Identification system and method for intrinsic semantic difference learning
By using multimodal signal learning and causal graph models, the intrinsic semantic features and component coupling relationships of baijiu are extracted, solving the problem of difficulty in identifying subtle differences between counterfeit and genuine baijiu in existing technologies, and achieving high-precision baijiu identification.
Patent Information
- Application Number
- CN202511519808.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-23
- Publication Date
- 2026-02-06
AI Technical Summary
Existing technologies struggle to effectively identify subtle but essential differences between highly realistic counterfeit liquor and genuine liquor, taking into account the natural differences within the liquor category. Furthermore, traditional methods neglect the complex interrelationships of the liquor's internal components.
An identification system employing intrinsic semantic difference learning learns the intrinsic semantic features of baijiu through multimodal signals, and uses physical neural networks and causal graph models to extract key difference features and component coupling relationships of the liquor, achieving high-level discrimination between genuine and counterfeit liquor.
It significantly improves the accuracy of identifying highly realistic counterfeit liquor, enhances the model's adaptability to real-world scenarios, and can accurately distinguish between genuine and counterfeit liquor even when the ingredients are highly similar, providing a more fundamental and reliable identification effect.
Smart Images
Figure CN121476550A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of liquor body identification and discrimination, and is applied to the process of authenticating real and fake liquor, and particularly relates to a discriminant system and method based on intrinsic semantic difference learning. BACKGROUND
[0002] As a traditional Chinese distilled liquor, the quality and authenticity of liquor is directly related to the consumer rights and brand credibility. In recent years, counterfeit liquor has highly imitated the appearance, packaging, label information and even the basic physicochemical indicators of the genuine product, making the traditional anti-counterfeiting methods face severe challenges. Especially at the level of liquor body composition, high-imitation counterfeit liquor often uses similar raw materials and processes, resulting in similar numerical characteristics in conventional testing between real and counterfeit liquor. It is difficult to effectively identify the subtle but key differences by relying on static comparison of ingredient content or single modal signal analysis.
[0003] On the other hand, even the same brand and model of liquor, there are natural fluctuations between different production batches. This intra-class difference is caused by the comprehensive influence of factors such as raw material origin and year, climate conditions, fermentation period, storage environment and artificial process control, making the sensory and physicochemical performance of real liquor not completely consistent. However, existing identification methods focus on the macro differences between different brands or real and fake categories, generally ignoring the reasonable natural variation within real liquor, resulting in misjudgment of normal batch fluctuations as abnormal in actual application, or loss of sensitivity to high-imitation counterfeit liquor due to excessive fitting of inter-class boundaries.
[0004] In addition, the flavor and quality of liquor are not simply the superposition of various chemical components, but are derived from the complex interaction and dynamic coupling between hundreds of trace components. This inherent correlation constitutes the essence of the uniqueness of liquor body, and is also the core of high-end liquor that is difficult to replicate. However, the current mainstream analysis method usually regards each component as an independent variable, and only observes the deviation of a single index through column chart, heat map and other visualization methods, ignoring the synergy and constraint relationship between component networks. This simplifying assumption is easily ineffective when faced with carefully prepared counterfeit liquor, as it may approach the real liquor in a single component, but cannot reproduce its overall coupling mechanism.
[0005] Therefore, there is an urgent need for a new identification method that can go beyond surface numerical comparison, take into account reasonable intra-class variation, and deeply explore the complex correlation structure within the liquor body. This method should be able to extract robust intrinsic features from multi-source perception signals, while having the ability to identify counterfeit traces from subtle abnormalities, thereby maintaining high discrimination in the face of highly simulated counterfeit liquor. SUMMARY
[0006] The purpose of the present application is to solve the problem that the prior art is difficult to effectively identify the weak but essential differences between the highly simulated fake liquor and the real liquor under the premise of considering the natural differences among the liquor category, therefore a discriminant system and method of intrinsic semantic difference learning are proposed. The present application learns the intrinsic semantic features of liquor perception signals from multi-modal signals, uses pseudo-learning to realize high discriminant difference feature learning of different liquor bodies and real and fake liquor bodies, and uses the representation learning ability of physical neural networks to learn the difference syndrome features of fake and real liquor from weak differences. At the same time, the coupling relationship between the components of the liquor body is learned by using the causal graph model, the cause and effect of the real and fake liquor is inferred, and finally the real and fake discrimination is realized.
[0007] The present application adopts the following technical solutions to achieve the purpose: A discriminant system of intrinsic semantic difference learning, the system comprises the following functional modules: An instrument module is used for data perception of the detected liquor body and the reference liquor body to form liquor body sample pairs of multi-source data in a unified reference environment and complete alignment of data signals; A feature embedding module is used for converting the liquor body sample pairs of multi-source data into high-level semantic features, so that the different modal features corresponding to the detected liquor body and the reference liquor body are respectively fused in the high-level semantic space; An intrinsic semantic extraction module is used for extracting the respective intrinsic semantic features of the detected liquor body and the reference liquor body from the fused corresponding features, wherein the intrinsic semantic features of the reference liquor body are used to remove the influence of different batches of reference liquor bodies on the discrimination result; A true and false difference syndrome extraction module is used to obtain the key difference features between the detected liquor body and the reference liquor body according to the respective intrinsic semantic features of the two, as the true and false wine segmentation boundary; A causal graph reasoning flavor extraction module is used to extract the component coupling relationship of the detected liquor body and the reference liquor body according to the respective intrinsic semantic features of the two; A true and false wine inference module is used to identify the true and false of the detected liquor body according to the key difference features between the detected liquor body and the reference liquor body, and the respective component coupling relationship of the two, and obtain the discrimination result.
[0008] Specifically, the signal types of the liquor body sample pairs of multi-source data perceived and formed by the instrument module include chromatographic signals, mass spectrometric signals and spectral signals, and each type of signal has one or more feature parameters to be embedded; the instrument module is also used to align the time sequence, intensity and sampling rate of the data signals respectively when the alignment of the data signals is completed.
[0009] Preferably, the intrinsic semantic extraction module is also used to build an intrinsic semantic extraction model of the wine body, which uses a twin neural network structure to extract corresponding intrinsic semantic features from the fused features; in the twin neural network structure, the encoder is a neural network in the form of LSTM+CNN, and the decoder is a fully convolutional neural network CNN, which is used for signal reconstruction in the self-supervised learning training process of the model.
[0010] The application also provides a discrimination method of the discrimination system of the intrinsic semantic difference learning, which comprises the following steps: S1, obtaining the chromatographic signal, mass spectrum signal and spectral signal of the detection wine body and the reference wine body respectively to form the multi-source data of the wine body sample and perform alignment and registration; S2, converting the multi-source data of the wine body sample into high-level semantic features, so that the corresponding different modal features of the detection wine body and the reference wine body are respectively fused in the high-level semantic space; S3, extracting the respective intrinsic semantic features of the detection wine body and the reference wine body from the fused corresponding features, and removing the influence of different batches of reference wine bodies on the discrimination result through the intrinsic semantic features of the reference wine body; S4, obtaining the key difference features between the detection wine body and the reference wine body according to the respective intrinsic semantic features of the detection wine body and the reference wine body, as the true and false wine segmentation boundary; S5, extracting the component coupling relationship of the detection wine body and the reference wine body according to the respective intrinsic semantic features of the detection wine body and the reference wine body; S6, identifying the true and false of the detection wine body according to the key difference features between the detection wine body and the reference wine body and the respective component coupling relationship of the detection wine body and the reference wine body, and obtaining the discrimination result.
[0011] Specifically, in step S1, the amplitude of the multi-source data of the wine body sample is aligned by using a data normalization method; in the multi-source data of the wine body sample, the sampling rate of the detection wine body signal data and the reference wine body signal data is matched and aligned by finding the shortest distance between the two; after the sampling rate matching and alignment, the time sequence of the detection wine body signal data and the reference wine body signal data is filled to the same length by linear or nonlinear interpolation, and the time sequence of the two is aligned.
[0012] Specifically, in step S2, a convolutional neural network is used to complete the embedding of the features, and one convolutional neural network is used to complete the embedding of the features of the chromatographic signal, mass spectrum signal and spectral signal respectively; in the embedding process, a scalar feature embedding strategy is used to uniformly complete the embedding of the scalar signal and the time sequence signal; for the chromatographic signal, mass spectrum signal and spectral signal, each type of signal has one or more feature parameters to be embedded, and after the embedding of all the feature parameters is completed, they are uniformly embedded into the semantic space, and different feature parameters are fused in the semantic space by using the feature connection method.
[0013] Preferably, in step S3, the wine body intrinsic semantic extraction model is pre-constructed, which adopts a twin neural network structure to extract corresponding intrinsic semantic features from the fused features; The self-supervised learning training of the wine body intrinsic semantic extraction model is as follows: Pre-acquire reference wine body data, including the corresponding brand and model of the reference wine body, and encode the classification to form chromatographic signals, mass spectrometric signals and spectral signals, thereby constructing an intrinsic semantic extraction training database D; During training, data is randomly taken from the training database D and input into the twin neural network of the model for training; the two extracted samples are denoted as and If and belong to the same model of wine body, then the corresponding intrinsic semantic features and are expected to be similar after training, otherwise they are expected to differ by more than a threshold value, thereby distinguishing different wine bodies; by repeatedly iterating the above training process, the wine body intrinsic semantic extraction model that meets the preset requirements is obtained.
[0014] Preferably, in step S4, according to the intrinsic semantic features of the detected wine body and the reference wine body, before obtaining the key difference features, the wine body intrinsic semantic extraction model is optimized and trained for the case where the detected wine body is fake wine.
[0015] Specifically, in step S5, a flavor extraction model is pre-constructed, which adopts a graph neural network structure; when training the flavor extraction model, if the input sample corresponds to the same wine body model, the contrast loss is expected to be small, otherwise the contrast loss is expected to be greater than a threshold value.
[0016] Specifically, in step S6, when identifying the authenticity of the detected wine body, a comparative analysis method is used for identification, and a multi-layer fully connected physical neural network is pre-constructed; when training the multi-layer fully connected physical neural network, two training scenarios corresponding to the actual identification process are divided, one is single detection wine body identification, and the other is identification of known detected wine body corresponding to reference wine body; For single detection wine body identification, samples are extracted one by one from the pre-constructed real wine fingerprint library and fake wine fingerprint library, and the multi-layer fully connected physical neural network is iteratively trained, so that for the input detected wine body, the corresponding label of real wine or fake wine is outputted; For identification of known detected wine body corresponding to reference wine body, a twin network sharing weights is trained with the two types of wine bodies; when the samples belong to the same real wine or fake wine, the corresponding label value is outputted.
[0017] In conclusion, by adopting the technical scheme, the application has the following beneficial effects: The application effectively integrates detection instrument parameters and environmental parameters by fusing multi-modal heterogeneous data, significantly reduces the influence of equipment fluctuation and external interference on the identification result, and improves the stability and reliability of the system in actual application. The liquor body intrinsic semantic extraction model constructed by the application can adaptively learn the inherent characteristics of liquor from complex perception signals, forming a highly representative liquor body fingerprint. The fingerprint not only reflects the essential properties of a specific type of liquor, but also is compatible with natural variations between different batches, enhancing the model's adaptability to real scenarios.
[0018] In view of the challenges of uneven distribution of real and fake liquor samples and weak differences, the neural network structure designed by the application focuses on the abnormal syndromes that are difficult to conceal in fake liquor through pseudo-learning mechanism, realizes high sensitivity identification of high-simulation fake liquor, and significantly improves the discrimination accuracy. At the same time, by introducing a causal graph neural network model, the application can actively learn the coupling and dependent relationship between each chemical component in the liquor body, surpassing the traditional independent variable assumption, and making causal inference from the mechanism level of flavor formation, so as to accurately distinguish real and fake liquor in the case of high similarity of components, and realize more essential and reliable identification effect. BRIEF DESCRIPTION OF DRAWINGS
[0019] The application further illustrates its embodiments and technical schemes by the following drawings, specifically including two drawings, as follows: Figure 1 The figure is a functional module diagram of the identification system of the intrinsic semantic difference learning of the application; Figure 2 The figure is a whole flowchart of the identification method of the intrinsic semantic difference learning of the application. DETAILED DESCRIPTION
[0020] To make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely in combination with the drawings in the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, not all the embodiments. The components of the embodiments of the application described and shown in the drawings can be arranged and designed in various different configurations.
[0021] Therefore, the following detailed description of the embodiments of the application provided in the drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the application.
[0022] Embodiment 1 As Figure 1As shown, a discriminant system for intrinsic semantic difference learning comprises the following functional modules: An instrument module is configured to perform data sensing on the detection liquor and the reference liquor to form liquor sample pair multi-source data in a unified reference environment and complete alignment of data signals. A feature embedding module is configured to convert the liquor sample pair multi-source data into high-level semantic features, so that different modal features corresponding to the detection liquor and the reference liquor are respectively fused in the high-level semantic space. An intrinsic semantic extraction module is configured to extract intrinsic semantic features of the detection liquor and the reference liquor from the fused corresponding features, wherein the intrinsic semantic features of the reference liquor are used to remove the influence of different batches of reference liquor on the discrimination result. A true-false difference syndrome extraction module is configured to obtain key difference features between the detection liquor and the reference liquor as a true-false wine segmentation boundary according to the intrinsic semantic features of the detection liquor and the reference liquor. A causal graph reasoning flavor extraction module is configured to extract ingredient coupling relationships of the detection liquor and the reference liquor according to the intrinsic semantic features of the detection liquor and the reference liquor. A true-false wine inference module is configured to identify the true or false of the detection liquor according to the key difference features between the detection liquor and the reference liquor and the ingredient coupling relationships of the detection liquor and the reference liquor, and obtain a discrimination result.
[0023] In this embodiment, the signal types of the liquor sample pair multi-source data sensed and formed by the instrument module include chromatographic signals, mass spectrometric signals and spectroscopic signals, and each type of signal has one or more feature parameters to be embedded. The instrument module is also configured to align the time sequence, intensity and sampling rate of the data signals when the data signals are aligned.
[0024] As a preferred embodiment, the intrinsic semantic extraction module is also configured to construct a liquor intrinsic semantic extraction model, which uses a twin neural network structure to extract corresponding intrinsic semantic features from the fused features. The model is as follows:
[0025] In the formula, represents an encoder, which is a neural network in the form of LSTM+CNN; represents the extracted intrinsic semantic features; represents a decoder, which is a fully convolutional neural network CNN used for signal reconstruction in the model self-supervised learning training process; represents the original input data of the model, represents the reconstructed signal data after model processing.
[0026] Embodiment 2 On the basis of Embodiment 1, the present embodiment provides a discrimination method of the discrimination system of the intrinsic semantic difference learning, which can refer to the schematic of Figure 2 , comprising the following steps: S1, obtaining the chromatographic signal, mass spectrum signal and spectrum signal of each of the detection wine body and the reference wine body, forming the multi-source data of the wine body sample and performing alignment registration; S2, converting the multi-source data of the wine body sample into high-level semantic features, so that the different modal features corresponding to each of the detection wine body and the reference wine body are respectively fused in the high-level semantic space; S3, extracting the respective intrinsic semantic features of the detection wine body and the reference wine body from the fused corresponding features, and removing the influence of different batches of reference wine bodies on the discrimination result through the intrinsic semantic features of the reference wine body; S4, obtaining the key difference features between the detection wine body and the reference wine body according to the respective intrinsic semantic features of the detection wine body and the reference wine body, as the true and false wine segmentation boundary; S5, extracting the component coupling relationship of the detection wine body and the reference wine body according to the respective intrinsic semantic features of the detection wine body and the reference wine body; S6, identifying the true and false of the detection wine body according to the key difference features between the detection wine body and the reference wine body and the respective component coupling relationship of the detection wine body and the reference wine body, and obtaining the discrimination result.
[0027] In the present embodiment, due to the differences in device signals and parameters, the signal data obtained by the original sensor has differences in intensity value and frequency of use. These data cannot be directly used for classification and discrimination. Therefore, in step S1, a data normalization method is used to realize the amplitude alignment of the multi-source data of the wine body sample, as follows:
[0028] In the formula, represents the th original data, represents the data expectation, represents the data standard deviation, represents the normalized data.
[0029] In the multi-source data of the wine body sample, the shortest distance between the detection wine body signal data and the reference wine body signal data is used to realize the matching and alignment of the sampling rates of the two; let , , and represent the number of data sampling of the two, respectively; the th detection wine body signal data and the The reference wine body signal data The matching point distance is represented as An optimization function is constructed as follows:
[0030] In the formula, represents the total number of matching patterns, represents the optimal matching pattern among all matching patterns, so as to realize matching alignment of the sampling rate; after completing the sampling rate matching alignment, the time sequence of the detected wine body signal data and the reference wine body signal data is filled to the same length through linear or nonlinear interpolation, so as to realize alignment of the time sequences of the two.
[0031] In this embodiment, step S2 solves the problems of structure and semantic difference of multi-source signal patterns. The first problem to be solved is the semantic embedding of chromatographic, mass spectrometric and spectral signals. The second problem to be solved is the semantic embedding of scalar information such as device environment information.
[0032] Convolutional neural networks are used to complete feature embedding, and the dimension of the high-level semantic feature to be converted is A convolutional neural network is used to complete feature embedding of chromatographic signals, mass spectrometric signals and spectral signals, respectively; the embedding mode is unified as follows:
[0033] In the formula, represents the embedded high-level semantic feature; represents the feature weight of the final output layer of the convolutional neural network; represents the nonlinear transformation in the first to the residual module; represents residual connection; represents the input of the convolutional neural network; the symbol represents a composite operation; represents the bias term of the final output layer of the convolutional neural network.
[0034] In the embedding process, scalar feature embedding strategy is used to complete the embedding of scalar signals and time sequence signals, as follows:
[0035] In the formula, represents the scalar signal collected by the instrument; according to the dimension number to constitute the above-mentioned time sequence data This time sequence data is represented by a convolutional neural network, i.e. using the above-mentioned unified embedding mode.
[0036] For the chromatographic signal, the mass spectrum signal and the spectrum signal, one or more characteristic parameters to be embedded are in each type of signal, and after the embedding of all the characteristic parameters is completed, they are unified into a semantic space, in which the fusion of different characteristic parameters is performed in a feature connection manner, as follows:
[0037] In the formula, represents the corresponding feature after fusion; represents the total number of characteristic parameters subjected to fusion.
[0038] In actual application, there are many types of fake wine, and it is difficult to fully obtain sample data. The embodiment proposes to preferentially use samples of real wine to realize the extraction of authentic semantic features of reaction wine. The design is applied to the training process of the related model in step S3, and in step S3, a wine body intrinsic semantic extraction model is constructed in advance, which uses a twin neural network structure to extract corresponding intrinsic semantic features from the fused features; the model is as follows:
[0039] In the formula, represents an encoder, which is a neural network in the form of LSTM+CNN; represents the extracted intrinsic semantic features; represents a decoder, which is a fully convolutional neural network CNN, used for signal reconstruction in the self-supervised learning training process of the model; represents the original input data of the model, represents the reconstructed signal data after model processing.
[0040] The self-supervised learning training of the wine body intrinsic semantic extraction model is as follows: Reference wine body data is collected in advance, including the brands and models corresponding to the reference wine bodies, and after coding classification, chromatographic signals, mass spectrum signals and spectrum signals are formed to construct an intrinsic semantic extraction training database D.
[0041] During training, data is randomly taken from the training database D and input into the twin neural network of the model for training; the two extracted samples are denoted as and If and belong to the same type of wine body, then the corresponding intrinsic semantic features and after training are expected to be similar, otherwise they are expected to differ by more than a threshold value, so as to distinguish different wine bodies; the loss function of the twin neural network is represented as follows:
[0042] In the formula, This represents a symbolic function, meaning that if two inputs are of the same type of liquor, the output is... Otherwise output ; The weights representing the control reconstruction error terms; This represents the weights of the loss term that control the semantic features compared to the loss term; and For the category labels of two samples, that is, when the two samples are of the same type of wine, By iterating the above training process repeatedly, a wine essence semantic extraction model that meets the preset requirements is obtained.
[0043] In step S4 of this embodiment, based on the intrinsic semantic features of the detected wine and the reference wine, before deriving the key difference features, the intrinsic semantic extraction model of the wine is optimized for counterfeit wine in the case where the detected wine is counterfeit.
[0044] In the counterfeit wine optimization training process, for each iteration, reference samples of genuine wine are taken from the training database D, and counterfeit wine samples are taken from the pre-set counterfeit wine database. These samples are then input into the Siamese neural network of the wine intrinsic semantic extraction model for training. At this time, the learning rate of the network is reduced to 10% of the original value. After the counterfeit wine optimization training is completed, the encoder obtains the true and false difference syndrome features of the detected wine and the reference wine for the existing samples, i.e., the key difference features. The features corresponding to the reference wine are stored in the genuine wine fingerprint database. In actual application and reasoning, the features of various reference wines are directly extracted from the genuine wine fingerprint database to realize the discrimination of the detected wine and obtain the corresponding key difference features.
[0045] The flavor of baijiu is determined by its components and their complex coupling relationships. An experienced bartender can use these subtle factors to identify baijiu. However, long-term research results show that this complex relationship is difficult to represent explicitly. Therefore, this embodiment proposes to use graph neural networks to learn and represent the flavor of baijiu.
[0046] In step S5, a flavor extraction model is pre-constructed. This model uses a graph neural network structure, where the encoder is represented as follows:
[0047] In the formula, Representative component characteristic matrix; The adjacency matrix represents the graph neural network; The encoding function representing a graph neural network; This represents the coupling relationship of the extracted components.
[0048] The decoder in a graph neural network structure is represented as follows:
[0049] In the formula, represents the reconstructed feature matrix; represents the decoding function formula of the graph neural network.
[0050] When training the flavor extraction model, when the input sample corresponds to the same wine body model, the contrast loss is expected to be small, otherwise the contrast loss is expected to be greater than the threshold, which is expressed as follows:
[0051] In the formula, represent two samples and belong to the same type of wine body, that is, the positive sample pair; represent two samples and belong to different types of wine body, or one is real wine and the other is fake wine, that is, the negative sample pair; represent the distance between the feature vectors of the two samples; represent the interval parameter, used to control the lower limit of the minimum distance between negative samples.
[0052] The overall loss function of the flavor extraction model is expressed as follows:
[0053] In the formula, is the weight loss, is the aforementioned contrast loss, is the contrast loss weight; is expressed as follows:
[0054] In the formula, represent the original feature vector of the model data group component, represent the corresponding vector after reconstruction; represent the total number of count values.
[0055] Finally, in step S6, when identifying and detecting the authenticity of the wine body, a comparative analysis method is used for identification, and a multi-layer fully connected physical neural network is constructed in advance, as follows:
[0056] In the formula, represent the identification result; represent the key difference features; represent the component coupling relationship; When training the multi-layer fully connected physical neural network, two training scenarios corresponding to the actual identification process are divided, one is single detection of wine body identification, and the other is identification of known detection wine body corresponding to reference wine body; For single detection wine body identification, samples are extracted from pre-constructed genuine wine fingerprint library and fake wine fingerprint library one by one, and a multi-layer fully connected physical neural network is iteratively trained, so that for the input detection wine body, it outputs the corresponding label of genuine wine or fake wine; For the identification of known detection wine body corresponding to reference wine body, a twin network sharing weights is trained, and when the samples belong to the same genuine wine or fake wine, the corresponding label value is output; here the loss function As follows:
[0057] In the formula, The sample label at the time of the th discrimination, that is, if it is genuine wine, it is 1, otherwise it is 0; The network predicts the probability that the given two samples belong to the same category when the th discrimination.
Claims
1. A discriminative system for inductive semantic discrepancy learning, characterized in that, The system comprises the following functional modules: An instrument module, configured to sense data of the detection liquor body and the reference liquor body to form liquor body sample multi-source data in a unified reference environment and complete alignment of data signals; A feature embedding module, configured to convert the liquor body sample multi-source data into high-level semantic features, so that different modal features corresponding to the detection liquor body and the reference liquor body are respectively fused in a high-level semantic space; An intrinsic semantic extraction module, configured to extract intrinsic semantic features of the detection liquor body and the reference liquor body from the fused corresponding features, wherein the intrinsic semantic features of the reference liquor body are used to remove the influence of different batches of reference liquor bodies on the identification result; A true-false difference syndrome extraction module, configured to obtain key difference features between the detection liquor body and the reference liquor body as a true-false liquor segmentation boundary according to the intrinsic semantic features of the detection liquor body and the reference liquor body; A causal graph reasoning flavor extraction module, configured to extract ingredient coupling relationships of the detection liquor body and the reference liquor body according to the intrinsic semantic features of the detection liquor body and the reference liquor body; A true-false liquor inference module, configured to identify the true or false of the detection liquor body according to the key difference features between the detection liquor body and the reference liquor body and the ingredient coupling relationships of the detection liquor body and the reference liquor body, and obtain an identification result.
2. The authentication system of claim 1, wherein: The signal types of the liquor body sample multi-source data sensed and formed by the instrument module include chromatographic signals, mass spectrometric signals and spectral signals, and each type of signal has one or more feature parameters to be embedded; the instrument module is also configured to align the time sequence, intensity and sampling rate of the data signals when the alignment of the data signals is completed.
3. The authentication system of claim 1, wherein: The intrinsic semantic extraction module is also configured to construct a liquor body intrinsic semantic extraction model, which adopts a twin neural network structure to extract corresponding intrinsic semantic features from the fused features; The model is as follows: wherein, represents an encoder, which is a neural network in the form of LSTM+CNN; represents the extracted intrinsic semantic features; represents a decoder, which is a fully convolutional neural network CNN, for signal reconstruction in the model self-supervised learning training process; represents the original input data of the model, represents the reconstructed signal data after model processing.
4. The method of claim 1, wherein the method comprises: The method comprises the following steps: S1, obtaining chromatographic signals, mass spectrometric signals and spectral signals of the detection liquor body and the reference liquor body, forming liquor body sample multi-source data and aligning and registering; S2, converting the liquor body sample multi-source data into high-level semantic features, so that different modal features corresponding to the detection liquor body and the reference liquor body are respectively fused in a high-level semantic space; S3, extracting intrinsic semantic features of the detection liquor body and the reference liquor body from the fused corresponding features, and removing the influence of different batches of reference liquor bodies on the identification result through the intrinsic semantic features of the reference liquor body; S4, obtaining key difference features between the detection liquor body and the reference liquor body as a true-false liquor segmentation boundary according to the intrinsic semantic features of the detection liquor body and the reference liquor body; S5, extracting ingredient coupling relationships of the detection liquor body and the reference liquor body according to the intrinsic semantic features of the detection liquor body and the reference liquor body; S6, identifying the true or false of the detection liquor body according to the key difference features between the detection liquor body and the reference liquor body and the ingredient coupling relationships of the detection liquor body and the reference liquor body, and obtaining an identification result.
5. The authentication method of claim 4, wherein: In step S1, a data normalization method is used to realize amplitude alignment of the liquor body sample multi-source data, as follows: wherein, represents the first original data, represents the data expectation, represents the data standard deviation, represents the normalized data; In the wine body sample to multi-source data, the matching alignment of the sampling rates of the detection wine body signal data and the reference wine body signal data is realized in the way of finding the shortest distance between them; let , , and represent the number of data sampling of the two respectively; the matching point distance of the first detection wine body signal data and the first reference wine body signal data is represented as , and the following optimization function is constructed: wherein represents the total number of matching patterns, represents the best matching pattern among all matching patterns, thereby achieving matching alignment of the sampling rates; After the sampling rate matching alignment is completed, the time sequence of the detection liquor body signal data and the reference liquor body signal data is filled to the same length by linear or nonlinear interpolation, and the time sequences of the two are aligned.
6. The authentication method of claim 4, wherein: In step S2, the embedding of the features is completed by using a convolutional neural network, and the dimension of the high-level semantic features to be converted is respectively using a convolutional neural network to complete the feature embedding of the chromatographic signal, the mass spectrum signal and the spectrum signal; the embedding mode is unified as follows: wherein, represents an embedded high-level semantic feature; represents a feature weight of a final output layer of a convolutional neural network; to represent nonlinear transformation in the 1st to the residual module; represents a residual connection; represents an input of a convolutional neural network; symbol represents a composite operation; represents a bias term of a final output layer of a convolutional neural network; In the embedding process, a scalar feature embedding strategy is used to uniformly complete the embedding of scalar signals and time sequence signals, as follows: wherein represents a scalar signal acquired by the instrument; according to the number of dimensions to constitute the above-mentioned time series data , this time series data is semantically represented by a convolutional neural network, i.e. by using the above-mentioned unified embedding schema; For chromatographic signals, mass spectrum signals and spectral signals, one or more feature parameters to be embedded are included in each type of signal, and after the embedding of all feature parameters is completed, they are uniformly embedded in a semantic space. In the semantic space, different feature parameters are fused by feature connection, as follows: In the formula, represent the corresponding features after fusion; represent the total number of feature parameters that are fused.
7. The authentication method of claim 4, wherein: In step S3, a liquor body intrinsic semantic extraction model is pre-constructed, which uses a twin neural network structure to extract corresponding intrinsic semantic features from the fused features; The model is as follows: In the formula, represents an encoder, which is a neural network in the form of LSTM+CNN; represents the extracted intrinsic semantic features; represents a decoder, which is a fully convolutional neural network CNN, for signal reconstruction in the model self-supervised learning training process; represents the original input data of the model, represents the reconstructed signal data after model processing; The self-supervised learning training of the liquor body intrinsic semantic extraction model is as follows: Reference liquor body data is pre-collected, including the corresponding brand and model of the reference liquor body, and after coding classification, chromatographic signals, mass spectrum signals and spectral signals are formed to construct an intrinsic semantic extraction training database D; During training, data is randomly taken from the training database D and input into the twin neural network of the model for training; the two samples extracted are respectively denoted as and If and belong to the same type of wine body, the corresponding intrinsic semantic features and after training are expected to be similar, otherwise, the difference is expected to be greater than a threshold, so as to distinguish different wine bodies; the loss function of the twin neural network is represented as follows: wherein, represents a symbol function, i.e. if the two inputs belong to the same wine body type, the output is , otherwise the output is ; represents the weight of the control reconstruction error term; represents the weight of the control semantic feature contrast loss term; and are the class labels of the two samples, i.e. the two samples are the same wine body type, ; through repeated iterations of the above training process, i.e. the wine body intrinsic semantic extraction model satisfying the preset requirements is obtained.
8. The authentication method of claim 7, wherein: In step S4, according to the intrinsic semantic features of the detection liquor body and the reference liquor body, before the key difference features are obtained, the liquor body intrinsic semantic extraction model is optimized and trained for the case that the detection liquor body is fake liquor; In the fake liquor optimization training process, for one iteration, a true liquor reference sample is taken from the training database D, and a fake liquor sample is taken from the pre-set fake liquor database, which is input into the twin neural network of the liquor body intrinsic semantic extraction model for training. At this time, the learning rate of the network is reduced by a pre-set percentage; after the fake liquor optimization training is completed, the true and false difference syndrome features, i.e. the key difference features, of the detection liquor body and the reference liquor body for the existing samples are obtained through the encoder, and the features corresponding to the reference liquor body are stored in the true liquor fingerprint library; in actual application and reasoning, the features of various reference liquor bodies are directly taken from the true liquor fingerprint library to realize the discrimination of the detection liquor body and obtain the corresponding key difference features.
9. The authentication method of claim 4, wherein: In step S5, a flavor extraction model is pre-constructed, which uses a graph neural network structure, and the encoder is as follows: In the formula, representing a component feature matrix; representing an adjacency matrix of a graph neural network; representing an encoding function formula of a graph neural network; representing an extracted component coupling relationship; The decoder in the graph neural network structure is as follows: In the formula, representing a reconstructed feature matrix; representing a decoding function formula of the graph neural network When training the flavor extraction model, if the sample input corresponds to the same liquor body model, the contrast loss is expected to be small, otherwise the contrast loss is expected to be greater than a threshold, represented as follows: wherein, represents two samples and belong to the same type of wine body, i.e. positive sample pair; represents two samples and belong to different types of wine body, or one is real wine and one is fake wine, i.e. negative sample pair; represents the distance between two sample feature vectors; represents an interval parameter, used to control the lower limit of the minimum distance between negative samples; The overall loss function of the flavor extraction model is represented as follows: wherein is a heavy loss, is the aforementioned contrast loss, is a contrast loss weight; is represented as follows: wherein a raw feature vector representing a model data set component, a reconstructed corresponding vector; a total number of counts.
10. The authentication method of claim 4, wherein: In step S6, when identifying the true and false of the detection liquor body, a comparative analysis method is used for identification, and a multi-layer fully connected physical neural network is pre-constructed, as follows: In the formula, represent the identification result; represent the key difference characteristics; represent the component coupling relationship; when training the multi-layer fully connected physical neural network, two training scenarios corresponding to the actual identification process are divided, one is single detection of wine body identification, and the other is identification of known detection of wine body corresponding to reference wine body; For single detection liquor body identification, samples are extracted from the pre-constructed true liquor fingerprint library and fake liquor fingerprint library one by one, and the multi-layer fully connected physical neural network is iteratively trained, so that for the input detection liquor body, the corresponding label of true liquor or fake liquor is output. For identifying a known wine type and its corresponding reference wine, a Siamese network with shared weights is trained between the two types of wine. When a sample pair belongs to either genuine or counterfeit wine, the corresponding label value is output; the corresponding loss function is... as follows: In the formula, Representing the The sample label during the second discrimination is 1 if both are genuine wines, and 0 otherwise. Representative of the first When given two samples, the network predicts the probability that they belong to the same category.