent sample to be identified, and each column corresponds to a specific collagen detection indicator.

[0020] All collected samples were clearly categorized. The categorization was mainly divided into two categories: "normal samples," which refer to samples where all test indicators fluctuate within the normal physiological range, with no obvious damage to collagen structure or metabolic abnormalities; and "abnormal samples," which refer to samples where one or more indicators significantly deviate from the normal range, such as excessive degradation, cross-linking impairment, oxidative damage, or a significant decrease in thermal stability—clear pathological or damaging characteristics. For abnormal samples, if conditions permit, they can be further subdivided into different abnormal subtypes, such as "degradation-type abnormalities" and "cross-linking impairment-type abnormalities," so that subsequent models can more precisely learn the typical characteristics of different abnormal patterns.

[0021] S2. Adaptive data preprocessing based on robust statistics and feature correlation Adaptive preprocessing was performed on the original collagen detection index data to obtain a normalized feature matrix with stable distribution and reduced correlation. The specific steps are as follows: S201. Calculate the robust statistics for each detection indicator. Traditional mean and standard deviation are easily skewed by outlier samples. This invention uses median and median absolute deviation to describe the central position and fluctuation range of each detection indicator, thereby improving the preprocessing's ability to resist interference from outliers. The specific operation is as follows: 1> Feature matrix of raw collagen Each column is statistically analyzed, and the median of all sample values ​​in that column is calculated to obtain a 12-dimensional robust center vector. , its first The element represents the first element. The median of a detection indicator represents the central level of that detection indicator; Original collagen feature matrix The size is ,in, This represents the number of samples, and 12 represents the number of collagen detection indicators. Each row in the matrix corresponds to one sample to be identified, and each column corresponds to one detection indicator. Collagen testing indicators include: hydroxyproline content, total collagen, proportion of soluble collagen, degree of cross-linking, degradation product fragments, thermal denaturation temperature, enzymatic resistance, fluorescence intensity, degree of glycosylation, oxidative damage markers, metal ion content, and moisture content, totaling 12 indicators.

[0022] 2> For each detection index, first calculate the absolute deviation of each sample value of that index from its median, then calculate the median of these absolute deviations to obtain a 12-dimensional robust scaling vector. , its first The element represents the first element. The median absolute deviation of a detection index represents the robustness of the index's dispersion.

[0023] It should be noted that when the absolute deviation of the median of a certain test indicator is too small or close to 0, a lower limit can be set for the absolute deviation of the median of that test indicator to avoid instability in subsequent division operations. For example, it can be set to... (Or set a minimum positive number based on the instrument resolution) to ensure that each detection index has an effective scale that can be used for scaling.

[0024] It should also be noted that in order to identify anomalies in 12 collagen detection indicators with high precision, high reliability and interpretability, the selected 12 collagen detection indicators have different dimensions, there are outliers (such as severely degraded samples), and there is a strong correlation (biological coupling) between the indicators. Moreover, the anomaly may be a slight shift of a single indicator or a drastic change in the nonlinearity and synergy of multiple indicators.

[0025] In one embodiment, the distribution of 12 detection indicators in the original data is shown, including the median, quartile range, and outliers. It can be seen that the different indicators have large differences in the units of measurement, and some indicators have obvious outliers.

[0026] S202, Implement robust centralization and robust scaling Using robust central vectors and robust scaling vector The feature matrix of the original collagen Centering and scaling are performed column by column to bring detection indicators of different dimensions into a more comparable numerical range. The specific operation is as follows: Feature matrix of raw collagen Subtracting the median of the corresponding detection index from each column yields a centered feature matrix, thereby eliminating the influence of the original baseline of different detection indices and making each detection index distributed around its own robust center. Furthermore, by dividing each column of the centered feature matrix by the median absolute deviation of the corresponding detection metric, a robustly scaled feature matrix is ​​obtained. The size is Each column corresponds to one detection metric, and each row corresponds to one sample, with robust scaling of the feature matrix. This will enable the effective unification of the fluctuation scale of different detection indicators and reduce the impact of dimensional differences on subsequent models.

[0027] It should be noted that when constructing the centered feature matrix, the original collagen feature matrix... Each column is subtracted from the median of the corresponding detection indicator. This is designed with special considerations in mind. For example, the mean cannot be used to replace the median to achieve the same technical effect because collagen samples may contain some extremely abnormal samples, such as samples with extremely high levels of degradation product fragments due to putrefaction. Conventional means and standard deviations would be severely "skewed" by these outliers, causing the distribution of normal samples to be distorted as well. The median and median absolute deviation are not sensitive to outliers and can more realistically reflect the center and fluctuation range of most normal samples. This is specifically designed to solve the problem of outlier interference that may occur in collagen detection.

[0028] S203, Adaptive decorrelation processing based on feature covariance matrix After robust scaling, the scales of the various detection indicators have been largely standardized, but some indicators still show correlations. For example, hydroxyproline content and total collagen may have a synergistic relationship, while cross-linking degree and soluble collagen ratio may have an inverse relationship.

[0029] This invention introduces a feature covariance matrix and performs an index correlation decoupling transformation on the robustly scaled feature matrix to obtain a normalized feature matrix with lower correlation. The specific operation is as follows: 1> Robust scaling of feature matrices based on training samples Calculate the covariance matrix among the 12 detection indicators. The size is This is used to describe the degree of linear correlation between different detection indicators; Furthermore, to avoid numerical instability of the covariance matrix when the sample size is small, the correlation between indicators is strong, or local features are repetitive, a regularization term is introduced into the covariance matrix to construct a stable indicator correlation matrix. The regularization parameter... This is used to control the strength of the covariance matrix correction, and the preferred value is 0.1.

[0030] In the specific implementation, when calculating the robust scaling feature matrix... covariance matrix (size To avoid matrix singularity or numerical instability caused by small sample sizes or strong correlations between indicators, a regularization term is introduced to construct a stable indicator correlation matrix. This is achieved by adding a small positive constant to the diagonal of the covariance matrix. ,in, for The identity matrix, Regularization parameter (preferred value) This ensures the stability of the indicator correlation matrix. The positive definiteness ensures that subsequent eigenvalue decomposition or singular value decomposition is stable and reliable, and the resulting collagen index decoupling transformation matrix can effectively weaken the linear correlation between features.

[0031] It's important to note that the 12 indicators of collagen are not biochemically independent. For example, "hydroxyproline content" is a component of "total collagen," and the two are naturally positively correlated. A higher "crosslinking degree" usually corresponds to a lower "soluble collagen ratio," and the two are negatively correlated. These inherent correlations can lead to information redundancy and may interfere with the model's identification of truly abnormal patterns. For instance, a sample with abnormal "crosslinking degree" and "soluble collagen ratio" but still maintaining the inverse relationship might just be experiencing normal fluctuations, while samples where the relationship is broken are the true anomalies. Therefore, when the training sample size is small, or when some indicators are nearly perfectly correlated due to biological characteristics (such as numerical collinearity), the covariance matrix... It may be "pathological," or even irreversible, leading to instability in the decoupling transformation process of collagen indicators, requiring the addition of regularization parameters. Adding noise to the diagonal is equivalent to adding a small amount of noise to the data, ensuring the numerical stability and positive definiteness of the matrix, and ensuring that the algorithm can run stably on a limited collagen dataset that may have strong correlations.

[0032] 2> Perform eigenvalue decomposition or singular value decomposition on the stability matrix of index correlation to obtain the collagen index decoupling transformation matrix, thereby weakening the linear correlation between various detection indicators and making the features more statistically independent.

[0033] 3. Multiply the robustly scaled feature matrix on the right by the collagen index decoupling transformation matrix to obtain the normalized feature matrix. The size is Each column corresponds to the collagen detection index characteristics after robust normalization and decorrelation processing, and has the characteristics of uniform dimensions, weak influence of outlier samples, and reduced correlation between features.

[0034] It should be noted that collagen detection indicators exhibit dimensional differences, statistical outliers, and cross-indicator correlations. Ordinary normalization alone cannot simultaneously address these issues. By using robust statistics to resist outliers and covariance whitening to weaken redundant correlations, the normalized feature matrix can be improved. It has stronger stability and divisibility.

[0035] It should also be noted that, in the face of the dual challenges of "outliers + strong correlations" in collagen data, "robust statistics" is a key and feasible solution to the outlier problem, while "regularization and whitening" is a key technical means to solve the algorithm instability caused by strong correlations.

[0036] In one embodiment, such as Figure 2 , Figure 3 As shown, the median and mean of each indicator are compared by a bar chart, which intuitively shows the impact of outliers on the mean. For indicators with extreme values, the mean deviates significantly from the median, while the median is more representative of the central level of most normal samples, providing a robust baseline for centering.

[0037] S3. Collagen Damage Perception Mapping Based on Adaptive Kernel Function Some abnormal collagen samples only show slight deviations in a few detection indicators, while other abnormal samples show significant linked changes in multiple detection indicators. If only a linear mapping of a single scale or a kernel function with a single bandwidth is used, it is often only sensitive to a certain type of change and it is difficult to take into account both slight and strong abnormalities at the same time.

[0038] This invention normalizes the feature matrix. By mapping to a high-dimensional multi-scale feature space, a collagen damage perception matrix is ​​obtained, thereby transforming the linearly inseparable pattern of the original collagen detection indicators into a linearly separable pattern in a high-dimensional space. The specific steps are as follows: S301. Obtaining typical abnormal collagen conformations To ensure the kernel function has a representative central location, we first start with the normalized characteristic matrix. Several abnormal typical conformations of collagen were extracted and used as the center points of the multi-scale kernel function. The specific operation is as follows: Normalized feature matrix As input for clustering, clustering is performed on all samples to obtain... Cluster centers, defined Indicates the first Each cluster center, representing different types of anomalies (such as degradation-related, cross-linking barrier-related, etc.), is a 12-dimensional vector used as the first cluster center. The center points of the Gaussian kernel function, where... This indicates the preset number of kernel functions; the example value is 8. Each cluster center corresponds to a typical collagen detection pattern, which may represent the common distribution of normal samples or the typical offset direction of a certain type of abnormal samples.

[0039] S302. Dynamically calculate the attention weights of each kernel function. Different samples have varying degrees of similarity to different typical patterns, therefore it is not advisable to use the same kernel weights for all samples. This invention uses an attention network to dynamically assign weights to each kernel function for each sample, enabling the model to adaptively select a more suitable scale matching value for the current sample. The specific operation is as follows: 1> From the normalized characteristic matrix Take any sample from the set, and define... This represents the normalized feature vector of a sample, which defaults to the current sample. It has a dimension of 12 and reflects the robust normalization results of the sample on 12 collagen detection indicators. Furthermore, the sample The input attention network can be composed of a linear projection layer, a non-linear activation layer, and a score output layer. Its function is to evaluate the importance of each kernel function to the current sample based on the sample features. For each kernel function branch, a raw score is calculated. This raw score indicates which cluster center and scale are more suitable to describe the current sample. The higher the score, the more important the corresponding kernel function is to the current sample.

[0040] In one implementation, the attention network consists of a multi-layered perceptron used to adjust the normalized feature vector of the current sample. The raw scores of each kernel function are dynamically calculated, and the input layer receives the normalized feature vector of the current sample. , dimension The hidden layer contains The layer contains neurons, and the activation function is ReLU (Rectified Linear Unit) to extract deep representations of the samples. The output layer contains... Each neuron has no activation function and directly outputs the raw score of each kernel function. The learnable parameters of the attention network include the weight matrix and bias from the input layer to the hidden layer, and the weight matrix and bias from the hidden layer to the output layer.

[0041] It should be noted that kernel function branches refer to pre-defined... There are 1 Gaussian kernel function, each corresponding to a cluster center and a bandwidth, used to measure the similarity between a sample and a specific typical pattern. For each kernel function branch, the attention network outputs a raw score, which reflects the degree to which the current sample matches the pattern represented by the kernel function in terms of content. The higher the score, the more important the kernel function is in describing the current sample.

[0042] 2> Normalize the original scores of all kernel function branches to convert them into weight allocation results, obtaining the attention weight set, and define... Indicates the first The attention weight, i.e. the th attention weight The attention weight of each kernel function for the current sample is an element in the set of attention weights, with a value ranging from 0 to 1.

[0043] It should be noted that the same set of Gaussian kernel functions are not equally important to different samples. By introducing a dynamic attention mechanism, slightly anomalous samples can preferentially rely on the more fine-grained kernel response (i.e., the matching degree calculated by the narrow-bandwidth kernel function for the local subtle differences of the sample), while strongly anomalous samples can preferentially rely on the wider range of kernel responses (i.e., the matching degree calculated by the wide-bandwidth kernel function for the local subtle differences of the sample), thereby avoiding the rigidity of expression caused by a single fixed weight.

[0044] S303, Collagen Damage Perception Mapping After obtaining the cluster centers and attention weights, multiple Gaussian kernel functions with different bandwidths are used to generate multi-scale feature representations for each sample. The specific operation is as follows: 1> Set bandwidth parameters for each kernel function branch, define Indicates the first The bandwidth of a Gaussian kernel function is used to control how sensitive the kernel function is to changes in sample distance. When the bandwidth is small, it is more sensitive to small local differences, while when the bandwidth is large, it pays more attention to the overall pattern similarity. In one implementation, a set of fixed bandwidth values ​​(such as 0.1, 0.5, 1, 2) are pre-selected to allocate different bandwidths to different kernel functions to cover different sensitivities ranging from subtle local changes to global modes.

[0045] 2> For the sample Calculate the distance between the current sample and each cluster center (using Euclidean distance), which represents the distance between the current sample and the cluster center. The degree of similarity between typical patterns; the smaller the distance, the closer the pattern is to the typical pattern. Furthermore, the sample The Gaussian kernel function corresponding to the distance to the cluster centers is used to obtain the first... Collagen abnormality pattern matching degree of each kernel function branch The calculation method is expressed as The matching degree of this collagen abnormality pattern reflects the matching strength of the current sample at the corresponding scale and in the corresponding pattern, between Between, when Take the maximum value at time The greater the distance, the smaller the matching degree, reflecting the sample's relevance to the first sample. Match strength of a typical pattern.

[0046] It should be noted that the matching degree of abnormal collagen patterns The calculation method uses the Gaussian radial basis kernel function to calculate the sample. With cluster center The distance between them is mapped to a similarity score, which measures the extent to which samples belong to the same order. In typical patterns, the closer the distance, the closer the scores. This indicates that the sample is highly similar to the pattern; the farther apart the sample is, the closer the score. This indicates that the sample deviates from the pattern, while also being measured by bandwidth. By controlling the decay rate, similarity measurements can be achieved at different scales.

[0047] 3> The first Collagen abnormality pattern matching degree of each kernel function branch With the attention weights Multiply to get the first Collagen damage perception value Characterizing samples In the The weighted matching degree on each kernel function branch is used to reflect the intensity of anomalous expression in a sample at a specific scale and under a specific pattern, and is expressed as: .

[0048] 4> All Each kernel function branch repeats the above calculation to obtain the sample. Collagen damage perception vector , dimension It is composed of the collagen damage perception values ​​of each nuclear function branch, i.e. .

[0049] 5> All The above process was repeated for each sample to obtain the collagen damage perception matrix. The size is Each row in the matrix corresponds to the collagen damage perception vector of one sample, and each column corresponds to the weighted matching degree of one kernel function branch.

[0050] It should be noted that the collagen damage perception matrix not only preserves the proximity between samples and typical patterns, but also reflects multi-scale change characteristics through different bandwidths and achieves sample-level adaptive adjustment through attention weights. It is more suitable for describing the continuous change process of "slight shift - moderate change - significant abnormality" in collagen detection indicators.

[0051] It should also be noted that in the field of anomaly detection, the use of multi-kernel functions (such as multi-kernel learning) is a known technique. However, the usual practice is to use the same fixed kernel weights for all samples, or to learn a globally optimal kernel combination. In this invention, the kernel weights are not globally fixed, but depend on the feature content of each collagen data sample. An "attention mechanism" used in sequence-to-sequence models is introduced into tabular biometric anomaly detection to achieve dynamic selection of the similarity matching scale. Moreover, this attention mechanism is not an attention enhancement method designed for variable-length sequence data, but is designed to solve the scale selection problem of biometrics. It transforms the continuity and diversity of collagen damage into a learnable and interpretable mathematical expression. Therefore, this invention does not achieve simple multi-kernel learning, but rather attention-guided adaptive multi-kernel perception.

[0052] In one embodiment, abnormal typical conformations of collagen were analyzed, and eight cluster centers were obtained through K-Means clustering, such as... Figure 4 As shown, each broken line represents a typical sample pattern, and these centers serve as the center points of the Gaussian kernel function, enabling the kernel function to focus on different abnormal patterns.

[0053] S4. Classification and identification of abnormal collagen detection indicators An anomaly detection model for collagen detection indicators was constructed. This model considers both data-driven nonlinear discriminant capabilities and the biophysical relationships between collagen indicators, thereby improving the accuracy, stability, and interpretability of anomaly detection results. The specific steps are as follows: S401, Collagen-specific characteristic interaction layer Collagen Damage Perception Matrix It excels at describing the multi-scale similarity relationship between samples and typical patterns, and normalizing the feature matrix. It excels at preserving the direct statistical expression of the original detection indicators. The two focus on different information, and if they are simply spliced ​​together, it is difficult to fully explore the deep coupling relationship between collagen indicators.

[0054] This invention uses projection and element-wise interaction to deeply fuse two types of features, generating an indicator coupling enhanced feature matrix. The specific operation is as follows: 1> Matrix for sensing the degree of collagen damage The first set of linear mapping layers has an input dimension of . The output dimension is (Example) The weights and biases are trainable parameters to obtain multi-scale projection features. This mapping layer is used to... Multi-scale feature transformation to a unified latent space; Normalized feature matrix The second set of linear mapping layers has an input dimension of 12 and an output dimension of 12. The weights and biases are trainable parameters to obtain normalized index projection features. This mapping layer is used to transform the 12-dimensional normalized index features into the same hidden space as the multi-scale features. At the same time, the collagen damage perception matrix and normalized characteristic matrix Two additional linear mapping layers (the third and fourth linear mapping layers) are input respectively to obtain two sets of hidden representations for interactive computation. The two sets of hidden representations (i.e., the hidden representations output by the third and fourth linear mapping layers) are multiplied element-wise at the same position to obtain interactive features. These interactive features can describe the coupling strength between "a certain type of multi-scale pattern response" and "a certain type of normalized detection index change". The input dimension of the third set of linear mapping layers is The output dimension is Its weights and biases are trainable parameters. The input dimension of the fourth linear mapping layer is 12, and the output dimension is... Its weights and biases are trainable parameters.

[0055] It should be noted that when those skilled in the art process multi-source features, the common operation is to concatenate or add them. This invention creatively introduces the idea of ​​feature crossover, but it is not a simple feature crossover. Instead, after mapping two features to the same space, element-wise multiplication is performed. This is a consideration of the synergistic nature of the product relationship between indicators. The element-wise multiplication operation can amplify the dimensions that are simultaneously active in the hidden representations output by the third and fourth linear mapping layers, highlighting the synergistic or coupling relationship between "specific multi-scale pattern response" and "specific normalized indicator change". This operation is similar to feature crossover and can model the nonlinear interaction between features. Therefore, it can effectively capture the joint abnormal patterns between collagen indicators.

[0056] 2> Add the multi-scale projection features, normalized index projection features, and interaction features element-wise, and then input them into a nonlinear activation function (such as the linear rectifier unit activation function) to obtain the index coupling enhancement feature matrix. The size is , This represents the hidden feature dimension, with an example value of 64. Each row in the matrix corresponds to the indicator coupling enhancement feature vector of a sample.

[0057] It should be noted that collagen abnormalities often manifest as synergistic changes in multiple indicators (such as imbalances in the ratio of hydroxyproline to total collagen, and inverse changes in cross-linking degree and soluble collagen). These relationships cannot be described independently by a single feature and require information on the combination of features. The feature interaction layer explicitly constructs product terms between features by multiplying element by element, which is equivalent to introducing a nonlinear combination of polynomial features into the hidden space. This allows for direct characterization of the dependencies between different feature dimensions, making it easier for the model to detect joint shifts between indicators, i.e., interaction anomalies.

[0058] In one embodiment, a heatmap is used to visualize the collagen damage perception matrix, such as... Figure 5 As shown, the matrix is ​​obtained by multiplying the attention weights by the Gaussian kernel matching degree. Each column corresponds to a kernel function, and the color intensity represents the damage perception value of the sample under that kernel. The matrix reveals the abnormal expression intensity of the sample in multi-scale space, providing rich nonlinear information for feature fusion.

[0059] S402, Anomaly Probability and Uncertainty Estimation Based on Evidence-Based Deep Learning In practical applications of anomaly identification for collagen detection indicators, simply outputting the anomaly probability is insufficient; it is also necessary to evaluate whether the probability is reliable. For example, some samples may be located near the boundary between normal and abnormal, or their combination of detection indicators may deviate from the training distribution. In such cases, even if the model provides a classification result, it should simultaneously provide an uncertainty warning.

[0060] This invention introduces the concept of evidence-based deep learning, enabling the model to output prediction uncertainty along with anomaly probability. The specific operation is as follows: 1> Enhance the feature matrix by coupling indicators Extract the indicator coupling enhancement feature vector from any sample, and define... Indicates the first The feature vector of index coupling enhancement for each sample, with dimension [missing information]. This is used to describe the comprehensive representation of the sample in the fused feature space; Furthermore, the indicators are coupled to enhance the feature vector. The input is the classification output layer, which calculates the output response for the normal category and the abnormal category respectively. Each category corresponds to a set of trainable weights and biases. The larger the output response, the stronger the supporting evidence accumulated by the model for that category. In the specific implementation, the classification output layer is a fully connected layer, and the input is the feature vector of index coupling enhancement. The output is The vectors are dimensional (corresponding to the normal and abnormal categories respectively), with their weights and biases being trainable parameters. The output response is the logits directly generated by the classification output layer, representing the model's original support for the two categories. The larger the value, the stronger the model's tendency to believe that the sample belongs to the corresponding category. Specifically, it is calculated through linear transformation.

[0061] 2. Convert the output responses for normal and abnormal categories into evidence values ​​(specifically using the Softplus function to convert the output responses into evidence), and define the evidence value for the normal category as... The value of abnormal evidence is The evidence value should be positive, and the larger the evidence value, the more fully the model supports that category; 3> Calculate the total evidence for each category based on the two types of evidence values. And further obtain the anomaly probability. The anomaly probability represents the likelihood that the sample belongs to an anomaly category; the larger the value, the more likely it is to be an anomaly. At the same time, the prediction uncertainty is calculated based on the total amount of evidence. (constant (where the total evidence is the number of categories). The larger the total evidence, the more information the model has and the lower the uncertainty; the smaller the total evidence, the less the model knows about the sample and the higher the uncertainty.

[0062] S403, Collagen Knowledge-Guided Multi-Objective Joint Optimization Function This invention, based on classification loss, introduces evidence regularization and domain knowledge constraint terms to construct a multi-objective joint optimization function guided by collagen knowledge. This enables the model to conform as closely as possible to the known relationships between collagen indicators while learning anomaly classification. The specific operation is as follows: 1> Establish a classification loss term, which measures the difference between the model's predicted class probability and the true label, enabling the model to learn the basic discrimination boundary between normal and abnormal samples. In one implementation, cross-entropy loss is used.

[0063] 2> Establish an evidence regularization term. The evidence regularization term is used to constrain the evidence distribution so that the model is not overconfident when it makes prediction errors, the samples are ambiguous or the support is insufficient, thereby improving the coordination between the abnormal probability output and the uncertainty output. In one implementation, based on the evidence value of each sample... and First, through comparison The value of 0.5 determines the class predicted by the model (or directly takes the class with the highest output probability). If the predicted class is inconsistent with the true label of the sample, it indicates that the model has made an incorrect judgment on that sample. In this case, the model is expected to have less evidence accumulation for that sample (i.e., higher uncertainty), therefore the total evidence... Apply a penalty; the penalty value can be set to... itself or The total loss is then multiplied by a tradeoff factor (e.g., 0.05); if the prediction is correct, no penalty is imposed.

[0064] 3. Establish domain knowledge constraints, which are based on the original collagen feature matrix. The system constructs real detection indicators to determine whether the current sample significantly violates known physical or biochemical relationships in the field of collagen.

[0065] In one implementation, from the original collagen feature matrix The following three types of relations are selected as constraint basis: a> Relationship between the ratio of hydroxyproline content to total collagen content This relationship is used to reflect whether the proportion of hydroxyproline in total collagen is close to the theoretical constant. If the ratio deviates significantly from the reasonable range, it indicates that the sample may have abnormal detection, abnormal degradation, or abnormal structure. b>The product of the degree of cross-linking and the proportion of soluble collagen This relationship is used to reflect the balance between the cross-linking state of collagen and the soluble components. If the product of the two deviates significantly from the empirical target value, it indicates that the sample may have abnormalities in terms of collagen maturity, degradation degree or structural stability. c> Construct a linear relationship between heat denaturation temperature and crosslinking degree, and calculate the heat denaturation temperature using a linear model. Assume that under normal conditions, there is a linear relationship between heat denaturation temperature and crosslinking degree. Furthermore, the slope and intercept are obtained by linear regression fitting of these two indicators on a large number of normal samples. This relationship is used to reflect the influence of changes in the degree of crosslinking on thermal stability. If there is a serious mismatch between the thermal denaturation temperature and the degree of crosslinking, it indicates that the sample may have structural damage, abnormal composition, or be affected by detection noise.

[0066] Furthermore, the domain knowledge constraint term sets target values ​​and tolerance thresholds for each type of domain knowledge relation. When the deviation of a sample on a certain relation does not exceed the tolerance threshold, no penalty is imposed. When the deviation exceeds the tolerance threshold, a constraint penalty is imposed on the excess part (using the Hinge loss calculation method, i.e., the penalty value is equal to the square of the deviation exceeding the threshold). This approach can avoid the model being overly sensitive to normal biological fluctuations, while ensuring that samples that clearly violate the domain rules will be given priority attention.

[0067] 5> The classification loss term, evidence regularization term, and domain knowledge constraint term are weighted and summed to obtain the overall composite loss, which is the collagen knowledge-guided multi-objective joint optimization function. The weight parameters are used to balance the relationship between classification performance, confidence expression, and domain consistency constraints. For example, the weights of the classification loss term, evidence regularization term, and domain knowledge constraint term are 1.0, 0.05, and 0.1, respectively. During the model training phase, the overall composite loss is used to jointly optimize all learnable parameters of the attention network, feature interaction layer, and classification output layer, so that the model can maintain consistency with the intrinsic laws of collagen while learning anomaly recognition.

[0068] It should be noted that collagen detection indicators are not independent of each other, but are jointly affected by structural composition, degree of cross-linking, degradation state and thermal stability. Directly incorporating this domain knowledge into the training objective can prevent the model from relying solely on statistical correlation for "surface classification". Instead, it can learn abnormal patterns that are more consistent with the collagen mechanism, so that the model can not only improve its generalization ability to unknown samples, but also enhance the credibility of the prediction results.

[0069] In one embodiment, 500 samples are used, of which approximately 70% are normal samples and 30% are abnormal samples. Figure 6 As shown, the evidence deep learning module of this invention can provide uncertainty estimation while outputting anomaly probabilities. In the figure, the horizontal axis represents anomaly probability (dimensionless), and the vertical axis represents uncertainty (dimensionless). Blue dots represent normal samples, and red dots represent anomaly samples. It can be seen that the vast majority of normal samples are concentrated in the low-probability and low-uncertainty region, while anomaly samples are concentrated in the high-probability and low-uncertainty region. A few boundary samples are scattered in the medium-probability and relatively high-uncertainty region, indicating that the model uses higher uncertainty to indicate the reliability of the prediction results, thus enhancing the credibility of the anomaly identification results.

[0070] S5. Model Training Based on Multi-Objective Joint Optimization Function The entire anomaly detection model is jointly trained end-to-end to optimize all learnable parameters, enabling the model to accurately identify anomalies in collagen detection indicators.

[0071] At the start of the training process, all labeled samples are first divided into a training set and a validation set. The training set is used to update model parameters, while the validation set is used to monitor the model's generalization ability and prevent overfitting. In each iteration, a batch of samples is randomly selected from the training set. First, the original collagen feature matrix is ​​obtained according to method S1. Then, it is processed sequentially through the adaptive data preprocessing procedure based on robust statistics and feature correlation described in S2 to obtain a normalized feature matrix. This matrix is ​​then input into the collagen damage perception mapping module based on adaptive kernel functions described in S3. The attention network dynamically calculates the attention weights for each kernel function based on the feature content of each sample and combines them with a multi-scale Gaussian kernel function to generate a collagen damage perception matrix. Finally, the normalized feature matrix and the damage perception matrix are input together into the collagen-specific feature interaction layer described in S401. Through multiple sets of linear mappings and element-wise multiplication, the two types of features are deeply fused to generate an indicator coupling enhancement feature matrix. The matrix is ​​then fed into the evidence-based deep learning-based anomaly probability and uncertainty estimation module described in S402. After passing through the classification output layer, output responses for two categories, normal and abnormal, are obtained and converted into corresponding evidence values. The anomaly probability and prediction uncertainty for each sample are then calculated. Based on the model's output and the sample's true label, a multi-objective joint optimization function guided by collagen knowledge constructed in S403 is used to calculate the overall composite loss. This composite loss includes a classification loss term that measures classification accuracy, an evidence regularization term that controls model confidence, and a domain knowledge constraint term that constrains the consistency between the model output and collagen domain knowledge (such as the ratio of hydroxyproline to total amount, the product of crosslinking degree and soluble collagen ratio, and the linear relationship between thermal denaturation temperature and crosslinking degree).

[0072] The gradient of the overall composite loss is propagated forward layer by layer using the backpropagation algorithm, while simultaneously updating all weights and bias parameters of the attention network, the four linear mapping layers in the feature interaction layer, and the classification output layer. After completing one iteration of all training set samples, the current model is used to predict the validation set samples, and the classification accuracy and composite loss value on the validation set are calculated. The stopping condition for model training is mainly based on the validation set performance. When the composite loss on the validation set no longer decreases over several consecutive iterations (e.g., 20 consecutive iterations), or when the classification accuracy on the validation set tends to stabilize or even shows a downward trend, it indicates that the model has fully converged and reached the optimal generalization state. At this point, training is immediately stopped, and the model parameters corresponding to the optimal validation set performance are saved as the final model used for anomaly detection.

[0073] S6. Identification of Abnormalities in Collagen Detection Indicators Once the model has been trained and its optimal parameters have been saved, it can be applied to actual tasks involving the identification of abnormal collagen detection indicators.

[0074] For a fresh test sample of unknown category, the raw data of its 12 collagen detection indicators are first obtained according to the standard detection process described in S1, forming a raw feature vector of 1 row and 12 columns. This raw feature vector is then input into the pre-trained anomaly recognition model. The model first performs the adaptive data preprocessing step in S2, which uses the robust center vector and robust scale vector calculated and saved during the training phase to robustly center and scale the raw features of the sample. Then, using the indicator correlation stability matrix and its corresponding decoupling transformation matrix calculated and saved during the training phase, the scaled features are decorrelated to obtain the normalized feature vector of the sample. Next, the model enters the damage level perception mapping stage in S3, where the normalized feature vector is simultaneously fed into the attention network and the multi-scale Gaussian kernel function module. The attention network dynamically calculates the attention weight of the sample on each preset kernel function based on the feature content of the sample; at the same time, it calculates the Euclidean distance between the sample and each cluster center (i.e., the typical conformation of collagen abnormality) and converts it into the corresponding collagen abnormality pattern matching degree through Gaussian kernel functions of different bandwidths. Then, the matching degree of each kernel function is multiplied by its corresponding attention weight to obtain the collagen damage perception vector for the sample. This perception vector, along with the original normalized feature vector, is then fed into the feature interaction layer of S401. After multiple sets of linear mappings, element-wise multiplications, and non-linear activations, an indicator coupling enhancement feature vector for the sample is generated. Finally, this feature vector is input into the evidence deep learning module of S402. The classification output layer calculates the output responses for normal and abnormal categories, which are then converted into evidence values ​​using the Softplus function. This conversion yields the final abnormality probability and prediction uncertainty for the sample.

[0075] The model outputs an anomaly probability, a value between 0 and 1, representing the likelihood that the sample belongs to an abnormal category. A probability closer to 1 indicates a higher probability of an abnormal sample. Simultaneously, the output prediction uncertainty reflects the model's confidence in the judgment; lower uncertainty indicates more comprehensive information and a more reliable judgment. Combining these two outputs allows for an accurate and reliable assessment of the collagen sample's health status. For example, a sample with an extremely high anomaly probability and extremely low uncertainty can be identified as a clearly abnormal sample. Conversely, a sample with a moderate anomaly probability and relatively high uncertainty suggests that the sample may be on the borderline between normal and abnormal, or that its combination of detection indicators is rare, requiring further comprehensive evaluation based on more clinical information or repeated testing.

[0076] Example 2 This embodiment provides an anomaly identification device for collagen detection indicators, which executes the anomaly identification method for collagen detection indicators described in Embodiment 1, including: Data acquisition module: used to collect raw collagen detection index data and form a raw collagen feature matrix; Adaptive preprocessing module: used to adaptively preprocess the raw collagen detection index data to obtain a normalized feature matrix; High-dimensional space perception mapping module: used to map the normalized feature matrix to a high-dimensional multi-scale feature space to obtain the collagen damage degree perception matrix; Anomaly Detection Module: This module is used to input the collagen damage perception matrix and normalized feature matrix into a pre-trained anomaly detection model for collagen detection indicators to obtain anomaly detection results.

[0077] Example 3 This embodiment provides a computer-readable storage medium storing a program that can be executed by a processor to implement the abnormal identification method for collagen detection indicators described in Embodiment 1.

[0078] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying abnormalities in collagen detection indicators, characterized in that, Includes the following steps: S1. Collect raw collagen detection index data to form a raw collagen feature matrix; S2. Adaptive preprocessing is performed on the original collagen detection index data to obtain the normalized feature matrix; S3. Map the normalized feature matrix to a high-dimensional multi-scale feature space to obtain the collagen damage perception matrix. S4. Input the collagen damage perception matrix and the normalized feature matrix into the pre-trained collagen detection index anomaly recognition model to obtain the anomaly recognition results.

2. The method for identifying abnormal collagen detection indicators according to claim 1, characterized in that, Step S1 specifically includes: The original collagen detection index data include hydroxyproline content, total collagen, soluble collagen ratio, cross-linking degree, degradation product fragments, thermal denaturation temperature, enzymatic resistance, fluorescence intensity, degree of glycosylation, oxidative damage markers, metal ion content, and moisture content. After completing the detection of all indicators, the 12 test results of each sample are summarized to form an original collagen characteristic matrix. The labeling categories are: normal samples and abnormal samples.

3. The method for identifying abnormal collagen detection indicators according to claim 1, characterized in that, Step S2 specifically includes: Feature matrix of raw collagen Each column is statistically analyzed, and the median of all sample values ​​in each column is calculated to obtain a 12-dimensional robust center vector. For each detection index, first calculate the absolute deviation of each sample value of each detection index from its median, then calculate the median of these absolute deviations to obtain a 12-dimensional robust scaling vector. Using robust center vectors and robust scale vectors to analyze the original collagen feature matrix By centering and scaling column by column, a robustly scaled feature matrix is ​​obtained. Based on the robust scaling feature matrix in the training samples, the covariance matrix among the 12 detection indicators is calculated. A regularization term is introduced to construct a stable matrix of index correlation. Perform eigenvalue decomposition or singular value decomposition on the stability matrix of the index correlation to obtain the collagen index decoupling transformation matrix; multiply the robustly scaled feature matrix on the right by the collagen index decoupling transformation matrix to obtain the normalized feature matrix. .

4. The method for identifying abnormal collagen detection indicators according to claim 1, characterized in that, Step S3 specifically includes: S31. Perform clustering on all samples in the normalized feature matrix to obtain... There are 1 cluster center, and each cluster center corresponds to a type of collagen detection mode; S32. Take any sample from the normalized feature matrix and input it into the attention network. Evaluate the importance of each kernel function to the current sample based on the sample's feature content, and obtain the raw scores of all kernel function branches. Normalize the raw scores of all kernel function branches and convert them into weight allocation results to obtain the attention weight set. Define Indicates the first The attention weight, i.e. the th attention weight The attention weights of each kernel function for the current sample; S33. Set bandwidth parameters for each kernel function branch, and apply them to the samples. Calculate the distance between each sample and each cluster center, and then... The Gaussian kernel function corresponding to the distance to the cluster centers is used to obtain the first... Collagen abnormality pattern matching degree of each kernel function branch ; will the first Collagen abnormality pattern matching degree of each kernel function branch With the attention weights Multiply to get the first... Collagen damage perception value ; for all Each kernel function branch repeats the above calculation to obtain the sample. Collagen damage perception vector ; for all The above process was repeated for each sample to obtain the collagen damage perception matrix. .

5. The method for identifying abnormal collagen detection indicators according to claim 1, characterized in that, In step S4, the abnormal identification model for collagen detection indicators includes a collagen-specific feature interaction layer: The collagen damage perception matrix The first set of linear mapping layers is input to obtain multi-scale projection features; the normalized feature matrix Input the second set of linear mapping layers to obtain the normalized index projection features; Collagen damage perception matrix and normalized characteristic matrix Two additional linear mapping layers are input to obtain two sets of hidden representations. Element-wise multiplication is performed on the two sets of hidden representations at the same positions to obtain the interaction features. The multi-scale projection features, normalized index projection features, and interaction features are then element-wise summed and input into a nonlinear activation function to obtain the index coupling enhancement feature matrix. .

6. The method for identifying abnormal collagen detection indicators according to claim 5, characterized in that, In step S4, the anomaly identification model for the collagen detection index further includes an anomaly probability and uncertainty estimation layer based on evidence deep learning: Enhance the feature matrix from the coupling of indicators The index coupling enhancement feature vector is extracted from any sample. Couple the indicators to enhance the feature vector Input the classification output layer and calculate the output responses for the normal and abnormal categories respectively; convert the output responses for the normal and abnormal categories into evidence values, defining the evidence value for the normal category as [value]. The value of abnormal evidence is ; Calculate the total evidence based on two types of evidence values. , This leads to the probability of anomalies. , ; Calculate the prediction uncertainty based on the total amount of evidence. , .

7. The method for identifying abnormal collagen detection indicators according to claim 1, characterized in that, In step S4, a multi-objective joint optimization function is used during training, and an evidence regularization term is introduced on top of the classification loss: Based on the evidence value for each sample, the predicted category is determined by comparing the anomaly probability with 0.

5. If the predicted category is inconsistent with the true label of the sample, it indicates that the model has made an incorrect judgment on the sample, and the total evidence is adjusted accordingly. Apply a penalty, with the penalty value set to... or The total loss is multiplied by a weighting factor and then added; if the prediction is correct, no penalty is imposed.

8. The method for identifying abnormal collagen detection indicators according to claim 7, characterized in that, In step S4, a domain knowledge constraint term is introduced in addition to the classification loss: Domain knowledge constraints are based on the original collagen feature matrix. The construction of real detection indicators in the matrix, starting from the original collagen feature matrix The following three types of relationships were selected as the basis for constraint: the ratio of hydroxyproline content to total collagen, the product of cross-linking degree and soluble collagen ratio, and the linear relationship between thermal denaturation temperature and cross-linking degree.

9. An anomaly identification device for collagen detection indicators, executing the anomaly identification method for collagen detection indicators as described in claim 1, characterized in that, include: Data acquisition module: used to collect raw collagen detection index data and form a raw collagen feature matrix; Adaptive preprocessing module: used to adaptively preprocess the raw collagen detection index data to obtain a normalized feature matrix; High-dimensional space perception mapping module: used to map the normalized feature matrix to a high-dimensional multi-scale feature space to obtain the collagen damage degree perception matrix; Anomaly Detection Module: This module is used to input the collagen damage perception matrix and normalized feature matrix into a pre-trained anomaly detection model for collagen detection indicators to obtain anomaly detection results.

10. A computer-readable storage medium, characterized in that, The medium stores a program that can be executed by a processor to implement an abnormal identification method for collagen detection indicators as described in any one of claims 1-8.

Images