Pressure-flow rate cooperative control method and system

Abstract

The invention relates to the field of medical instruments, and discloses a pressure-flow rate cooperative control method and system, and the method comprises the steps: obtaining standardized operation data of a flow cytometer fluid path; constructing a liquid path sparse Bayesian causal graph structure; performing feature screening and weight addition on the sparse Bayesian causal graph structure of the liquid path to obtain a multi-level risk probability of each time window; adjusting the importance coefficient of nodes in the liquid path sparse Bayesian causality graph structure by using the multi-level risk probability; using a liquid path sparse Bayesian causal graph structure to constrain the reinforcement learning strategy space to obtain a liquid path parameter adjustment suggestion; a liquid path parameter adjustment suggestion is combined with PID control to form a pressure and flow rate dual-target adaptive control instruction set; and executing the adaptive control instruction set, and driving continuous evolution of the liquid path parameter adjustment suggestions. According to the invention, the flow velocity is stable, the repeatability of cell concentration, absolute counting and signal intensity is improved, and high-quality quantitative analysis is supported.

Description

Technical Field

[0001] This invention relates to the field of medical devices, and more specifically, to a pressure-flow rate coordinated control method and system. Background Technology

[0002] Flow cytometry is a technology for high-speed, multi-parameter quantitative analysis of cells arranged in a single file in a fluid stream. Its core principle utilizes sheath fluid focusing technology, causing cells in a cell suspension to arrange themselves in a single file at the center of the flow and pass through a laser detection point at high speed one by one. When cells are irradiated by the laser, they produce scattered light related to their size and internal structure. If the cells are labeled with fluorescent dyes, they will also emit fluorescence at specific wavelengths. These light signals are received by the detector and converted into electrical signals, which are then processed by a computer to achieve cell population counting and phenotypic analysis.

[0003] In this system, stable sample flow rate is crucial, serving as the cornerstone of all quantitative data. Stable flow rate first ensures a constant number of cells detected per unit time, thereby guaranteeing the accuracy of cell concentration and absolute count. Second, it fixes the time each cell spends passing through the laser beam, ensuring the reliability and comparability of fluorescence and scattered light signal intensities. Unstable flow rate leads to signal fluctuations, decreased resolution, and fluid flow jitter, directly impairing the accuracy and reproducibility of the data and making effective comparisons between different batches impossible. Therefore, maintaining stable flow rate is a critical prerequisite for obtaining reliable, high-quality data in flow cytometry analysis.

[0004] In addition, the following specific issues are also included:

[0005] 1. How to maintain the synergistic relationship between pressure and flow rate in hydraulic circuit control, and dynamically adapt to changes in variables under different operating conditions.

[0006] 2. How to avoid the policy deviating from the real physical causal relationship in reinforcement learning fluid circuit control training, thereby improving its usability in actual fluid circuits.

[0007] 3. Traditional standalone PID can only handle short-term errors and cannot optimize globally, while standalone reinforcement learning (RL) is slow to respond to instantaneous changes. How can the two be combined?

[0008] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0009] In view of the problems in related technologies, this invention proposes a pressure-flow rate coordinated control method and system to overcome the above-mentioned technical problems existing in the existing related technologies.

[0010] Therefore, the specific technical solution adopted by the present invention is as follows:

[0011] According to a first aspect of the present invention, a pressure-flow rate coordinated control method is provided, comprising:

[0012] S1. Collect the operating data of the flow cytometer's fluid circuit, including flow chamber pressure data, waste liquid tank pressure data, sheath fluid pump speed, waste liquid pump speed, valve opening ratio, and sample channel flow rate;

[0013] S2. Perform time synchronization and unit unification processing on the operational data to obtain standardized operational data;

[0014] S3. Based on time series samples of standardized operating data, construct a sparse Bayesian causal graph structure for the liquid path to ensure that the relationship between target variables is preserved under different operating conditions. Perform feature filtering and weighting on the sparse Bayesian causal graph structure to improve the sensitivity of the risk prediction network to the coordinated pressure-flow rate pattern and obtain multi-level risk probabilities for each time window. In the next time window, use the multi-level risk probabilities to adjust the importance coefficients of nodes in the sparse Bayesian causal graph structure for the liquid path.

[0015] S4. Use the sparse Bayesian causal graph structure of the fluid path to constrain the reinforcement learning policy space. Combine the importance coefficient of the nodes to make the policy optimization focus on the target variable and improve the stability of the control payoff, and obtain suggestions for adjusting the fluid path parameters.

[0016] S5. Combine the fluid circuit parameter adjustment suggestions with PID control to form an adaptive control instruction set with dual objectives of pressure and flow rate;

[0017] S6. Execute the adaptive control instruction set and drive the continuous evolution of hydraulic circuit parameter adjustment suggestions to update the adaptive control instruction set.

[0018] Furthermore, based on time-series samples of standardized operational data, the sparse Bayesian causal graph structure of the liquid path is constructed, including:

[0019] Pre-set time windows and construct time series samples from standardized operational data;

[0020] The start and end times of the sequence are appended to the data blocks of each time series sample to ensure the temporal causality of the causal analysis;

[0021] Initialize a temporal Bayesian graph containing all variable connections, apply a sparsity prior to each edge in the temporal Bayesian graph, and calculate the significance probability.

[0022] Edges with a significance probability below a probability threshold in the temporal Bayes graph are pruned to preserve the target causal path;

[0023] The conditional probability table of the remaining edges in the temporal Bayesian graph is updated using Bayesian posterior method. The updated temporal Bayesian graph is then saved to obtain the liquid path sparse Bayesian causal graph structure, which focuses on the relationship of the target variable.

[0024] Furthermore, feature filtering and weighting are applied to the sparse Bayesian causal graph structure of the liquid path to improve the sensitivity of the risk prediction network to the cooperative pressure-flow velocity pattern, and to obtain the multi-level risk probability for each time window, including:

[0025] The edge weights in the sparse Bayesian causal graph structure of the liquid path are assigned according to the significance probability; the feature weight vector is assigned according to the edge weights, and the features in the time series samples are weighted to obtain the weighted feature vector.

[0026] The weighted feature vector is input into the risk prediction neural network to reduce interference from irrelevant variables; the risk prediction neural network outputs multi-level risk probabilities for each time window.

[0027] In the risk prediction neural network, during backpropagation updates, the weight decay mechanism is correlated with the criticality of the liquid-channel sparse Bayesian causal graph structure.

[0028] Furthermore, the weighted feature vector is input into the risk prediction neural network to reduce interference from irrelevant variables; the risk prediction neural network outputs multi-level risk probabilities for each time window, including:

[0029] Obtain the weighted feature vector, and adjust the values ​​of each dimension of the weighted feature vector according to the edge weight or node importance to reduce the influence of irrelevant variables on the prediction results.

[0030] The weighted feature vectors are processed according to the input requirements and then fed into the input layer of the risk prediction neural network.

[0031] The risk prediction neural network performs weighted summation, nonlinear activation, and feature mapping on the input vector layer by layer to extract deep feature patterns related to the risk status of the liquid circuit, and generates the initial value of the probability distribution of the corresponding risk status in the output layer.

[0032] The output layer of the risk prediction neural network generates multi-level risk probability values ​​according to the set risk level categories; the output multi-level risk probability values ​​are then normalized.

[0033] Furthermore, adjusting the importance coefficients of nodes in the sparse Bayesian causal graph structure of the liquid path using multi-level risk probabilities includes:

[0034] Obtain multi-level risk probabilities, including first risk probability, second risk probability, and third risk probability, and treat the target variable node as a liquid circuit risk node;

[0035] Based on the sparse Bayesian causal graph structure of the liquid path, find the required set of nodes that have causal paths with the risk nodes of the liquid path;

[0036] Calculate the risk weight factor based on the multi-level risk probability vector, and use the risk weight factor to update the node importance of the required node set;

[0037] The adjusted node importance is written back to the node attributes of the sparse Bayesian causal graph.

[0038] Furthermore, after adjusting the importance coefficients of nodes in the sparse Bayesian causal graph structure of the liquid path using multi-level risk probabilities, the following is also included:

[0039] By incorporating the node importance coefficient as part of the sparsity prior, nodes with higher importance coefficients are more difficult to prune during saliency calculation.

[0040] When selecting features from a sparse Bayesian causal graph structure in a liquid path, the node importance coefficient determines the priority.

[0041] By adding the node importance coefficient to the weighting term of the reinforcement learning reward function, reinforcement learning prioritizes the control of the fluid circuit parameters corresponding to the target node.

[0042] Furthermore, the sparse Bayesian causal graph structure of the fluid path is used to constrain the reinforcement learning policy space. Combined with the importance coefficients of nodes, this allows policy optimization to focus on the target variable and improves the stability of control returns. The resulting fluid path parameter tuning suggestions include:

[0043] Importance coefficients and edge weights related to risk nodes in the sparse Bayesian causal graph structure of the fluid path are extracted and combined with the standardized operation data of the fluid path to form a reinforcement learning state vector, and the importance coefficients of the nodes are used as weighting terms of the reinforcement learning state vector.

[0044] When setting the actor network to make action decisions, explore the combination of pump speed, valve, and flushing actions associated with the target variable to reduce the dimensionality of the strategy space;

[0045] A causal consistency penalty term is added to the reward function, and when the value function calculated by the evaluator network is updated, the return discount rate is adjusted using a liquid-path sparse Bayesian causal graph structure to make the return of the target risk path more sensitive.

[0046] The policy is updated using a pre-configured reinforcement learning architecture, the optimal action set is output, and the optimal action set is mapped to suggestions for adjusting the fluid circuit parameters.

[0047] Furthermore, a causal consistency penalty term is added to the reward function, and when the value function calculated by the evaluator network is updated, the reward discount rate is adjusted using a liquid-path sparse Bayesian causal graph structure, making the reward of the target risk path more sensitive, including:

[0048] If the current action deviates from the causal structure of the target risk path, the causal consistency penalty value is calculated based on the edge weight and conditional probability of the corresponding target risk path, and the degree of deviation of the causal relationship is reflected by the causal consistency penalty value.

[0049] When calculating immediate rewards, the causal consistency penalty is combined with the control target reward so that the reward value negatively adjusts for causal deviation behavior.

[0050] During the process of updating the value function in the evaluator network, the cumulative probability of the target risk path is read, and the discount rate is adjusted using the cumulative probability so that the discount rate is greater on paths with higher cumulative probability.

[0051] The value function of the evaluator network is trained and updated using a reward value that includes causal consistency adjustment and a discount rate that corrects for causal probability structure.

[0052] Furthermore, by combining the fluid path parameter adjustment recommendations with PID control, an adaptive control instruction set with dual pressure and flow rate objectives is formed, including:

[0053] The parameters of PID control are updated by updating the variables in the target risk path in the sparse Bayesian causal graph structure of the liquid path;

[0054] PID control uses the sample channel flow velocity as the final feedback quantity and the flow chamber pressure data and waste liquid tank pressure data as state parameters. It comprehensively calculates and outputs the speed command for controlling the flow chamber sheath pump and the on / off command for the waste liquid tank pump.

[0055] The recommended adjustment of the fluid circuit parameters is combined with the speed command of the sheath pump in the flow chamber and the on / off command of the waste liquid pump in the waste liquid pool to generate an adaptive control command set with dual objectives of pressure and flow rate.

[0056] According to a second aspect of the present invention, a pressure-flow rate coordinated control system is provided, comprising:

[0057] The fluid circuit operation data acquisition and standardization processing module is used to acquire the operation data of the flow cytometer fluid circuit, including flow chamber pressure data, waste liquid tank pressure data, sheath fluid pump speed, waste liquid pump speed, valve opening ratio, and sample channel flow rate; and to perform time synchronization and dimension unification processing on the operation data to obtain standardized operation data.

[0058] The causal structure construction and risk probability assessment module is used to construct a sparse Bayesian causal graph structure for the liquid path based on time series samples of standardized operating data, so as to ensure that the relationship of the target variable is preserved under different operating conditions. Feature selection and weighting are performed on the sparse Bayesian causal graph structure to improve the sensitivity of the risk prediction network to the coordinated pressure-flow rate pattern and obtain multi-level risk probabilities for each time window. In the next time window, the importance coefficient of the nodes in the sparse Bayesian causal graph structure is adjusted by using the multi-level risk probabilities.

[0059] The causal constraint reinforcement learning policy optimization module is used to constrain the reinforcement learning policy space using the sparse Bayesian causal graph structure of the fluid path. By combining the importance coefficients of the nodes, the policy optimization focuses on the target variable and improves the stability of control returns, and obtains suggestions for adjusting the fluid path parameters.

[0060] The dual-objective adaptive fluid circuit control execution module is used to combine fluid circuit parameter adjustment suggestions with PID control to form an adaptive control instruction set with dual objectives of pressure and flow rate; execute the adaptive control instruction set and drive the continuous evolution of fluid circuit parameter adjustment suggestions to update the adaptive control instruction set.

[0061] According to a third aspect of the present invention, a computer device is provided.

[0062] The computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the above method.

[0063] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.

[0064] The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the above method.

[0065] The beneficial effects of this invention are as follows:

[0066] 1. This invention makes predictions based on multi-level risk probabilities before fluctuations occur significantly, uses causal constraints to ensure that control actions conform to the physical laws of the liquid path, PID ensures instantaneous correction, and reinforcement learning framework ensures global optimization, making the flow rates of different batches more comparable; after the flow rate stabilizes, the repeatability of cell concentration, absolute count, and signal strength is improved, supporting high-quality quantitative analysis.

[0067] 2. The real-time synergy between the causal structure and risk prediction in this invention improves the response speed and prediction accuracy of fluid circuit anomalies. Dynamic adjustment of node importance enables the causal structure to quickly reflect changes in the distribution of fluid circuit risks, and the input feature weight values ​​of the prediction network always conform to the current causal relationship, reducing detection delays or misjudgments caused by interference from irrelevant features.

[0068] 3. This invention prioritizes high-risk paths, optimizing strategy convergence speed and long-term return stability. The causal consistency penalty term in the reward function ensures the strategy does not deviate from the true causal relationship of the physical-fluid pathway; the discount rate is adjusted according to the probability of the risk path, enabling reinforcement learning to "focus on key issues further away" during training, accelerating convergence, while maintaining better reward stability during operation.

[0069] 4. This invention integrates reinforcement learning and PID control to achieve closed-loop regulation of the hydraulic circuit that balances global optimization and instantaneous precise control. The dual-objective adaptive control instruction set can simultaneously satisfy the two major objectives of pressure stability and flow rate accuracy. Reinforcement learning optimizes the overall hydraulic circuit operating efficiency, while PID control quickly corrects instantaneous deviations. The combination of these two methods improves the robustness and adaptability of the system. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 This is a flowchart of a pressure-flow rate coordinated control method according to an embodiment of the present invention;

[0072] Figure 2 This is a principle block diagram of a pressure-flow rate coordinated control system according to an embodiment of the present invention;

[0073] Figure 3 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0074] Figure 4 This is a schematic diagram of the flow rate control architecture according to an embodiment of the present invention;

[0075] Figure 5 This is one of the temperature and pressure fitting diagrams according to an embodiment of the present invention;

[0076] Figure 6 This is the second temperature and pressure fitting diagram according to an embodiment of the present invention. Detailed Implementation

[0077] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0078] According to an embodiment of the present invention, a pressure-flow rate coordinated control method and system are provided.

[0079] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a pressure-flow rate coordinated control method is provided, comprising:

[0080] S1. Collect the operating data of the flow cytometer's fluid circuit, including flow chamber pressure data, waste liquid tank pressure data, sheath fluid pump speed, waste liquid pump speed, valve opening ratio, and sample channel flow rate.

[0081] S2. Perform time synchronization and unit unification processing on the operating data to obtain standardized operating data.

[0082] S3. Based on time series samples of standardized operating data, construct a sparse Bayesian causal graph structure for the liquid path to ensure that the relationship between target variables is preserved under different operating conditions. Perform feature filtering and weighting on the sparse Bayesian causal graph structure to improve the sensitivity of the risk prediction network to the coordinated pressure-flow rate pattern and obtain multi-level risk probabilities for each time window. In the next time window, use the multi-level risk probabilities to adjust the importance coefficients of nodes in the sparse Bayesian causal graph structure for the liquid path.

[0083] S4. Using the sparse Bayesian causal graph structure of the fluid path to constrain the reinforcement learning policy space, combined with the importance coefficient of the nodes, the policy optimization focuses on the target variable and improves the stability of control returns, and obtains suggestions for adjusting the fluid path parameters.

[0084] S5. Combine the fluid path parameter adjustment suggestions with PID control to form an adaptive control instruction set with dual objectives of pressure and flow rate.

[0085] S6. Execute the adaptive control instruction set and drive the continuous evolution of hydraulic circuit parameter adjustment suggestions to update the adaptive control instruction set.

[0086] In one embodiment, constructing a sparse Bayesian causal graph structure for the liquid path based on time-series samples of standardized operational data includes:

[0087] A time window is pre-set, and standardized operational data is constructed into time series samples. The start and end times of the sequence are appended to the data blocks of each time series sample to ensure the temporal causality of the causal analysis. A temporal Bayesian graph containing all variable connections is initialized, and a sparsity prior is applied to each edge in the temporal Bayesian graph to calculate the significance probability. Edges in the temporal Bayesian graph with significance probabilities below a probability threshold are pruned to preserve the target causal path. The conditional probability table of the remaining edges in the temporal Bayesian graph is updated using Bayesian posterior, and the updated temporal Bayesian graph is saved to obtain the liquid path sparse Bayesian causal graph structure, focusing on the relationship of the target variables.

[0088] In one embodiment, feature filtering and weighting are applied to the sparse Bayesian causal graph structure of the liquid path to enhance the sensitivity of the risk prediction network to the cooperative pressure-flow rate pattern, and to obtain the multi-level risk probability for each time window, including:

[0089] Edge weights are assigned to the liquid path sparse Bayesian causal graph structure based on significance probability; feature weight vectors are allocated based on edge weights, and the features in the time series samples are weighted to obtain a weighted feature vector; the weighted feature vector is input into a risk prediction neural network to reduce interference from irrelevant variables; the risk prediction neural network outputs multi-level risk probabilities for each time window; when the risk prediction neural network is updated via backpropagation, the weight decay mechanism is correlated with the criticality of the liquid path sparse Bayesian causal graph structure.

[0090] In one embodiment, the weighted feature vector is input into the risk prediction neural network to reduce interference from irrelevant variables; the risk prediction neural network outputs multi-level risk probabilities for each time window, including:

[0091] Obtain the weighted feature vector, and adjust the values ​​of each dimension of the weighted feature vector according to the edge weight or node importance to reduce the influence of irrelevant variables on the prediction results; process the weighted feature vector according to the input requirements and feed it into the input layer of the risk prediction neural network; the risk prediction neural network performs weighted summation, nonlinear activation and feature mapping on the input vector layer by layer to extract deep feature patterns related to the risk status of the liquid circuit, and generates the initial value of the probability distribution of the corresponding risk status in the output layer; the output layer of the risk prediction neural network generates multi-level risk probability values ​​according to the set risk level categories (such as high risk, medium risk, low risk); and normalize the output multi-level risk probability values.

[0092] In one embodiment, adjusting the importance coefficient of nodes in a sparse Bayesian causal graph structure of a liquid path using multi-level risk probabilities includes:

[0093] Obtain multi-level risk probabilities, including first risk probability, second risk probability, and third risk probability, and designate the target variable node as a liquid path risk node; based on the liquid path sparse Bayesian causal graph structure, find the required set of nodes with causal paths to the liquid path risk nodes; calculate risk weight factors based on the multi-level risk probability vectors, and update the node importance of the required node set using the risk weight factors; write the adjusted node importance back to the node attributes of the liquid path sparse Bayesian causal graph.

[0094] In one embodiment, after adjusting the importance coefficients of nodes in the sparse Bayesian causal graph structure of the fluid path using multi-level risk probabilities, the following is also included:

[0095] By incorporating node importance coefficients as part of the sparsity prior, nodes with higher importance coefficients are more difficult to prune during saliency calculation. When selecting features from the sparse Bayesian causal graph structure of the fluid path, the node importance coefficients determine the priority. The node importance coefficients are added to the weighting term of the reinforcement learning reward function, enabling reinforcement learning to prioritize the control of fluid path parameters corresponding to the target node.

[0096] In one embodiment, a sparse Bayesian causal graph structure of the fluid path is used to constrain the reinforcement learning policy space. Combined with the importance coefficients of nodes, this allows policy optimization to focus on the target variable and improves the stability of control gains, resulting in fluid path parameter tuning suggestions including:

[0097] Importance coefficients and edge weights related to risk nodes in the sparse Bayesian causal graph structure of the fluid path are extracted and combined with standardized operational data of the fluid path to construct a reinforcement learning state vector, with the importance coefficients of the nodes used as weighting terms in the reinforcement learning state vector. When making action decisions, the actor network explores combinations of pump speed, valve, and flushing actions associated with the target variable to reduce the dimensionality of the policy space. A causal consistency penalty term is added to the reward function, and when the value function calculated by the evaluator network is updated, the reward discount rate is adjusted using the sparse Bayesian causal graph structure of the fluid path to make the reward of the target risk path more sensitive. The configured reinforcement learning architecture is used for policy updates, outputting the optimal action set, which is then mapped to fluid path parameter adjustment suggestions.

[0098] In one embodiment, a causal consistency penalty term is added to the reward function, and when the value function calculated by the evaluator network is updated, the reward discount rate is adjusted using a liquid-path sparse Bayesian causal graph structure to make the reward of the target risk path more sensitive, including:

[0099] If the current action deviates from the causal structure of the target risk path, a causal consistency penalty value is calculated based on the edge weights and conditional probabilities of the corresponding target risk path, and the degree of deviation from the causal relationship is reflected by the causal consistency penalty value. When calculating the immediate reward, the causal consistency penalty term is combined with the control target reward so that the reward value negatively adjusts for causal deviation behavior. During the process of updating the value function of the evaluator network, the cumulative probability of the target risk path is read, and the discount rate is adjusted using the cumulative probability so that the discount rate is larger on paths with higher cumulative probabilities. The value function of the evaluator network is trained and updated using the reward value including the causal consistency adjustment and the discount rate corrected by the causal probability structure.

[0100] In one embodiment, combining fluid path parameter adjustment recommendations with PID control to form an adaptive control instruction set with dual pressure and flow rate objectives includes:

[0101] The parameters of the PID control are updated by the variables in the target risk path in the sparse Bayesian causal graph structure of the liquid path. The PID control uses the sample channel flow velocity as the final feedback quantity and the flow chamber pressure data and waste liquid tank pressure data as state parameters. It comprehensively calculates and outputs the speed command for controlling the flow chamber sheath pump and the switching command for the waste liquid tank pump. The liquid path parameter adjustment suggestions are combined with the speed command for the flow chamber sheath pump and the switching command for the waste liquid tank pump to generate an adaptive control command set with dual objectives of pressure and flow velocity.

[0102] like Figure 2 As shown, according to another embodiment of the present invention, a pressure-flow rate coordinated control system is also provided, comprising:

[0103] The liquid circuit operation data acquisition and standardization processing module 1 is used to acquire the operation data of the flow cytometer liquid circuit, including flow chamber pressure data, waste liquid tank pressure data, sheath liquid pump speed, waste liquid pump speed, valve opening ratio, and sample channel flow rate; and to perform time synchronization and dimension unification processing on the operation data to obtain standardized operation data.

[0104] The causal structure construction and risk probability assessment module 2 is used to construct a sparse Bayesian causal graph structure for the liquid path based on time series samples of standardized operating data, so as to ensure that the relationship of the target variable is preserved under different operating conditions. Feature selection and weighting are performed on the sparse Bayesian causal graph structure to improve the sensitivity of the risk prediction network to the coordinated pressure flow rate pattern and obtain the multi-level risk probability for each time window. In the next time window, the importance coefficient of the nodes in the sparse Bayesian causal graph structure is adjusted by using the multi-level risk probability.

[0105] The Causal Constraint Reinforcement Learning Policy Optimization Module 3 is used to constrain the reinforcement learning policy space using the sparse Bayesian causal graph structure of the fluid path. By combining the importance coefficients of the nodes, the policy optimization focuses on the target variable and improves the stability of control returns, and obtains suggestions for adjusting the fluid path parameters.

[0106] The dual-objective adaptive fluid circuit control execution module 4 is used to combine fluid circuit parameter adjustment suggestions with PID control to form an adaptive control instruction set with dual objectives of pressure and flow rate; execute the adaptive control instruction set and drive the continuous evolution of fluid circuit parameter adjustment suggestions to update the adaptive control instruction set.

[0107] To facilitate understanding of the above technical solutions of the present invention, the working principle of the present invention in actual process will be described in detail below.

[0108] This invention relates to flow rate control in a flow cytometer fluidization system based on negative pressure sample loading. Through multi-sensor fusion feedback, it overcomes the problems of insufficient control precision, slow control speed, and poor anti-interference capability in existing technologies. It integrates a high-precision pressure sensor module, a real-time flow monitoring module, an intelligent pump and valve control system, and a PID control algorithm with pre-adjustment. By dynamically adjusting the opening and closing states of the pump and valves, the flow curve, and the rotational speed parameters, the system can achieve precise synchronous control of the pressure balance between the flow chamber and the waste liquid tank, thereby stably, accurately, and quickly achieving and maintaining the fluid velocity within the set range. The system adopts a standardized and scalable modular architecture design, offering significant advantages such as flexible system architecture, economical implementation costs, simple and efficient control algorithms, and stable and reliable operation. It can be easily ported and adapted to fluid control scenarios on various product platforms. This invention is not only applicable to structural parameter optimization, performance verification, and process exploration in the product development stage, but also effectively meets the stringent requirements for production consistency, process controllability, and quality stability in large-scale commercial production, possessing broad application prospects and industrialization value.

[0109] The specific process is as follows:

[0110] 1. Sensors are deployed at multiple key locations in the flow cytometer's liquid path to collect multi-source operating parameters in real time, including flow chamber pressure, waste liquid tank pressure, sheath fluid pump speed, waste liquid pump speed, valve opening ratio, and sample channel flow rate. A time synchronization algorithm is used to align asynchronous data collected from different sensors, and dimensional unification processing converts all parameters to a unified physical unit and numerical range, forming standardized operating data that can be directly used for multivariate correlation analysis and modeling. Synchronous acquisition and dimensional unification eliminate data misalignment and proportional distortion caused by time differences and unit variations between sensors, resulting in a more complete and accurate description of the liquid path's operating status.

[0111] 2. Standardized operational data, after time synchronization and dimensional unification, is divided into time series samples according to a preset time window, and start and end time labels are added to each data block to ensure the accuracy of the time series in causal analysis; a time series Bayesian graph containing all variable connections is initialized, and the significance probability of each edge is calculated based on the sparsity prior, thereby quantifying the significance of the causal relationship between variables; edges with significance probabilities below a set threshold are pruned, retaining only causal paths that are significant to the target variable; the conditional probability table of the retained edges is updated using Bayesian posterior, so that the graph structure parameters are consistent with the latest data distribution, forming a sparse Bayesian causal graph structure of the liquid path that focuses on the relationship between the target variable, such as pressure and flow rate.

[0112] By leveraging the saliency probabilities of each edge in the sparse Bayesian causal graph structure of the liquid path, corresponding causal connection weights are assigned, quantifying the correlation strength between each variable and the target variable across the entire graph. A feature weight vector is constructed based on the edge weights, and the features in the time series are weighted to ensure that features highly correlated with target stability account for a larger proportion in risk prediction. The weighted feature vector is input into the risk prediction neural network to weaken the interference of irrelevant variables and enhance the network's sensitivity to the cooperative pressure-flow rate pattern. During the backpropagation update process, a weight decay mechanism related to the criticality of the causal graph is introduced, keeping the weights of highly correlated features stable while gradually decaying the weights of low-correlation features, outputting multi-level risk probabilities for each time window, achieving hierarchical prediction from low to high risk levels.

[0113] The sparse Bayesian causal graph for liquid circuits is a probabilistic graph based on Bayesian networks. By statistically learning from liquid circuit operation data and historical events, it infers the conditional dependencies between variables. The sparsification mechanism removes low-relevance edges through significance testing, information criteria, or physical constraints, resulting in a sparse structure that retains only the core causal paths. During inference, the posterior probability of any node can be calculated using Bayes' theorem, and combined with causal path analysis, risk sources can be located or changes in liquid circuit status can be predicted. This structure not only preserves the causal logic between liquid circuit variables but also improves inference efficiency and robustness through sparsification.

[0114] The numerical values ​​of each dimension of the feature vector are adjusted according to the edge weights or node importance coefficients to reduce the impact of irrelevant variables on risk prediction. The risk prediction neural network adopts a multi-layer feedforward structure, consisting of an input layer, several hidden layers, and an output layer. The input layer receives the weighted fluid flow feature vector and maps it to the first hidden layer node. The hidden layers perform weighted summation operations through a fully connected weight matrix and combine activation functions, such as ReLU or sigmoid, to achieve nonlinear transformation, capturing the complex coupling relationships and patterns between features. Multiple hidden layers are stacked sequentially, enabling the model to progressively abstract features and express higher-order risk states. The output layer uses normalization functions such as softmax to map the linear combination result of the last layer to a probability distribution of multi-level risk categories, achieving quantitative prediction of different risk levels. During backpropagation training, the network adjusts the weights of each layer according to the prediction error, combined with a weight decay mechanism related to the criticality of the fluid flow causal graph, to keep the weights of highly correlated features stable, thus balancing prediction accuracy and long-term reliability.

[0115] A multi-level risk probability vector is used, including a first risk probability (high risk), a second risk probability (medium risk), and a third risk probability (low risk). Target variable nodes, such as fluid path pressure nodes or sample channel flow rate nodes, are defined as fluid path risk nodes. The system searches for the required set of nodes that have direct or indirect causal paths with these risk nodes. A risk weight factor is calculated based on the multi-level risk probability vector, which can be based on weighted sum, information entropy, or a hazard level function. This factor is used to dynamically adjust the node importance coefficients of the required node set, increasing the importance of nodes on high-risk paths and decreasing the importance of nodes on low-risk paths. Finally, the updated node importance values ​​are written back to the attributes of the corresponding nodes in the fluid path sparse Bayesian causal graph, providing real-time causal weight feedback for subsequent feature selection, weighted prediction, and fluid path control.

[0116] Node importance coefficients can serve as part of a sparsity prior: nodes with high importance have edges that are more difficult to prune during saliency calculation. When selecting input features from a causal graph, importance coefficients determine which nodes are prioritized for inclusion in the input set. Node importance coefficients are then weighted in the reward function, causing the policy to focus more on highly important fluid variables.

[0117] 3. Extract the importance coefficients and associated edge weights of risk nodes in the liquid path from the sparse Bayesian causal graph structure of the liquid path, and combine them with standardized operational data to form a state vector for reinforcement learning:

[0118] ;

[0119] In the formula, This is the reinforcement state vector at time step; The first in the liquid circuit A running variable in time Standardized values, such as pump speed, pressure, flow rate, etc. Let be the node importance coefficient extracted from the sparse Bayesian causal graph of the liquid path, and let represent the variable. Relevance weights on the target risk path This represents the total number of variables involved in the operation of the hydraulic circuit.

[0120] By limiting the action search scope in the actor network, only exploring hydraulic control combinations significantly related to the target variable, such as pump speed, valve opening, and flushing actions, the irrelevant dimensions of the policy space are reduced. A causal consistency penalty term is introduced into the reward function. If the current action deviates from the causal structure on the target risk path, the causal consistency penalty value is calculated based on edge weights and conditional probabilities, and the deviating action is negatively adjusted in the immediate reward. When updating the value function, the evaluator network dynamically adjusts the discount rate according to the cumulative probability of the target risk path, so that the reward weight of the value function on high-risk paths is greater. The reinforcement learning network is trained using the reward value and discount rate including causal consistency adjustment to output the optimal action set, and this set is mapped to hydraulic parameter adjustment suggestions, realizing a closed-loop optimization process from causal constraints to directly executable control commands.

[0121] Instant rewards At time step Defined as:

[0122] ;

[0123] Causal consistency penalty item The calculation formula is:

[0124] ;

[0125] In the formula, Total instant reward value; In time The control target reward value, such as maintaining flow rate and pressure within a set range; The causal consistency penalty coefficient controls the proportion of the penalty's impact on the total reward. For the current action The penalty value for causal consistency; Let be the set of all causal edges on the target risk path; For nodes in a sparse Bayesian causal graph of a liquid path To the node Edge weights; The indicator function is defined if the current action conforms to the set of causal constraints. Time to take Otherwise take .

[0126] Discount rate Adjusted based on the cumulative probability of the target risk path:

[0127] ;

[0128] The base discount rate, the default value when there is no causal adjustment; This is a discount rate amplification factor used to adjust the return sensitivity of high-risk paths; The cumulative probability of the target risk path is calculated based on the conditional probability of the causal graph.

[0129] 4. Utilizing the variables and their weights related to the target risk path in the sparse Bayesian causal graph structure of the liquid path, the proportional coefficient, integral coefficient, and derivative coefficient of the PID controller are dynamically adjusted to match the PID control parameter settings with the current liquid path risk state. The PID controller uses the sample channel flow velocity as the final feedback quantity and the flow chamber pressure data and waste liquid pool pressure data as state parameters to calculate the speed command of the flow chamber sheath pump and the on / off command of the waste liquid pool pump in real time to meet the dual-objective control requirements of pressure and flow velocity. The liquid path parameter adjustment suggestions output by the reinforcement learning module are integrated with the above PID commands, and the final pressure-flow velocity dual-objective adaptive control command set is generated through comprehensive judgment and weighted selection.

[0130] Figure 4 The diagram shows one architecture of the present invention, and the control flow is as follows:

[0131] Power-on initialization. After the user sets the target sample channel flow rate, the controller initiates the initial strong drive phase, setting the flow chamber sheath pump to its maximum speed to rapidly increase system pressure and shorten settling time. The controller monitors the error between the pressure value fed back by the flow chamber pressure sensor and the target pressure in real time. When this difference decreases and falls below a preset first threshold, such as 5 kPa, it indicates that the system has passed the rapid pressurization period and entered the pre-stabilization transition phase. In this phase, the controller switches the sheath pump speed to a fixed value preset based on historical stability data, close to the expected final speed, such as 27,000 rpm, allowing the system to smoothly transition to near the stabilization point. Simultaneously, the system independently monitors the waste liquid tank pressure. Based on the readings from the waste liquid tank pressure sensor, the controller controls the operation of the waste liquid tank pump in switch mode to ensure that the gas pressure in the waste liquid tank remains stable within a necessary operating window, providing a stable back pressure environment for the downstream liquid flow.

[0132] When the system reaches a stable critical state, i.e., the change in flow chamber pressure falls within a small preset range, such as the second threshold of 0.2 kp / 100 ms, the system enters the precise PID control stage. In this stage, the PID control algorithm is officially activated. Based on the difference between the actual measured value of the sample channel flowmeter and the target flow velocity, it performs proportional, integral, and derivative calculations and outputs precise speed adjustment commands to the flow chamber sheath pump, ultimately achieving high-precision, zero-static-error stable control of the sample channel flow velocity. The specific calculation is as follows:

[0133] 1. Error calculation and dead zone determination:

[0134] Error calculation: Calculate the pressure error value e0 at the current sampling time = actual value - setpoint. Dead zone judgment: When the absolute value of the error exceeds 0.005, PID regulation is initiated; otherwise, the output is zero to prevent the system from oscillating near the equilibrium point.

[0135] 2. Calculation of error difference term:

[0136] Calculate the three key difference terms:

[0137] The proportional term ep: e(k) - e(k-1) reflects the trend of error change. The integral term ei: e(k) is the current error value, used to eliminate steady-state error. The differential term ed: e(k) - 2e(k-1) + e(k-2) reflects the acceleration of error change.

[0138] 3. Incremental output calculation:

[0139] Adopting the standard incremental PID format:

[0140] Δu(k)=Kp×[e(k)-e(k-1)]+Ki×e(k)+Kd×[e(k)-2e(k-1)+e(k-2)];

[0141] Where: Kp is the proportional coefficient, with a base value of 200, which can be adjusted according to the stage; Ki is the integral coefficient, with a value of 40; Kd is the differential coefficient, with a value of 200.

[0142] 4. Error history update:

[0143] Establish an error queue to prepare historical data for the next cycle of calculation:

[0144] e2←e1, the previous error becomes the error before e1; e1←e0, the current error becomes the previous error.

[0145] 5. Speed ​​integration and limiting processing:

[0146] Speed ​​integral: The increment of the PID output is added to the current speed value; Output limiting: Ensures that the control quantity is within a reasonable range and prevents actuator saturation.

[0147] In this invention, a flow chamber pressure sensor is configured in the inlet liquid path of the flow chamber for real-time measurement of the flow chamber hydraulic pressure; a flow chamber sheath pump is associated with the liquid path where the flow chamber pressure sensor is located, and its rotation speed is adjusted to control the flow chamber pressure; a waste liquid tank pressure sensor is configured in the waste liquid tank for real-time measurement of the waste liquid tank gas pressure; a waste liquid tank pump is connected to the waste liquid tank, and its opening or closing operation is used to adjust the waste liquid tank pressure; a sample channel flow meter is configured in the sample flow path for real-time measurement of the sample channel flow velocity; and a controller is electrically connected to the flow chamber pressure sensor, waste liquid tank pressure sensor, sample channel flow meter, flow chamber sheath pump, and waste liquid tank pump. The controller incorporates a PID control algorithm configured to: use the reading of the sample channel flow meter as the final feedback quantity, and use the readings of the flow chamber pressure sensor and the waste liquid tank pressure sensor as system state parameters, comprehensively calculate and output commands to control the rotation speed of the flow chamber sheath pump and the on / off command of the waste liquid tank pump, thereby achieving precise control of the target flow velocity in the sample channel. Flow chamber: A sample flow channel used to form sheath fluid focusing. Sheath fluid pump: Used to provide the flow force for the sample. Waste liquid tank: Used to collect waste liquid after analysis.

[0148] The controller is configured to execute a closed-loop control algorithm that uses the target channel flow velocity as a setpoint and simultaneously uses measured values ​​of the flow chamber pressure and waste liquid tank pressure as feedback signals. By comparing the target flow velocity with the actual flow velocity in real time, the controller derives or estimates the output of the sheath pump from the two pressure signals, thereby achieving precise and stable control of the channel flow velocity. The control algorithm is a PID control algorithm or a variant thereof, which fuses the feedback signals from the two pressure sensors to calculate precise drive commands for the sheath pump.

[0149] The control logic for the waste liquid tank pump is as follows: when the reading of the waste liquid tank pressure sensor exceeds the preset pressure range, such as -20 to -10 kPa, the waste liquid tank pump is started or stopped to maintain the waste liquid tank pressure within the preset range. The system indirectly controls the sample channel flow rate by controlling the flow chamber pressure and the waste liquid tank pressure, where the relationship between pressure and sample channel flow rate is a function of temperature. The controller's operating logic includes three stages:

[0150] Initial strong drive phase: The initial speed of the flow chamber sheath pump is set to the maximum speed to quickly build up pressure; Pre-stabilization transition phase: When the system pressure change rate is detected to be lower than the first set threshold, the speed of the flow chamber sheath pump is switched to a fixed speed close to the final stable value, i.e., 27,000~28,000 rpm; Precise PID adjustment phase: When the error between the reading of the flow chamber pressure sensor and the target pressure value is less than the second set threshold, the PID control algorithm is activated to perform closed-loop precise adjustment of the speed of the flow chamber sheath pump.

[0151] It should be noted that the pressure-flow velocity relationship model in the system can be obtained through calibration experiments at different temperatures and stored in the controller for feedforward compensation and environmental adaptation. The experimental steps are as follows:

[0152] Experimental Environment Setup: A constant-temperature experimental environment was provided, and the experimental apparatus was placed inside the constant-temperature chamber. The temperature testing range was set from 1℃ to 39℃, with multiple temperature testing points established at 2℃ temperature gradient intervals. Temperature Stabilization: The temperature of the constant-temperature chamber was adjusted to the first target temperature value and maintained at this temperature condition for at least 20 minutes to ensure that the temperature of all components of the experimental apparatus reached a uniform and stable state. The actual measured values ​​of the temperature sensor were recorded as the ambient temperature baseline. Sheath Fluid Flow Rate Baseline Establishment: The sheath fluid flow rate was controlled to be stable within the range of 9-10 mL / min, and the measured data of the sheath fluid flow rate were monitored and recorded in real time using a sheath fluid flow rate sensor.

[0153] Sample flow rate adjustment and data acquisition:

[0154] Procedure 1: Set the target sample flow rate to 10 μL / min and manually adjust the speed of the sheath pump to achieve this target flow rate.

[0155] Procedure 2: Record the actual rotational speed of the sheath pump and the actual measured value of the sample flow rate at this time, as shown in Table 1 and Table 2.

[0156] Step 3: Simultaneously record the measured pressure values ​​of the waste liquid tank and the flow chamber, calculate the pressure difference between the two, and record it as shown in Tables 1 and 2. It should be noted that both the flow chamber and waste liquid tank pressures are negative, i.e., vacuum; the flow chamber pressure sensor is located near the flow chamber in the sheath fluid pipeline; a standard flow cytometer is used, with a sample volume of 1 mL; Tables 1 and 2 include the pressure and rotation speed corresponding to different sheath fluid flow rates.

[0157] Multi-sample flow rate test: The target flow rate value of the sample is increased in increments of 10 μL / min to the next target value. The above procedure is repeated until the test of the target flow rate values ​​of all samples in the range of 10 μL / min to 100 μL / min is completed.

[0158] Multi-temperature point test: Adjust the temperature of the constant temperature chamber to the next temperature test point, and repeat the above operation until experimental data collection is completed for all temperature test points within the range of 1℃ to 39℃.

[0159] Data processing and model building: Based on the actual measurement data collected at all temperature and sample flow rate test points, a functional relationship between pressure and temperature was established using numerical fitting methods.

[0160] ;

[0161] In the formula, P is the pressure in the flow chamber, and a i For the fitting coefficients, T i The ambient temperature is used. The coefficients of the pressure-temperature fitting polynomial are calculated, forming a complete functional expression of pressure with respect to temperature, as shown in Table 2. Figure 5 , Figure 6 As shown, Figure 5 In the diagram, y represents the pressure in the flow chamber, x represents the ambient temperature, and R... 2 The coefficient of determination measures how well the model fits the data.

[0162] Figure 4 In the diagram, A1: flow chamber; A2: waste liquid tank; A4: sample channel; P1: sheath liquid pump; P2: waste liquid pump; F2: degassing device; R1: flow damping device; S1: flow sensor; S2: sheath liquid tank pressure sensor; S3: waste liquid tank pressure sensor.

[0163] Table 1. Compilation of Temperature Test Data (Part 1)

[0164] Table 2. Compilation of Temperature Test Data (Part 2)

[0165]

[0166]

[0167] Table 3 Summary of Temperature Tests

[0168] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0169] Those skilled in the art will understand that Figure 3The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0170] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0171] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0172] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0173] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A pressure-flow rate coordinated control method, characterized in that, include: S1. Collect the operating data of the flow cytometer's fluid circuit, including flow chamber pressure data, waste liquid tank pressure data, sheath fluid pump speed, waste liquid pump speed, valve opening ratio, and sample channel flow rate; S2. Perform time synchronization and unit unification processing on the operational data to obtain standardized operational data; S3. Based on time series samples of standardized operating data, construct a sparse Bayesian causal graph structure for the liquid circuit to ensure that the relationship between the target variables is preserved under different operating conditions; Feature filtering and weighting are applied to the sparse Bayesian causal graph structure of the liquid path to improve the sensitivity of the risk prediction network to the cooperative pressure-flow velocity pattern and obtain the multi-level risk probability for each time window. In the next time window, the importance coefficient of the nodes in the sparse Bayesian causal graph structure of the liquid path is adjusted by using multi-level risk probabilities. S4. Use the sparse Bayesian causal graph structure of the fluid path to constrain the reinforcement learning policy space. Combine the importance coefficient of the nodes to make the policy optimization focus on the target variable and improve the stability of the control payoff, and obtain suggestions for adjusting the fluid path parameters. S5. Combine the fluid circuit parameter adjustment suggestions with PID control to form an adaptive control instruction set with dual objectives of pressure and flow rate; S6. Execute the adaptive control instruction set and drive the continuous evolution of hydraulic circuit parameter adjustment suggestions to update the adaptive control instruction set.

2. The pressure-flow rate coordinated control method according to claim 1, characterized in that, The construction of the sparse Bayesian causal graph structure for the liquid path based on time-series samples of standardized operational data includes: Pre-set time windows and construct time series samples from standardized operational data; The start and end times of the sequence are appended to the data blocks of each time series sample to ensure the temporal causality of the causal analysis; Initialize a temporal Bayesian graph containing all variable connections, apply a sparsity prior to each edge in the temporal Bayesian graph, and calculate the significance probability. Edges with a significance probability below a probability threshold in the temporal Bayes graph are pruned to preserve the target causal path; The conditional probability table of the remaining edges in the temporal Bayesian graph is updated using Bayesian posterior method. The updated temporal Bayesian graph is then saved to obtain the liquid path sparse Bayesian causal graph structure, which focuses on the relationship of the target variable.

3. The pressure-flow rate coordinated control method according to claim 2, characterized in that, The sparse Bayesian causal graph structure of the liquid path is subjected to feature filtering and weighting to improve the sensitivity of the risk prediction network to the cooperative pressure-flow rate pattern, and to obtain the multi-level risk probability for each time window, including: The edge weights in the sparse Bayesian causal graph structure of the liquid path are assigned according to the significance probability; the feature weight vector is assigned according to the edge weights, and the features in the time series samples are weighted to obtain the weighted feature vector. The weighted feature vector is input into the risk prediction neural network to reduce interference from irrelevant variables; the risk prediction neural network outputs multi-level risk probabilities for each time window. In the risk prediction neural network, during backpropagation updates, the weight decay mechanism is correlated with the criticality of the liquid-channel sparse Bayesian causal graph structure.

4. The pressure-flow rate coordinated control method according to claim 3, characterized in that, The process of inputting the weighted feature vector into the risk prediction neural network reduces interference from irrelevant variables; the multi-level risk probability for each time window is output through the risk prediction neural network, including: Obtain the weighted feature vector, and adjust the values ​​of each dimension of the weighted feature vector according to the edge weight or node importance to reduce the influence of irrelevant variables on the prediction results. The weighted feature vectors are processed according to the input requirements and then fed into the input layer of the risk prediction neural network. The risk prediction neural network performs weighted summation, nonlinear activation, and feature mapping on the input vector layer by layer to extract deep feature patterns related to the risk status of the liquid circuit, and generates the initial value of the probability distribution of the corresponding risk status in the output layer. The output layer of the risk prediction neural network generates multi-level risk probability values ​​according to the set risk level categories; the output multi-level risk probability values ​​are then normalized.

5. The pressure-flow rate coordinated control method according to claim 1, characterized in that, The method of adjusting the importance coefficient of nodes in the sparse Bayesian causal graph structure of the fluid flow using multi-level risk probabilities includes: Obtain multi-level risk probabilities, including first risk probability, second risk probability, and third risk probability, and treat the target variable node as a liquid circuit risk node; Based on the sparse Bayesian causal graph structure of the liquid path, find the required set of nodes that have causal paths with the risk nodes of the liquid path; Calculate the risk weight factor based on the multi-level risk probability vector, and use the risk weight factor to update the node importance of the required node set; The adjusted node importance is written back to the node attributes of the sparse Bayesian causal graph.

6. The pressure-flow rate coordinated control method according to claim 1, characterized in that, After adjusting the importance coefficients of nodes in the sparse Bayesian causal graph structure of the fluid flow using multi-level risk probabilities, the following is also included: By incorporating the node importance coefficient as part of the sparsity prior, nodes with higher importance coefficients are more difficult to prune during saliency calculation. When selecting features from a sparse Bayesian causal graph structure in a liquid path, the node importance coefficient determines the priority. By adding the node importance coefficient to the weighting term of the reinforcement learning reward function, reinforcement learning prioritizes the control of the fluid circuit parameters corresponding to the target node.

7. The pressure-flow rate coordinated control method according to claim 1, characterized in that, The method of using a sparse Bayesian causal graph structure in the fluid path to constrain the reinforcement learning policy space, combined with the importance coefficients of nodes, enables policy optimization to focus on the target variable and improves the stability of control returns, resulting in fluid path parameter adjustment suggestions including: Importance coefficients and edge weights related to risk nodes in the sparse Bayesian causal graph structure of the fluid path are extracted and combined with the standardized operation data of the fluid path to form a reinforcement learning state vector, and the importance coefficients of the nodes are used as weighting terms of the reinforcement learning state vector. When setting the actor network to make action decisions, explore the combination of pump speed, valve, and flushing actions associated with the target variable to reduce the dimensionality of the strategy space; A causal consistency penalty term is added to the reward function, and when the value function calculated by the evaluator network is updated, the return discount rate is adjusted using a liquid-path sparse Bayesian causal graph structure to make the return of the target risk path more sensitive. The policy is updated using a pre-configured reinforcement learning architecture, the optimal action set is output, and the optimal action set is mapped to suggestions for adjusting the fluid circuit parameters.

8. The pressure-flow rate coordinated control method according to claim 7, characterized in that, The addition of a causal consistency penalty term to the reward function, and the use of a liquid-path sparse Bayesian causal graph structure to adjust the reward discount rate when the value function calculated by the evaluator network is updated, to make the reward of the target risk path more sensitive, includes: If the current action deviates from the causal structure of the target risk path, the causal consistency penalty value is calculated based on the edge weight and conditional probability of the corresponding target risk path, and the degree of deviation of the causal relationship is reflected by the causal consistency penalty value. When calculating immediate rewards, the causal consistency penalty is combined with the control target reward so that the reward value negatively adjusts for causal deviation behavior. During the process of updating the value function in the evaluator network, the cumulative probability of the target risk path is read, and the discount rate is adjusted using the cumulative probability so that the discount rate is greater on paths with higher cumulative probability. The value function of the evaluator network is trained and updated using a reward value that includes causal consistency adjustment and a discount rate that corrects for causal probability structure.

9. The pressure-flow rate coordinated control method according to claim 7, characterized in that, The adaptive control instruction set that combines the fluid path parameter adjustment suggestions with PID control to form a dual-objective pressure and flow rate control includes: The parameters of PID control are updated by updating the variables in the target risk path in the sparse Bayesian causal graph structure of the liquid path; PID control uses the sample channel flow velocity as the final feedback quantity and the flow chamber pressure data and waste liquid tank pressure data as state parameters. It comprehensively calculates and outputs the speed command for controlling the flow chamber sheath pump and the on / off command for the waste liquid tank pump. The recommended adjustment of the fluid circuit parameters is combined with the speed command of the sheath pump in the flow chamber and the on / off command of the waste liquid pump in the waste liquid pool to generate an adaptive control command set with dual objectives of pressure and flow rate.

10. A pressure-flow rate coordinated control system for implementing the pressure-flow rate coordinated control method according to any one of claims 1-9, characterized in that, include: The fluid circuit operation data acquisition and standardization processing module is used to acquire the operation data of the flow cytometer fluid circuit, including flow chamber pressure data, waste liquid tank pressure data, sheath fluid pump speed, waste liquid pump speed, valve opening ratio, and sample channel flow rate; and to perform time synchronization and dimension unification processing on the operation data to obtain standardized operation data. The causal structure construction and risk probability assessment module is used to construct a sparse Bayesian causal graph structure for the liquid circuit based on time series samples of standardized operating data, so as to ensure that the relationship of the target variable is preserved under different operating conditions. Feature filtering and weighting are applied to the sparse Bayesian causal graph structure of the liquid path to improve the sensitivity of the risk prediction network to the cooperative pressure-flow rate pattern and obtain multi-level risk probabilities for each time window; in the next time window, the importance coefficient of the nodes in the sparse Bayesian causal graph structure of the liquid path is adjusted by using the multi-level risk probabilities. The causal constraint reinforcement learning policy optimization module is used to constrain the reinforcement learning policy space using the sparse Bayesian causal graph structure of the fluid path. By combining the importance coefficients of the nodes, the policy optimization focuses on the target variable and improves the stability of control returns, and obtains suggestions for adjusting the fluid path parameters. The dual-objective adaptive fluid circuit control execution module is used to combine fluid circuit parameter adjustment suggestions with PID control to form an adaptive control instruction set with dual objectives of pressure and flow rate; execute the adaptive control instruction set and drive the continuous evolution of fluid circuit parameter adjustment suggestions to update the adaptive control instruction set.

Images