Cell metabolite detection method based on infrared spectrum multi-scale feature fusion network

Abstract

The invention provides a cell metabolite detection method based on an infrared spectrum multi-scale feature fusion network, which comprises the following steps: collecting time sequence infrared spectrum data and corresponding content labels of cell metabolite, and constructing a data set after preprocessing; designing a deep learning architecture composed of feature coding, double-flow feature extraction, a fusion network and an output network; after the feature coding network extracts spectral local features, the double-flow network obtains global spectral features through remote dependence modeling of spectral branches and analyzes and extracts long-term time-varying features through time dependence analysis of time sequence branches; introducing a spectrum channel importance evaluation matrix into the feature fusion network, performing channel enhancement and redundancy suppression on time-varying features, and extracting deep relation features in combination with a product adjacency matrix; and finally, mapping into a metabolite type and concentration prediction value through a multi-stage linkage output network. The traditional manual analysis bottleneck is broken through, and high-precision quantitative detection is realized through spectrum-time sequence cross-scale feature collaborative optimization.

Description

Technical Field

[0001] The present invention belongs to the technical field of automated monitoring of metabolites in a cell culture process, and in particular relates to a method for detecting cell metabolites based on an infrared spectroscopy multi-scale feature fusion network. Background Art

[0002] Accurate monitoring of metabolites during cell culture is crucial for studying cellular metabolism, playing a key role in maintaining culture system stability, optimizing cell status, and enhancing product expression efficiency. During in vitro culture, cells continuously consume nutrients and produce a variety of metabolites. These metabolites directly reflect cell growth and functional activity, but can also have a feedback effect on the culture environment, thereby regulating cell behavior. Therefore, real-time, precise monitoring of metabolite types and levels in the culture medium is essential for assessing cellular metabolic status and determining whether cells are in optimal growth or expression stages. For example, the consumption rates of glucose and glutamine, and the accumulation levels of lactate and ammonia, are often closely correlated with cell proliferation rate, metabolic stress, and the risk of apoptosis. Failure to promptly monitor and regulate metabolite levels can lead to acidification, toxicity, or nutritional imbalance in the culture environment, compromising cell viability and target product expression efficiency. In large-scale cell culture and bioreactor operations, metabolite monitoring data can also be integrated with process control systems to dynamically optimize parameters such as feeding schedules, gas supply, and pH regulation. Furthermore, in the culture of specialized cell types such as stem cells or immune cells, certain metabolites may serve as key signaling molecules for inducing differentiation, activation, or maintaining function. Therefore, their precise detection is crucial for ensuring the consistency and functionality of cell products. Therefore, qualitative and quantitative analysis of metabolites is not only an important tool for process monitoring and optimization during cell culture, but also the core foundation for achieving high-quality cell products and stable process conditions.

[0003] Currently, metabolite detection relies primarily on highly sensitive and high-resolution instrumental analytical techniques. The most commonly used method is mass spectrometry (MS) combined with liquid chromatography (LC) / gas chromatography (GC), known as LC-MS / GC-MS. This combination enables simultaneous qualitative and quantitative analysis of multiple metabolites in complex biological samples. High-performance liquid chromatography (HPLC) uses a liquid as the mobile phase. A high-pressure infusion system pumps a single solvent of varying polarity, or a mixture of solvents or buffers in varying proportions, through a chromatographic column containing a stationary phase. After separation within the column, the components are detected by a detector, enabling analysis of the sample. In liquid chromatography-mass spectrometry (LC-MS), the individual components of a sample separated by liquid chromatography are sequentially introduced into the mass spectrometer detector. These components are ionized in the ion source, generating ions with distinct charges and mass numbers. Different ions behave differently in an electromagnetic field. The mass analyzer separates these ions according to their mass-to-charge ratio (m / z), producing a mass spectrum arranged in order of mass-to-charge ratio. Analysis and processing of the mass spectrum yields qualitative and quantitative results. Gas chromatography (GC) is a common separation and analysis technique primarily used to detect and quantify the composition of mixed gases or volatile organic compounds. Its operating principle is to separate sample components due to their different distribution coefficients as they pass through a stationary phase (chromatographic column) while being carried by a flowing inert gas (such as nitrogen or helium). A detector then performs qualitative and quantitative analysis of each component. Gas chromatography-mass spectrometry (GC-MS) is a high-end detection technology that combines gas chromatography with mass spectrometry. In this system, the sample is first separated by gas chromatography, where components are separated into individual components based on their retention times on the chromatographic column. These individual components are then individually ionized and mass analyzed by a mass spectrometer, providing accurate mass and molecular structure information for each component. However, these physical and chemical analysis-based metabolite detection methods suffer from long identification times and destructive processing. Therefore, the challenge of non-destructive and efficient monitoring of cellular metabolites has become an urgent issue. Summary of the Invention

[0004] In response to the above technical problems, the present invention provides a method for detecting cell metabolites based on an infrared spectroscopy multi-scale feature fusion network.

[0005] The technical solution adopted by the present invention to solve the technical problem is: A method for detecting cell metabolites based on an infrared spectroscopy multi-scale feature fusion network, the method comprising the following steps: S100: collecting spectral values ​​and contents of cell metabolites at preset time intervals, using the sampling time and spectral values ​​as source data and the corresponding contents as labels, constructing an original monitoring data set, and preprocessing the original monitoring data set to obtain a preprocessed data set; S200: Build an infrared spectrum multi-scale fusion network, including a feature encoding network, a spectral-temporal dual-stream feature extraction network, a feature fusion network, and a hierarchical linkage task result output network; S300: inputting the preprocessed data set into the feature encoding network for processing to obtain the spectral local features after feature aggregation; S400: The local spectral features after feature aggregation are input into the spectral-temporal dual-stream feature extraction network. The spectral branch network generates spectral global features by modeling long-range dependencies. The temporal branch network combines spectral values ​​and sampling time to analyze time dependencies and extract long-term time-varying features of metabolites. S500: Input the spectral global features and time-varying features into the feature fusion network, construct a spectral channel importance evaluation matrix based on the spectral global features, enhance the differential spectral channels and time steps and suppress redundant channels for the time-varying features to obtain fused features; establish a product adjacency matrix based on the fused features to obtain spectral deep relationship features; S600: The multi-level linkage task result output network is used to map the deep spectral relationship features into the corresponding metabolite types and quantitative prediction values ​​of their concentrations; S700: Based on the quantitative prediction values ​​of metabolite types and their concentrations and a preset loss function, the infrared spectroscopy multi-scale fusion network is trained, the network parameters are updated by backpropagation, and the network performance is evaluated to obtain a trained infrared spectroscopy multi-scale fusion network model. Cellular metabolites are monitored based on the trained infrared spectroscopy multi-scale fusion network model.

[0006] Preferably, in S100, the original monitoring data set is preprocessed to obtain a preprocessed data set including: Normalization is used to correct the original cell metabolite monitoring data to reduce the impact of data fluctuations on the detection, specifically: ; in represents the corrected data, Represents the original data, Represents the minimum value in the original data, Indicates the maximum value in the original data.

[0007] Preferably, S300 includes: The spectral values ​​of cell metabolites are extracted along the spectral dimension through a two-dimensional convolution layer to obtain local spectral features. At the same time, feature aggregation is performed to obtain the local spectral features after feature aggregation. The two-dimensional convolution layer includes four consecutive two-dimensional convolution operations, specifically: ; in, is the local feature of the spectrum after feature aggregation, is the spectral value of cell metabolites, Represents a 2D convolution.

[0008] Preferably, the spectral branch network in S400 includes 5 consecutive layers of Transformer blocks, and the spectral branch network in S400 generates spectral global features by long-range dependency modeling, specifically: ; in Indicates the The output of the layer Transformer; Each layer of Transformer block structure is an encoding and decoding structure, specifically: ; ; ; ; ; ; in, are the query, key, and value of attention respectively, for activation function, , are the weight parameters obtained by learning the input features through the linear layer, represents a linear layer, is the obtained global characteristic of the spectrum, is the local feature of the spectrum.

[0009] Preferably, the temporal branch network in S400 includes five consecutive layers of Mamba blocks. The temporal branch network in S400 combines spectral values ​​and sampling time to analyze time dependency and extract long-term time-varying characteristics of metabolites, specifically: ; in Indicates the Layer Mamba's output; The structure of the Mamba network is as follows: ; ; ) ; ; ; in , Represent the discretized mapping matrix, Indicates the length of the walk. Transformation matrix, Represents time-varying features.

[0010] Preferably, in S500, a spectral channel importance evaluation matrix is ​​constructed based on the global spectral features, and the time-varying features are enhanced by different spectral channels and time steps and suppressed by redundant channels to obtain fusion features, including: S510: Global spectral characteristics use Perform shape transformation, then perform global pooling and average pooling to obtain the spectral importance matrix ; S520: Time-varying features use Perform shape transformation, then perform global pooling and average pooling to obtain the time step importance matrix ; S530: After shape transformation and the time step importance matrix Perform matrix multiplication and residual connection to obtain spectral aggregation features ; After shape transformation and the spectral importance matrix Perform matrix multiplication and residual connection to obtain time-varying aggregation features ; Finally and Perform weighted feature fusion to obtain Fusion features, specifically: ; ; ; ; ; in is the global spectral feature, is a time-varying feature, is the spectral importance matrix, is the time step importance matrix, is shape transformation, MaxPool is global pooling, AvgPool is average pooling, Sigmoid is Sigmoid activation function, For fusion features.

[0011] Preferably, in S500, the product adjacency matrix is ​​established based on the fusion features to obtain the deep spectral relationship features, including: S540: Fusion features The shallow metabolite relationship matrix is ​​obtained using global pooling, average pooling and Sigmoid activation function. , and fusion features and shallow metabolite relationship matrix Perform matrix multiplication to obtain shallow relational features ; S550: Shallow relationship features Using the mapping layer and Sigmoid activation function to obtain the deep metabolite relationship matrix , and shallow relationship features and deep metabolite relationship matrix Perform matrix multiplication to obtain deep relational features , specifically: ; ; ; ; in To fusion features, It is a shallow relationship feature. is the shallow metabolite relationship matrix, is the deep metabolite relationship matrix, It is the characteristic of deep spectral relationship.

[0012] Preferably, the operations in S600 are specifically: ; Where out is the infrared spectrum multi-scale fusion network, which represents the quantitative prediction value of metabolite types and their concentrations, and F is the deep relationship feature of the spectrum. For the mapping layer.

[0013] Preferably, S700 includes: S710: Input the preprocessed data set into the infrared spectrum multi-scale fusion model to obtain the product category label prediction and product category content prediction corresponding to the input data; S720: Applying a preset loss function to the class label prediction probability of the network model and the class label in the true label, calculating the difference between the model prediction and the true label, and evaluating the network performance; S730: Based on the difference between the model prediction and the true label, calculate the gradient of the difference with respect to each parameter in each layer of the network. The specific formula is as follows: ; Among them, L is the preset loss function, is the calculated difference value, is the derivative of the activation function, is the output value of the network at this layer, To seek guidance; S740: Update the network weight. The specific formula is as follows: ; in For the The original weights of the layer network, For the The original weight of the layer network is the derivative of the loss function, is the preset learning rate; S750: Repeat the above steps for 300 rounds and the training is completed.

[0014] Preferably, the preset loss function in S700 is specifically: ; ; ; in is the loss value, Predicting loss for metabolite labels, Metabolites The label value of The infrared spectrum multi-scale fusion model Class product label prediction value, Predicting losses for metabolite content, For the Category content of product, The infrared spectrum multi-scale fusion model Predicted value of product category content, Indicates summation, represents the total number of metabolite categories; The network performance is evaluated using the root mean square error (RMSE) and F1-score, where F1-score is: ; ; ; ; in To correctly predict The number of metabolites, The wrong prediction is The number of metabolites, For the actual Metabolites but not predicted quantities, is the total number of metabolite categories, Indicates summation, Indicates the F1-score of metabolite-like products; The RMSE is specifically: ; in It is The true value of the metabolite content, It is the first infrared spectrum multi-scale fusion model Class product label prediction value, Represents the total number of metabolite categories.

[0015] The above-mentioned cell metabolite detection method based on the infrared spectral multi-scale feature fusion network innovatively integrates the spectral spatiotemporal features to achieve end-to-end mapping of infrared data to metabolic parameters, breaking through the bottleneck of traditional manual analysis. Through dual-stream feature collaborative extraction, three-dimensional attention dynamic fusion and multi-task linkage output design, it significantly improves the detection accuracy of the product, provides real-time monitoring and abnormal warning for the biopharmaceutical process, and is suitable for high-value scenarios such as vaccine and anti-cancer drug development. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of a method for detecting cell metabolites based on an infrared spectroscopy multi-scale feature fusion network in one embodiment of the present invention; Figure 2 This is a schematic diagram of a multi-scale fusion network based on infrared spectroscopy in one embodiment of the present invention; Figure 3 Schematic diagram of a feature fusion network in one embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below with reference to the accompanying drawings.

[0018] In one embodiment, Figure 1 and Figure 2 As shown, a method for detecting cell metabolites based on an infrared spectroscopy multi-scale feature fusion network comprises the following steps: S100: Collecting spectral values ​​and contents of cell metabolites at preset time intervals, using the sampling time and spectral values ​​as source data and the corresponding contents as labels to construct an original monitoring data set, and preprocessing the original monitoring data set to obtain a preprocessed data set; in this embodiment, the contents and spectral values ​​of cell metabolites are acquired every 0.5 hours; S200: Build an infrared spectrum multi-scale fusion network, including a feature encoding network, a spectral-temporal dual-stream feature extraction network, a feature fusion network, and a hierarchical linkage task result output network; S300: inputting the preprocessed data set into the feature encoding network for processing to obtain the spectral local features after feature aggregation; S400: The local spectral features after feature aggregation are input into the spectral-temporal dual-stream feature extraction network. The spectral branch network generates spectral global features by modeling long-range dependencies. The temporal branch network combines spectral values ​​and sampling time to analyze time dependencies and extract long-term time-varying features of metabolites. S500: Input the spectral global features and time-varying features into the feature fusion network, construct a spectral channel importance evaluation matrix based on the spectral global features, enhance the differential spectral channels and time steps and suppress redundant channels for the time-varying features to obtain fused features; establish a product adjacency matrix based on the fused features to obtain spectral deep relationship features; S600: The multi-level linkage task result output network is used to map the deep spectral relationship features into the corresponding metabolite types and quantitative prediction values ​​of their concentrations; S700: Based on the quantitative prediction values ​​of metabolite types and their concentrations and a preset loss function, the infrared spectroscopy multi-scale fusion network is trained, and the network parameters are updated by backpropagation to obtain a trained infrared spectroscopy multi-scale fusion network model. Cellular metabolites are monitored based on the trained infrared spectroscopy multi-scale fusion network model.

[0019] The core innovation of this method for detecting cellular metabolites, based on a multi-scale infrared spectroscopy feature fusion network, lies in the construction of an end-to-end model for spatiotemporal feature fusion, replacing the traditional manual analysis process. Through spectral-temporal dual-stream collaborative modeling, three-dimensional attention dynamic fusion, and a multi-task linkage mechanism, high-precision detection of products such as antibodies and recombinant proteins is achieved. This method supports zero-sample loss monitoring during bioreactor cultivation and provides real-time warnings of metabolic anomalies. It is suitable for biomanufacturing scenarios such as vaccine and anticancer drug development, significantly improving production efficiency and quality control.

[0020] In one embodiment, preprocessing the original monitoring data set in S100 to obtain the preprocessed data set includes: correcting the original cell metabolite monitoring data using normalization to reduce the impact of data fluctuations on detection, specifically: ; in represents the corrected data, Represents the original data, Represents the minimum value in the original data, Indicates the maximum value in the original data.

[0021] In one embodiment, S300 includes: The spectral values ​​of cell metabolites are extracted along the spectral dimension through a two-dimensional convolution layer to obtain local spectral features. At the same time, feature aggregation is performed to obtain the local spectral features after feature aggregation. The two-dimensional convolution layer includes four consecutive two-dimensional convolution operations, specifically: ; in, is the local feature of the spectrum after feature aggregation, is the spectral value of cell metabolites, Represents 2D convolution. This can further improve the efficiency of subsequent feature extraction modules and reduce the amount of subsequent processing computation.

[0022] Furthermore, the spectral branch network in the spectral-temporal dual-stream module is used to model the long-range dependency of the spectral local features of cellular metabolites to obtain the spectral global features; the temporal branch network in the spectral-temporal dual-stream module is used to extract the long-term dependency of the metabolites over time based on the spectral values ​​and sampling times of the cellular metabolites.

[0023] In one embodiment, the spectral branch network in S400 includes five consecutive layers of Transformer blocks. The spectral branch network in S400 generates spectral global features by long-range dependency modeling, specifically: ; in Indicates the The output of the layer Transformer; Each layer of Transformer block structure is an encoding and decoding structure, specifically: ; ; ; ; ; ; in, are the query, key, and value of attention respectively, for activation function, , are the weight parameters obtained by learning the input features through the linear layer, represents a linear layer, is the obtained global characteristic of the spectrum, is the local feature of the spectrum.

[0024] In one embodiment, the temporal branch network in S400 includes five consecutive layers of Mamba blocks. The temporal branch network in S400 combines spectral values ​​and sampling time to analyze time dependencies and extract long-term time-varying features of metabolites, specifically: ; in Indicates the Layer Mamba's output; The structure of the Mamba network is as follows: ; ; ) ; ; ; in , Represent the discretized mapping matrix, Indicates the length of the walk. Transformation matrix, Represents time-varying features.

[0025] Furthermore, a spectral channel importance evaluation matrix is ​​constructed based on the global spectral features. Based on this importance evaluation matrix, the time-varying features are enhanced by different spectral channels and time steps, and redundant channels are suppressed to obtain fused features. A product adjacency matrix is ​​then established based on the fused features, in which each metabolite corresponds to a eigenvector. The inner product of the eigenvectors represents the degree of interaction or association between metabolites. This allows for learning the interactions and importance between metabolites and deriving deep features.

[0026] In one embodiment, Figure 3 As shown, in S500, a spectral channel importance evaluation matrix is ​​constructed based on the global spectral features, and the time-varying features are enhanced by different spectral channels and time steps and redundant channels are suppressed to obtain fusion features, including: S510: Global spectral characteristics use Perform shape transformation, then perform global pooling and average pooling to obtain the spectral importance matrix ; S520: Time-varying features use Perform shape transformation, then perform global pooling and average pooling to obtain the time step importance matrix ; S530: After shape transformation and the time step importance matrix Perform matrix multiplication and residual connection to obtain spectral aggregation features ; After shape transformation and the spectral importance matrix Perform matrix multiplication and residual connection to obtain time-varying aggregation features ; Finally and Perform weighted feature fusion to obtain Fusion features, specifically: ; ; ; ; ; in is the global spectral feature, is a time-varying feature, is the spectral importance matrix, is the time step importance matrix, is shape transformation, MaxPool is global pooling, AvgPool is average pooling, Sigmoid is Sigmoid activation function, For fusion features.

[0027] In one embodiment, Figure 3 As shown, in S500, a product adjacency matrix is ​​established based on the fusion features to obtain the deep relationship features of the spectrum, including: S540: Fusion features The shallow metabolite relationship matrix is ​​obtained using global pooling, average pooling and Sigmoid activation function. , and fusion features and shallow metabolite relationship matrix Perform matrix multiplication to obtain shallow relational features ; S550: Shallow relationship features Using the mapping layer and Sigmoid activation function to obtain the deep metabolite relationship matrix , and shallow relationship features and deep metabolite relationship matrix Perform matrix multiplication to obtain deep relational features , specifically: ; ; ; ; in To fusion features, It is a shallow relationship feature. is the shallow metabolite relationship matrix, is the deep metabolite relationship matrix, It is the characteristic of deep spectral relationship.

[0028] In one embodiment, the operations in S600 are specifically as follows: ; Where out is the infrared spectrum multi-scale fusion network, which represents the quantitative prediction value of metabolite types and their concentrations, and F is the deep relationship feature of the spectrum. For the mapping layer.

[0029] In one embodiment, S700 includes: S710: Input the preprocessed data set into the infrared spectrum multi-scale fusion model to obtain the product category label prediction and product category content prediction corresponding to the input data; S720: Applying a preset loss function to the class label prediction probability of the network model and the class label in the true label, calculating the difference between the model prediction and the true label, and evaluating the network performance; S730: Based on the difference between the model prediction and the true label, calculate the gradient of the difference with respect to each parameter in each layer of the network. The specific formula is as follows: ; Among them, L is the preset loss function, is the calculated difference value, is the derivative of the activation function, is the output value of the network at this layer, To seek guidance; S740: Update the network weight. The specific formula is as follows: ; in For the The original weights of the layer network, For the The original weight of the layer network is the derivative of the loss function, is the preset learning rate; S750: Repeat the above steps for 300 rounds and the training is completed.

[0030] In this embodiment, during training, use The hierarchical neural model was trained using an optimizer with a learning rate of 0.001, a momentum parameter of 0.9, a weight decay of 0.001, a batch size of 1, and a regularization lambda of 0.5. During testing, the data was preprocessed and fed into the trained feature extraction model to convert the hyperspectral image into a feature vector representation, which was then fed into the classifier for prediction. The classifier loss was calculated using Binary Cross-Entropy (BCE) and Mean Squared Error (MSE), and evaluated using Root Mean Squared Error (RMSE) and F1-score.

[0031] In one embodiment, the preset loss function in S700 is specifically: ; ; ; in is the loss value, Predicting loss for metabolite labels, Metabolites The label value of The infrared spectrum multi-scale fusion model Class product label prediction value, Predicting losses for metabolite content, For the Category content of product, The infrared spectrum multi-scale fusion model Predicted value of product category content, Indicates summation, represents the total number of metabolite categories; The network performance is evaluated using the root mean square error (RMSE) and F1-score, where F1-score is: ; ; ; ; in To correctly predict The number of metabolites, The wrong prediction is The number of metabolites, For the actual Metabolites but not predicted quantities, is the total number of metabolite categories, Indicates summation, Indicates the F1-score of metabolite-like products; The RMSE is specifically: ; in It is The true value of the metabolite content, It is the first infrared spectrum multi-scale fusion model Class product label prediction value, Represents the total number of metabolite categories.

[0032] Furthermore, after the model is established, the type and content of metabolites can be monitored only with the spectral value and sampling time of the cell product to be tested. It has the advantages of being objective, simple, fast and non-destructive, and can be used for batch identification.

[0033] The core innovation of this method for detecting cellular metabolites based on a multi-scale infrared spectral feature fusion network lies in the construction of a multi-scale infrared spectral fusion network model, which achieves a direct end-to-end mapping from raw infrared spectral data to cellular metabolites, eliminating the cumbersome manual monitoring steps required in traditional methods. The invention comprises the following design: first, a spectral-temporal dual-stream feature extraction network is constructed, in which a 2D convolution combined with a Transformer captures global and local features of molecular bond vibrations; a Mamba temporal network extracts time-varying spectral features during the culture phase; second, a three-dimensional attention network is introduced, which, through the synergistic effect of a spectral channel importance evaluation matrix and a metabolite association module, achieves adaptive focusing on feature regions and intelligent suppression of interfering signals; finally, a hierarchical task result output network is designed to simultaneously output the type and content of products (such as monoclonal antibodies and recombinant proteins). A metabolic balance prediction algorithm is embedded in the output layer, providing early warning of metabolic abnormalities. This method provides a high-precision, fully automated, and zero-sample-loss metabolite monitoring solution for biomanufacturing fields such as vaccine production and anticancer drug development.

[0034] The above is a detailed introduction to the cell metabolite detection method based on the infrared spectroscopy multi-scale feature fusion network provided by the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the core idea of ​​the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, the present invention can also be improved and modified in several ways, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting cell metabolites based on an infrared spectroscopy multi-scale feature fusion network, characterized in that: The method comprises the following steps: S100: collecting the spectral values, contents, and sampling times of cell metabolites at preset time intervals, constructing an original monitoring data set with the spectral-time data as input and the contents as labels, and obtaining a preprocessed data set after preprocessing; S200: Build an infrared spectrum multi-scale fusion network, including a feature encoding network, a spectral-temporal dual-stream feature extraction network, a feature fusion network, and a hierarchical linkage task result output network; S300: inputting the preprocessed data set into the feature encoding network to extract local spectral features; S400: The local spectral features are input into the spectral-temporal dual-stream feature extraction network. The spectral branch network generates spectral global features by modeling long-range dependencies. The temporal branch network combines spectral values ​​and sampling time to analyze temporal dependencies and extract long-term time-varying features of metabolites. S500: Input the spectral global features and time-varying features into the feature fusion network, construct a spectral channel importance evaluation matrix based on the spectral global features, enhance the differential spectral channels and time steps and suppress redundant channels for the time-varying features to obtain fused features; establish a product adjacency matrix based on the fused features to obtain spectral deep relationship features; S600: The multi-level task result output network maps the deep spectral relationship features into the corresponding metabolite types and quantitative prediction values ​​of their concentrations; S700: Based on the quantitative prediction values ​​of metabolite types and their concentrations and a preset loss function, the infrared spectroscopy multi-scale fusion network is trained, the network parameters are updated by backpropagation, and the network performance is evaluated to obtain a trained infrared spectroscopy multi-scale fusion network model. Cellular metabolites are monitored based on the trained infrared spectroscopy multi-scale fusion network model.

2. The method according to claim 1, characterized in that In S100, the original monitoring data set is preprocessed to obtain a preprocessed data set including: Normalization is used to correct the original cell metabolite monitoring data to reduce the impact of data fluctuations on the detection, specifically: ； in represents the corrected data, Represents the original data, Represents the minimum value in the original data, Indicates the maximum value in the original data.

3. The method according to claim 2, characterized in that S300 includes: The spectral values ​​of cell metabolites are extracted along the spectral dimension through a two-dimensional convolution layer to obtain local spectral features. At the same time, feature aggregation is performed to obtain the local spectral features after feature aggregation. The two-dimensional convolution layer includes four consecutive two-dimensional convolution operations, specifically: ； in, is the local feature of the spectrum after feature aggregation, is the spectral value of cell metabolites, Represents a 2D convolution.

4. The method according to claim 3, characterized in that The spectral branch network in S400 consists of five consecutive layers of Transformer blocks. The spectral branch network in S400 generates spectral global features by modeling long-range dependencies, specifically: ； in Indicates the The output of the layer Transformer; Each layer of Transformer block structure is an encoding and decoding structure, specifically: ； ； ； ； ； ； in, are the query, key, and value of attention respectively, for activation function, , , are the weight parameters obtained by learning the input features through the linear layer, represents a linear layer, is the obtained global characteristic of the spectrum, is the local feature of the spectrum.

5. The method according to claim 4, characterized in that The temporal branch network in S400 consists of five consecutive layers of Mamba blocks. The temporal branch network in S400 combines spectral values ​​and sampling time to analyze time dependencies and extract long-term time-varying features of metabolites. Specifically: ； in Indicates the Layer Mamba's output; The structure of the Mamba network is as follows: ； ； )； ； ； in , and Represent continuous mapping matrices, , Represent the discretized mapping matrix, Indicates the length of the walk. represents the input spectral value of cell metabolites, Transformation matrix, Represents time-varying features.

6. The method according to claim 5, characterized in that In S500, a spectral channel importance evaluation matrix is ​​constructed based on the global spectral features. The time-varying features are enhanced by different spectral channels and time steps and redundant channels are suppressed to obtain fusion features, including: S510: Global spectral characteristics use Perform shape transformation, then perform global pooling and average pooling to obtain the spectral importance matrix ; S520: Time-varying features use Perform shape transformation, then perform global pooling and average pooling to obtain the time step importance matrix ; S530: After shape transformation and the time step importance matrix Perform matrix multiplication and residual connection to obtain spectral aggregation features ; After shape transformation and the spectral importance matrix Perform matrix multiplication and residual connection to obtain time-varying aggregation features ; Finally and Perform weighted feature fusion to obtain Fusion features, specifically: ； ； ； ； ； in is the global spectral feature, is a time-varying feature, is the spectral importance matrix, is the time step importance matrix, is the spectral aggregation feature, is the time-varying aggregation feature, is shape transformation, MaxPool is global pooling, AvgPool is average pooling, Sigmoid is Sigmoid activation function, For fusion features.

7. The method according to claim 6, characterized in that In S500, the product adjacency matrix is ​​established based on the fusion features to obtain the deep relationship features of the spectrum, including: S540: Fusion features The shallow metabolite relationship matrix is ​​obtained using global pooling, average pooling and Sigmoid activation function. , and fusion features and shallow metabolite relationship matrix Perform matrix multiplication to obtain shallow relational features ; S550: Shallow relationship features Using the mapping layer and Sigmoid activation function to obtain the deep metabolite relationship matrix , and shallow relationship features and deep metabolite relationship matrix Perform matrix multiplication to obtain deep relational features , specifically: ； ； ； ； in To fusion features, It is a shallow relationship feature. is the shallow metabolite relationship matrix, is the deep metabolite relationship matrix, It is the characteristic of deep spectral relationship.

8. The method according to claim 7, characterized in that The specific operations in S600 are: ； Where out is the infrared spectrum multi-scale fusion network, which represents the quantitative prediction value of metabolite types and their concentrations, and F is the deep relationship feature of the spectrum. For the mapping layer.

9. The method according to claim 8, characterized in that The S700 includes: S710: Input the preprocessed data set into the infrared spectrum multi-scale fusion model to obtain the product category label prediction and product category content prediction corresponding to the input data; S720: Applying a preset loss function to the class label prediction probability of the network model and the class label in the true label, calculating the difference between the model prediction and the true label, and evaluating the network performance; S730: Based on the difference between the model prediction and the true label, calculate the gradient of the difference with respect to each parameter in each layer of the network. The specific formula is as follows: ； Among them, L is the preset loss function, is the calculated difference value, is the derivative of the activation function, is the output value of the network at this layer, To seek guidance; S740: Update the network weight. The specific formula is as follows: ； in For the The original weights of the layer network, For the The original weight of the layer network is the derivative of the loss function, is the preset learning rate; S750: Repeat the above steps for 300 rounds and the training is completed.

10. The method according to claim 9, characterized in that The preset loss function in S700 is specifically: ； ； ； in is the loss value, Predicting loss for metabolite labels, Metabolites The label value of The infrared spectrum multi-scale fusion model Class product label prediction value, Predicting losses for metabolite content, For the Category content of product, The infrared spectrum multi-scale fusion model Predicted value of product category content, Indicates summation, represents the total number of metabolite categories; The network performance is evaluated using the root mean square error (RMSE) and F1-score, where F1-score is: ； ； ； ； in To correctly predict The number of metabolites, The wrong prediction is The number of metabolites, For the actual Metabolites but not predicted quantities, is the total number of metabolite categories, Indicates summation, Indicates the F1-score of metabolite-like products; The RMSE is specifically: ； in It is The true value of the metabolite content, It is the first infrared spectrum multi-scale fusion model Class product label prediction value, Represents the total number of metabolite categories.

Images