CN119001031B

CN119001031B - A method and system for identifying olfactory characteristics in water based on electronic nose topological fingerprint

Info

Publication number: CN119001031B
Application number: CN202411469292.3A
Authority: CN
Inventors: 周石庆; 黄苑曦; 卜令君; 张可佳; 郭洪光; 伍洋涛; 楚文海
Original assignee: Hunan University
Current assignee: Hunan University
Priority date: 2024-10-21
Filing date: 2024-10-21
Publication date: 2025-01-28
Anticipated expiration: 2044-10-21
Also published as: CN119001031A

Abstract

The invention discloses a method and a system for identifying smell characteristics in water based on an electronic nose topology fingerprint, and relates to the technical field of smell characteristic identification in water; the method comprises the following steps of data acquisition, data preprocessing, topology fingerprint construction, namely converting the preprocessed electrical signal data into topology fingerprints, wherein nodes in the topology fingerprints represent sensors, edges represent interrelations and interaction modes among the sensors, the graph neural network model is processed, the topology fingerprint data is input into the graph neural network model, the graph neural network model extracts and fuses the characteristics of the nodes and neighbor nodes thereof layer by layer through multi-layer graph convolution operation to generate high-level characteristic representation, and the odor category identification and the odor threshold prediction of the odor substances are carried out based on the characteristic representation in the graph neural network model. The invention improves the detection capability of complex odor and low-concentration odor substances.

Description

Method and system for identifying smell characteristics in water based on electronic nose topology fingerprints

Technical Field

The invention belongs to the technical field of recognition of smell characteristics in water, and particularly relates to a recognition method and a recognition system of smell characteristics in water based on an electronic nose topology fingerprint.

Background

The problem of smell in drinking water is one of the environmental problems of general concern throughout the world. The olfactory material not only affects the organoleptic quality of the water, reducing user satisfaction, but may also imply potential health risks. Thus, identifying the odor characteristics of the odorous substances in water is critical to ensuring drinking water safety.

The smell characteristics of the smell material comprise a smell threshold value and a smell category. The olfactory threshold refers to the concentration of a substance that causes minimal stimulation of human olfaction by a olfactory substance. In the treatment of an abnormal odor event in water, it is often necessary to quickly and accurately determine the specific odorogenic substances causing the odor in order to take corresponding measures for the treatment and resolution. By determining the odor threshold of volatile organic compounds in water, potential key odor-causing substances can be rapidly positioned, and targeted treatment measures can be taken. Therefore, the determination of the odor threshold of the odorogenic substances in water has important significance for solving the abnormal odor event, can help to rapidly locate and identify the problem, and provides effective guidance and reference for subsequent water quality management and treatment. In addition to treating off-odor events in water, determining the odor threshold is also widely used in the fields of food, beverages, cosmetics, fragrances, and the like. In these fields, odor thresholds are used to evaluate the quality and safety of products, guiding the formulation and improvement of the products.

In addition, when the problem of smell in drinking water is studied, it is also necessary to classify and identify the type of smell in detail. The smell category refers to the classification of smell substances into different categories according to their organoleptic properties, such as soil mildew smell, fishy smell, chemical taste, etc. Such classification is crucial for understanding the origin and nature of the odorous substances, as the different classes of odors are usually directed to different causes and sources. For example, the soil mildew taste is often produced by microorganisms such as Bacillus licheniformis and is the result of the breakdown of organic matter in water, while the fishy smell may be derived from algae at high concentrations in water and some types of bacterial activity. Identifying these odor categories can not only help to locate problems quickly, but can also more effectively select appropriate water treatment techniques or improvements.

By accurately identifying the odor characteristics, the water quality problem can be monitored and managed more systematically, the pertinence and the efficiency of water treatment work are improved, the relevant departments are helped to quickly respond when an abnormal odor event occurs, and appropriate emergency measures are taken to ensure public health and satisfaction. The existing methods for analyzing the smell substances comprise the following steps:

Common methods for chemical analysis include gas chromatography-mass spectrometry (GC-MS), liquid chromatography (HPLC), and the like. Among them, gas chromatography-mass spectrometry is a commonly used analytical method, which can be used for quantifying and identifying Volatile Organic Compounds (VOCs) in water. By extracting and concentrating organic matters in a water sample, and analyzing the organic matters by using a GC-MS technology, the odor threshold value of different organic matters can be determined, and the correlation of the concentration of the organic matters in the water and human perception can be evaluated. The gas chromatography-sniffing combined technology combines the characteristics of gas chromatography and human perception sniffing, and can directly connect volatile organic compounds in water with human smell. This technique combines GC-separated compounds with the nasal cavity in the sniffing chamber, and the odor threshold of each compound is determined by a trained discriminator. These methods have high sensitivity and high selectivity, but generally require complicated pretreatment steps and expensive equipment, have long operation time and complicated operation, and are not suitable for real-time monitoring.

The sensory evaluation method is characterized in that a professional evaluator performs odor sensory evaluation on the water sample, and the odor intensity and type can be directly reflected, but the influence of subjective factors is large, the standardization is difficult, and the method is not suitable for large-scale monitoring.

Electronic nose technology electronic nose (E-nose) is an intelligent sensor system that simulates the human olfactory system for detecting and identifying complex odors. The electronic nose is composed of a plurality of chemical sensors, each of which responds to a specific odor molecule, and by analyzing the overall response pattern of the sensor array, a different odor can be identified.

The electronic nose mainly comprises a sensor array, a signal acquisition system and a pattern recognition algorithm. The sensor array comprises a plurality of chemical sensors sensitive to different odors, the signal acquisition system is used for acquiring and processing response signals of the sensors, and the pattern recognition algorithm is used for analyzing and classifying odor data. When the odor molecules are in contact with the surface of the sensor, the sensor generates electric signal responses, the response intensities of the different sensors to the odor molecules are different, and the characteristic information of the odor can be obtained and classified by analyzing the response signals. However, the traditional electronic nose has poor adaptability under different environments and background odors, is easily interfered by environmental factors, and causes unstable detection results. And the data processing and pattern recognition algorithms (such as principal component analysis, linear discriminant analysis and the like) of the traditional electronic nose have limited performances when processing high-dimensional and nonlinear data, and are difficult to accurately classify and recognize. The traditional electronic nose has the defects of sensitivity and accuracy in the aspects of detecting complex odor mixtures and low-concentration odor substances, and is difficult to effectively identify and classify.

A graph neural network is a type of machine learning model used to process graph data. The method can effectively capture complex relations among nodes in the graph and perform tasks such as classification, link prediction, attribute prediction and the like on the nodes in the graph. The graph neural network generally builds a deeper model by stacking a plurality of graph convolution layers and full connection layers to improve the performance and learning ability of the model. Meanwhile, some advanced graph neural network models may also include techniques of attention mechanisms, residual connection, graph attention network, etc. to further improve the performance and generalization ability of the model.

The invention provides a method and a system for identifying the smell characteristics in water based on an electronic nose topology fingerprint, which are used for realizing the improvement of sensitivity and accuracy by combining an electronic nose with a Graph Neural Network (GNN), realizing real-time online monitoring and improving environmental adaptability.

Disclosure of Invention

The invention aims to provide a method and a system for identifying the smell characteristics in water based on the topological fingerprint of an electronic nose, which are used for solving the problems that the traditional chemical analysis method in the prior art provided in the background art cannot realize real-time monitoring, the traditional electronic nose has poor environmental adaptability, the limitation of data processing and pattern identification, the insufficient detection sensitivity and accuracy of smell substances and the like.

In order to achieve the above purpose, the invention adopts the following technical scheme:

The first aspect of the invention provides a method for identifying the smell characteristics in water based on the topological fingerprint of an electronic nose, which comprises the following steps:

s1, data acquisition, namely generating corresponding electric signal response by utilizing the reaction of the electronic nose sensor array and the smell substances, and acquiring electric signal data;

S2, preprocessing the data, namely preprocessing the electric signal data, including denoising, filtering and normalizing;

s3, constructing a topology fingerprint, namely converting the preprocessed electric signal data into a specific topology data structure, namely a topology fingerprint, wherein nodes in the topology fingerprint represent sensors, and edges in the topology fingerprint represent the relation among the sensors;

S4, processing the graph neural network model, namely inputting the topological fingerprint data into the graph neural network model, extracting and fusing the characteristics of the nodes and the neighbor nodes thereof layer by layer through multi-layer graph convolution operation by the graph neural network model, and generating high-level characteristic representation;

s5, identifying the smell category and predicting the smell threshold, and identifying the smell category and predicting the smell threshold of the smell substance based on the characteristic representation in the graph neural network model.

Preferably, the topology fingerprint in S3 is constructed as follows:

And constructing edges among the nodes according to the response modes among the sensors, wherein each sensor in the electronic nose sensor array is used as a node in the topological fingerprint, a response signal of the sensor is used as a node characteristic vector, and the edges represent the association among the sensors.

Preferably, the processing of the neural network model in S4 specifically includes the following steps:

s401, constructing a graph neural network model;

The graph neural network model comprises an input layer, a graph roll layer, a pooling layer, a full-connection layer and an output layer which are sequentially connected;

The image convolution layer is used for updating the representation of the nodes by aggregating the neighbor node characteristics of each node so as to capture the local connection mode among the nodes, and the aggregation process comprises weighted aggregation, wherein the aggregated characteristics are then processed through a linear transformation (weight matrix multiplication) to adapt to the input format of the next layer;

The pooling layer is used for downsampling the graph, adopts a multi-scale pooling strategy to capture topological fingerprint information of different levels, so as to better understand and express various smell characteristics, and aggregates node subsets in the topological fingerprint into smaller representations, and retains the overall structure and characteristics of the graph;

The full-connection layer comprises a plurality of neurons, each neuron is connected with the output of the pooling layer, the characteristic transformation and the nonlinear mapping are carried out through the learning weight and the bias, meanwhile, a random inactivation strategy is applied in the full-connection layer to prevent overfitting, and the self-adaptive learning rate adjustment technology is adopted to improve the learning efficiency and the model stability;

The output layer is responsible for specific smell category identification and smell threshold prediction tasks based on the processing of the layers, introduces a multi-task learning framework, and shares bottom layer characteristic representation to optimize classification and prediction tasks simultaneously so as to improve the accuracy and efficiency of prediction;

S402, training a graph neural network model and optimizing;

s403, processing the topological fingerprint data by using the trained graph neural network model.

Preferably, the recognition of the smell category and the prediction of the smell threshold in S5 are specifically as follows:

The output layer of the graph neural network model is utilized to identify the smell category of the smell material and predict the smell threshold value, a classifier and a regression model are adopted in the output layer, and a multi-task learning framework is introduced;

classifying the characteristics processed by the full-connection layer through a classifier, and identifying the odor category of the odor substance in water;

carrying out regression analysis on the characteristics processed by the full-connection layer through a regression model, and predicting the odor threshold value of the odor substances in water;

the underlying feature representation is shared through the multitasking learning framework while optimizing the smell category recognition and smell threshold prediction tasks.

Preferably, the training and optimizing the graph neural network model in S402 specifically includes the following steps:

s4021, acquiring electric signal data of an electronic nose sensor array, processing the electric signal data into topology fingerprint data, and marking smell categories and smell thresholds of smell substances in the topology fingerprint data to form a data set;

s4022, dividing the data set into a training set, a testing set and a verification set, and training and optimizing the neural network model based on the training set.

Preferably, in S403, the topology fingerprint data is processed, specifically as follows:

Carrying out convolution operation on the characteristics of each node and neighbor nodes thereof through a graph convolution layer to generate new node characteristics;

feature aggregation is carried out through the pooling layer, and features of all nodes are aggregated to form a global feature representation of the whole graph;

And carrying out nonlinear transformation through the full connection layer, and transforming the aggregated features through a nonlinear activation function for enhancing the feature expression capability.

The second aspect of the invention provides an in-water smell characteristic identification system based on electronic nose topology fingerprints, which is applied to the method, and comprises the following steps:

The electronic nose sensor array comprises ten different types of sensors and is used for detecting smelly substances in water and generating corresponding electric signal responses;

The signal acquisition and preprocessing module is used for acquiring the electric signals of the sensor and preprocessing data, wherein the data preprocessing comprises denoising, filtering and normalizing operations;

The topology fingerprint construction module is used for constructing the preprocessed sensor data into topology fingerprints, wherein nodes represent sensors, and edges represent the relation among the sensors;

The graph neural network model comprises an input layer, a graph convolution layer, a pooling layer, a full connection layer and an output layer, wherein the graph neural network model performs feature extraction and information fusion on topology fingerprint data through graph convolution operation to generate high-level feature representation, the output layer comprises a classifier and a regression model, the feature representation is subjected to smell category identification of smell substances through the classifier, the feature representation is subjected to regression analysis through the regression model, and smell threshold values of the smell substances are predicted.

Compared with the prior art, the invention has the beneficial effects that:

(1) The analysis system can improve detection sensitivity and accuracy, and the detection capability of low concentration and complex mixed smell is improved by extracting high-level features in sensor response data through multi-layer graph rolling operation through the feature extraction and data fusion of the GNN, wherein the data fusion is to effectively fuse response data of different sensors by utilizing a GNN model, so that the detection capability of weak smell signals is improved, and the detection capability of complex smell and low concentration smell substances is improved.

(2) The analysis system of the invention realizes real-time on-line monitoring, and combines the rapid response characteristic of the electronic nose and the high-efficiency computing capability of the GNN, thereby realizing the real-time on-line monitoring of the odor substances.

(3) The analysis system has high robustness, the electronic nose sensor array has high stability and reliability, can stably work in different environments, the GNN model has high noise resistance, can extract useful information from noise data, and can process high-dimensional and nonlinear data through nonlinear modeling of the GNN, and capture complex topological fingerprint information, so that accuracy and robustness of pattern recognition are improved, and environmental adaptability is further improved.

(4) The analysis system can carry out diversified detection, the sensor array comprises a plurality of sensors, and can detect a plurality of types of smelling substances, including aromatic components, oxynitride, ammonia, sulfide and the like, the GNN model can effectively process a plurality of sensor data, and the smell category and the smell threshold of the smelling substances are comprehensively analyzed.

Drawings

FIG. 1 is a flow chart of a method for identifying the smell characteristics in water based on the topology fingerprint of an electronic nose;

FIG. 2 is a schematic diagram of a topology fingerprint in the present invention;

FIG. 3 is a training diagram of the neural network model of the present invention, wherein (a) is a schematic diagram of the effect of the verification index R2 of the model, and (b) is a schematic diagram of the effect of the verification index RMSE of the model;

FIG. 4 is a graph of model training effect (ODT: odor threshold) for the neural network model of the present invention;

FIG. 5 is a graph of model effectiveness (ODT: odor threshold) of the neural network model of the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Example 1:

the system for identifying the smell characteristics in the water based on the topological fingerprint of the electronic nose comprises the following parts:

The electronic nose sensor array comprises a plurality of chemical sensors of different types, and is used for detecting smelly substances in water and generating corresponding electric signal responses.

In the embodiment, the model of the electronic nose is PEN3 (AIRSENSE ANALYTICS, schwerin, germany), and the sensor array of the electronic nose is specifically shown in table 1.

Table 1 names of sensors in an electronic nose sensor array

Through reasonable combination and configuration of the sensor array, the high-efficiency and accurate detection of various smelly substances in water can be realized. The sensors cooperate to form a comprehensive detection network for the odor substances, can cover a wide range of pollutant types, and provide reliable detection results.

And the signal acquisition and preprocessing module acquires the electric signals of the sensor and performs data preprocessing such as filtering, normalization and the like.

And the topology fingerprint construction module is used for constructing the preprocessed sensor data into topology fingerprints, wherein nodes represent sensors, and edges represent interrelationships and interaction modes among the sensors.

And (3) carrying out feature extraction and information fusion on topology fingerprint data through a graph rolling operation to generate a high-level feature representation by a Graph Neural Network (GNN) model.

In this embodiment, the graph neural network includes the following layers:

An Input Layer (Input Layer) receives a representation of the graph data, and a topology fingerprint is formed by nodes and edges. Each node carries a feature vector of the joint point attribute, and each edge carries a feature vector of the related edge attribute.

And the graph convolution layer (Graph Convolutional Layer) is one of core layers of the graph neural network and is responsible for information transmission and feature extraction on the topological fingerprint. It updates the representation of the nodes by aggregating the neighboring node features of each node in order to capture the local connection patterns between the nodes. The aggregation process involves weighted aggregation, and the aggregated features are then processed through a linear transformation (weight matrix multiplication) to adapt to the input format of the next layer. This structure allows the graph convolution layer to efficiently integrate the information of the surrounding environment while maintaining the local structural information of each node.

Pooling layer (Pooling Layer) the pooling layer is used for downsampling of the graph, aggregating the subset of nodes in the graph into a smaller representation. This helps to reduce computational complexity and improve the efficiency of the model while preserving the overall structure and features of the graph. A multi-scale pooling strategy is employed to capture topological fingerprint information at different levels to better understand and express various smell characteristics.

The fully connected layer (Fully Connected Layer) receives the output of the pooling layer and maps it to one or more target variables. The method can comprise a plurality of neurons, each neuron is connected with the output of the pooling layer, the characteristic transformation and the nonlinear mapping are carried out through the learning weight and the bias, meanwhile, the random inactivation strategy is applied in the full-connection layer to prevent overfitting, and the self-adaptive learning rate adjustment technology is adopted to improve the learning efficiency and the model stability.

Output Layer (Output Layer) the Output Layer is responsible for generating the final prediction result. Based on the processing of the layers, the output layer is responsible for specific smell category identification and smell threshold prediction tasks. In addition, a multi-task learning framework is introduced, and the underlying feature representation is shared to simultaneously optimize classification and prediction tasks so as to improve the accuracy and efficiency of prediction. For different tasks, different activation functions and loss functions can be adopted for the output layer, the sigmoid function is adopted for the two-class tasks, the softmax function is used for the multi-class tasks, the softmax function is used for outputting the probability of each class and is used for smell classification, the linear function is used for regression tasks and the like, the output layer has only one neuron, and a continuous numerical value is output and is used for smell threshold prediction.

The method comprises the steps of classifying the characteristics output by the GNN through a classifier, identifying the smell category of the smell substances in water, carrying out regression analysis on the characteristics output by the GNN through a regression model, and predicting the smell threshold value of the smell substances.

Specifically, the electrical signal of the electronic nose is input into the convolution layer and the full connection layer. In the convolution layer, the image neural network performs feature extraction and updating on the electronic nose signal so as to capture important information of the smell feature. And then, mapping the extracted features through a full connection layer, classifying and predicting the concentration of the smelling substances by utilizing the output features in the GNN model, and finally outputting the predicted smelling features.

And the result output and display module outputs the detection result of the smelling substances, wherein the detection result comprises information such as smell category, smell threshold value and the like.

Referring to fig. 1, the method for analyzing the smelling substances in water by adopting the system comprises the following steps:

Step one, collecting sensor data;

The electronic nose sensor reacts with the smelling substances to generate corresponding electric signals. These electrical signals reflect the characteristic information of the odorous substances.

Step two, signal pretreatment;

The signal acquisition module acquires the electric signal data of the sensor and performs preprocessing including denoising, filtering, normalization and other operations to obtain clear and consistent data.

Step three, constructing topology fingerprints;

The preprocessed sensor data is converted into a specific topology data structure, i.e. "topology fingerprint", as shown in fig. 2. The specific method comprises the following steps:

And each sensor is used as one node in the graph, and the node characteristic vector represents the response signal of the sensor.

Edge-building edges according to response modes among sensors, and representing the association among the sensors.

Fourthly, processing the graph neural network;

The topology fingerprint data is input into a Graph Neural Network (GNN) model. The GNN extracts and fuses the characteristics of the nodes and the neighbor nodes thereof layer by layer through a multi-layer graph rolling operation, and high-level characteristic representation is generated. The specific processing steps of the GNN are as follows:

And (3) carrying out graph convolution, namely carrying out convolution operation on the characteristics of each node and the neighbor nodes thereof, and generating new node characteristics.

Feature aggregation, namely aggregating the features of all nodes to form a global feature representation of the whole graph.

And the nonlinear transformation is to transform the aggregated features through a nonlinear activation function, so as to enhance the feature expression capability.

Step five, classifying and predicting;

And carrying out odor classification and concentration prediction of the odor substances by using the output characteristics of the GNN model. The specific method comprises the following steps:

And (4) classifying the odor category, namely classifying the characteristics output by the GNN through a classifier and identifying the odor category of the odorous substances in the water.

And predicting the smell threshold value of the smell substances by carrying out regression analysis on the characteristics output by the GNN through a regression model.

Step six, outputting results

And outputting the classification and prediction results to a result display module, and displaying the detected smell substance type, concentration and smell intensity thereof.

And (3) experimental verification:

the invention performs training optimization on a Graph Neural Network (GNN) model. The training set, the test set and the verification set effects are gradually improved along with the increase of training rounds to reflect the effective learning and optimization of the model in the training process, the improvement of the training set effects indicates that the model fits training data better, the improvement of the test set and the verification set effects indicates that the generalization capability of the model to unknown data is enhanced, and meanwhile, the improvement of the verification set effects also implies the effectiveness of the super-parameter adjustment of the model. In this case, the model has a stronger adaptability and predictive ability, can predict new data more accurately, and can achieve more superior performance on different data sets. As shown in fig. 3.

The R2 scores of the model on the training set, test set and validation set were 0.82, 0.72 and 0.70, respectively, as shown in fig. 4, indicating good performance of the model. Although the R2 score on the test set and the verification set is slightly lower than that on the training set, a higher level is still achieved, which indicates that the model has good generalization capability and can accurately predict unknown data.

In the external verification process, the invention additionally collects 200 smell threshold data from the literature, wherein the data are completely different from the data used in the modeling process, and aim to verify the generalization capability of the model on brand new data. By preprocessing the data and inputting the data into the trained model, the invention evaluates the prediction performance of the model on new data, and finally, R ² reaches 0.71, as shown in fig. 5, so that the robustness and reliability of the model in practical application are ensured.

The in-water smell characteristic recognition system based on the electronic nose topology fingerprint has the following advantages:

The system rapidly collects data through the electronic nose sensor array, and efficiently processes and analyzes the data through the GNN model, so that the detection time is obviously shortened. The automatic detection process reduces manual operation and intervention, and improves the working efficiency.

The system is simple and easy to use in design, and a user can finish detection of the smelling substances only by performing basic operation. The result output and display module provides visual detection results and detailed analysis reports, and is convenient for users to understand and apply.

The method has the advantages that the combination of the electronic nose sensor array and the GNN model reduces the detection cost, and compared with the traditional laboratory detection method, the method is more economical and practical. The system can be reused, the maintenance cost of the sensor array is low, and the long-term use cost is controllable.

The innovation is that the system combines the electronic nose technology with a Graphic Neural Network (GNN) model for the first time, and provides a brand-new method for detecting the smelly substances in water. The system design has innovativeness in the aspects of sensor data preprocessing, topology fingerprint construction, GNN model training, result output and the like, and fills the blank of the prior art.

The foregoing is only for aiding in understanding the method and the core of the invention, but the scope of the invention is not limited thereto, and it should be understood that the technical scheme and the inventive concept according to the invention are equivalent or changed within the scope of the invention by those skilled in the art. In view of the foregoing, this description should not be construed as limiting the invention.

Claims

1. The method for identifying the smell characteristics in the water based on the topological fingerprint of the electronic nose is characterized by comprising the following steps of:

s3, constructing a topology fingerprint, namely converting the preprocessed electric signal data into the topology fingerprint, wherein nodes in the topology fingerprint represent sensors, and edges represent interrelationships and interaction modes among the sensors;

taking each sensor in the electronic nose sensor array as a node in the topological fingerprint, and taking a response signal of the sensor as a node characteristic vector; constructing edges between nodes according to response modes between the sensors, and representing the association between the sensors;

S401, constructing a graph neural network model, wherein the graph neural network model comprises an input layer, a graph roll layer, a pooling layer, a full-connection layer and an output layer which are sequentially connected;

S402, training a graph neural network model and optimizing;

s4022, dividing the data set into a training set, a testing set and a verification set, and training and optimizing the neural network model based on the training set;

s403, processing topology fingerprint data by using the trained graph neural network model;

Non-linear transformation is carried out through the full connection layer, and the aggregated characteristics are transformed through a non-linear activation function, so that the characteristic expression capacity is enhanced;

S5, identifying the smell category and predicting the smell threshold value, wherein the smell category of the smell material and the smell threshold value are identified based on the characteristic representation in the graph neural network model;

2. The method for identifying the smell characteristics in water based on the topological fingerprint of the electronic nose according to claim 1, wherein the graph neural network model comprises an input layer, a graph roll layer, a pooling layer, a full-connection layer and an output layer which are sequentially connected;

The graph convolution layer is used for updating the representation of the nodes by aggregating the neighbor node characteristics of each node so as to capture the local connection mode among the nodes, wherein the aggregation process comprises weighted aggregation;

A pooling layer for downsampling of the graph, employing a multi-scale pooling strategy to capture topological fingerprint information of different levels, and aggregating node subsets in the topological fingerprint into smaller representations;

And the full-connection layer comprises a plurality of neurons, each neuron is connected with the output of the pooling layer, the characteristic transformation and the nonlinear mapping are carried out through the learning weight and the bias, and meanwhile, a random inactivation strategy is applied in the full-connection layer to prevent overfitting and an adaptive learning rate adjustment technology is adopted.

3. An in-water smell characteristic recognition system based on an electronic nose topology fingerprint for use in the method of claim 1 or 2, comprising: