Cell culture monitoring method and system based on artificial intelligence

Abstract

The invention relates to the technical field of cell culture monitoring, in particular to a cell culture monitoring method and system based on artificial intelligence, and the method comprises the steps: dividing a target incubator into a plurality of grids, monitoring environmental parameters in each grid, collecting microscope images in the grids according to a preset first sampling frequency, monitoring solution parameters in each grid according to a preset second sampling frequency; performing topological modeling on the target incubator according to the environmental parameters and the plurality of grids to obtain a topological graph structure of the target incubator; and according to an environment offset detection model, environment offset detection is carried out on the target incubator through the topological graph structure so as to obtain an environment offset label of the grid. According to the invention, the incubator is subdivided into a plurality of grids, and microscope images and solution parameters are acquired in each grid in real time, so that the cell growth environment and the dynamic change of cells can be monitored more accurately.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cell culture monitoring, and in particular to a cell culture monitoring method and system based on artificial intelligence. BACKGROUND

[0002] Traditional methods usually rely on manual or low-frequency monitoring, which may result in lagging monitoring data and fail to reflect the rapid changes in cell growth state in real time, which may lead to the failure to discover the deviation or abnormality of the cell culture environment in time, thereby affecting the culture effect. In addition, traditional methods usually lack detailed environmental monitoring, which may fail to identify the subtle differences in each area of the cell incubator. The environmental changes may be different in different grids, but traditional methods are difficult to capture these changes and may ignore the environmental deviation in local areas. Moreover, traditional methods usually require manual intervention to adjust the culture environment. When an abnormality is detected, the operator may need to manually adjust the temperature, humidity and other parameters. This manual adjustment method is not only inefficient but also prone to human errors, which may lead to unstable cell culture environment. Furthermore, the monitoring of solution contamination and cell density by traditional methods is usually not accurate enough to enable in-depth data analysis and risk assessment, which makes it difficult to identify potential solution contamination or abnormal cell growth, and may lead to quality problems in the cell culture process. In addition, traditional methods rely on manual monitoring and management of the cell culture process, which cannot identify risk areas in real time and take corresponding measures. This manual management method is tedious and lacks intelligent decision support, which may lead to untimely risk handling and failure to ensure the optimal state of the culture process. SUMMARY

[0003] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art and provide a cell culture monitoring method and system based on artificial intelligence.

[0004] The technical solution adopted to solve the above technical problems is a cell culture monitoring method based on artificial intelligence, comprising: dividing a target incubator into multiple grids, monitoring the environmental parameters in each grid, collecting microscope images in the grid according to a preset first sampling frequency, and monitoring the solution parameters in each grid according to a preset second sampling frequency; topologically modeling the target incubator according to the environmental parameters and the multiple grids to obtain a topological graph structure of the target incubator; detecting the environmental deviation of the target incubator through the topological graph structure according to an environmental deviation detection model to obtain the environmental deviation label of the grid; perform cell detection on the microscope images according to a pre-trained target detection model to obtain cell detection results of the microscope images, and perform statistics on the cell detection results to obtain cell densities of each grid in the target incubator; perform feature extraction on solution parameters in a plurality of grids in the target incubator to obtain a solution feature set, and perform probability regression on the grids according to the solution feature set to obtain solution contamination probabilities of the grids; weight nodes of the topological graph structure according to the cell densities and the solution contamination probabilities to obtain a weighted topological graph structure, perform risk detection on the target incubator through the weighted topological graph structure according to a risk detection model to obtain risk labels of the grids, and perform environmental adjustment on the grids according to the environmental offset labels and the risk labels.

[0005] Preferably, the environmental parameters include temperature and humidity, the solution parameters include pH value and dissolved oxygen, the solution feature set includes time series features, spatial features, and composite features, and the environmental offset detection model includes a graph convolution network, a graph attention mechanism, and a Softmax function.

[0006] Preferably, the target incubator is topologically modeled according to the environmental parameters and the plurality of grids to obtain a topological graph structure of the target incubator, including: each grid in the target incubator is taken as a topological node of the topological graph structure; determining environmental parameter differences between adjacent grids in the target incubator, wherein the environmental parameter differences include temperature differences and humidity differences, and if the environmental parameter differences are all less than a preset environmental threshold, it indicates that the adjacent grids have spatial correlations; connecting two topological nodes corresponding to the adjacent grids having spatial correlations to obtain connection edges of the topological graph structure, and obtaining a topological node set and a connection point set corresponding to the topological graph structure when all grids in the target incubator are processed; determining the topological graph structure of the target incubator according to the topological node set and the connection point set.

[0007] Preferably, environmental offset detection is performed on the target incubator through the topological graph structure according to an environmental offset detection model to obtain environmental offset labels of the grids, including: determining an adjacency matrix of the topological graph structure according to the topological node set and the connection point set, and normalizing the adjacency matrix to obtain a normalized adjacency matrix; convolving the normalized adjacency matrix according to a graph convolution network to pass spatial information to each topological node to obtain spatial features of the grid; determining attention weights between adjacent grids according to a graph attention mechanism, weighting spatial features of the adjacent grids according to the attention weights to obtain spatial topological features of the grid; performing environment bias classification on the spatial topological features according to a Softmax function to obtain an environment bias probability of the grid, and determining an environment bias label of the grid according to the environment bias probability.

[0008] Preferably, solution parameters in the plurality of grids in the target incubator are extracted to obtain a solution feature set, including: sequencing the solution parameters in the plurality of grids in the target incubator in chronological order to obtain a solution parameter sequence; determining time sequence features of the solution parameter sequence, wherein the time sequence features include statistical features, trend features and dynamic features, wherein the statistical features include mean, variance, maximum and minimum of the solution parameter sequence, the trend features include linear regression slope of the solution parameter sequence, and the dynamic features include solution parameter variation between adjacent solution parameters in the solution parameter sequence; determining spatial features according to solution parameters between adjacent grids in the target incubator; determining a composite feature corresponding to a sampling time determined by a solution parameter variation between adjacent solution parameters in the solution parameter sequence and a second sampling frequency.

[0009] Preferably, the target detection model adopts a Faster R-CNN model.

[0010] Preferably, the risk detection model includes an embedding layer, a graph attention layer, a propagation aggregation layer, and a prediction layer. The embedding layer embeds node features in the weighted topological graph structure to obtain a node feature embedding matrix. The graph attention layer calculates attention weights between topological nodes in the weighted topological graph structure using an attention mechanism. Based on these attention weights, it captures the first-order information of the weighted topological graph structure through a weighted representation and fuses this first-order information using a fusion gate to obtain a fused embedding vector corresponding to the weighted topological graph structure. The propagation aggregation layer uses a graph neural network to perform message passing on the fused embedding vector corresponding to the weighted topological graph structure to obtain layer representations corresponding to the weighted topological graph structure. The prediction layer concatenates each layer representation with a single vector and performs an inner product to obtain the risk probability of the grid. The risk label of the grid is then determined based on the risk probability.

[0011] Preferably, the expression for the attention weight is as follows: ; in, Indicates attention weights, express Activation function , Indicates that the node is composed of Projecting dimensional node space to 3D relation space, Indicates the first Feature embedding matrix of each node via edge Transformation matrix Perform a linear transformation, then add edges. relation vector offset, Represents the hyperbolic tangent activation function; The expression for the first-order information of the weighted topological graph structure is as follows: ; in, Indicate category node First-order information.

[0012] Preferably, the expression for the fused embedding vector is as follows: ; in, Represents the fused embedding vector. This represents the activation function. and This represents the learnable transformation parameters. Indicate category node First-order information; The expressions for each layer corresponding to the weighted topology graph structure are as follows: ; in, In a weighted topological graph structure, the first... Layer representation, express Activation function In a weighted topological graph structure, the first... Layer fusion embedding vector, In a weighted topological graph structure, the first... The fused embedding vector after message passing in the layer.

[0013] The technical solution adopted to solve the above-mentioned technical problems is: an artificial intelligence-based cell culture monitoring system, which is applicable to the aforementioned artificial intelligence-based cell culture monitoring method, comprising: The data monitoring unit is used to divide the target incubator into multiple grids, monitor the environmental parameters in each grid, acquire microscope images in each grid at a preset first sampling frequency, and monitor the solution parameters in each grid at a preset second sampling frequency. A topology modeling unit is used to perform topology modeling on the target incubator based on the environmental parameters and the multiple grids to obtain the topology structure of the target incubator; An offset detection unit is used to perform environmental offset detection on the target incubator according to the environmental offset detection model through the topology graph structure, so as to obtain the environmental offset label of the grid. A cell detection unit is used to perform cell detection on the microscope image according to a pre-trained target detection model to obtain the cell detection results of the microscope image, and to statistically analyze the cell detection results to obtain the cell density of each grid in the target incubator. A contamination regression unit is used to extract features of solution parameters in multiple grids in the target incubator to obtain a solution feature set, and to perform probability regression on the grids based on the solution feature set to obtain the solution contamination probability of the grids. a risk detection unit configured to weight nodes of the topological graph structure according to the cell density and the solution contamination probability to obtain a weighted topological graph structure, and to perform risk detection on the target incubator according to a risk detection model and the weighted topological graph structure to obtain a risk label of the grid, and to perform environmental adjustment on the grid according to the environmental deviation label and the risk label.

[0014] The present application has the following advantages: (1) The present application subdivides the incubator into multiple grids, and collects microscope images and solution parameters in each grid in real time, which can more accurately monitor the cell growth environment and the dynamic changes of cells. The frequent collection of microscope images and solution data ensures real-time understanding of each grid, avoids the monitoring gap or lag problem that may exist in traditional methods, and through the environmental deviation detection model, the incubator environment is intelligently monitored and identified according to the topological graph structure, which can automatically find the environmental deviation, so as to take measures in advance to prevent the negative impact of incubator environmental abnormalities on cell culture; (2) The present application uses the target detection model to detect cells in the microscope image, which, combined with cell density statistics, helps to quantitatively evaluate the growth state of cells, which not only improves the accuracy of cell state analysis, but also provides data support for subsequent cell growth optimization, and through feature extraction of solution parameters and probability regression analysis, the contamination risk of the solution in the incubator can be evaluated, which ensures that the quality of the solution in the cell culture process is effectively monitored, avoiding the interference of solution contamination on cell growth; (3) The present application weights the grid nodes by combining cell density and solution contamination probability information, and detects risks through the weighted topological graph structure, which can intelligently identify potential risk areas and label risk labels, which helps to real-time manage and optimize the cell culture process, ensures that risk points are timely processed, and based on the environmental deviation label and the risk label, the grid can be adjusted, the system can automatically adjust the parameters in the incubator when environmental abnormalities or potential risks are found, realize accurate adjustment of the cell culture environment, and further improve the stability and reliability of cell culture. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A schematic diagram of the step flow of the overall method in an embodiment of the present application is shown in the figure. Figure 2 A schematic diagram of the system architecture of the overall system in an embodiment of the present application is shown in the figure.

[0016] Reference signs: 1, data monitoring unit; 2, topological modeling unit; 3, deviation detection unit; 4, cell detection unit; 5, contamination regression unit; 6, risk detection unit. DETAILED DESCRIPTION

[0017] Embodiment one, as shown in Figure 1 The application provides a cell culture monitoring method based on artificial intelligence, which comprises the following steps: S1, dividing a target incubator into multiple grids, monitoring environmental parameters in each grid, collecting microscope images in the grid according to a preset first sampling frequency, and monitoring solution parameters in each grid according to a preset second sampling frequency; S2, topological modeling of the target incubator according to the environmental parameters and the multiple grids, so as to obtain a topological graph structure of the target incubator; S3, environmental offset detection of the target incubator according to an environmental offset detection model and the topological graph structure, so as to obtain an environmental offset label of the grid; S4, cell detection of the microscope images according to a pre-trained target detection model, so as to obtain a cell detection result of the microscope images, and statistical processing of the cell detection result, so as to obtain cell densities of the grids in the target incubator; S5, feature extraction of the solution parameters in the multiple grids in the target incubator, so as to obtain a solution feature set, and probability regression of the grids according to the solution feature set, so as to obtain solution pollution probabilities of the grids; S6, weighting of nodes of the topological graph structure according to the cell densities and the solution pollution probabilities, so as to obtain a weighted topological graph structure, risk detection of the target incubator according to a risk detection model and the weighted topological graph structure, so as to obtain risk labels of the grids, and environmental adjustment of the grids according to the environmental offset labels and the risk labels.

[0018] In the present application, according to the environmental offset label and the risk label, the grid is properly adjusted, for example, for the area with high cell density or high solution pollution probability, the temperature and humidity, solution composition and other conditions can be adjusted to help improve the environment and optimize cell growth; in the target incubator, the space is divided into multiple small areas (grids) for cell culture and monitoring, each grid represents a sub-area for cell growth; the first sampling frequency refers to the frequency of collecting microscope images, in order to monitor the morphological changes of cells during the culture process, the microscope images are taken at a set frequency; the second sampling frequency refers to the frequency of collecting solution parameters, which helps to ensure that the chemical composition in the solution is suitable for cell growth; the environmental offset detection model is a model based on a topological graph, which is used to detect changes in environmental parameters in the target incubator, to determine whether the environment of a certain area deviates from the normal range, so as to discover potential problems in advance; the cell density refers to the number of cells per unit area or volume in a certain area, and the cell density is an important indicator for measuring cell growth in cell culture, and too high or too low cell density can affect the normal growth of cells; probability regression is a statistical method, which establishes a regression model by using a known solution feature set to predict the solution pollution probability of a certain grid, and the regression model can output a numerical value representing the possibility of solution pollution in the grid.

[0019] In an optional embodiment, the environmental parameters include temperature and humidity, the solution parameters include pH value and dissolved oxygen, the solution feature set includes time series features, spatial features and composite features, and the environmental offset detection model includes a graph convolution network, a graph attention mechanism and a Softmax function.

[0020] In embodiment two, the present application proposes a cell culture monitoring method based on artificial intelligence, compared with embodiment one, the present embodiment further comprises: topological modeling of the target incubator according to the environmental parameters and the plurality of grids to obtain the topological graph structure of the target incubator, comprising: A1, each grid in the target incubator is taken as a topological node of the topological graph structure; A2, determine the environmental parameter difference between adjacent grids in the target incubator, wherein the environmental parameter difference includes temperature difference and humidity difference, and if the environmental parameter difference is less than the preset environmental threshold, it indicates that the adjacent grids have spatial correlation; A3, connect the two topological nodes corresponding to the adjacent grids with spatial correlation to obtain the connection edges of the topological graph structure, and when all the grids in the target incubator are processed, the topological node set and the connection point set corresponding to the topological graph structure are obtained; A4, determine the topological graph structure of the target incubator according to the topological node set and the connection point set.

[0021] In this embodiment, the temperature difference refers to the temperature difference between two adjacent grids. If the temperature difference is large, it means that the environmental temperature difference between the two grids is large. The humidity difference refers to the humidity difference between two adjacent grids. If the humidity difference is large, it means that the humidity conditions of the two grids are different.

[0022] In an optional embodiment, the environmental drift detection model detects the environmental drift of the target incubator through the topology graph structure to obtain the environmental drift label of the grid, including: B1, determining the adjacency matrix of the topology graph structure according to the topology node set and the connection point set, and normalizing the adjacency matrix to obtain the normalized adjacency matrix; B2, performing convolution operation on the normalized adjacency matrix according to the graph convolution network to pass the spatial information to each topology node to obtain the spatial feature of the grid; B3, determining the attention weight between adjacent grids according to the graph attention mechanism, and weighting the spatial features of adjacent grids according to the attention weight to obtain the spatial topology feature of the grid; B4, classifying the spatial topology feature according to the Softmax function to obtain the environmental drift probability of the grid, and determining the environmental drift label of the grid according to the environmental drift probability.

[0023] It should be noted that the graph convolution network is a deep learning model for processing graph data. Through convolution operation, the neighbor information of the node is passed to the target node to realize feature extraction of graph structure data. In the graph convolution network, the adjacency matrix is used as the input, and the convolution operation is used to aggregate the neighborhood information of the node. The graph attention mechanism is an improved graph convolution method. By introducing the attention mechanism, each node can adaptively adjust the weight of each neighbor node when aggregating the information of the neighbor nodes. The attention mechanism usually uses self-attention to calculate the correlation between each pair of nodes. The Softmax function is a commonly used activation function, which is usually used in multi-classification problems. It maps the input vector to a probability distribution, so that the sum of all output values is 1. Here, the Softmax function is used to convert the spatial topology feature into a probability distribution of environmental drift.

[0024] In an optional embodiment, the solution parameters in the plurality of grids in the target incubator are extracted to obtain a solution feature set, including: C1, sorting the solution parameters in the plurality of grids in the target incubator in time sequence to obtain a solution parameter sequence; C2. Determine the time series characteristics of the solution parameter series, where the time series characteristics include statistical characteristics, trend characteristics, and dynamic characteristics. The statistical characteristics include the mean, variance, maximum and minimum values ​​of the solution parameter series. The trend characteristics include the linear regression slope of the solution parameter series. The dynamic characteristics include the change in solution parameters between adjacent solution parameters in the solution parameter series. C3. Determine spatial characteristics based on the solution parameters between adjacent grids in the target incubator; C4. Determine the composite feature by identifying the change in solution parameters between adjacent solution parameters in the solution parameter sequence and the sampling time corresponding to the second sampling frequency.

[0025] It should be noted that the calculation formula for the spatial features is as follows: ; in, and Indicates spatial characteristics, and Indicates the first incubator in the target incubator Solution parameters for each grid Indicates the target incubator and the first The number of adjacent grids of a grid. Indicates the target incubator and the first A set of adjacent grids; The calculation formula for the composite feature is as follows: ; in, Indicates composite features, This represents the change in solution parameter between adjacent solution parameters in a solution parameter sequence. Indicates the sampling time.

[0026] In an optional embodiment, the object detection model uses the Faster R-CNN model.

[0027] It should be noted that Faster R-CNN is a deep learning model for object detection tasks. It improves the efficiency and accuracy of object detection by combining a region proposal network and a CNN. The "Faster" in Faster R-CNN refers to the fact that this model generates region proposals faster and has higher detection accuracy than previous models (such as R-CNN and Fast R-CNN).

[0028] In an optional embodiment, the risk detection model includes an embedding layer, a graph attention layer, a propagation aggregation layer, and a prediction layer. The embedding layer embeds node features in a weighted topological graph structure to obtain a node feature embedding matrix. The graph attention layer calculates attention weights between topological nodes in the weighted topological graph structure using an attention mechanism. Based on these attention weights, it captures first-order information of the weighted topological graph structure through a weighted representation and fuses this first-order information using a fusion gate to obtain a fused embedding vector corresponding to the weighted topological graph structure. The propagation aggregation layer uses a graph neural network to perform message passing on the fused embedding vector corresponding to the weighted topological graph structure to obtain layer representations corresponding to the weighted topological graph structure. The prediction layer concatenates each layer representation with a single vector and performs an inner product to obtain the risk probability of the grid. The risk label of the grid is then determined based on the risk probability.

[0029] It's important to note that in graph neural networks, the embedding layer maps the features of each node in the graph to a low-dimensional vector space. These vectors are called node feature embeddings. Through the embedding layer, the original node features (such as numerical values ​​and categories) are transformed into vector form, enabling the subsequent graph neural network to process these features more efficiently. The graph attention layer is a variant of graph convolutional networks based on the attention mechanism. It adaptively adjusts the way information is propagated by calculating the attention weights between each node in the graph. That is, the influence of each node's neighboring nodes on it is weighted according to the calculated attention weights. The propagation aggregation layer is responsible for message passing in the graph. Each node updates its own representation by performing aggregation operations (such as weighted summation, averaging, etc.) based on the information of its neighboring nodes. This propagation mechanism helps nodes continuously update their own features through the information of their neighbors, thereby better capturing the global information in the graph structure.

[0030] In an optional embodiment, the expression for the attention weights is as follows: ; in, Indicates attention weights, express Activation function , Indicates that the node is composed of Projecting dimensional node space to 3D relation space, Indicates the first Feature embedding matrix of each node via edge Transformation matrix Perform a linear transformation, then add edges. relation vector offset, represents a hyperbolic tangent activation function; The expression of the first-order information of the weighted topological graph structure is as follows: ; wherein, represents the category of the node .

[0031] In an optional embodiment, the expression of the fusion embedding vector is as follows: ; wherein, represents the fusion embedding vector, represents an activation function, and represents a learnable transformation parameter, represents the category of the node . The expression of the representation of each layer corresponding to the weighted topological graph structure is as follows: ; wherein, represents the representation of the layer in the weighted topological graph structure, represents an activation function, represents the fusion embedding vector of the layer in the weighted topological graph structure, represents the fusion embedding vector after message passing of the layer in the weighted topological graph structure.

[0032] Embodiment three, as shown in Figure 2 , the present application proposes a cell culture monitoring system based on artificial intelligence, which is suitable for the cell culture monitoring method based on artificial intelligence, comprising: A data monitoring unit 1 is used to divide the target incubator into multiple grids, monitor the environmental parameters in each grid, collect microscope images in the grid according to a preset first sampling frequency, and monitor solution parameters in each grid according to a preset second sampling frequency. A topological modeling unit 2 is used to topologically model the target incubator according to the environmental parameters and the multiple grids to obtain a topological graph structure of the target incubator. An offset detection unit 3 is used to detect the environmental offset of the target incubator through the topological graph structure according to an environmental offset detection model to obtain an environmental offset label of the grid. The cell detection unit 4 is configured to perform cell detection on the microscope image according to a pre-trained target detection model to obtain a cell detection result of the microscope image, and perform statistics on the cell detection result to obtain cell densities of the grids in the target incubator. The pollution regression unit 5 is configured to perform feature extraction on solution parameters in the grids in the target incubator to obtain a solution feature set, and perform probability regression on the grids according to the solution feature set to obtain solution pollution probabilities of the grids. The risk detection unit 6 is configured to weight nodes of the topological graph structure according to the cell densities and the solution pollution probabilities to obtain a weighted topological graph structure, perform risk detection on the target incubator according to a risk detection model through the weighted topological graph structure to obtain risk labels of the grids, and perform environmental adjustment on the grids according to the environmental offset labels and the risk labels.

[0033] The embodiments of the present application are described in detail above with reference to the drawings, but the present application is not limited thereto, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present application.

Claims

1. A method of cell culture monitoring based on artificial intelligence, characterized by, The method comprises the following steps: dividing a target incubator into multiple grids, monitoring environmental parameters in each of the grids, collecting microscope images in the grids according to a preset first sampling frequency, and monitoring solution parameters in each of the grids according to a preset second sampling frequency; topologically modeling the target incubator according to the environmental parameters and the multiple grids to obtain a topological graph structure of the target incubator; detecting environmental deviation of the target incubator through the topological graph structure according to an environmental deviation detection model to obtain an environmental deviation label of the grid; detecting cells in the microscope images according to a pre-trained target detection model to obtain a cell detection result of the microscope images, and statistically processing the cell detection result to obtain a cell density of each grid in the target incubator; extracting features of the solution parameters in the multiple grids in the target incubator to obtain a solution feature set, and performing probability regression on the grid according to the solution feature set to obtain a solution pollution probability of the grid; weighting nodes of the topological graph structure according to the cell density and the solution pollution probability to obtain a weighted topological graph structure, detecting risks of the target incubator through the weighted topological graph structure according to a risk detection model to obtain a risk label of the grid, and adjusting the environment of the grid according to the environmental deviation label and the risk label.

2. The method of claim 1, wherein the method is based on artificial intelligence. The environmental parameters include temperature and humidity, the solution parameters include PH value and dissolved oxygen, the solution feature set includes time sequence features, spatial features and composite features, and the environmental deviation detection model includes a graph convolution network, a graph attention mechanism and a Softmax function.

3. The method of claim 2, wherein the method is based on artificial intelligence. The method comprises the following steps: taking each of the grids in the target incubator as a topological node of the topological graph structure; determining an environmental parameter difference value between adjacent grids in the target incubator, wherein the environmental parameter difference value includes a temperature difference value and a humidity difference value, and if the environmental parameter difference value is less than a preset environmental threshold value, it indicates that the adjacent grids have spatial correlation; connecting two topological nodes corresponding to the adjacent grids having spatial correlation to obtain a connection edge of the topological graph structure, and obtaining a topological node set and a connection point set corresponding to the topological graph structure when all the grids in the target incubator are processed; determining the topological graph structure of the target incubator according to the topological node set and the connection point set.

4. The method of claim 3, wherein the method is based on artificial intelligence. The method comprises the following steps: determining an adjacency matrix of the topological graph structure according to the topological node set and the connection point set, and normalizing the adjacency matrix to obtain a normalized adjacency matrix; performing convolution operation on the normalized adjacency matrix according to a graph convolution network to transfer spatial information to each topological node to obtain spatial features of the grid; determine the attention weight between adjacent grids according to the graph attention mechanism, and weight the spatial features of the adjacent grids according to the attention weight to obtain the spatial topology features of the grid; classify the spatial topology features according to the Softmax function to obtain the environmental deviation probability of the grid, and determine the environmental deviation label of the grid according to the environmental deviation probability.

5. The method of claim 4, wherein the method is based on artificial intelligence. feature extraction is performed on the solution parameters in the plurality of grids in the target incubator to obtain a solution feature set, including: sequentially sorting the solution parameters in the plurality of grids in the target incubator according to time to obtain a solution parameter sequence; determining the time sequence features of the solution parameter sequence, wherein the time sequence features include statistical features, trend features and dynamic features, wherein the statistical features include the mean, variance, maximum and minimum of the solution parameter sequence, the trend features include the linear regression slope of the solution parameter sequence, and the dynamic features include the solution parameter change amount between adjacent solution parameters in the solution parameter sequence; determining the spatial features according to the solution parameters between adjacent grids in the target incubator; determining the composite features corresponding to the solution parameter change amount between adjacent solution parameters in the solution parameter sequence and the second sampling frequency.

6. The method of claim 5, wherein the method is based on artificial intelligence. The target detection model adopts a Faster R-CNN model.

7. The method of claim 6, wherein the method is based on artificial intelligence. The risk detection model includes an embedding layer, a graph attention layer, a propagation aggregation layer and a prediction layer, wherein the embedding layer is used to embed the node features in the weighted topological graph structure to obtain a node feature embedding matrix, the graph attention layer is used to calculate the attention weight between each topological node in the weighted topological graph structure through an attention mechanism, capture the first-order information of the weighted topological graph structure based on the attention weight between each topological node in the weighted topological graph structure, and fuse the first-order information of the weighted topological graph structure through a fusion gate to obtain a fusion embedding vector corresponding to the weighted topological graph structure, the propagation aggregation layer is used to perform message passing on the fusion embedding vector corresponding to the weighted topological graph structure by using a graph neural network to obtain each layer representation corresponding to the weighted topological graph structure, and the prediction layer is used to connect each layer representation with a single vector and perform inner product to obtain the risk probability of the grid, and determine the risk label of the grid according to the risk probability.

8. The method of claim 7, wherein the method is based on artificial intelligence. The expression of the attention weight is as follows: ； in, Indicates attention weights, express Activation function , Indicates that the node is composed of Dimensional node space projection to 3D relation space, Indicates the first Feature embedding matrix of each node via edge Transformation matrix Perform a linear transformation, then add edges. relation vector offset, Represents the hyperbolic tangent activation function; The expression of the first-order information of the weighted topological graph structure is as follows: ； wherein representing a category node first order information.

9. The method of claim 8, wherein the method is based on artificial intelligence. The expression of the fusion embedding vector is as follows: ； wherein, denotes a fusion embedding vector, denotes an activation function, and denotes a learnable transformation parameter, denotes a class node first-order information; The expression of each layer representation corresponding to the weighted topological graph structure is as follows: ； in, In a weighted topological graph structure, the first... Layer representation, express Activation function In a weighted topological graph structure, the first... Layer fusion embedding vector, In a weighted topological graph structure, the first... The fused embedding vector after message passing in the layer.

10. An artificial intelligence-based cell culture monitoring system suitable for use in an artificial intelligence-based cell culture monitoring method according to claim 9, characterized by including: A data monitoring unit (1) is used to divide a target incubator into a plurality of grids, monitor the environmental parameters in each grid, collect microscope images in the grid according to a preset first sampling frequency, and monitor solution parameters in each grid according to a preset second sampling frequency. a topology modeling unit (2) configured to perform topology modeling on the target incubator according to the environment parameters and the plurality of grids to obtain a topology graph structure of the target incubator; an offset detection unit (3) configured to perform environment offset detection on the target incubator according to the topology graph structure and an environment offset detection model to obtain an environment offset label of the grid; a cell detection unit (4) configured to perform cell detection on the microscope image according to a pre-trained target detection model to obtain a cell detection result of the microscope image, and to perform statistics on the cell detection result to obtain a cell density of each grid in the target incubator; a pollution regression unit (5) configured to perform feature extraction on solution parameters in a plurality of grids in the target incubator to obtain a solution feature set, and to perform probability regression on the grid according to the solution feature set to obtain a solution pollution probability of the grid; a risk detection unit (6) configured to weight nodes of the topology graph structure according to the cell density and the solution pollution probability to obtain a weighted topology graph structure, to perform risk detection on the target incubator according to the weighted topology graph structure and a risk detection model to obtain a risk label of the grid, and to perform environment adjustment on the grid according to the environment offset label and the risk label.

Images