Method and device for determining content of starch molecules in fermented grains, equipment and storage medium

Abstract

The application discloses a kind of determination methods, device, equipment and storage medium of wine lees starch molecule content, relate to brewing technology field.The method comprises the following steps: obtaining the multimodal spectral data of the wine lees to be measured;Feature extraction is carried out to the multimodal spectral data, and the multimodal data features extracted are dimensionally unified, to obtain the multimodal data features of the same dimension;Attention weight processing is carried out to the multimodal data features of the same dimension, to obtain the weighted multimodal data features;According to the weighted multimodal data features, multi-scale convolution processing is carried out, and the starch molecule content of the wine lees to be measured is determined according to the obtained multi-scale features.The application does not need artificial detection, realizes the quick, automatic determination of wine lees starch content, and simultaneously, due to the comprehensive effect of dimension unification, attention weighting and multi-scale convolution, the accuracy of wine lees starch content detection is effectively improved.

Description

Technical Field

[0001] This application belongs to the field of brewing technology, specifically relating to a method, apparatus, equipment, and storage medium for determining the starch molecule content of brewing mash. Background Technology

[0002] Fermented mash is an important intermediate product in the fermentation process of alcoholic beverages, and its physicochemical indicators are directly related to the metabolic activities of microorganisms and the quality and yield of the final alcoholic beverage. In existing alcoholic beverage production processes, monitoring the starch content of fermented mash is a key step in regulating the fermentation process. The relevant technology involves manually sampling the fermented mash and transporting the samples to a laboratory for routine biochemical analysis using chemical reagents, such as titration or colorimetry, to determine the starch content.

[0003] However, the manual testing process for starch content in fermented mash in related technologies is not only labor-intensive and time-consuming, but also prone to inaccurate determination of starch content due to non-standard manual sampling operations. Summary of the Invention

[0004] This application provides a method, apparatus, equipment, and storage medium for determining the starch content of fermented grains, which solves the technical problems of traditional manual detection of starch content in fermented grains being time-consuming, labor-intensive, and inaccurate. This application provides the following technical solution: In a first aspect, embodiments of this application provide a method for determining the starch molecule content of fermented mash, comprising: extracting features from the multimodal spectral data, and unifying the dimensions of the extracted multimodal data features to obtain multimodal data features of the same dimension; performing attention weight processing on the multimodal data features of the same dimension to obtain weighted multimodal data features; performing multi-scale convolution processing on the weighted multimodal data features, and determining the starch molecule content of the fermented mash to be tested based on the obtained multi-scale features.

[0005] Optionally, feature extraction of the multimodal spectral data includes: convolving the multimodal spectral data to obtain initial multimodal data features; encoding the initial multimodal data features to obtain encoded features corresponding to the initial multimodal data features; generating multimodal data features based on the encoded features and the initial multimodal data features; and unifying the dimensions of the extracted multimodal data features to obtain multimodal data features of the same dimension, including: aligning the lengths of features with different sequence lengths in the multimodal data features to obtain multimodal data features of the same dimension.

[0006] Optionally, attention weighting is applied to the multimodal data features of the same dimension to obtain weighted multimodal data features, including: mapping each modal data feature in the multimodal data features of the same dimension to a first tensor, a second tensor, and a third tensor, respectively; calculating the similarity between the first tensor and the second tensor corresponding to each modal data feature, and normalizing the calculated similarity matrix to obtain an attention weight matrix; weighting the third tensor based on the attention weight matrix, and using the weighted modal data features obtained after the processing as the weighted multimodal data features.

[0007] Optionally, performing multi-scale convolution processing on the weighted multimodal data features and determining the starch molecule content of the mash to be tested based on the obtained multi-scale features includes: performing multi-scale convolution processing on the weighted multimodal data features to obtain multi-scale features; fusing the multi-scale features and the weighted multimodal data features to obtain fused modal data features; and determining the starch molecule content of the mash to be tested based on the fused modal data features.

[0008] Optionally, the multi-scale features and the weighted multimodal data features are fused to obtain fused modal data features, including: concatenating the features in the multi-scale features to obtain concatenated features; mapping the channel dimensions of the weighted multimodal data features to the same channel dimensions as the concatenated features to obtain mapped features; and superimposing the mapped features and the concatenated features to obtain fused modal data features.

[0009] Optionally, the weighted multimodal data features are subjected to multi-scale convolution processing to obtain multi-scale features, including: performing convolution processing on the weighted multimodal data features through convolution branches with different dilation rates, and using the different scale features obtained as multi-scale features.

[0010] Optionally, the convolutional branches with different dilation rates belong to the trained analysis network, which is obtained by preprocessing the multimodal spectral data of the sample mash to obtain training data, and preprocessing the starch content of the sample mash to obtain data labels; wherein, the multimodal spectral data includes near-infrared spectral data and Raman spectral data; the analysis network to be trained is trained based on the training data and the data labels to obtain the trained analysis network; wherein, the basic prototype of the analysis network is the U-net model.

[0011] Secondly, embodiments of this application provide an apparatus for determining the starch molecule content of fermented mash, comprising: an acquisition module for acquiring multimodal spectral data of the fermented mash to be tested; a feature extraction module for extracting features from the multimodal spectral data and unifying the dimensions of the extracted multimodal data features to obtain multimodal data features of the same dimension; a weighting processing module for performing attention weighting processing on the multimodal data features of the same dimension to obtain weighted multimodal data features; and a content determination module for performing multi-scale convolution processing based on the weighted multimodal data features and determining the starch molecule content of the fermented mash to be tested based on the obtained multi-scale features.

[0012] Thirdly, embodiments of this application provide an electronic device, which includes a processor; a memory and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the determination method as described in the first aspect.

[0013] Fourthly, embodiments of this application provide a storage medium storing a program or instructions, which, when executed by a processor, implement the determination method described in the first aspect.

[0014] In summary, one or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: This application enables automated acquisition of multimodal spectral data of the mash to be tested without manual sampling, saving manpower and time. By extracting features and unifying dimensions of the multimodal spectral data, the spectral data of different modes are aligned in the feature space, eliminating data barriers caused by differences in wavelength range. By applying attention weights to multimodal data features of the same dimension, weights can be adaptively assigned to different modal data features, highlighting key features and suppressing redundant information. By performing multi-scale convolution processing on the weighted multimodal data features, features can be extracted from different scales, fully mining the deep information in the spectral data. The starch content can be determined based on the obtained multi-scale features without manual detection, realizing rapid and automated determination of starch content in mash. At the same time, due to the combined effect of dimension unification, attention weighting, and multi-scale convolution, the accuracy of starch content detection in mash is effectively improved.

[0015] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more easily understood, specific embodiments of this application are given below. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0017] Figure 1 This is a flowchart illustrating an exemplary embodiment of the present application of a method for determining the starch molecule content in fermented mash.

[0018] Figure 2 Based on Figure 1 The exemplary embodiment shown illustrates a flowchart of another method for determining the starch molecule content of fermented mash.

[0019] Figure 3 This is a schematic diagram illustrating the scaling dot product attention mechanism and the multi-head attention mechanism in an exemplary embodiment of this application.

[0020] Figure 4 This is a schematic diagram of the structure of an analysis network shown in an exemplary embodiment of this application.

[0021] Figure 5 yes Figure 4 The diagram shows the structure of the multi-scale parallel convolution module in the analyzed network structure.

[0022] Figure 6 This is a schematic diagram of the structure of a device for determining the starch molecule content of fermented mash, as shown in an exemplary embodiment of this application.

[0023] Figure 7 This is a schematic diagram of the structure of a computer system for an electronic device illustrated in an exemplary embodiment of this application. Detailed Implementation

[0024] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0025] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0026] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0027] In this application, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0028] Fermented mash is an important intermediate product in the fermentation process of alcoholic beverages. Its starch content and other parameters directly affect the growth, reproduction, and metabolism of microorganisms, thus determining the yield and quality of alcohol. Related technologies typically employ manual sampling combined with conventional chemical testing methods to determine the starch content in the fermented mash. However, this manual testing method is not only labor-intensive and time-consuming, but also prone to significant errors in determining the starch content due to the lack of standardization in manual sampling procedures, making it difficult to meet the real-time and accuracy requirements of modern brewing production.

[0029] Therefore, one aspect of this application provides a method for determining the starch content of fermented mash, thereby solving the technical problems of traditional manual detection of starch content in fermented mash being time-consuming, labor-intensive, and inaccurate. Please refer to [link / reference] for details. Figure 1 , Figure 1 This is a schematic flowchart illustrating a method for determining the starch molecule content of fermented mash according to an exemplary embodiment of this application. The method includes at least steps S110 to S140, which are detailed below: S110 acquires multimodal spectral data of the mash to be tested.

[0030] Multimodal spectral data refers to the collection of data containing multiple spectral types obtained by scanning the mash under test using spectral acquisition equipment. For example, multimodal spectral data includes near-infrared spectral data and Raman spectral data. Near-infrared spectral data mainly reflects the overtone and combination frequency absorption information of hydrogen-containing groups (such as CH, OH, NH) in the mash, while Raman spectral data provides fingerprint information of molecular vibrational and rotational energy levels. The two are complementary in characterizing the structure of starch molecules.

[0031] S120 performs feature extraction on multimodal spectral data and unifies the dimensions of the extracted multimodal data features to obtain multimodal data features with the same dimensions.

[0032] Since spectral data of different modalities (such as near-infrared and Raman spectra) have different wavelength ranges and sequence lengths, it is difficult to directly perform joint analysis on both. Therefore, this embodiment extracts features from the spectral data of each modality and then unifies the dimensions of the extracted multimodal data features, so that the data features of different modalities have the same dimensions.

[0033] Feature extraction is the process of converting raw spectral data into high-dimensional feature vectors. For example, convolution processing is performed on multimodal spectral data to map single-channel spectral sequences into multi-channel hidden layer features, and the number of output channels is set as the hidden layer dimension d_model. Subsequently, a position encoder is introduced, which generates fixed-position codes through sine and cosine functions and injects them into the feature sequence to preserve the order information of the spectral bands. The feature tensors of different modes (such as near-infrared and Raman) are aligned at the sequence length level.

[0034] Dimensionality unification refers to the process of transforming feature tensors from different modalities with different sequence lengths into a uniform shape. Since near-infrared and Raman spectra differ in wavelength range and number of sampling points, their initial feature tensor shapes are (batch, d_model, length1) and (batch, d_model, length2), respectively, with length1 ≠ length2. Through interpolation or projection transformation, the sequence lengths of both are adjusted to a preset target length_target, thus obtaining multimodal data features with the same dimension and a shape of (batch, d_model, length_target).

[0035] S120 is illustrated as follows: The multimodal spectral data is convolved to obtain initial multimodal data features; the initial multimodal data features are encoded to obtain the encoded features corresponding to the initial multimodal data features; multimodal data features are generated based on the encoded features and the initial multimodal data features; features with different sequence lengths in the multimodal data features are length aligned to obtain multimodal data features of the same dimension.

[0036] Convolutional processing enables the transformation and preliminary extraction of feature dimensions from different modal spectral data. For example, the multimodal spectral data of the mash to be tested includes near-infrared and Raman spectra, whose original shape can be represented as (batch, 1, length), i.e., batch size, single channel, and sequence length. By using a 1×1 convolutional kernel to process the multimodal spectral data, the original spectral data features are transformed from a single channel to a multi-channel hidden layer representation without changing the input sequence length. For example, if the hidden layer dimension of the model is set to d_model, after this convolutional processing, the shape of the output vector becomes (batch, d_model, length), thus completing the dimensional adaptation from input features to model features. This 1×1 convolutional embedding method effectively eliminates feature barriers caused by differences in the original number of channels between different modalities.

[0037] The encoding process involves generating fixed-position codes and adding them to the initial multimodal data features. The positional codes are obtained using the following formula: ; ; Where pos represents the band position in the spectral sequence, and i represents the encoding dimension. This represents the hidden layer dimension. After encoding, the generated positional encoding (i.e., encoded features) is element-wise added to the convolutional features (i.e., initial multimodal data features) to obtain the multimodal data features, i.e., different modal data features. The significance of positional encoding is to assign a specific frequency sine / cosine wave factor to the feature vector of each band position to represent the sequential order of each band in the spectral sequence, preserving the intensity information of the original spectrum while embedding the relative positional relationship of the bands. For example, after convolutional processing of near-infrared spectral data and Raman spectral data, the resulting multimodal data features have shapes (batch, d_model, length1) and (batch, d_model, length2), respectively; where length1 and length2 may be unequal due to differences in the wavelength range and number of sampling points of the two spectra. Adding the encoded features to the initial multimodal data features constructs a unified feature representation space for data of different modalities, facilitating subsequent cross-modal alignment and interactive modeling.

[0038] To eliminate the feature dimension mismatch caused by wavelength range differences and enable parallel processing of spectral data from different modalities at the same scale, the sequence length of both is adjusted to a preset target length (length_target) by interpolating and expanding shorter sequences or by compressing and sampling longer sequences. For example, if the near-infrared spectral feature length is 150 and the Raman spectral feature length is 200, and the target length is set to 200, the near-infrared spectral feature sequence is extended to 200 points through linear interpolation, so that both data streams ultimately become multimodal data features of the same dimension with shape (batch, d_model, length_target).

[0039] S130 performs attention weighting on multimodal data features of the same dimension to obtain weighted multimodal data features.

[0040] Attention weighting is an adaptive process of calculating and assigning weights to features in each band to capture long-range dependencies within the spectral sequence. It assigns higher weights to key bands highly correlated with starch content while suppressing interference from noisy bands or redundant information. For example, in near-infrared spectroscopy, bands corresponding to the starch CH bond absorption peak are given higher attention weights, while in Raman spectroscopy, bands corresponding to characteristic peaks of the starch crystal structure are significantly enhanced. For instance, modal data features of the same dimension are mapped to query tensors, key tensors, and value tensors through three independent linear transformation layers; the dot product similarity between the query tensor and the key tensor is calculated and adjusted using a scaling factor. Normalization is performed, and then the attention weight matrix is ​​obtained by processing with the Softmax function. This matrix quantifies the correlation strength between features at any two positions in the sequence. The value tensor is weighted and summed based on the attention weight matrix, and the outputs of multiple attention heads are concatenated and fused to obtain the weighted multimodal data features.

[0041] For example, each modal data feature in the multimodal data features of the same dimension is mapped to a first tensor, a second tensor, and a third tensor, respectively. The similarity between the first tensor and the second tensor corresponding to each modal data feature is calculated, and the calculated similarity matrix is ​​normalized to obtain an attention weight matrix. The third tensor is weighted based on the attention weight matrix, and the weighted modal data features are used as the weighted multimodal data features. Here, the first tensor, the second tensor, and the third tensor correspond to the query, key, and value tensors in the attention mechanism, respectively; the first tensor is used to initiate the query request, the second tensor is used to provide the key-value information being queried, and the third tensor carries the actual feature content information.

[0042] Multimodal data features after dimensional unification in the preceding steps (e.g., batch × d shape) model A tensor of length × is projected through three independent linear transformation layers. These three linear transformation layers have learnable weight matrices, which are used to map the input features to different subspaces, thereby generating a first tensor, a second tensor, and a third tensor, namely the query tensor, the key tensor, and the value tensor. This mapping process aims to transform the raw spectral features into a representation suitable for calculating the correlation strength.

[0043] Similarity calculation typically employs dot product operations, with a scaling factor introduced to prevent numerical overflow. Specifically, the transposes of the first and second tensors are multiplied to obtain an initial similarity matrix reflecting the correlation strength between positions in each band. To stabilize the gradient and prevent the dot product from becoming too large when the feature dimension is high, thus avoiding saturation of the Softmax function, the similarity matrix is ​​divided by the square root of the feature dimension of the second tensor as a scaling factor. Subsequently, Softmax normalization is performed on the scaled matrix, ensuring that the sum of the elements in each row is 1, thus obtaining the attention weight matrix. Each element in this attention weight matrix represents the degree of attention or dependence of a feature at a certain position in the sequence on features at other positions. For example, during calculation, if two bands have a strong chemical coupling relationship, the dot product of their corresponding first and second tensors will be large, resulting in a higher weight value after normalization; conversely, the weight values ​​of unrelated bands tend to be close to zero. The resulting attention weight matrix can adaptively identify and quantify long-range dependencies and local correlations between different bands within a spectral sequence, effectively suppressing interference from redundant noise bands and highlighting key feature regions sensitive to starch molecule content.

[0044] The weighted processing involves performing matrix multiplication between the attention weight matrix obtained above and the generated third tensor. Through this operation, the feature vectors at each position in the third tensor are recombinated and weighted according to the corresponding weight coefficients in the attention weight matrix. Feature regions with higher weights are enhanced in the output, while regions with lower weights are suppressed, thus achieving selective feature focusing. The final weighted multimodal data features not only retain the rich information of the original spectrum but also incorporate the intramodal global correlation structure mined through the attention mechanism. For example, after weighted processing, the starch feature peak information originally scattered at different wavelengths is aggregated and enhanced, forming a feature vector with stronger representational power. This vector is directly used as the input for subsequent multi-scale convolution processing. In this way, the signal-to-noise ratio of the feature sequence is significantly improved, enabling more accurate capture of features highly correlated with the starch content of the mash.

[0045] For example, when the input features are fused features of near-infrared and Raman spectra aligned, they are multiplied by the weight matrix W through three linear transformation layers. Q W K and W V This yields the corresponding Q, K, and V tensors. This mapping method decomposes high-dimensional spectral data into a query and key space for similarity calculation, and a value space for final feature aggregation, laying a data foundation for subsequent accurate extraction of key band features within modes. Furthermore, the query tensor, key tensor, and value tensor are further split dimensionally by the number of heads (num_heads), such as... Figure 3 As shown, a multi-head feature structure is formed to achieve parallel feature weight learning across multiple subspaces. Figure 3 In this context, Scaled Dot-Product Attention represents scaled dot-product attention; MatMul represents matrix multiplication; Scale represents scaling; Mask (opt.) represents masking (optional); SoftMax represents SoftMax normalization; Q represents the query tensor; K represents the key tensor; V represents the value tensor; Multi-Head Attention represents multi-head attention; Linear represents a linear transformation layer; Scaled Dot-Product Attention represents scaled dot-product attention; and Concat represents concatenation. The relevant calculation formulas involved in the attention mechanism are shown below: Scaling dot product attention formula: ; Multi-head attention formula: ; ; Where Q represents the query tensor; K represents the key tensor; and V represents the value tensor. Represents the normalization function; This represents the dot product of the query tensor and the key tensor. This represents the scaling factor, i.e., the dimension of the key tensor, which is usually equal to the size of the last dimension of K, and is used to prevent the gradient from vanishing due to excessively large dot product values. This indicates a splicing operation; h represents the number of attention heads; This represents the transformation matrix of the output; This represents the output of the i-th head; This represents the Q-weight matrix of the i-th head; This represents the K-weight matrix of the i-th head; V represents the weight matrix of the i-th head.

[0046] Within each attention head, a similarity matrix is ​​obtained by performing a dot product operation between the query tensor and the key tensor. This matrix quantifies the association strength between features at any two positions in the sequence. To avoid numerical overflow caused by feature dimensionality, a factor is used. The similarity matrix is ​​scaled and then normalized using the Softmax function to obtain the final attention weight matrix, where each element represents the association weight percentage of the corresponding position pair. This attention weight matrix is ​​then used to weight the value tensor to obtain the output of the attention head. The outputs of all attention heads are concatenated along the channel dimension and then transformed using an output transformation matrix. A linear transformation is performed to obtain the final output with the same dimension as the input features, thus completing the fusion of multi-head information.

[0047] The attention weight matrix directly encodes the dependencies between features at different positions in the sequence. Higher weight values ​​indicate stronger correlation between the corresponding position pairs in the feature space. The multi-head parallel structure enables the model to learn multiple correlation patterns simultaneously in different feature subspaces (such as short-term correlations between local neighboring features and long-term correlations between cross-regional features). The global computational characteristics of self-attention are not limited by the sequence length and can capture feature correlations globally. The final output feature vector integrates intra-modal correlation information of the entire sequence, providing a structured representation of correlation features for subsequent feature analysis.

[0048] S140 performs multi-scale convolution processing based on the weighted multimodal data features, and determines the starch molecule content of the mash to be tested based on the obtained multi-scale features.

[0049] Multi-scale convolution processing refers to the process of performing convolutional operations with different dilation rates to extract deep features from multiple receptive field perspectives. For example, consider an analysis network modified from the U-net architecture. Its encoder section contains multiple cascaded units, each with a multi-scale parallel convolution module. This module comprises four parallel branches: one branch uses 1×1 convolution for channel transformation; the other three branches use dilated convolutions with dilation rates of 2, 4, and 8, respectively, to expand the receptive field while maintaining the feature map resolution. Small dilation rates are used to capture fine local spectral fluctuations, while large dilation rates are used to capture global macroscopic spectral trend features.

[0050] Determining starch molecular content involves upsampling multi-scale features through a decoder, feature fusion, and a mapping layer to output the final regression prediction. Specifically, the features extracted by multi-scale convolution are fused (e.g., concatenated or superimposed) with weighted multimodal data features to restore the spatial dimensions of the features. A mapping layer consisting of 1×1 convolutions and a fully connected network maps the high-dimensional features to a single starch content value. The fully connected network can be configured with a three-layer neuron structure (e.g., 36-18-1), with each layer followed by a ReLU activation function and a Dropout layer to prevent overfitting. For example, if the input weighted feature tensor has a shape of (16, 16, 200), after multi-scale parallel convolution, the features output from the four branches are concatenated along the channel dimension and superimposed with the original input residual to form a comprehensive feature containing multi-scale information. Finally, the regression head outputs the predicted starch content percentage, such as 18.5%.

[0051] This step fully exploits the local details and global contextual information in high-dimensional spectral data through multi-scale parallel convolution, and combined with the encoding and decoding capabilities of deep networks, it achieves accurate quantitative prediction of starch molecule content in fermented mash.

[0052] This embodiment can automatically acquire multimodal spectral data of the mash to be tested without manual sampling, saving manpower and time. By extracting features and unifying dimensions of the multimodal spectral data, the spectral data of different modes are aligned in the feature space, eliminating data barriers caused by differences in wavelength range, etc. By applying attention weights to the multimodal data features of the same dimension, weights can be adaptively assigned to different modal data features, highlighting key features and suppressing redundant information. By performing multi-scale convolution processing on the weighted multimodal data features, features can be extracted from different scales, fully mining the deep information in the spectral data. The starch content can be determined based on the obtained multi-scale features without manual detection, realizing rapid and automated determination of starch content in mash. At the same time, due to the combined effect of dimension unification, attention weighting, and multi-scale convolution, the accuracy of starch content detection in mash is effectively improved.

[0053] In another exemplary embodiment of this application, it is described how to perform multi-scale convolution processing based on weighted multimodal data features, and determine the starch molecule content of the mash to be tested based on the obtained multi-scale features. Please refer to [link to relevant documentation] for details. Figure 2 , Figure 2 Based on Figure 1 The exemplary embodiment shown illustrates a flowchart of another method for determining the starch molecule content in fermented mash. This method, in accordance with... Figure 1 The S140 shown includes S210 to S230, which are described in detail below: S210 performs multi-scale convolution processing on the weighted multimodal data features to obtain multi-scale features.

[0054] The multi-scale convolution processing here can be parallelized, that is, feature extraction at different granularities can be performed simultaneously.

[0055] For example, the weighted multimodal data features are convolved by convolution branches with different dilation rates, and the resulting features at different scales are used as multi-scale features.

[0056] Convolutional branches with different dilation rates belong to the trained analysis network, which is built based on the U-net model. During execution, the input feature sequence simultaneously enters four parallel convolutional branches (i.e., convolutional branches with different dilation rates): the first branch directly transforms the feature channels using a 1×1 convolutional kernel, and its dilation rate is equivalent to 1; the second, third, and fourth branches employ dilation strategies with dilation rates of 2, 4, and 8, respectively, and each branch contains a cascaded structure of 1×1 convolution + batch normalization + ReLU activation + dilated convolution + batch normalization. By setting different dilation rates, each branch can cover spectral response regions of varying widths, from local subtle peaks to global baseline drift, without increasing the number of parameters or reducing resolution. For example, when processing a spectral feature sequence of length 200, the branch with a dilation rate of 2 mainly captures short-term correlation features between adjacent bands, while the branch with a dilation rate of 8 can capture long-range dependencies over longer distances. The feature tensors output from these four branches are concatenated along the channel dimension to obtain initial multi-scale features containing rich multi-scale information. Subsequently, to enhance the propagation capability of the features, the concatenated features are element-wise superimposed with the original input features that have undergone channel mapping, ultimately outputting complete multi-scale features.

[0057] S220 fuses multi-scale features and weighted multimodal data features to obtain fused modal data features.

[0058] This step is used to integrate the extracted multi-scale semantic information with the original contextual information to prevent information loss and improve the integrity of the feature representation.

[0059] For example, the features in the multi-scale features are concatenated to obtain concatenated features; the channel dimensions of the weighted multimodal data features are mapped to the same channel dimensions as the concatenated features to obtain mapped features; the mapped features and the concatenated features are superimposed to obtain fused modal data features.

[0060] The features obtained from convolutions with different dilation rates (e.g., 1, 2, 4, 8) are concatenated to form a high-dimensional concatenated feature with an increased number of channels. For example, assuming the input multi-scale features contain outputs from four branches, each with four channels, the concatenated feature will have 16 channels while maintaining the same sequence length. The channel dimension (8 channels) of the weighted multimodal data features is mapped to the same channel dimension (16 channels) of the concatenated features, expanding the number of channels from 8 to 16, thus obtaining a channel-dimensional aligned mapping feature. The mapped feature and the concatenated feature are then element-wise added to obtain the fused modal data features, forming a residual connection structure that fuses the deep features extracted from multi-scale features with the shallow features that retain the original information.

[0061] S230 determines the starch molecule content of the mash to be tested based on the characteristics of the fused modal data.

[0062] Based on the fused high-dimensional feature vector, the analysis network outputs the predicted value of starch molecule content in the fermented mash. Specifically, the fused modal data features are fed into the decoder or mapping layer of the analysis network for processing. In one optional implementation, the fused features, after upsampling and feature fusion by the decoder, enter the top-level mapping layer. This mapping layer consists of a 1×1 convolutional layer and a fully connected network, where the fully connected network contains three layers of neurons with the number of neurons decreasing sequentially (e.g., set to 36-18-1). Each layer of neurons is followed by a ReLU activation function and a Dropout layer (the Dropout rate can be set to 0.1) to prevent overfitting. The final output of a single neuron in this network is the predicted starch molecule content value. In practical applications, if the label data is logarithmically transformed during the training phase to balance the data distribution, then in this step, the network output value also needs to be inversely logarithmically transformed to restore the actual starch content physical quantity.

[0063] The convolutional branches with different dilation rates mentioned above belong to the trained analysis network, which is obtained as follows: the multimodal spectral data of the sample mash is preprocessed to obtain training data, and the starch content of the sample mash is preprocessed to obtain data labels; wherein, the multimodal spectral data includes near-infrared spectral data and Raman spectral data; the analysis network to be trained is trained based on the training data and data labels to obtain the trained analysis network.

[0064] By combining different types of spectral data (near-infrared and Raman spectral data) for modeling, complementary modal information is achieved, improving the detection accuracy of target physicochemical components and enabling rapid detection of starch molecule content in the mash during baijiu fermentation. The basic prototype of the analysis network is the U-net model, and its structure is as follows: Figure 4 As shown. Among them, Encoder represents the encoder; Decoder represents the decoder; DownSample represents downsampling; UpSample represents upsampling; Multi-scale Conv represents multi-scale convolution; Concatenation represents concatenation; Mapping represents mapping. The analysis network adopts an encoder-decoder symmetric architecture, consisting of an encoder module, a decoder module and a mapping layer, used to process feature extraction and transformation of sequence data. The encoder part contains 7 cascaded units, all of which are composed of multi-scale parallel convolution blocks and downsampling blocks connected in sequence to realize feature downsampling and channel diffusion. The input feature sequence passes through 7 cascaded units in sequence. Each time it passes through a unit, the number of channels increases by three times and the sequence length is halved. For a single cascaded unit, the input sequence first enters the multi-scale parallel convolution module, which realizes two functions: (1) realizes multi-scale feature extraction of the sequence by setting four parallel convolution branches to extract sequence features with different receptive fields; (2) increases the number of channels of the feature sequence, with each channel corresponding to a specific high-level representation. Next, we enter the downsampling module, which consists of "one-dimensional convolutional layer + batch normalization + ReLU" with a convolutional kernel size of 1×3. This module realizes local feature aggregation and expansion of the receptive field.

[0065] The decoder consists of seven cascaded units, each composed of an upsampling module and a feature concatenation module, to achieve feature upsampling and channel fusion. The input feature sequence passes through the seven cascaded units sequentially. Each time it passes through a unit, the number of channels is reduced by a factor of three, and the sequence length is doubled. For a single cascaded unit, the input sequence first enters the upsampling module, which performs two functions: (1) summarizing the multi-channel features of the input sequence; (2) restoring the spatial size of the feature map. Subsequently, the upsampled low-level features and the high-level features from the encoder are concatenated through the feature concatenation module to achieve cross-scale feature fusion of high and low level features. Finally, the top-level output of the decoder is processed by the mapping layer, which maps the features to the target output through a combination of 1×1 convolution and a fully connected network. The fully connected network consists of three layers of neurons, with the number of neurons being 36-18-1 respectively. Each layer of neurons is followed by a ReLU activation function and a Dropout layer, with the Dropout rate set to 0.1.

[0066] The multi-scale parallel convolution module achieves multi-scale feature extraction through multi-branch parallel convolution and enhances feature propagation by combining residual connections. The structure of the multi-scale parallel convolution module is as follows: Figure 5As shown, it mainly consists of four parallel convolutional branches and a residual concatenation layer. Here, input represents the input features, i.e., the weighted multimodal data features; 1×1 Conv represents a 1×1 convolution; BN represents batch normalization; ReLU represents the rectified linear activation function; 1×3 Conv represents a 1×3 convolution; FeatureGraph represents the concatenated features; Feature-wise addition represents element-wise addition; Dilation rate represents the dilation rate; and out represents the output features, i.e., the fused modal data features.

[0067] Parallel convolution branches: Branch 1 first uses 1×1 convolution to perform channel dimension transformation on the input features, and then obtains the branch output after batch normalization (BatchNorm1d) and ReLU activation; Branches 2, 3, and 4 are dilated convolution branches, all of which are composed of a cascaded structure of "1×1 convolution + batch normalization + ReLU + dilated convolution + batch normalization". The 1×1 convolution first performs channel dimension transformation on the input features. The four branches use convolution strategies with the same kernel (kernel_size=3) but different dilation rates (dilation=2, 4, 8).

[0068] Multi-scale feature fusion: The outputs of the four branches are concatenated along the channel dimension. Meanwhile, the input features are mapped to the sum of the number of channels of the four branches through the feature mapping module (1×1 convolution + batch normalization) to ensure that they match the dimension of the fused features. Finally, the fused features and the mapped input features are superimposed element by element through the residual concatenation layer to output a comprehensive feature containing multi-scale information and the residual of the original features.

[0069] Assuming the shape of the input tensor tensorinput is (batch, channel, length) = (16, 8, 200), then the shapes of the four branch output tensors tensora, tensorsorb, tensorc, and tensord are all (batch, channel, length) = (16, 4, 200). The shape of the concatenated tensor tensorconcat is (batch, channel, length) = (16, 16, 200). The shape of the input tensor tensorinput' after channel mapping is (batch, channel, length) = (16, 16, 200), and the shape of the output tensor tensoroutput is (batch, channel, length) = (16, 16, 200).

[0070] This design captures features of different receptive fields (small dilation rate corresponds to local features, and large dilation rate corresponds to global features) through parallel convolutions with different dilation rates, while maintaining computational efficiency, thus achieving parallel extraction and fusion of multi-scale information.

[0071] The training method of the analysis network mainly includes three steps: dataset construction, data preprocessing and network training. The first step is to construct a dataset of samples to be analyzed. The dataset uses near-infrared-Raman spectra as input data and physicochemical components (starch) content as dataset labels. The second step is to preprocess the spectral data and label data of the samples respectively: (1) Savitzky-Golay filtering and airPLS filtering are applied to the Raman spectral data and near-infrared spectral data of the samples respectively. Savitzky-Golay filtering is used to smooth and reduce noise of the signal and remove high-frequency noise; airPLS filtering is used to remove fluorescence noise of the spectral signal. The above two preprocessing methods can reduce the difficulty of model learning and avoid overfitting by highlighting the effective features related to physicochemical components and eliminating redundant interference information; (2) Logarithmic transformation is applied to the label data of the samples. Logarithmic transformation is applied to labels with sparse distribution and large extreme value range. This can compress the numerical range to reduce the impact of extreme values ​​on model training, adjust the skewed distribution to a distribution closer to normal, make the data distribution more balanced, and thus improve the accuracy and stability of regression prediction. The third step involves loading the preprocessed dataset in batches into a regression analysis network composed of a multimodal data fusion module and a multi-scale parallel convolutional network for model training. This ultimately yields the trained multimodal spectral data analysis network.

[0072] The prediction and analysis of starch molecule content in fermented mash can be directly performed through the host computer system, mainly including three steps: data collection of the sample to be tested, data preprocessing, and prediction of physicochemical components. The first step is data collection. Unlike the deployment phase, which requires collecting both spectral data and physicochemical component content, this step only requires collecting near-infrared-Raman spectral data of the sample to be tested. The second step is data preprocessing. In the host computer's data preprocessing interface, after setting the relevant configuration parameters, Savitzky-Golay filtering and airPLS filtering operations on the spectral data can be completed with one click. It also supports real-time viewing of the before-and-after effect comparison and the ability to undo and redo the preprocessing operation. The third step is physicochemical component prediction. Adding the preprocessed spectral data as input data to the prediction and analysis interface of the host computer allows the network trained in the deployment phase to predict the physicochemical components of the sample.

[0073] The host computer system can acquire multimodal spectral data of the mash to be tested in real time for predictive analysis, and display the determined starch molecule content of the mash on the display interface. The host computer system has a simple and highly interactive display interface, and also supports the management of historical data. The host computer includes four main core functions: data reading, preprocessing, quantitative prediction, and data storage. It can provide the backend analysis network with simple one-click operation and intuitive visualization effects, lowering the usage threshold of the analysis network for starch molecule of mash.

[0074] Another aspect of this application provides a device for determining the starch molecule content of fermented mash, such as... Figure 6 As shown, Figure 6 This is a schematic diagram illustrating the structure of a device for determining the starch molecule content of fermented mash, as shown in an exemplary embodiment of this application. The determining device 600 includes: The acquisition module 610 is used to acquire multimodal spectral data of the mash to be tested.

[0075] The feature extraction module 630 is used to extract features from multimodal spectral data and unify the dimensions of the extracted multimodal data features to obtain multimodal data features with the same dimensions.

[0076] The weighting module 650 is used to perform attention weighting on multimodal data features of the same dimension to obtain weighted multimodal data features.

[0077] The content determination module 670 is used to perform multi-scale convolution processing based on the weighted multimodal data features, and determine the starch molecule content of the mash to be tested based on the obtained multi-scale features.

[0078] In another exemplary embodiment, the feature extraction module 630 includes: The convolution processing unit is used to perform convolution processing on multimodal spectral data to obtain initial multimodal data features.

[0079] The encoding unit is used to encode the initial multimodal data features to obtain the encoded features corresponding to the initial multimodal data features.

[0080] The generation unit is used to generate multimodal data features based on the encoded features and the initial multimodal data features.

[0081] Alignment units are used to align features of different sequence lengths in multimodal data features to obtain multimodal data features of the same dimension.

[0082] In another exemplary embodiment, the weight processing module 650 includes: The mapping unit is used to map each modal data feature in the multimodal data features of the same dimension into the first tensor, the second tensor, and the third tensor, respectively.

[0083] The computation unit is used to calculate the similarity between the first tensor and the second tensor corresponding to the features of each modality data, and to normalize the calculated similarity matrix to obtain the attention weight matrix.

[0084] The weighting unit is used to weight the third tensor based on the attention weight matrix, and the weighted modal data features obtained are used as the weighted multimodal data features.

[0085] In another exemplary embodiment, the content determination module 670 includes: Multi-scale convolutional units are used to perform multi-scale convolution processing on weighted multimodal data features to obtain multi-scale features.

[0086] The feature fusion unit is used to fuse multi-scale features and weighted multimodal data features to obtain fused modal data features.

[0087] The content determination unit is used to determine the starch molecule content of the mash to be tested based on the characteristics of the fusion modal data.

[0088] In another exemplary embodiment, the feature fusion unit includes: The splicing module is used to splice together the features from multiple scales to obtain spliced ​​features.

[0089] The mapping module is used to map the channel dimensions of the weighted multimodal data features to the same channel dimensions as the concatenated features, thus obtaining the mapped features.

[0090] Overlay blocks are used to overlay mapping features and splicing features to obtain fused modal data features.

[0091] In another exemplary embodiment, the multi-scale convolutional unit includes: The multi-scale convolution module is used to perform convolution processing on the weighted multimodal data features through convolution branches with different dilation rates, and the resulting features at different scales are used as multi-scale features.

[0092] In another exemplary embodiment, convolutional branches with different dilation rates belong to the trained analysis network, which is obtained in the following way: The multimodal spectral data of the sample mash were preprocessed to obtain training data, and the starch content of the sample mash was preprocessed to obtain data labels. The multimodal spectral data included near-infrared spectral data and Raman spectral data. The analysis network to be trained was trained based on the training data and data labels to obtain the trained analysis network. The basic prototype of the analysis network was the U-net model.

[0093] This application's device can automatically acquire multimodal spectral data of the mash to be tested without manual sampling, saving manpower and time. By extracting features and unifying dimensions of the multimodal spectral data, the spectral data of different modes are aligned in the feature space, eliminating data barriers caused by differences in wavelength range. By applying attention weights to multimodal data features of the same dimension, weights can be adaptively assigned to different modal data features, highlighting key features and suppressing redundant information. By performing multi-scale convolution processing on the weighted multimodal data features, features can be extracted from different scales, fully mining the deep information in the spectral data. The starch content can be determined based on the obtained multi-scale features without manual detection, realizing rapid and automated determination of starch content in mash. At the same time, due to the combined effect of dimension unification, attention weighting, and multi-scale convolution, the accuracy of starch content detection in mash is effectively improved.

[0094] It should be noted that the apparatus provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The determining apparatus and implementation process provided in the above embodiments have been described in detail in the method embodiments, and will not be repeated here.

[0095] Another aspect of this application provides an electronic device, including: a processor; a memory and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement any of the determined methods described above.

[0096] Please see Figure 7 , Figure 7 This is a schematic diagram of the structure of a computer system for an electronic device according to an exemplary embodiment of this application, illustrating a schematic diagram of the structure of a computer system suitable for implementing the embodiments of this application.

[0097] It should be noted that, Figure 7 The computer system 700 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0098] like Figure 7As shown, the computer system 700 includes a Central Processing Unit (CPU) 701, which can perform various appropriate actions and processes, such as executing the methods described in the above embodiments, based on programs stored in Read-Only Memory (ROM) 702 or programs loaded from storage portion 708 into Random Access Memory (RAM) 703. The RAM 703 also stores various programs and data required for system operation. The CPU 701, ROM 702, and RAM 703 are interconnected via a bus 704. An Input / Output (I / O) interface 705 is also connected to the bus 704.

[0099] The following components are connected to I / O interface 705: an input section 706 including a keyboard, mouse, etc.; an output section 707 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 708 including a hard disk, etc.; and a communication section 709 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to I / O interface 705 as needed. A removable medium 711, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 710 as needed so that computer programs read from it can be installed into storage section 708 as needed.

[0100] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 709, and / or installed from removable medium 711. When the computer program is executed by central processing unit (CPU) 701, it performs various functions defined in the system of this application.

[0101] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable computer program. The transmitted data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0102] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation that may be implemented in systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0103] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0104] Another aspect of this application provides a storage medium on which a program or instructions are stored, which, when executed by a processor, implement any of the determined methods described above. This storage medium may be included in the electronic device described in the above embodiments, or it may exist independently and not assembled into the electronic device.

[0105] The above description is merely a preferred exemplary embodiment of this application and is not intended to limit the implementation of this application. Those skilled in the art can easily make corresponding modifications or alterations based on the main concept and spirit of this application. Therefore, the scope of protection of this application should be determined by the scope of protection claimed in the claims.

Claims

1. A method for determining the starch molecule content of fermented mash, characterized in that, include: Acquire multimodal spectral data of the mash to be tested; Feature extraction is performed on the multimodal spectral data, and the extracted multimodal data features are dimension-unified to obtain multimodal data features with the same dimension. Attention weighting is applied to the multimodal data features of the same dimension to obtain weighted multimodal data features; Multi-scale convolution processing is performed on the weighted multimodal data features, and the starch molecule content of the mash to be tested is determined based on the obtained multi-scale features.

2. The determination method according to claim 1, characterized in that, Feature extraction of the multimodal spectral data includes: The multimodal spectral data is convolved to obtain initial multimodal data features; The initial multimodal data features are encoded to obtain the encoded features corresponding to the initial multimodal data features; Based on the encoded features and the initial multimodal data features, multimodal data features are generated; The extracted multimodal data features are unified in dimensionality to obtain multimodal data features with the same dimensionality, including: The features with different sequence lengths in the multimodal data features are length aligned to obtain multimodal data features of the same dimension.

3. The determination method according to claim 1, characterized in that, Attention weighting is applied to the multimodal data features of the same dimension to obtain weighted multimodal data features, including: Each modal data feature in the multimodal data features of the same dimension is mapped to a first tensor, a second tensor, and a third tensor, respectively. Calculate the similarity between the first tensor and the second tensor corresponding to each modality data feature, and normalize the calculated similarity matrix to obtain the attention weight matrix; The third tensor is weighted based on the attention weight matrix, and the weighted modal data features are used as the weighted multimodal data features.

4. The determination method according to claim 1, characterized in that, Multi-scale convolution processing is performed on the weighted multimodal data features, and the starch molecule content of the mash to be tested is determined based on the obtained multi-scale features, including: The weighted multimodal data features are subjected to multi-scale convolution processing to obtain multi-scale features; The multi-scale features and the weighted multimodal data features are fused to obtain fused modal data features; The starch molecule content of the mash to be tested is determined based on the characteristics of the fused modal data.

5. The determination method according to claim 4, characterized in that, The multi-scale features and the weighted multimodal data features are fused to obtain fused modal data features, including: The features in the multi-scale features are concatenated to obtain the concatenated features; The channel dimensions of the weighted multimodal data features are mapped to the same channel dimensions as the concatenated features to obtain the mapped features; The mapping features and the splicing features are superimposed to obtain the fused modal data features.

6. The determination method according to claim 4, characterized in that, The weighted multimodal data features are subjected to multi-scale convolution processing to obtain multi-scale features, including: The weighted multimodal data features are convolved by convolution branches with different dilation rates, and the resulting features at different scales are used as multi-scale features.

7. The determination method according to claim 6, characterized in that, The convolutional branches with different dilation rates belong to the trained analysis network, which is obtained through the following method: The multimodal spectral data of the sample mash are preprocessed to obtain training data, and the starch content of the sample mash is preprocessed to obtain data labels; wherein, the multimodal spectral data includes near-infrared spectral data and Raman spectral data; The analysis network to be trained is trained based on the training data and the data labels to obtain the trained analysis network; wherein the basic prototype of the analysis network is the U-net model.

8. A device for determining the starch molecule content of fermented mash, characterized in that, include: The acquisition module is used to acquire multimodal spectral data of the mash to be tested; The feature extraction module is used to extract features from the multimodal spectral data and unify the dimensions of the extracted multimodal data features to obtain multimodal data features of the same dimension. The weighting module is used to perform attention weighting on the multimodal data features of the same dimension to obtain weighted multimodal data features; The content determination module is used to perform multi-scale convolution processing based on the weighted multimodal data features, and determine the starch molecule content of the mash to be tested based on the obtained multi-scale features.

9. An electronic device, characterized in that, include: processor; A memory and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the determination method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a program or instructions that, when executed by a processor, implement the determination method as described in any one of claims 1 to 7.

Images