A mixed STR spectrum denoising method and system based on multi-scale wavelet domain decoupling representation

Abstract

The present application relates to the technical field of forensic genetic analysis, and in particular to a mixed STR pattern denoising method and system based on multi-scale wavelet domain decoupling characterization. The denoising method comprises the following steps: performing one-dimensional wavelet transform on the original mixed STR pattern data, decomposing into sub-band signals of multiple scales; data enhancement is performed; then inverse wavelet transform is performed to reconstruct, generate a first pattern view and a second pattern view which are positive sample pairs of each other, input the first pattern view and the second pattern view into the multi-scale decoupling encoder respectively, and map to obtain a high-dimensional latent representation tensor; the latent representation tensor is cut and first and second feature vectors are extracted; an unsupervised contrast learning is used to construct a joint loss function to iteratively train the multi-scale decoupling encoder and the reconstruction decoder; the original mixed STR pattern data to be tested is input into the trained multi-scale decoupling encoder, and the first feature vector is extracted and output.

Description

Technical fields: This invention relates to the field of forensic genetic analysis technology, specifically to a method and system for denoising hybrid STR maps based on multi-scale wavelet domain decoupling representation. Background technology: In the field of forensic DNA analysis, accurate denoising and genotyping information extraction from mixed STR profiles are crucial prerequisites for subsequent individual identification and evidence interpretation. However, current technologies face the following challenges: Environmental interference components are highly coupled with genotyping signals in terms of morphology. Environmental interferences such as baseline drift, sequence degradation trends, and instrument random noise coexist and overlap with true allele peaks in the signal waveform. Traditional global noise reduction methods, while filtering out noise, are prone to misfiltering weak allele peaks of low template amounts or minor contributors, resulting in the loss of genotyping information.

[0003] There is a lack of effective means to structurally separate genotyping information from environmental interference in the feature space. Conventional encoders typically map the spectrum as a single latent variable, failing to explicitly model the distribution characteristics of different types of interference. This results in a large number of non-genotyping variations remaining in the extracted feature vectors, limiting the accuracy of subsequent genotyping inference. Summary of the Invention: The purpose of this invention is to provide a hybrid STR spectral denoising method and system based on multi-scale wavelet domain decoupling representation. By using a hierarchical orthogonal penalty term, the correlation between feature subspaces is eliminated from both same-layer and cross-layer dimensions, preventing crosstalk between signal components of different scales during the encoding process, ensuring the strictness of decoupling, and enabling the first feature vector output by the mesoscale layer to truly reflect the intrinsic structure of the fractal signal without being contaminated by global drift trends or local noise variance.

[0005] In a first aspect, the present invention provides a hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation, comprising the following steps: S1: Obtain unlabeled raw mixed STR map data; perform one-dimensional wavelet transform on the raw mixed STR map data to decompose it into sub-band signals of multiple scales; then perform data enhancement based on the differential perturbation superposition of wavelet sub-band characteristics; then reconstruct the data through inverse wavelet transform to generate a first map view and a second map view that are positive sample pairs, and construct negative sample pairs by combining non-homologous data within the batch. S2: Construct a multi-scale decoupled encoder, input the first and second map views into the multi-scale decoupled encoder respectively, and map to obtain a high-dimensional latent representation tensor; divide the latent representation tensor into layers according to the receptive field scale, and extract the first feature vector representing the pure map classification information and the second feature vector representing the environmental interference features. S3: Based on unsupervised contrastive learning, a joint loss function is constructed to perform end-to-end iterative training on the multi-scale decoupled encoder and the reconstruction decoder; wherein, the multi-scale decoupled encoder optimizes its parameters by calculating cross-view contrastive loss on the extracted first feature vector using the positive and negative samples, and the reconstruction decoder optimizes its parameters by calculating reconstruction loss based on signal strength weighting; and the joint loss function also incorporates variational information bottleneck regularization and hierarchical orthogonality penalty term to achieve orthogonal decoupling of the latent representation tensor and lossless purification of fractal information; S4: Input the original mixed STR map data to be tested into the trained multi-scale decoupled encoder, extract and output the first feature vector, which is the pure mixed map typing information after filtering out environmental interference features.

[0006] As one possible implementation, the differential perturbation superposition in S1 specifically includes: In a large-scale envelope subband characterizing the global trend, random walk sequences are superimposed to simulate the degradation trend of the map sequence and the global baseline drift distribution; In the mesoscale detail subband characterizing the local topological receptive field, a scaling perturbation is applied to simulate the fluctuating distribution of allele peaks and the relative relationships of local peak heights. Gaussian white noise is injected into small-scale detail subbands that characterize fine-grained variations to simulate the random noise distribution of the instrument.

[0007] As one possible implementation, the latent representation tensor in S2 is hierarchically segmented according to the receptive field scale, specifically employing a multi-scale feature decoupling pyramid structure: Large-scale global envelope layer: Based on the large global receptive field, feature vectors representing the degradation trend of the characterization map sequence and the global baseline drift are extracted; Mesoscale structural semantic layer: Based on the local topological receptive field, the first feature vector containing the relative relationship between the core allele peak and the local peak height is segmented; Small-scale noise perturbation layer: Based on fine-grained receptive field segmentation, feature vectors characterizing random noise in the instrument are extracted; The feature vectors output by the large-scale global envelope layer and the small-scale noise perturbation layer together constitute the second feature vector characterizing the environmental interference features.

[0008] As one possible implementation, the cross-view contrast loss calculation method in S3 is as follows: After normalizing the first feature vector of all views, the cosine similarity is calculated. The normalized temperature cross-entropy loss is used to make two views of the same sample close to each other in the feature space, and views of different samples far apart.

[0009] As one possible implementation, the variational information bottleneck regularization term in S3 is calculated as follows: The mean and variance of the first feature vector are obtained by reparameterized sampling, and the posterior probability distribution of its latent space is constructed. While maximizing the information retention of the first feature vector, KL divergence is introduced to force the posterior probability distribution to approximate the preset standard normal prior distribution, which is used as a regularization penalty term to filter out non-fractal interference information with high uncertainty in the latent space.

[0010] As one possible implementation, the hierarchical orthogonality penalty term in S3 includes intra-level orthogonal constraints and cross-level orthogonal constraints as follows: Same-layer orthogonality constraint: Calculate the matrix inner product of the first feature vector in the mesoscale structural semantic layer and the local interference vector segmented in the same layer, and minimize this inner product during backpropagation; Cross-layer orthogonality constraint: Force the first feature vector output by the mesoscale structural semantic layer to be orthogonal to the feature vector representing the global degradation trend output by the large-scale global envelope layer and the feature vector representing random noise output by the small-scale noise perturbation layer in the latent space. The hierarchical orthogonal constraints prevent cross-scale feature aliasing or information leakage from occurring during the extraction of multi-scale spectral signals.

[0011] As one possible implementation, the reconstructed graph in S3 is: The first and second feature vectors after segmentation are concatenated and recombined along the channel dimension, and then input into the reconstruction decoder to obtain the reconstructed map.

[0012] As one possible implementation, the signal strength-weighted reconstruction loss in S3 is calculated as follows: Weak peak regions with signal intensities below an adaptive threshold in the original mixed STR spectral data are penalized with a higher weight than the main peak regions to ensure that the pure spectral classification information of extremely low template amounts or minor contributors is preserved during the dimensionality reduction and decoupling process.

[0013] As one possible implementation, the joint loss function in S3 is calculated as follows: in, This is the balance coefficient; For cross-view contrast loss; This is a variational information bottleneck regularization term; This is a hierarchical orthogonal penalty term; Reconstruction loss based on signal strength weighting; joint loss The multi-scale decoupled encoder and reconstruction decoder are iteratively trained end-to-end via backpropagation.

[0014] Secondly, the present invention provides a hybrid STR map denoising system based on multi-scale wavelet domain decoupling representation, comprising: The data augmentation module is used to acquire unlabeled raw mixed STR map data; perform a one-dimensional wavelet transform on the raw mixed STR map data to decompose it into sub-band signals of multiple scales; then perform data augmentation based on the differential perturbation superposition of wavelet sub-band characteristics; then reconstruct the data through inverse wavelet transform to generate a first map view and a second map view that are positive sample pairs, and construct negative sample pairs by combining non-homologous data within the batch. A construction and feature segmentation module is used to construct a multi-scale decoupled encoder. The first and second map views are respectively input into the multi-scale decoupled encoder to map and obtain a high-dimensional latent representation tensor. The latent representation tensor is segmented hierarchically according to the receptive field scale to extract a first feature vector representing the pure map classification information and a second feature vector representing the environmental interference features. The joint loss optimization module constructs a joint loss function based on unsupervised contrastive learning to perform end-to-end iterative training on the multi-scale decoupled encoder and the reconstruction decoder. Specifically, the multi-scale decoupled encoder optimizes its parameters by calculating a cross-view contrastive loss on the extracted first feature vector using the positive and negative samples, and the reconstruction decoder optimizes its parameters by calculating a reconstruction loss weighted by signal strength. Furthermore, the joint loss function incorporates a variational information bottleneck regularization term and a hierarchical orthogonality penalty term to achieve orthogonal decoupling of the latent representation tensor and lossless purification of fractal information. The information extraction and output module is used to input the raw mixed STR map data to be tested into the trained multi-scale decoupled encoder, extract and output the first feature vector, which is the pure mixed map typing information after filtering out environmental interference features.

[0015] The beneficial effects of this invention are as follows: 1. This invention adopts an unsupervised contrastive learning framework, which automatically generates positive and negative sample pairs by utilizing wavelet domain differential perturbation. This eliminates the dependence on manually labeled classification information and enables large-scale training on a large amount of unlabeled measured data, significantly reducing data preparation costs.

[0016] 2. This invention uses a multi-scale decoupled encoder and pyramid-shaped feature segmentation to structurally separate global trends such as baseline drift and environmental interferences such as instrument random noise from the classification information in the latent space, forming independent representations of classification features and interference features, making the denoising process more transparent and controllable.

[0017] 3. This invention eliminates the correlation between feature subspaces from two dimensions, namely, the same layer and cross layers, through a hierarchical orthogonal penalty term. This prevents information crosstalk between signal components of different scales during the encoding process, ensures the strictness of decoupling, and enables the first feature vector output by the mesoscale layer to truly reflect the intrinsic structure of the fractal signal without being contaminated by global drift trends or local noise variance. Attached image description: The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of the hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation of the present invention; Figure 2 This is a diagram illustrating the overall framework of the hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation of the present invention. Detailed implementation method: To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0020] It should be noted that when a component is referred to as "fixed to" or "set on" another component, it can be directly on or indirectly on that other component. When a component is referred to as "connected to" another component, it can be directly connected to or indirectly connected to that other component.

[0021] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0022] In the description of this invention, it should be understood that the terms "upper" and "lower" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0023] Example 1: This invention provides a hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation, see [link to relevant documentation]. Figure 1 and Figure 2 The noise reduction method includes the following steps: S1: Obtain unlabeled raw mixed STR map data; perform one-dimensional wavelet transform on the raw mixed STR map data to decompose it into sub-band signals of multiple scales; then perform data augmentation based on the differential perturbation superposition of wavelet sub-band characteristics; then reconstruct the data through inverse wavelet transform to generate a first map view and a second map view that are positive sample pairs, and construct negative sample pairs by combining non-homologous data within the batch. S2: Construct a multi-scale decoupled encoder, input the first and second map views into the multi-scale decoupled encoder respectively, and map to obtain a high-dimensional latent representation tensor; divide the latent representation tensor into layers according to the receptive field scale, and extract the first feature vector representing the pure map classification information and the second feature vector representing the environmental interference features. S3: Based on unsupervised contrastive learning, a joint loss function is constructed to perform end-to-end iterative training on the multi-scale decoupled encoder and the reconstruction decoder. The multi-scale decoupled encoder optimizes its parameters by calculating the cross-view contrastive loss on the extracted first feature vector using positive and negative samples. The reconstruction decoder optimizes its parameters by calculating the reconstruction loss based on signal strength. Furthermore, the joint loss function incorporates variational information bottleneck regularization and hierarchical orthogonality penalty terms to achieve orthogonal decoupling of the latent representation tensor and lossless purification of fractal information. S4: Input the original mixed STR map data to be tested into the trained multi-scale decoupled encoder, extract and output the first feature vector, which is the pure mixed map classification information after filtering out environmental interference features.

[0024] As one possible implementation, the differential perturbation superposition in step S1 specifically includes: In a large-scale envelope subband characterizing the global trend, random walk sequences are superimposed to simulate the degradation trend of the map sequence and the global baseline drift distribution; In the mesoscale detail subband characterizing the local topological receptive field, a scaling perturbation is applied to simulate the fluctuating distribution of allele peaks and the relative relationships of local peak heights. Gaussian white noise is injected into small-scale detail subbands that characterize fine-grained variations to simulate the random noise distribution of the instrument.

[0025] Example: Wavelet domain differential perturbation and multi-view generation. Obtaining unlabeled raw mixed STR map data. x Using one-dimensional discrete wavelet transform to... x Decomposed into sub-band signals of multiple scales, let the number of decomposition levels be . J A set of approximation coefficients was obtained. With detail coefficients of each layer satisfy: in This is the inverse wavelet transform operator. Subbands are divided into three categories based on scale: large-scale approximation subbands (corresponding to...). ), mesoscale detail sub-bands (corresponding to intermediate layers) ) and small-scale detail sub-bands (corresponding to high-rise buildings) ).

[0026] Perform differentiated perturbations on various subbands: superimpose random walk sequences on large-scale approximation subbands. Simulate global baseline drift and degradation trends; multiply mesoscale detail subbands by a scaling factor. Simulate the relative height fluctuations of allele peaks; inject Gaussian white noise into small-scale detail subbands. Random noise in the simulation instrument.

[0027] Let the coefficients after the disturbance be The first atlas view is obtained after reconstruction by inverse wavelet transform. Repeat the above process to generate the second atlas view. . Positive sample pairs are formed by combining samples from different sources within the same batch, while negative sample pairs are formed by combining other non-homologous samples within the same batch.

[0028] As one possible implementation, step S2 involves hierarchically segmenting the latent representation tensor according to the receptive field scale, specifically employing a multi-scale feature decoupling pyramid structure: Large-scale global envelope layer: Based on the large global receptive field, feature vectors representing the degradation trend of the characterization map sequence and the global baseline drift are extracted; Mesoscale structural semantic layer: Based on the local topological receptive field, the first feature vector containing the relative relationship between the core allele peak and the local peak height is segmented; Small-scale noise perturbation layer: Based on fine-grained receptive field segmentation, feature vectors characterizing random noise in the instrument are extracted; Among them, the feature vectors output by the large-scale global envelope layer and the small-scale noise perturbation layer together constitute the second feature vector characterizing the environmental interference features.

[0029] Example: Multi-scale decoupled encoder and feature hierarchical segmentation. Constructing a multi-scale decoupled encoder. , respectively and Mapped to a high-dimensional latent representation tensor ,in For the number of channels, For the temporal dimension. The encoder employs a multi-receptive-field convolutional structure, enabling... Each channel exhibits a differentiated response along the receptive field scale. (The sentence is incomplete and requires further context.) It is divided into three subsets along the channel dimension: large-scale global envelope layer Extracting feature vectors representing global degradation trends and baseline drift; mesoscale structural semantic layer. Extract the first feature vector containing the relative relationship between the core allele peak and the local peak height. Small-scale noise disturbance layer Extract the feature vectors that characterize high-frequency random noise.

[0030] in, and Together they constitute the second eigenvector characterizing the features of environmental disturbance. : in, This indicates a cascading operation performed on the channel dimension of the feature tensor.

[0031] As one possible implementation, the cross-view contrast loss calculation method in step S3 is as follows: After normalizing the first feature vector of all views, the cosine similarity is calculated. The normalized temperature cross-entropy loss is used to make two views of the same sample close to each other in the feature space, and views of different samples far apart.

[0032] For example, the cosine similarity of the first feature vectors of all views is calculated after normalization, and the normalized temperature-scaling cross-entropy loss function is used for optimization constraints. Among them, two views generated from the same original mixed STR map constitute positive sample pairs. By constraining positive sample pairs to be close to each other in the feature space, the model is prompted to filter out external environmental noise such as baseline drift and sequence degradation, and extract stable and pure map typing features. Views generated from different original mixed STR maps constitute negative sample pairs. By forcing negative sample pairs to be far apart from each other in the feature space, the feature representation collapse of the multi-scale decoupled encoder that maps all inputs to a single vector is prevented. At the same time, the repulsive force generated by the distance between negative sample pairs amplifies the weak allelic differences between non-homologous maps, thereby delineating clear individual DNA typing boundaries in the latent space.

[0033] As one possible implementation, the variational information bottleneck regularization term in step S3 is calculated as follows: The mean and variance of the first feature vector are obtained by reparameterized sampling, and the posterior probability distribution of its latent space is constructed. While maximizing the information retention of the first feature vector, KL divergence is introduced to force the posterior probability distribution to approximate the preset standard normal prior distribution. This is used as a regularization penalty term to filter out non-fractal interference information with high uncertainty in the latent space.

[0034] As one possible implementation, the hierarchical orthogonality penalty term in step S3 includes intra-level orthogonal constraints and cross-level orthogonal constraints as follows: Same-layer orthogonality constraint: Calculate the matrix inner product of the first feature vector in the mesoscale structural semantic layer and the local interference vector segmented in the same layer, and minimize this inner product during backpropagation; Cross-layer orthogonality constraint: Force the first feature vector output by the mesoscale structural semantic layer to be orthogonal to the feature vector representing the global degradation trend output by the large-scale global envelope layer and the feature vector representing random noise output by the small-scale noise perturbation layer in the latent space. Hierarchical orthogonal constraints are used to prevent cross-scale feature aliasing or information leakage from multi-scale spectral signals during the extraction process.

[0035] As one possible implementation, the reconstructed graph in step S3 is as follows: The first and second feature vectors after segmentation are concatenated and recombined along the channel dimension, and then input into the reconstruction decoder to obtain the reconstructed map.

[0036] As one possible implementation, the signal strength-weighted reconstruction loss in step S3 is calculated as follows: Weak peak regions with signal intensities below an adaptive threshold in the original mixed STR spectral data are penalized with a higher weight than the main peak regions to ensure that the pure spectral classification information of extremely low template amounts or minor contributors is preserved during the dimensionality reduction and decoupling process.

[0037] As one possible implementation, the joint loss function in step S3 is calculated as follows: in, This is the balance coefficient; For cross-view contrast loss; This is a variational information bottleneck regularization term; This is a hierarchical orthogonal penalty term; Reconstruction loss based on signal strength weighting; joint loss The multi-scale decoupled encoder and reconstruction decoder are iteratively trained end-to-end via backpropagation.

[0038] Example: Based on an unsupervised contrastive learning framework and combined with a graph reconstruction task, a joint loss function is constructed for the encoder. and decoder Perform end-to-end optimization. Joint loss function. It consists of four items: in This is the balance coefficient.

[0039] (1) Cross-view contrast loss : For each sample i, let its first feature vector be . and Normalized temperature cross-entropy loss is used: in For cosine similarity, For temperature parameters, For batch size, superscript This indicates that the view is from any view.

[0040] (2) Variational information bottleneck regularization term : Using reparameterization techniques, the mean of the first feature vector is calculated from the output of the mesoscale layer. and logarithmic variance Constructing an approximate posterior By minimizing its prior with respect to the standard normal. KL divergence, compressibility uncertainty: in, The feature dimension is defined by this regularization term, which forces the fractal features to be distributed in a low-information bottleneck space, automatically filtering out non-fractal variations with high uncertainty.

[0041] (3) Hierarchical orthogonal penalty term : Hierarchical orthogonal penalty term It includes two parts: intra-layer orthogonal and inter-layer orthogonal.

[0042] Orthogonality within the same layer: Ensures that the fractal subspace and the local disturbance subspace within the mesoscale layer are uncorrelated. Assume that within the mesoscale layer, excluding the first eigenvector... In addition, local interference vectors are also segmented. ,but: Cross-layer orthogonality: forcing the first feature vector output by the mesoscale layer to be orthogonal to the feature vector of the large-scale layer. Small-scale layer feature vectors They are pairwise orthogonal within the potential space: The first feature vector of the forced mesoscale layer output is compared with the feature vector of the large-scale layer that constitutes the second feature vector. Small-scale layer feature vectors They are pairwise orthogonal within the potential space: The total orthogonal penalty is: (4) Reconstruction loss based on signal strength weighting : The segmented multi-scale feature vectors , , , Cascaded along the channel dimension, input decoder To generate reconstructed maps : In the formula, This indicates a cascading operation performed on the channel dimension of the feature tensor. The nonlinear mapping function of the reconstructed decoder.

[0043] When calculating the point-by-point reconstruction error, the original spectral signal strength is used as the basis. and adaptive threshold Assign difference weights Adaptive threshold Typically, a certain quantile of the peak intensity of the spectrum is chosen. The weighted reconstruction loss is defined as: In the formula, This represents the total number of signal points in the spectral map. By assigning higher weights to weak peak regions, the decoder is forced to faithfully preserve information from low-signal-intensity regions during reconstruction, thus achieving fidelity in fractal information.

[0044] Joint losses Through the backpropagation algorithm, along the attached Figure 2 Data flow to the reverse update encoder and decoder Complete end-to-end joint training.

[0045] In a specific instance; S4: Inference and Feature Output. After training, the decoder and perturbation module are discarded, and only the multi-scale decoupled encoder is retained. Input the mixed STR spectrum to be tested. The latent representation tensor is segmented according to step S2, and the first feature vector is extracted and output. This feature vector represents the pure genotyping information after filtering out environmental interferences such as global degradation trends, baseline drift, and high-frequency random noise, and can be directly used for subsequent genotype inference.

[0046] Example 2: This invention provides a hybrid STR map denoising system based on multi-scale wavelet domain decoupling representation, comprising: The data augmentation module is used to acquire unlabeled raw mixed STR map data; perform one-dimensional wavelet transform on the raw mixed STR map data to decompose it into sub-band signals of multiple scales; then perform data augmentation by superimposing differential perturbations based on the characteristics of wavelet sub-bands; then reconstruct the data by inverse wavelet transform to generate a first map view and a second map view that are positive sample pairs, and construct negative sample pairs by combining non-homologous data within the batch. The construction and feature segmentation module is used to construct a multi-scale decoupled encoder. The first and second map views are input into the multi-scale decoupled encoder respectively to obtain a high-dimensional latent representation tensor. The latent representation tensor is segmented hierarchically according to the receptive field scale to extract the first feature vector representing the pure map classification information and the second feature vector representing the environmental interference features. The joint loss optimization module is used to construct a joint loss function based on unsupervised contrastive learning to perform end-to-end iterative training on the multi-scale decoupled encoder and the reconstruction decoder. The multi-scale decoupled encoder optimizes its parameters by calculating the cross-view contrastive loss on the extracted first feature vector using positive and negative samples. The reconstruction decoder optimizes its parameters by calculating the reconstruction loss based on signal strength. Furthermore, the joint loss function incorporates variational information bottleneck regularization and hierarchical orthogonality penalty terms to achieve orthogonal decoupling of the latent representation tensor and lossless purification of fractal information. The information extraction and output module is used to input the raw mixed STR map data to be tested into the trained multi-scale decoupled encoder, extract and output the first feature vector, which is the pure mixed map classification information after filtering out environmental interference features.

[0047] This system is used to execute the hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation in Embodiment 1 above.

[0048] Example 3: This invention discloses a hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation. The specific implementation steps are as follows: Step S1: Wavelet domain differential perturbation and multi-view generation. A set of unlabeled mixed STR map data is collected. Each map is a one-dimensional signal sequence without any annotation information such as contributor attribution, allele type, or mixing ratio.

[0049] A one-dimensional discrete wavelet transform was performed on each spectrum. In this embodiment, the db4 wavelet basis was selected, and the decomposition level was set to 5. After the transform, a set of approximation coefficients and five sets of detail coefficients were obtained. These coefficients were divided into three categories according to scale: large-scale approximation subbands reflect global baseline drift and degradation trends; medium-scale detail subbands correspond to the width range of allele peaks, carrying the core peak morphology and peak height relationship; small-scale detail subbands correspond to high-frequency components, mainly characterizing instrument random noise.

[0050] Differential perturbations are performed on various subbands to generate two enhanced views. For large-scale approximation subbands, a random walk sequence with a step size following a normal distribution is superimposed to simulate baseline variations between different samples. For mesoscale detail subbands, a scaling factor based on sampling from a uniform distribution is applied to simulate peak height fluctuations caused by different mixing ratios or amplification efficiencies, while preserving peak shape characteristics. For small-scale detail subbands, zero-mean Gaussian white noise with a standard deviation equal to one-tenth of the standard deviation of the corresponding subband coefficients is injected to simulate random variations in instrument noise.

[0051] The first spectrogram view is obtained by reconstructing all perturbation coefficients using inverse wavelet transform; the perturbation parameters are then resampled and the above process is repeated to obtain the second spectrogram view. The two views originate from the same sample and are positive sample pairs. Within a training batch, several spectrograms from different sources are randomly selected, and the views of non-similar sources within the batch are combined to form negative sample pairs.

[0052] Step S2: Multi-scale decoupled encoder and feature hierarchical segmentation. A multi-scale decoupled encoder is constructed, primarily using a three-layer one-dimensional convolutional neural network. Each layer employs a different kernel size to form a multi-receptive field representation, followed by batch normalization and ReLU activation. The encoder outputs a high-dimensional latent representation tensor.

[0053] The latent representation tensor is divided into three subsets along the channel dimension, forming a multi-scale feature pyramid. The large-scale global envelope layer uses fewer channels, and due to its deep receptive field, it tends to capture global trends. The mesoscale structural semantic layer uses more channels and is further subdivided into two parts: the primary part serves as the first feature vector, carrying allelic genotyping information; the secondary part serves as the local perturbation vector within the same layer. The small-scale noise perturbation layer uses the remaining channels to encode high-frequency noise features.

[0054] The outputs of the large-scale and small-scale layers, after being flattened, yield the global feature vector and the noise feature vector, respectively. These two vectors, together with the local disturbance vector in the mesoscale layer, constitute the second feature vector, i.e., the environmental disturbance feature. The first feature vector of the mesoscale layer, after being flattened, serves as the core fractal representation.

[0055] Step S3: Multi-objective joint loss function and training. Construct a reconstruction decoder using a transposed convolutional structure symmetric to the encoder to reconstruct the original map from the cascaded multi-scale feature vectors.

[0056] The joint loss function consists of four terms: cross-view contrast loss, variational information bottleneck regularization term, hierarchical orthogonality penalty term, and signal strength-weighted reconstruction loss.

[0057] The cross-view comparison loss is calculated using only the first feature vector. After normalizing the first feature vectors of all views, the cosine similarity is calculated, and the normalized temperature cross-entropy loss is used to make two views of the same sample closer in the feature space, and views of different samples farther apart.

[0058] The variational information bottleneck regularization term adds two parallel fully connected heads to the output of the first feature vector in the mesoscale layer, which output the mean and variance respectively. The first feature vector is obtained through reparameterized sampling, and the KL divergence between its posterior distribution and the standard normal prior is calculated, which forces the distribution of the genotyping features to be smooth and compact, and suppresses non-genotyping variation.

[0059] The hierarchical orthogonality penalty term comprises two parts: intra-layer and cross-layer. Intra-layer orthogonality calculates the squared Frobenius norm of the transpose of the first eigenvector and the local disturbance vector within the midscale layer. Cross-layer orthogonality calculates the sum of the squared Frobenius norms between each pair of eigenvectors: the first eigenvector and the global eigenvector, the first eigenvector and the noisy eigenvector, and the global eigenvector and the noisy eigenvector. By minimizing the orthogonality penalty, the correlation between eigenspaces is explicitly eliminated.

[0060] In the signal strength-weighted reconstruction loss, an adaptive threshold is first set, assigning lower weights to the main peak regions with signal strength above the threshold and higher weights to the weak peak regions with signal strength below the threshold. Then, the weighted mean square error is calculated. This amplifies the reconstruction error in weak regions, forcing the decoder to accurately reconstruct low-signal regions and ensuring that the fractal information of secondary contributors is not lost during dimensionality reduction and decoupling.

[0061] During training, the Adam optimizer is used for end-to-end optimization until all losses converge.

[0062] Step S4: Inference and Feature Output. After training, the decoder, reparameterization sampling module, and data augmentation perturbation process are discarded, retaining only the encoder and its internal segmentation structure. The mixed STR map to be tested is directly input into the encoder to extract the first feature vector of the mesoscale layer. The mean value is then taken as the deterministic output. This feature vector represents the pure genotyping information after filtering out global baseline drift, degradation trends, and high-frequency random noise, and can be directly used for subsequent genotype inference or likelihood ratio calculation.

[0063] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the description of the drawings, in carrying out the claimed invention. In this specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple components. A single processor or other unit can implement several of the functions listed in the specification. While certain measures are described in different embodiments, this does not mean that these measures cannot be combined to produce good results.

[0064] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely illustrative of the invention and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation, characterized in that, Includes the following steps: S1: Obtain unlabeled raw mixed STR map data; perform one-dimensional wavelet transform on the raw mixed STR map data to decompose it into sub-band signals of multiple scales; then perform data enhancement based on the differential perturbation superposition of wavelet sub-band characteristics; Subsequently, after inverse wavelet transform reconstruction, a first and second spectrogram views with positive sample pairs are generated, and negative sample pairs are constructed by combining non-homologous data within the batch. S2: Construct a multi-scale decoupled encoder, input the first and second map views into the multi-scale decoupled encoder respectively, and map to obtain a high-dimensional latent representation tensor; divide the latent representation tensor into layers according to the receptive field scale, and extract the first feature vector representing the pure map classification information and the second feature vector representing the environmental interference features. S3: Based on unsupervised contrastive learning, a joint loss function is constructed to perform end-to-end iterative training on the multi-scale decoupled encoder and the reconstruction decoder; wherein, the multi-scale decoupled encoder optimizes its parameters by calculating cross-view contrastive loss on the extracted first feature vector using the positive and negative samples, and the reconstruction decoder optimizes its parameters by calculating reconstruction loss based on signal strength weighting; and the joint loss function also incorporates variational information bottleneck regularization and hierarchical orthogonality penalty term to achieve orthogonal decoupling of the latent representation tensor and lossless purification of fractal information; S4: Input the original mixed STR map data to be tested into the trained multi-scale decoupled encoder, extract and output the first feature vector, which is the pure mixed map typing information after filtering out environmental interference features.

2. The hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation according to claim 1, characterized in that, The differential perturbation superposition in S1 specifically includes: In a large-scale envelope subband characterizing the global trend, random walk sequences are superimposed to simulate the degradation trend of the map sequence and the global baseline drift distribution; In the mesoscale detail subband characterizing the local topological receptive field, a scaling perturbation is applied to simulate the fluctuating distribution of allele peaks and the relative relationships of local peak heights. Gaussian white noise is injected into small-scale detail subbands that characterize fine-grained variations to simulate the random noise distribution of the instrument.

3. The hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation according to claim 1, characterized in that, The latent representation tensor in S2 is hierarchically segmented according to the receptive field scale, specifically employing a multi-scale feature decoupling pyramid structure: Large-scale global envelope layer: Based on the large global receptive field, feature vectors representing the degradation trend of the characterization map sequence and the global baseline drift are extracted; Mesoscale structural semantic layer: Based on the local topological receptive field, the first feature vector containing the relative relationship between the core allele peak and the local peak height is segmented; Small-scale noise perturbation layer: Based on fine-grained receptive field segmentation, feature vectors characterizing random noise in the instrument are extracted; The feature vectors output by the large-scale global envelope layer and the small-scale noise perturbation layer together constitute the second feature vector characterizing the environmental interference features.

4. The hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation according to claim 1, characterized in that, The cross-view comparison loss calculation method in S3 is as follows: After normalizing the first feature vector of all views, the cosine similarity is calculated. The normalized temperature cross-entropy loss is used to make two views of the same sample close to each other in the feature space, and views of different samples far apart.

5. The hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation according to claim 1, characterized in that, The variational information bottleneck regularization term in S3 is calculated as follows: The mean and variance of the first feature vector are obtained by reparameterized sampling, and the posterior probability distribution of its latent space is constructed. While maximizing the information retention of the first feature vector, KL divergence is introduced to force the posterior probability distribution to approximate the preset standard normal prior distribution, which is used as a regularization penalty term to filter out non-fractal interference information with high uncertainty in the latent space.

6. The hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation according to claim 3, characterized in that, The hierarchical orthogonal penalty term in S3 includes intra-level orthogonal constraints and cross-level orthogonal constraints as follows: Same-layer orthogonality constraint: Calculate the matrix inner product of the first feature vector in the mesoscale structural semantic layer and the local interference vector segmented in the same layer, and minimize this inner product during backpropagation; Cross-layer orthogonality constraint: Force the first feature vector output by the mesoscale structural semantic layer to be orthogonal to the feature vector representing the global degradation trend output by the large-scale global envelope layer and the feature vector representing random noise output by the small-scale noise perturbation layer in the latent space. The hierarchical orthogonal constraints prevent cross-scale feature aliasing or information leakage from occurring during the extraction of multi-scale spectral signals.

7. The hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation according to claim 1, characterized in that, The processing of the reconstruction decoder is as follows: The first and second feature vectors after segmentation are concatenated and recombined in the channel dimension and input into the reconstruction decoder to obtain the reconstruction spectrum. The reconstruction loss based on signal strength weighting is then calculated using the reconstruction spectrum.

8. The hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation according to claim 1 or 7, characterized in that, The reconstruction loss based on signal strength weighting in S3 is calculated as follows: Weak peak regions with signal intensities below an adaptive threshold in the original mixed STR spectral data are penalized with a higher weight than the main peak regions to ensure that the pure spectral classification information of extremely low template amounts or minor contributors is preserved during the dimensionality reduction and decoupling process.

9. The hybrid STR map denoising method based on multi-scale wavelet domain decoupling representation according to claim 1, characterized in that, The joint loss function in S3 is calculated as follows: in, This is the balance coefficient; For cross-view contrast loss; This is a variational information bottleneck regularization term; This is a hierarchical orthogonal penalty term; Reconstruction loss based on signal strength weighting; joint loss The multi-scale decoupled encoder and reconstruction decoder are iteratively trained end-to-end via backpropagation.

10. A hybrid STR map denoising system based on multi-scale wavelet domain decoupling representation, characterized in that, include: The data augmentation module is used to acquire unlabeled raw mixed STR map data; The original mixed STR map data is subjected to a one-dimensional wavelet transform to decompose it into sub-band signals of multiple scales; then, data enhancement is performed by superimposing differential perturbations based on the wavelet sub-band characteristics. Subsequently, after inverse wavelet transform reconstruction, a first and second spectrogram views with positive sample pairs are generated, and negative sample pairs are constructed by combining non-homologous data within the batch. A construction and feature segmentation module is used to construct a multi-scale decoupled encoder. The first and second map views are respectively input into the multi-scale decoupled encoder to map and obtain a high-dimensional latent representation tensor. The latent representation tensor is segmented hierarchically according to the receptive field scale to extract a first feature vector representing the pure map classification information and a second feature vector representing the environmental interference features. The joint loss optimization module is used to construct a joint loss function based on unsupervised contrastive learning to perform end-to-end iterative training on the multi-scale decoupled encoder and the reconstruction decoder. The multi-scale decoupled encoder optimizes its parameters by calculating a cross-view contrastive loss on the extracted first feature vector using the positive and negative samples. The reconstruction decoder optimizes its parameters by calculating a reconstruction loss weighted by signal strength. Furthermore, the joint loss function incorporates a variational information bottleneck regularization term and a hierarchical orthogonality penalty term to achieve orthogonal decoupling of the latent representation tensor and lossless purification of fractal information. The information extraction and output module is used to input the raw mixed STR map data to be tested into the trained multi-scale decoupled encoder, extract and output the first feature vector, which is the pure mixed map typing information after filtering out environmental interference features.

Images