Data preprocessing method based on multi-frequency domain attention mechanism, medium and system

Abstract

The application discloses a data preprocessing method, medium and system based on a multi-frequency domain attention mechanism, wherein the method comprises the following steps: acquiring multi-source unstructured data and preprocessing the multi-source unstructured data to generate standardized data; performing frequency domain transformation to generate a frequency domain feature tensor and performing reversible frequency domain decomposition to generate corresponding multi-frequency sub-band features; performing adaptive modeling and feature fusion to generate multi-frequency fusion features; performing dynamic quantization and coding processing on the multi-frequency fusion features to generate corresponding multi-frequency feature representations; constructing a multi-scale frequency domain pyramid structure to perform layered modeling and cross-scale collaborative fusion and generate multi-scale frequency domain collaborative feature representations; and performing inverse frequency domain transformation operations to effectively reduce noise and redundant features of the final preprocessing data.

Description

Technical Field

[0001] This application relates to the field of input data preprocessing technology for detection and control systems, and in particular to a data preprocessing method, medium and system based on a multi-frequency domain attention mechanism. Background Technology

[0002] With the continuous development of the electronic information industry and computer technology, the scale of unstructured data generated during engineering design, intelligent manufacturing, industrial testing, and information system operation is constantly expanding. This data comes from more diverse sources, exhibits more complex forms, and generally possesses characteristics such as inconsistent structure, significant scale differences, and the superposition of noise and redundant information. Such unstructured data typically contains key structural information, functional semantics, or operational status features, and its parsing and processing results directly affect the accuracy and reliability of subsequent system modeling, performance evaluation, fault diagnosis, and safety decisions. Automated parsing and processing of unstructured data using computer systems can not only effectively reduce the uncertainty brought about by manual analysis and experience-based judgment, but also improve the consistency, objectivity, and efficiency of the data processing process, which is of great significance in enhancing the intelligence level of information systems and the reliability of engineering projects.

[0003] In related technologies, when preprocessing input data (especially unstructured data) of information systems and detection and control systems, statistical analysis or feature extraction is often performed directly in the time domain or the original spatial domain. However, this approach is poorly adaptable to noise interference, inconsistent data scales, and heterogeneous multi-source characteristics, and is prone to causing aliasing of processed feature information. Summary of the Invention

[0004] This invention aims to at least partially address one of the technical problems in related technologies. To this end, one objective of this invention is to propose a data preprocessing method based on a multi-frequency domain attention mechanism, which can effectively reduce noise and redundancy features in the final preprocessed data.

[0005] In a first aspect, embodiments of the present invention propose a data preprocessing method based on a multi-frequency domain attention mechanism, comprising the following steps: acquiring multi-source unstructured data and preprocessing the multi-source unstructured data to generate standardized data; performing frequency domain transformation on the standardized data to generate a frequency domain feature tensor, and performing reversible frequency domain decomposition on the frequency domain feature tensor to generate corresponding multi-frequency sub-band features; performing adaptive modeling and feature fusion on the multi-frequency sub-band features to generate multi-frequency fused features; performing dynamic quantization and encoding processing on the multi-frequency fused features to generate corresponding multi-frequency feature representations; constructing a multi-scale frequency domain pyramid structure to perform hierarchical modeling and cross-scale collaborative fusion on the multi-frequency feature representations to generate multi-scale frequency domain collaborative feature representations; performing an inverse frequency domain transformation operation on the multi-scale frequency domain collaborative feature representations to achieve high-fidelity reconstruction of the multi-source unstructured data, thereby obtaining the final preprocessing result corresponding to the multi-source unstructured data. This effectively reduces noise and redundant features in the final preprocessed data.

[0006] In some embodiments, the multi-source unstructured data includes high-resolution data and low-resolution multidimensional data. Preprocessing the multi-source unstructured data to generate standardized data includes: converting the format of the multi-source unstructured data to generate corresponding floating-point data; normalizing the format-converted data; and denoising the normalized data to generate standardized data.

[0007] In some embodiments, the standardized data is subjected to frequency domain transformation to generate a frequency domain feature tensor, and the frequency domain feature tensor is subjected to reversible frequency domain decomposition to generate corresponding multi-frequency sub-band features. This includes: applying a frequency domain transformation operator to the standardized data to map the standardized data to a frequency domain space to generate a frequency domain feature tensor; performing reversible frequency domain decomposition on the frequency domain feature tensor to divide the frequency domain space into multiple sub-band regions, and calculating the corresponding multi-frequency sub-band features based on a frequency mask function; wherein the multi-frequency sub-band features include: high-frequency sub-band features, mid-frequency sub-band features, and low-frequency sub-band features.

[0008] In some embodiments, adaptive modeling and feature fusion are performed on the multi-frequency sub-band features to generate multi-frequency fused features, including: constructing a frequency-aware query vector based on the multi-frequency sub-band features, and mapping the multi-frequency sub-band features to an attention query space based on the frequency-aware query vector; constructing a spatial domain key vector based on the multi-frequency sub-band features to describe the overall structure and contextual relationships of the data in the spatial dimension; constructing a fusion value vector based on the multi-frequency sub-band features; calculating the attention weight corresponding to each multi-frequency sub-band feature according to the query vector, the spatial domain key vector, and the fusion value vector, and performing weighted fusion of the multi-frequency sub-band features based on the attention weights to generate multi-frequency fused features.

[0009] In some embodiments, dynamic quantization and encoding processing is performed on the multi-frequency fusion features to generate corresponding multi-frequency feature representations, including: performing statistical characteristic analysis on the multi-frequency fusion features to obtain statistical distribution information corresponding to the multi-frequency fusion features; calculating the frequency importance weight corresponding to each frequency sub-band in the multi-frequency fusion features, and determining the quantization step size corresponding to each frequency sub-band based on the statistical distribution information and the frequency importance weight; performing quantization mapping operation on the multi-frequency fusion features based on the quantization step size to obtain corresponding quantization results; and encoding the quantization results based on a bit allocation strategy to generate corresponding multi-frequency feature representations.

[0010] In some embodiments, a multi-scale frequency domain pyramid structure is constructed to perform hierarchical modeling and cross-scale collaborative fusion of the multi-frequency feature representations to generate a multi-scale frequency domain collaborative feature representation, including: generating multiple frequency domain feature representations at different scale levels in the frequency domain space based on a scale transformation operator to form multi-scale frequency domain features; and performing cross-scale mapping and fusion of the multi-scale frequency domain features based on a cross-scale mapping operator to generate a multi-scale frequency domain collaborative feature representation.

[0011] In some embodiments, an inverse frequency domain transformation operation is performed on the multi-scale frequency domain collaborative feature representation to achieve high-fidelity reconstruction of the multi-source unstructured data and obtain the final preprocessing result corresponding to the multi-source unstructured data. This includes: decoding and inverse quantizing the multi-scale frequency domain collaborative feature representation based on decoding and inverse quantization operators to obtain recovered frequency domain features at each scale level; performing inverse scale mapping and fusion processing on the recovered frequency domain features at each scale level based on inverse scale mapping operators to obtain aligned frequency domain feature representations; performing cross-scale aggregation on the aligned frequency domain feature representations to generate an aggregation result; and performing an inverse frequency domain transformation operation on the aggregation result to generate the final preprocessing result.

[0012] Secondly, embodiments of the present invention provide a computer-readable storage medium storing a data preprocessing program based on a multi-frequency domain attention mechanism, which, when executed by a processor, implements the data preprocessing method based on the multi-frequency domain attention mechanism as described above.

[0013] Thirdly, this invention proposes a data preprocessing system based on a multi-frequency domain attention mechanism, comprising: a multi-source unstructured data acquisition and preprocessing module, which acquires multi-source unstructured data and preprocesses it to generate standardized data; a reversible frequency domain decomposition and multi-frequency feature construction module, which performs frequency domain transformation on the standardized data to generate a frequency domain feature tensor and performs reversible frequency domain decomposition on the frequency domain feature tensor to generate corresponding multi-frequency sub-band features; and a multi-frequency domain attention fusion module, which adaptively models and performs feature fusion on the multi-frequency sub-band features. The system comprises: a feature fusion module to generate multi-frequency fusion features; an adaptive quantization and encoding module to perform dynamic quantization and encoding processing on the multi-frequency fusion features to generate corresponding multi-frequency feature representations; a multi-scale frequency domain pyramid collaborative modeling module to construct a multi-scale frequency domain pyramid structure for hierarchical modeling and cross-scale collaborative fusion of the multi-frequency feature representations to generate multi-scale frequency domain collaborative feature representations; and a frequency domain reconstruction and output module to perform inverse frequency domain transformation on the multi-scale frequency domain collaborative feature representations to achieve high-fidelity reconstruction of the multi-source unstructured data and obtain the final preprocessing result corresponding to the multi-source unstructured data.

[0014] In some embodiments, the multi-source unstructured data includes high-resolution data and low-resolution multidimensional data. The multi-source unstructured data acquisition and preprocessing module is further configured to perform format conversion on the multi-source unstructured data to generate corresponding floating-point data; normalize the format-converted data; and denoise the normalized data to generate standardized data.

[0015] According to the data preprocessing method based on a multi-frequency domain attention mechanism of the present invention, firstly, multi-source unstructured data is acquired and preprocessed to generate standardized data; then, the standardized data undergoes frequency domain transformation to generate a frequency domain feature tensor, and the frequency domain feature tensor is reversibly decomposed in the frequency domain to generate corresponding multi-frequency sub-band features; thus, by performing reversible frequency domain decomposition on the frequency domain feature tensor and adaptive modeling of the multi-frequency sub-band features, the global structural information and local detail information in the data can be effectively decoupled, reducing the interference of noise and redundant features on the processing results; then, the multi-frequency sub-band features are adaptively modeled and feature fused to generate multi-frequency fused features. The system first identifies the multi-frequency fusion features and then performs dynamic quantization and encoding to generate corresponding multi-frequency feature representations. Next, a multi-scale frequency domain pyramid structure is constructed to perform hierarchical modeling and cross-scale collaborative fusion of the multi-frequency feature representations, generating a multi-scale frequency domain collaborative feature representation. This multi-scale frequency domain pyramid structure enables cross-scale feature collaborative modeling, preventing unstable output results due to excessively large scale variations in multi-source unstructured data. Then, an inverse frequency domain transformation operation is performed on the multi-scale frequency domain collaborative feature representation to achieve high-fidelity reconstruction of the multi-source unstructured data, obtaining the final preprocessed result. Thus, through reversible frequency domain transformation and high-fidelity reconstruction, a clear mathematical mapping relationship is maintained between the processing result and the multi-source unstructured data, enhancing its verifiability and interpretability.

[0016] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of a data preprocessing method based on a multi-frequency domain attention mechanism according to an embodiment of the present invention;

[0018] Figure 2 This is a block diagram of a data preprocessing system based on a multi-frequency domain attention mechanism according to an embodiment of the present invention. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0020] The data preprocessing method based on a multi-frequency domain attention mechanism according to an embodiment of the present invention is described below with reference to the accompanying drawings.

[0021] Please see Figure 1 , Figure 1 This is a flowchart illustrating a data preprocessing method based on a multi-frequency domain attention mechanism according to an embodiment of the present invention. Figure 1 As shown, this data preprocessing method based on a multi-frequency domain attention mechanism includes the following steps:

[0022] S101: Acquire multi-source unstructured data and preprocess the multi-source unstructured data to generate standardized data.

[0023] In some embodiments, the multi-source unstructured data includes high-resolution data and low-resolution multidimensional data. Preprocessing the multi-source unstructured data to generate standardized data includes: converting the format of the multi-source unstructured data to generate corresponding floating-point data; normalizing the format-converted data; and denoising the normalized data to generate standardized data.

[0024] As an example, firstly, unstructured data is obtained from at least two different data sources, i.e., multi-source unstructured data is obtained. This multi-source unstructured data can be divided into two categories: high-resolution data and low-resolution multidimensional data. High-resolution data can be represented as:

[0025]

[0026] in, Represents high-resolution data. and This represents the spatial resolution dimension of high-resolution data. Indicates the number of channels.

[0027] Low-resolution multidimensional data is represented as follows:

[0028]

[0029] in, Represents low-resolution multidimensional data. , , and This represents the spatial resolution dimension of low-resolution multidimensional data. This represents the feature dimensions of low-resolution multidimensional data. It addresses the differences in sampling density, dimensional definitions, and structural forms among multi-source data by performing unified preprocessing and modeling on unstructured multi-source data.

[0030] Next, at the data representation level, the multi-source unstructured data is first processed for format normalization, converting the original data into floating-point form, which is expressed as follows:

[0031]

[0032] in, Represents floating-point data. Represents data type conversion operators. This represents multi-source unstructured data.

[0033] Then, a unified constraint is applied to the differences in dimensional arrangement across different data sources, ensuring that all data adopts a "spatial dimension – channel dimension" arrangement structure. When the number of channels differs between different data sources, a channel mapping function is used. Mapping low-resolution data to a unified feature space, i.e.:

[0034]

[0035] This achieves a consistent representation of the feature dimensions.

[0036] Next, at the numerical scale level, scale normalization or standardization is performed on the multi-source data to eliminate differences in numerical range, dynamic range, and physical dimensions between different data sources. This is achieved by minimizing the statistical characteristic parameters. Maximum value mean and standard deviation Linear normalization can be used:

[0037]

[0038] Alternatively, a standardized approach can be adopted:

[0039]

[0040] This is to ensure that multi-source data maintains consistency in statistical distribution.

[0041] Then, at the spatial scale level, to address the spatial inconsistency between high-resolution and low-resolution data, this module uses an interpolation mapping function. Scale alignment of low-resolution data is mathematically expressed as follows:

[0042]

[0043] This enables unified alignment of multi-source unstructured data in the spatial dimension.

[0044] S102, perform frequency domain transformation on the standardized data to generate a frequency domain feature tensor, and perform reversible frequency domain decomposition on the frequency domain feature tensor to generate the corresponding multi-frequency sub-band features.

[0045] In some embodiments, a frequency domain transformation is performed on the standardized data to generate a frequency domain feature tensor, and a reversible frequency domain decomposition is performed on the frequency domain feature tensor to generate corresponding multi-frequency sub-band features. This includes: applying a frequency domain transformation operator to the standardized data to map the standardized data to the frequency domain space to generate a frequency domain feature tensor; performing a reversible frequency domain decomposition on the frequency domain feature tensor to divide the frequency domain space into multiple sub-band regions, and calculating the corresponding multi-frequency sub-band features based on a frequency mask function; wherein the multi-frequency sub-band features include: high-frequency sub-band features, mid-frequency sub-band features, and low-frequency sub-band features.

[0046] As an example, firstly, suppose the preprocessed... Multi-source unstructured data (i.e., the first type) Normalized data is represented as:

[0047]

[0048] in, Indicates the first Classified data, and This represents the unified spatial resolution. This represents the aligned feature channel dimension. To characterize the frequency characteristics of the data at different scales, such as global structure, regional variations, and local details, a frequency domain transformation operator is applied to the data. By mapping it from the spatial domain to the frequency domain, we obtain the corresponding frequency domain representation:

[0049]

[0050] in, Indicates the first The frequency representation corresponding to the normalized data is a frequency domain feature tensor in complex or real number form, used to characterize the energy distribution and structural characteristics of different frequency components in the overall data. As an invertible transform operator, it can take the form of two-dimensional discrete Fourier transform, discrete cosine transform, or invertible wavelet transform, etc., and its general expression can be represented as:

[0051]

[0052] in, Indicates spatial coordinate position, Indicates the corresponding frequency coordinates. This serves as the channel index. Thus, through the above mapping relationship, the smooth structure, texture variations, and local abrupt changes in the original normalized data are transformed into the amplitude and phase distributions of different frequency components in the frequency domain.

[0053] Then, after completing the frequency domain mapping, a reversible frequency domain decomposition operation is performed on the frequency domain features. According to a preset frequency partitioning strategy, the frequency domain space is divided into multiple sub-band regions with clear frequency semantics, and corresponding frequency mask functions are constructed for each. Let the low-frequency, mid-frequency, and high-frequency mask functions be respectively... , and The corresponding multi-frequency subband characteristics are expressed as:

[0054]

[0055] in, This represents element-wise multiplication. Indicates the characteristics of low-frequency subbands. Represents the low-frequency masking function. Indicates the characteristics of the intermediate frequency subband. Represents the intermediate frequency masking function. Indicates the characteristics of high-frequency subbands. This represents the high-frequency masking function. It should be noted that low-frequency subband features are used to characterize global structural information and slow-changing trends in the data, mid-frequency subband features are used to depict regional transitions and mid-scale texture features, and high-frequency subband features are used to describe edges, details, and rapidly changing local information.

[0056] In some embodiments, to ensure the reversibility and information completeness of the frequency domain decomposition process, the frequency mask function satisfies the following completeness constraint:

[0057]

[0058] This ensures that the features of each frequency sub-band do not overlap in the frequency domain and collectively cover the entire spectrum. Under this constraint, multi-frequency sub-band features can be transformed using the inverse frequency domain transform operator. To achieve lossless reconstruction, the relationship can be represented as follows:

[0059]

[0060] In this way, the lossless nature of information can be guaranteed in the frequency domain decomposition and multi-frequency feature construction process, both from a theoretical and engineering implementation perspective.

[0061] S103 performs adaptive modeling and feature fusion on multi-frequency sub-band features to generate multi-frequency fused features.

[0062] In some embodiments, adaptive modeling and feature fusion of multi-frequency sub-band features are performed to generate multi-frequency fused features, including: constructing a frequency-aware query vector based on the multi-frequency sub-band features, and mapping the multi-frequency sub-band features to an attention query space based on the frequency-aware query vector; constructing a spatial domain key vector based on the multi-frequency sub-band features to describe the overall structure and contextual relationships of the data in the spatial dimension; constructing a fusion value vector based on the multi-frequency sub-band features; calculating the attention weight corresponding to each multi-frequency sub-band feature according to the query vector, the spatial domain key vector, and the fusion value vector, and performing weighted fusion of the multi-frequency sub-band features based on the attention weights to generate multi-frequency fused features.

[0063] As an example, by constructing a frequency-semantic-oriented attention mechanism, we can achieve dynamic evaluation and weighting of the contribution of different frequency components in the target task, thereby enhancing the discriminative power and robustness of the fused features while maintaining the frequency decoupling characteristics.

[0064] First, let the low-frequency subband characteristics, mid-frequency subband characteristics, and high-frequency subband characteristics be represented as follows:

[0065]

[0066] The features of each frequency sub-band are consistent in both spatial and channel dimensions, with differences only at the frequency semantic level.

[0067] To achieve differentiated modeling of features from different frequency sub-bands, a frequency-aware query vector is constructed based on multi-frequency sub-band features. Specifically, for each frequency sub-band feature, a query mapping function is used... Mapping to a unified attention query space yields...

[0068]

[0069]

[0070] in, This represents the mapping value corresponding to the sub-band feature. This represents the attention embedding dimension, used to characterize the discriminative features of the corresponding frequency sub-band at the frequency domain level.

[0071] To incorporate spatial structure information into the frequency selection process, spatial domain key vectors are constructed based on multi-frequency sub-band features to describe the overall structure and contextual relationships of the data in the spatial dimension. The generation method of the spatial domain key vectors is as follows:

[0072]

[0073] in, Represents the spatial domain key vector. This represents the spatial domain key vector mapping function. By jointly modeling multi-frequency features, the attention calculation process takes into account both frequency semantic differences and spatial consistency constraints.

[0074] Meanwhile, to achieve effective aggregation of multi-frequency feature information, a fusion value vector is constructed as the carrier of attention weighting, and its definition is:

[0075]

[0076] in, Represents the fused value vector. This represents the fusion value vector mapping function, which is used to map features from different frequency subbands to a unified feature representation space to ensure numerical consistency in subsequent weighted fusion operations.

[0077] After obtaining the frequency domain query vector, spatial domain key vector, and fused value vector, weights are assigned to features of different frequency sub-bands using an attention calculation mechanism. For the first... The attention weight coefficients for each frequency sub-band feature are defined as follows:

[0078]

[0079] in, Indicates the first Attention weights corresponding to each frequency sub-band feature This represents the vector dot product operation. This represents the scaling factor, used to mitigate numerical instability in high-dimensional spaces. Normalization ensures that the sum of attention weights for each frequency subband is 1, thus forming an interpretable frequency contribution distribution.

[0080] After obtaining the attention weights corresponding to each frequency sub-band, a weighted fusion operation is performed on the multi-frequency sub-band features to construct a multi-frequency fusion feature representation, the expression of which is:

[0081]

[0082] in, This indicates a multi-frequency fusion feature, in which the fusion process gives higher weight to frequency components that contribute more to the target task, while redundant or noisy frequency components are effectively suppressed.

[0083] S104 performs dynamic quantization and encoding processing on the multi-frequency fusion features to generate the corresponding multi-frequency feature representation.

[0084] In some embodiments, dynamic quantization and encoding processing is performed on the multi-frequency fusion features to generate corresponding multi-frequency feature representations, including: performing statistical characteristic analysis on the multi-frequency fusion features to obtain statistical distribution information corresponding to the multi-frequency fusion features; calculating the frequency importance weight corresponding to each frequency sub-band in the multi-frequency fusion features, and determining the quantization step size corresponding to each frequency sub-band based on the statistical distribution information and the frequency importance weight; performing quantization mapping operation on the multi-frequency fusion features based on the quantization step size to obtain the corresponding quantization result; and encoding the quantization result based on a bit allocation strategy to generate the corresponding multi-frequency feature representations.

[0085] As an example, dynamic quantization and encoding are performed on multi-frequency feature representations to reduce the storage overhead and computational complexity of feature representations while ensuring the fidelity of key semantic information and structural features.

[0086] Specifically, firstly, let the first part be the number of frequencies obtained after reversible frequency domain decomposition or multi-frequency domain attention fusion. Multi-frequency features are represented as

[0087]

[0088] in, , and These represent the low-frequency subband characteristics, mid-frequency subband characteristics, and high-frequency subband characteristics, respectively.

[0089] The characteristics of each frequency sub-band are first analyzed using statistical properties to obtain numerical distribution characteristics. For any frequency sub-band... Calculate its mean and variance:

[0090]

[0091] in, This represents the mean value corresponding to the frequency sub-band characteristics. This represents the variance corresponding to the frequency sub-band characteristics. This represents the total number of characteristic elements in this frequency sub-band. Statistical information is used to characterize the differences in numerical dynamic range and energy distribution among different frequency sub-bands.

[0092] Frequency importance weights are introduced to modulate the quantization strategy. These weights can be directly derived from the frequency attention coefficients obtained in the multi-frequency domain attention fusion step, or obtained through an independent frequency evaluation function. Let the... The weights of each frequency sub-band are And satisfy:

[0093]

[0094] Among them, the larger ones This indicates that the frequency subband has a higher semantic contribution to the target task.

[0095] Based on statistical distribution information and frequency importance weights, the quantization step size of each frequency sub-band is adaptively determined. The quantization step size of each frequency sub-band is defined as:

[0096]

[0097] in, Indicates the first The quantization step size of each frequency sub-band This is a global quantization adjustment coefficient used to control the overall compressive strength. To prevent the use of tiny constants with zero denominators, this method employs a smaller quantization step size for frequency subbands with higher semantic weights or greater statistical fluctuations, thereby preserving more key feature information.

[0098] Quantization mapping is applied to features in each frequency sub-band. For any frequency sub-band feature element... The quantitative results are as follows:

[0099]

[0100] The corresponding inverse quantization reconstruction form is

[0101]

[0102] In this way, the quantization process can be guaranteed to have reversible or nearly reversible characteristics within a controllable error range.

[0103] To further reduce storage and transmission overhead, a bit allocation strategy is introduced to encode the quantization results. Based on frequency importance weights and system resource constraints, different coding bit widths are allocated to each frequency sub-band. Let the total available bit budget of the system be... Then the first The bit allocation strategy for each frequency sub-band is as follows:

[0104]

[0105] in Indicates assignment to the first The number of coded bits for each frequency subband allows for higher coding accuracy in low-frequency and key semantic frequency subbands, while high-frequency noise or redundant frequency subbands are represented by more compact coding.

[0106] The multi-frequency features after quantization and encoding are represented as follows:

[0107]

[0108] in This represents an encoding operator used to encapsulate quantization results into a data format that is efficient for storage or transmission.

[0109] S105. Construct a multi-scale frequency domain pyramid structure to perform hierarchical modeling and cross-scale collaborative fusion of multi-frequency feature representations, so as to generate multi-scale frequency domain collaborative feature representations.

[0110] In some embodiments, a multi-scale frequency domain pyramid structure is constructed to perform hierarchical modeling and cross-scale collaborative fusion of multi-frequency feature representations to generate multi-scale frequency domain collaborative feature representations. This includes: generating multiple frequency domain feature representations at different scale levels in the frequency domain space based on scale transformation operators to form multi-scale frequency domain features; and performing cross-scale mapping and fusion of multi-scale frequency domain features based on cross-scale mapping operators to generate multi-scale frequency domain collaborative feature representations.

[0111] As an example, hierarchical modeling and cross-scale collaborative fusion of features at different scale levels in the frequency domain space are performed. By constructing a multi-scale frequency domain pyramid structure, a unified expression of unstructured data in terms of global structure, local details, and cross-scale semantic relationships can be achieved.

[0112] Specifically, let the frequency domain features obtained after frequency domain decomposition, multi-frequency domain attention fusion, and adaptive quantization coding be represented as follows:

[0113]

[0114] in, , This represents the frequency domain feature resolution at the current scale. Indicates the feature channel dimension.

[0115] Constructing a multi-scale frequency domain pyramid structure using a scaling operator Multiple frequency domain feature representations at different scale levels are generated in the frequency domain space, forming a scale set. The scales, from high to low, correspond to feature representations from fine to coarse. The frequency domain features at each scale level are represented as follows:

[0116]

[0117] The scaling operator can be implemented through frequency domain downsampling, bandpass filtering, spectral remapping, or a combination thereof, so that larger scales correspond to global structural information dominated by low frequencies, while smaller scales correspond to local detail information dominated by high frequencies.

[0118] The multi-scale frequency domain pyramid is represented as follows:

[0119] Adjacent scales satisfy a progressive relationship between spatial resolution and frequency coverage. To achieve information synergy between scales, a step-by-step fusion and information transfer operation is performed on multi-scale frequency domain features. Let the... Layer and First The frequency domain characteristics of the layers are respectively and Through cross-scale mapping operators Mapping low-scale features to a high-scale space:

[0120]

[0121] in, It can represent frequency domain interpolation, spectral alignment, or learnable linear mapping functions.

[0122] The mapped features are fused with the current-scale frequency domain features, and the fusion relationship is as follows:

[0123]

[0124] in, This represents the scale-frequency domain features after fusion. The fusion operator can employ weighted summation, linear transformation after concatenation, or attention-weighted fusion methods to achieve collaborative modeling of cross-scale frequency domain information. The hierarchical fusion mechanism gradually injects low-scale global frequency semantics into high-scale features, enabling high-resolution frequency domain representations to retain detailed information while possessing stronger global consistency.

[0125] At each scale level of the multi-scale frequency domain pyramid, an adaptive quantization coding mechanism is introduced to perform compressed representation of the fused frequency domain features. Let the... The frequency domain features after layer fusion are The quantization encoding process is then represented as:

[0126]

[0127] in, Representation and Scale An appropriate quantization operator is used to dynamically adjust the quantization precision based on the frequency distribution characteristics and semantic importance at that scale. Different quantization strategies are adopted for different scale levels, enabling lower-scale global features to achieve higher structural fidelity, while higher-scale local features achieve efficient compression while preserving key details.

[0128] After completing the hierarchical fusion and quantization encoding of multi-scale frequency domain features, cross-scale and cross-frequency domain joint modeling results are output through inverse-scale mapping and frequency domain reconstruction operations. The final multi-scale frequency domain collaborative feature representation is as follows:

[0129]

[0130] in, This represents a multi-scale frequency domain collaborative feature representation. This represents the inverse quantization and frequency domain reconstruction operator for the corresponding scale, used to map frequency domain features at different scales back to a unified feature representation space.

[0131] S106 performs an inverse frequency domain transformation operation on the multi-scale frequency domain collaborative feature representation to achieve high-fidelity reconstruction of multi-source unstructured data and obtain the final preprocessing result corresponding to the multi-source unstructured data.

[0132] In some embodiments, an inverse frequency domain transformation operation is performed on the multi-scale frequency domain collaborative feature representation to achieve high-fidelity reconstruction of multi-source unstructured data and obtain the final preprocessing result corresponding to the multi-source unstructured data. This includes: decoding and inverse quantizing the multi-scale frequency domain collaborative feature representation based on decoding and inverse quantization operators to obtain the recovered frequency domain features at each scale level; performing inverse scale mapping and fusion processing on the recovered frequency domain features at each scale level based on inverse scale mapping operators to obtain the aligned frequency domain feature representation; performing cross-scale aggregation on the aligned frequency domain feature representation to generate the aggregation result; and performing an inverse frequency domain transformation operation on the aggregation result to generate the final preprocessing result.

[0133] As an example, an inverse frequency domain transformation operation is performed on the frequency domain features after multi-frequency domain attention fusion, multi-scale frequency domain pyramid collaborative modeling, and adaptive quantization coding, thereby completing the high-fidelity reconstruction of unstructured data or forming an intermediate feature representation that can be directly used for subsequent computer data processing tasks, and the result is output to the storage unit or computing module of the computer system.

[0134] Specifically, firstly, let the set of multi-scale frequency domain collaborative feature representations obtained after multi-frequency fusion and quantization coding be represented as:

[0135]

[0136] in, This represents a set of multi-scale frequency domain collaborative feature representations. Indicates the first Multi-scale frequency domain collaborative feature representation at each scale level.

[0137] First, decoding and inverse quantization operations are performed on the encoded frequency domain features using the decoding operator. With inverse quantization operator Recovering the frequency domain feature representations at each scale level:

[0138]

[0139] in, Indicates the number after recovery Frequency domain features at each scale level Indicates the inverse quantization operator. This represents the decoding operator. It should be noted that the dequantization operation, based on the quantization step size and bit allocation strategy determined in the adaptive quantization coding step, keeps the reconstruction error controllable across frequency subbands and scale levels.

[0140] Subsequently, inverse scaling and fusion processing are performed on the frequency domain features at different scale levels. This is achieved through the inverse scaling operator. The frequency domain features at each scale level are uniformly mapped to the target scale space to obtain the scale-aligned frequency domain feature representation:

[0141]

[0142] in, Indicates the aligned first Frequency domain features at each scale level This represents the inverse scaling operator.

[0143] Based on this, frequency domain features at all scales are aggregated across scales, and the aggregation result is expressed as follows:

[0144]

[0145] in, Indicates the aggregation result. Indicates the first The fusion weights corresponding to each scale are used to adjust the contribution ratio of scale features in the final reconstruction, and satisfy the following conditions: .

[0146] After completing the cross-scale frequency domain feature aggregation, an inverse frequency domain transformation operation is performed on the aggregated frequency domain features using the inverse frequency domain transformation operator. Mapping the frequency domain representation back to the spatial domain or the original data representation space, the reconstruction process is expressed as follows:

[0147]

[0148] in, This indicates the final preprocessing result. This represents the inverse frequency domain transform operator. Alternatively, it can be used as an intermediate feature representation for downstream computational tasks. The inverse frequency domain transform operator and the aforementioned frequency domain transform operator form a one-to-one reversible mapping relationship, thereby ensuring that the overall processing flow theoretically possesses high-fidelity reconstruction capabilities.

[0149] It should be noted that in practical applications, when the target task is data reconstruction or enhanced display, data reconstruction... It can be directly output and stored in the storage unit of the computer system; when the target task is feature analysis, recognition, prediction or other intelligent computing tasks, the reconstruction result or its intermediate frequency domain features can be directly passed to the downstream computing module as input.

[0150] Through the above frequency domain reconstruction and output steps, stable reconstruction and efficient output of multi-source unstructured data after multi-frequency domain and multi-scale processing can be achieved. While ensuring structural consistency and fidelity of key details, it effectively suppresses quantization errors and error accumulation caused by multiple fusions, providing a reliable data foundation and engineering feasibility support for unstructured data processing tasks in complex computer systems.

[0151] In summary, the data preprocessing method based on a multi-frequency domain attention mechanism according to embodiments of the present invention first acquires multi-source unstructured data and preprocesses it to generate standardized data; then, it performs frequency domain transformation on the standardized data to generate a frequency domain feature tensor, and performs reversible frequency domain decomposition on the frequency domain feature tensor to generate corresponding multi-frequency sub-band features; then, it performs adaptive modeling and feature fusion on the multi-frequency sub-band features to generate multi-frequency fused features; next, it performs dynamic quantization and encoding processing on the multi-frequency fused features to generate corresponding multi-frequency feature representations; then, it constructs a multi-scale frequency domain pyramid structure to perform hierarchical modeling and cross-scale collaborative fusion on the multi-frequency feature representations to generate multi-scale frequency domain collaborative feature representations; finally, it performs an inverse frequency domain transformation operation on the multi-scale frequency domain collaborative feature representations to achieve high-fidelity reconstruction of the multi-source unstructured data, obtaining the final preprocessing result corresponding to the multi-source unstructured data; thereby effectively reducing noise and redundant features in the final preprocessed data.

[0152] Secondly, embodiments of the present invention provide a computer-readable storage medium storing a data preprocessing program based on a multi-frequency domain attention mechanism, which, when executed by a processor, implements the data preprocessing method based on the multi-frequency domain attention mechanism as described above.

[0153] Thirdly, embodiments of the present invention propose a data preprocessing system based on a multi-frequency domain attention mechanism, such as... Figure 2 As shown, the data preprocessing system based on the multi-frequency domain attention mechanism includes: a multi-source unstructured data acquisition and preprocessing module 10, a reversible frequency domain decomposition and multi-frequency feature construction module 20, a multi-frequency domain attention fusion module 30, an adaptive quantization coding module 40, a multi-scale frequency domain pyramid collaborative modeling module 50, and a frequency domain reconstruction and output module 60.

[0154] The multi-source unstructured data acquisition and preprocessing module 10 is used to acquire multi-source unstructured data and preprocess the multi-source unstructured data to generate standardized data.

[0155] The reversible frequency domain decomposition and multi-frequency feature construction module 20 is used to perform frequency domain transformation on standardized data to generate a frequency domain feature tensor, and to perform reversible frequency domain decomposition on the frequency domain feature tensor to generate corresponding multi-frequency sub-band features.

[0156] The multi-frequency domain attention fusion module 30 is used to adaptively model and fuse multi-frequency sub-band features to generate multi-frequency fused features;

[0157] The adaptive quantization and encoding module 40 is used to perform dynamic quantization and encoding processing on the multi-frequency fusion features to generate the corresponding multi-frequency feature representation;

[0158] The multi-scale frequency domain pyramid collaborative modeling module 50 is used to construct a multi-scale frequency domain pyramid structure to perform hierarchical modeling and cross-scale collaborative fusion of multi-frequency feature representations to generate multi-scale frequency domain collaborative feature representations.

[0159] The frequency domain reconstruction and output module 60 is used to perform inverse frequency domain transformation on the multi-scale frequency domain collaborative feature representation in order to achieve high-fidelity reconstruction of multi-source unstructured data and obtain the final preprocessing result corresponding to the multi-source unstructured data.

[0160] In some embodiments, the multi-source unstructured data includes high-resolution data and low-resolution multidimensional data. The multi-source unstructured data acquisition and preprocessing module 10 is also used to perform format conversion on the multi-source unstructured data to generate corresponding floating-point data; normalize the format-converted data; and denoise the normalized data to generate standardized data.

[0161] It should be noted that the above description of the data preprocessing method based on the multi-frequency domain attention mechanism also applies to this data preprocessing system based on the multi-frequency domain attention mechanism, and will not be repeated here.

[0162] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0163] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0164] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0165] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and are not intended to 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 limitations on this invention.

[0166] 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 at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0167] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0168] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0169] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A data preprocessing method based on a multi-frequency domain attention mechanism, characterized in that, Includes the following steps: Acquire multi-source unstructured data and preprocess the multi-source unstructured data to generate standardized data; The standardized data is subjected to frequency domain transformation to generate a frequency domain feature tensor, and the frequency domain feature tensor is subjected to reversible frequency domain decomposition to generate corresponding multi-frequency sub-band features. Adaptive modeling and feature fusion are performed on the multi-frequency sub-band features to generate multi-frequency fused features; Dynamic quantization and encoding processing are performed on the multi-frequency fusion features to generate corresponding multi-frequency feature representations; A multi-scale frequency domain pyramid structure is constructed to perform hierarchical modeling and cross-scale collaborative fusion of the multi-frequency feature representations, so as to generate multi-scale frequency domain collaborative feature representations. An inverse frequency domain transformation operation is performed on the multi-scale frequency domain collaborative feature representation to achieve high-fidelity reconstruction of the multi-source unstructured data, thereby obtaining the final preprocessing result corresponding to the multi-source unstructured data. The process of adaptively modeling and fusing the multi-frequency sub-band features to generate multi-frequency fused features includes: A frequency-aware query vector is constructed based on the multi-frequency sub-band features, and the multi-frequency sub-band features are mapped to the attention query space based on the frequency-aware query vector. Based on the multi-frequency subband features, a spatial domain key vector is constructed to describe the overall structure and contextual relationship of the data in the spatial dimension; A fusion value vector is constructed based on the multi-frequency sub-band features; The attention weights corresponding to each multi-frequency sub-band feature are calculated based on the query vector, the spatial domain key vector, and the fusion value vector. The multi-frequency sub-band features are then weighted and fused based on the attention weights to generate multi-frequency fusion features. The process of performing dynamic quantization and encoding on the multi-frequency fusion features to generate corresponding multi-frequency feature representations includes: Statistical characteristic analysis is performed on the multi-frequency fusion features to obtain the statistical distribution information corresponding to the multi-frequency fusion features; Calculate the frequency importance weight corresponding to each frequency sub-band in the multi-frequency fusion feature, and determine the quantization step size corresponding to each frequency sub-band based on the statistical distribution information and the frequency importance weight; Based on the quantization step size, the multi-frequency fusion feature is quantized and mapped to obtain the corresponding quantization result; The quantization results are encoded based on a bit allocation strategy to generate corresponding multi-frequency feature representations; Specifically, a multi-scale frequency domain pyramid structure is constructed to perform hierarchical modeling and cross-scale collaborative fusion of the multi-frequency feature representations, thereby generating multi-scale frequency domain collaborative feature representations, including: Based on the scaling transformation operator, multiple frequency domain feature representations at different scale levels are generated in the frequency domain space to form multi-scale frequency domain features; The multi-scale frequency domain features are mapped and fused across scales based on the cross-scale mapping operator to generate a multi-scale frequency domain collaborative feature representation. Specifically, an inverse frequency domain transformation operation is performed on the multi-scale frequency domain collaborative feature representation to achieve high-fidelity reconstruction of the multi-source unstructured data, obtaining the final preprocessing result corresponding to the multi-source unstructured data, including: The multi-scale frequency domain collaborative feature representation is decoded and dequantized based on the decoding operator and the dequantization operator to obtain the recovered frequency domain features at each scale level. The restored frequency domain features at each scale level are subjected to inverse scale mapping and fusion processing based on the inverse scale mapping operator to obtain the aligned frequency domain feature representation. The aligned frequency domain feature representations are aggregated across scales to generate an aggregation result; The aggregation result is subjected to an inverse frequency domain transformation operation to generate the final preprocessed result.

2. The data preprocessing method based on a multi-frequency domain attention mechanism as described in claim 1, characterized in that, The multi-source unstructured data includes high-resolution data and low-resolution multidimensional data. The multi-source unstructured data undergoes preprocessing to generate standardized data, including: The multi-source unstructured data is format-converted to generate corresponding floating-point data; The converted data is normalized, and the normalized data is then denoised to generate standardized data.

3. The data preprocessing method based on a multi-frequency domain attention mechanism as described in claim 1, characterized in that, The standardized data is subjected to frequency domain transformation to generate a frequency domain feature tensor, and the frequency domain feature tensor is then subjected to invertible frequency domain decomposition to generate corresponding multi-frequency sub-band features, including: A frequency domain transformation operator is applied to the standardized data to map the standardized data to the frequency domain space, generating a frequency domain feature tensor; Perform invertible frequency domain decomposition on the frequency domain feature tensor to divide the frequency domain space into multiple sub-band regions, and calculate the corresponding multi-frequency sub-band features based on the frequency mask function; The multi-frequency sub-band features include: high-frequency sub-band features, mid-frequency sub-band features, and low-frequency sub-band features.

4. A computer-readable storage medium, characterized in that, It stores a data preprocessing program based on a multi-frequency domain attention mechanism, which, when executed by the processor, implements the data preprocessing method based on a multi-frequency domain attention mechanism as described in any one of claims 1-3.

5. A data preprocessing system based on a multi-frequency domain attention mechanism, characterized in that, include: A multi-source unstructured data acquisition and preprocessing module is used to acquire multi-source unstructured data and preprocess the multi-source unstructured data to generate standardized data. A reversible frequency domain decomposition and multi-frequency feature construction module is used to perform frequency domain transformation on the standardized data to generate a frequency domain feature tensor, and to perform reversible frequency domain decomposition on the frequency domain feature tensor to generate corresponding multi-frequency sub-band features. A multi-frequency domain attention fusion module is used to adaptively model and fuse the multi-frequency sub-band features to generate multi-frequency fused features. An adaptive quantization and encoding module is used to perform dynamic quantization and encoding processing on the multi-frequency fusion features to generate corresponding multi-frequency feature representations; A multi-scale frequency domain pyramid collaborative modeling module is used to construct a multi-scale frequency domain pyramid structure to perform hierarchical modeling and cross-scale collaborative fusion of the multi-frequency feature representation to generate a multi-scale frequency domain collaborative feature representation. The frequency domain reconstruction and output module is used to perform inverse frequency domain transformation on the multi-scale frequency domain collaborative feature representation to achieve high-fidelity reconstruction of the multi-source unstructured data and obtain the final preprocessing result corresponding to the multi-source unstructured data. The process of adaptively modeling and fusing the multi-frequency sub-band features to generate multi-frequency fused features includes: A frequency-aware query vector is constructed based on the multi-frequency sub-band features, and the multi-frequency sub-band features are mapped to the attention query space based on the frequency-aware query vector. Based on the multi-frequency subband features, a spatial domain key vector is constructed to describe the overall structure and contextual relationship of the data in the spatial dimension; A fusion value vector is constructed based on the multi-frequency sub-band features; The attention weights corresponding to each multi-frequency sub-band feature are calculated based on the query vector, the spatial domain key vector, and the fusion value vector. The multi-frequency sub-band features are then weighted and fused based on the attention weights to generate multi-frequency fusion features. The process of performing dynamic quantization and encoding on the multi-frequency fusion features to generate corresponding multi-frequency feature representations includes: Statistical characteristic analysis is performed on the multi-frequency fusion features to obtain the statistical distribution information corresponding to the multi-frequency fusion features; Calculate the frequency importance weight corresponding to each frequency sub-band in the multi-frequency fusion feature, and determine the quantization step size corresponding to each frequency sub-band based on the statistical distribution information and the frequency importance weight; Based on the quantization step size, the multi-frequency fusion feature is quantized and mapped to obtain the corresponding quantization result; The quantization results are encoded based on a bit allocation strategy to generate corresponding multi-frequency feature representations; Specifically, a multi-scale frequency domain pyramid structure is constructed to perform hierarchical modeling and cross-scale collaborative fusion of the multi-frequency feature representations, thereby generating multi-scale frequency domain collaborative feature representations, including: Based on the scaling transformation operator, multiple frequency domain feature representations at different scale levels are generated in the frequency domain space to form multi-scale frequency domain features; The multi-scale frequency domain features are mapped and fused across scales based on the cross-scale mapping operator to generate a multi-scale frequency domain collaborative feature representation. Specifically, an inverse frequency domain transformation operation is performed on the multi-scale frequency domain collaborative feature representation to achieve high-fidelity reconstruction of the multi-source unstructured data, obtaining the final preprocessing result corresponding to the multi-source unstructured data, including: The multi-scale frequency domain collaborative feature representation is decoded and dequantized based on the decoding operator and the dequantization operator to obtain the recovered frequency domain features at each scale level. The restored frequency domain features at each scale level are subjected to inverse scale mapping and fusion processing based on the inverse scale mapping operator to obtain the aligned frequency domain feature representation. The aligned frequency domain feature representations are aggregated across scales to generate an aggregation result; The aggregation result is subjected to an inverse frequency domain transformation operation to generate the final preprocessed result.

6. The data preprocessing system based on a multi-frequency domain attention mechanism as described in claim 5, characterized in that, The multi-source unstructured data includes high-resolution data and low-resolution multidimensional data. The multi-source unstructured data acquisition and preprocessing module is also used to convert the format of the multi-source unstructured data to generate corresponding floating-point data; normalize the format-converted data; and denoise the normalized data to generate standardized data.

Images