A Multimodal Agent Perception, Memory, and Autonomous Decision-Making Method and System

Abstract

This invention provides a multimodal agent perception memory and autonomous decision-making method and system, relating to the field of reinforcement learning technology. The method includes: dividing the exhaust membrane of a microfluidic system into regions to obtain multiple pore cluster units; calculating the filament connectivity based on the structural connectivity between each pore cluster unit; obtaining topological memory units corresponding to the operating stage based on the filament connectivity; sorting all topological memory units to obtain sorted topological memory units; calculating the average quality of neighbor detections based on the sorted topological memory units; generating a filament topological state vector based on the average quality of neighbor detections; obtaining updated action value data based on the filament topological state vector; and executing control operations based on the updated action value data. This invention improves the adaptive adjustment capability of microfluidic devices under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of reinforcement learning technology, and in particular to a multimodal agent perception, memory, and autonomous decision-making method and system. Background Technology

[0002] With the widespread application of microfluidic technology in fields such as biological detection, chemical analysis, and medical diagnostics, microfluidic devices are gradually developing towards higher integration and higher reliability. During the actual operation of microfluidic devices, fluid transport in microscale fluid channels is prone to bubble generation. To ensure stable fluid transport, porous venting membranes are typically installed in the fluid channels for gas discharge. However, the venting membrane is composed of numerous micropores, and the liquid occupancy and connectivity within it continuously evolve during operation with changes in pressure, flow rate, and control operations. Especially under repeated operation or complex conditions, liquid may gradually enter the venting membrane and form continuous pathways between multiple pore regions. Since the venting membrane is usually encapsulated inside the device, the liquid state inside the membrane is difficult to observe directly. Existing operational control often relies on external information such as microscopic images, flow data, and control commands for comprehensive judgment. This limits the ability of microfluidic devices to perceive the internal state of the venting membrane during long-term operation.

[0003] In existing technologies, control methods for the operational stability of microfluidic devices often employ rule-driven or single-state-judgment-based control strategies. This involves adjusting flow rate and pressure based on the current image display or sensor readings. Such methods struggle to characterize the evolution of the liquid connectivity within the exhaust membrane over multiple control cycles and cannot systematically store and utilize historical operational state changes and detection results. This can easily lead to misjudgments of the operational state and continued use of inappropriate control operations when the liquid within the exhaust membrane has gradually formed a continuous pathway but the external performance remains relatively stable, thus adversely affecting the stability of the detection results. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a multimodal agent perception memory and autonomous decision-making method, which can solve the technical problem that the prior art has difficulty in characterizing the evolution process of the internal liquid connectivity state of the exhaust membrane in multiple control cycles.

[0005] A first aspect of this invention proposes a multimodal agent perception memory and autonomous decision-making method, comprising: S1: Divide the exhaust membrane of the microfluidic system into regions to obtain multiple pore cluster units; S2: Calculate the liquid filament connectivity based on the structural connectivity between each pore cluster unit; S3: Based on the liquid filament connectivity, obtain the topological memory unit corresponding to the operation stage; S4: Sort all topological memory units to obtain sorted topological memory units; S5: Calculate the average neighbor detection quality based on the sorted topological memory units; S6: Generate the liquid filament topology state vector based on the average quality of neighboring detections; S7: Based on the liquid filament topological state vector, the updated action value data is obtained; S8: Perform control operations based on the updated action value data.

[0006] A second aspect of this invention provides a multimodal agent perception memory and autonomous decision-making system, comprising: a processor and a memory; The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the multimodal agent perception memory and autonomous decision-making method as described in the first aspect.

[0007] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. By dividing the exhaust membrane into regions and introducing a description method using pore cluster units, the distribution and connectivity of the liquid inside the membrane, which were originally difficult to observe and quantify directly, are transformed into calculable and recordable structured information. Furthermore, by calculating the structural connectivity between pore clusters and the liquid filament connectivity, the evolution process of the liquid inside the exhaust membrane from local occupation to continuous connectivity can be continuously characterized. This allows the state changes hidden inside the device during microfluidic operation to be effectively perceived and included in the operational analysis, solving the problem that existing technologies rely solely on external images and instantaneous sensing data, which are insufficient to reflect the long-term evolution state inside the exhaust membrane.

[0008] 2. By associating the liquid filament connectivity sequence, connectivity change sequence, control action sequence, and corresponding detection quality indicators, a topological memory unit corresponding to each operational stage is constructed. This allows the liquid connectivity evolution process, control behavior, and final results of different operational stages to be stored as a whole and recalled as historical experience. Furthermore, the topological memory units are sorted and filtered by absolute difference distance to achieve accurate positioning of historical operational stages that are highly similar to the current operational state. This avoids the problem of fragmented use or complete neglect of historical operational information in existing technologies, enabling operational decisions to be based on historical experience with similar evolutionary backgrounds.

[0009] 3. By generating action value data based on the liquid filament topology state vector and control actions, and using detection quality indicators as rewards to continuously update the action value data, the control decision-making process can continuously incorporate real operating results for correction and optimization in multiple runs. In subsequent runs, it prioritizes control actions that perform better in similar liquid connectivity states, thereby getting rid of control methods that rely on fixed rules or single judgments, realizing closed-loop collaboration between perception, memory and decision-making, and improving the adaptive adjustment capability of microfluidic devices under complex operating conditions and the stability of detection results. Attached Figure Description

[0010] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0011] Figure 1 This is a flowchart illustrating a multimodal agent perception, memory, and autonomous decision-making method provided in an embodiment of the present invention.

[0012] Figure 2 This is a schematic diagram of the structure of a multimodal agent perception memory and autonomous decision-making system provided in an embodiment of the present invention. Detailed Implementation

[0013] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0014] The multimodal agent perception, memory, and autonomous decision-making method provided by the present invention will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.

[0015] Reference manual attached Figure 1 The diagram shows a flowchart of a multimodal agent perception memory and autonomous decision-making method provided by an embodiment of the present invention.

[0016] This invention provides a multimodal agent perception memory and autonomous decision-making method, which may include the following steps: S1: Divide the exhaust membrane of the microfluidic system into regions to obtain multiple pore cluster units.

[0017] In one possible implementation, S1 specifically includes sub-steps S101 to S103: S101: Acquire multimodal observation data of the microfluidic system, including microscopic image data, sensor data, and control command data.

[0018] It should be noted that a microfluidic system refers to an integrated system used for the transport, processing, and detection of trace fluids, which contains fluid channels to guide fluid flow. Multimodal observation data refers to the collection of various types of data acquired during the operation of a microfluidic system.

[0019] It should be noted that microscopic image data refers to data obtained by imaging the fluid state within a fluid channel using a microscopic imaging device, used to reflect fluid morphology, bubble distribution, and interface changes. Sensor data refers to numerical data reflecting the fluid flow state collected by flow sensors, pressure sensors, etc., installed in the microfluidic system. Control command data refers to the control setting data used to drive actuators such as pumps and valves in the microfluidic system.

[0020] Specifically, during the operation of the microfluidic system, a microscopic imaging device installed on the system continuously or intermittently images the fluid channel region to acquire microscopic image data reflecting fluid morphology, bubble distribution, and interface changes. Simultaneously, flow sensors and pressure sensors installed in the microfluidic system collect real-time data on the fluid flow state within the channel to obtain sensor data characterizing changes in fluid velocity and pressure within the channel. Furthermore, in the control unit of the microfluidic system, control commands for driving the pump and valve devices are synchronously read or recorded to acquire control command data reflecting the current operational and regulatory state of the microfluidic system. During data acquisition, the microscopic image data, sensor data, and control command data are organized in a unified temporal order, ensuring that each type of data corresponds to the same operational process, thus forming multimodal observation data reflecting the operational status of the microfluidic system.

[0021] S102: Determine the venting membrane installed on the fluid channel in the microfluidic system.

[0022] It should be noted that the exhaust membrane refers to a porous thin film structure set on the fluid channel of a microfluidic system. It is used to communicate with the fluid channel and allow gas to pass through. Its interior is composed of a large number of interconnected micropores, which can provide an exhaust path for gas during fluid operation.

[0023] It should be noted that fluid channels refer to structural paths set inside a microfluidic system to guide fluid flow. Their shape consists of microscale grooves, cavities, or channels, which can provide a confined flow space for fluid during the operation of the microfluidic system, allowing the fluid to be transported, distributed, and collected within the system according to a predetermined path, thereby achieving continuous or intermittent flow control of trace fluids.

[0024] Specifically, assembly relationship data related to fluid channels is read from the structural information of the microfluidic system. Combined with the design drawings or structural descriptions of the microfluidic system, the functional components connected to the fluid channels along their extension direction are identified one by one. From this, the membrane structure communicating with the fluid channels and used for gas exhaust is determined. During the identification process, by analyzing the spatial position and connectivity between the membrane structure and the fluid channels, a direct fluid or gas exchange path is confirmed, thus identifying the membrane structure as an exhaust membrane. Simultaneously, a correlation is established between the identified exhaust membrane and the corresponding fluid channel for subsequent structural analysis and data processing of the exhaust membrane, ensuring that the identified exhaust membrane is a component actually involved in gas exhaust during the operation of the microfluidic system.

[0025] S103: Based on the structural parameters of the exhaust film on the fluid channel, the exhaust film on the fluid channel is divided into regions to obtain multiple pore cluster units.

[0026] It should be noted that structural parameters refer to the parameter information describing the morphology of the porous structure inside the exhaust membrane, including the distribution, connectivity, and spatial relationship of the pores. Region division refers to dividing the overall area of ​​the exhaust membrane into multiple local regions based on its structural parameters. Each local region corresponds to a set of interconnected pore structures, thus forming multiple pore cluster units. A pore cluster unit refers to a local structural unit obtained based on the spatial distribution and connectivity between micropores within the exhaust membrane. Each pore cluster unit consists of several interconnected micropores and is used to characterize the connectivity state of the porous structure inside the exhaust membrane within a local region.

[0027] Specifically, structural parameter information describing the porous structure inside the exhaust membrane is obtained. This information includes the spatial distribution of micropores within the exhaust membrane, the connectivity between micropores, and their positional relationship within the overall exhaust membrane. Based on this information, the entire exhaust membrane region is analyzed. A group of spatially adjacent and interconnected micropores is defined as a local region, and different local regions are distinguished by their connectivity and spatial location. After dividing the overall exhaust membrane region, each local region consisting of several interconnected micropores is defined as a pore cluster unit, thus forming multiple pore cluster units within the exhaust membrane. Each pore cluster unit can independently characterize the connectivity state of the porous structure within the corresponding region, which is used for subsequent analysis of the liquid occupancy and connectivity within the exhaust membrane.

[0028] In this embodiment of the invention, by dividing the exhaust membrane into regions and obtaining multiple pore cluster units, the originally continuous and complex porous structure inside the exhaust membrane is decomposed into several structural units with clear spatial ranges and connectivity relationships. This facilitates the partitioned description and analysis of the liquid occupancy inside the exhaust membrane, and helps to transform the overall connectivity state inside the exhaust membrane into multiple local connectivity states for characterization. This improves the ability to identify the liquid connectivity changes inside the exhaust membrane, and provides a foundation for subsequent calculation of liquid filament connectivity and state analysis based on pore cluster units. This allows the structural changes inside the exhaust membrane to participate in subsequent data processing and decision-making processes in a clearer and more manageable manner.

[0029] In this embodiment of the invention, multiple pore cluster units are required because the exhaust membrane is composed of a large number of interconnected micropores. The liquid wetting inside the membrane does not cover the entire membrane all at once, but rather occupies local micropore regions first. Subsequently, multiple occupied local regions gradually connect to form a continuous channel from the liquid side to the gas side, forming a transpore cluster liquid filament. By dividing the exhaust membrane into multiple pore cluster units, the micropore structure inside the membrane can be numbered and managed one by one according to the locally connected regions. This allows the occupancy state of each pore cluster unit and whether connections occur between pore cluster units to be recorded and processed item by item. Thus, even with only external data such as channel images, flow rate, pressure, and control commands, the process of liquid occupancy inside the membrane expanding from local to interconnected can be expressed accordingly. This avoids the inability to distinguish between local occupancy and interconnectedness when only the overall membrane state is described, thereby reducing the possibility of multimodal agents mistakenly writing bubble reduction and flow stability into long-term memory as safe states.

[0030] It should be noted that the transpore cluster through-fluid filament refers to the continuous liquid passage formed by the wetting and occupation of the micropores inside the hydrophobic porous venting membrane by the liquid. This passage is formed by multiple liquid-occupied pore cluster units interconnected in the three-dimensional pore cluster network of the venting membrane, and it penetrates the liquid side and the gas side along the thickness direction of the venting membrane, so that a thin and long connected path occupied by liquid appears inside the venting membrane. This connected path may not necessarily show obvious leakage immediately on a macroscopic scale, but it forms a through connection of liquid from the liquid side to the gas side in terms of structural connectivity.

[0031] S2: Calculate the liquid filament connectivity based on the structural connectivity between each pore cluster unit.

[0032] In one possible implementation, S2 specifically includes sub-steps S201 to S204: S201: Determine the structural connectivity between each pore cluster unit.

[0033] It should be noted that structural connectivity refers to whether different pore cluster units are interconnected through microporous channels within the internal spatial structure of the exhaust membrane, reflecting the actual connectivity between pore cluster units within the exhaust membrane.

[0034] Specifically, the structural description information or manufacturing parameter information of the exhaust membrane is obtained, and the relative spatial positions of each pore cluster unit within the exhaust membrane are determined accordingly. The existence of direct or indirect micropore connectivity between any two pore cluster units is analyzed. By determining whether two pore cluster units are connected through a continuous micropore structure within the exhaust membrane, the structural connectivity between them is determined. When there is a continuous pathway composed of micropores between two pore cluster units, they are identified as interconnected pore cluster units; when there is no continuous micropore pathway between two pore cluster units, they are identified as disconnected pore cluster units. By analyzing the pairwise connectivity of all pore cluster units in the exhaust membrane, a complete set of connectivity relationships describing the structural connectivity between each pore cluster unit is formed, which is used for subsequent representation and processing of the overall pore cluster network structure.

[0035] S202: Construct an adjacency matrix based on the structural connectivity.

[0036] It should be noted that the adjacency matrix refers to a data set established based on structural connectivity, where each item is used to indicate whether there is a connectivity relationship and the degree of connectivity between two corresponding pore cluster units, and is used to describe the overall connectivity structure of the pore cluster network inside the exhaust membrane.

[0037] Specifically, using the defined pore cluster units as row and column indices of the matrix, a data structure describing the connectivity relationships between pore cluster units is established. For any pair of pore cluster units, based on the aforementioned structural connectivity relationships, it is determined whether the pair of pore cluster units are connected within the exhaust membrane through micropore pathways. When a connectivity relationship exists, the corresponding row and column indices in the adjacency matrix record data indicating connectivity; when no connectivity relationship exists, the corresponding positions in the adjacency matrix record data indicating non-connectivity. By filling the matrix content with the structural connectivity relationships between all pore cluster units in the exhaust membrane one by one, a complete adjacency matrix is ​​formed, enabling the adjacency matrix to systematically reflect the connectivity between each pore cluster unit within the exhaust membrane, which is used for subsequent calculations and analyses based on the connectivity relationships of the pore cluster units.

[0038] S203: Take the microscopic image data, sensor data and control command data as input data and input them into the trained multimodal computing model to obtain the pore cluster occupancy vector.

[0039] It should be noted that the pore cluster occupancy vector refers to the data set used to describe the liquid occupancy state in each pore cluster unit inside the exhaust membrane. Each component in the vector corresponds to a pore cluster unit and is used to represent the degree to which the micropores inside the pore cluster unit are occupied by liquid in the current state, thereby reflecting the distribution of liquid in different local areas inside the exhaust membrane.

[0040] Specifically, during the operation of the microfluidic system, the microscopic image data, sensor data, and control command data acquired within the same control cycle are time-correspondingly processed to ensure that all types of data reflect the same operating state. Subsequently, the processed microscopic image data is input to a computing unit for image information processing to extract image information representations reflecting fluid morphology and interface changes. The processed sensor data is input to a computing unit for numerical information processing to extract numerical information representations reflecting changes in fluid flow state. The processed control command data is input to a computing unit for control information processing to extract control information representations reflecting the system's regulation state. Based on this, the data representations from these different sources are fused and input into a trained multimodal computing model. The multimodal computing model performs overall calculations on the fused data and outputs a data set describing the degree of liquid occupancy of each pore cluster unit within the exhaust film, thus obtaining a pore cluster occupancy vector. This pore cluster occupancy vector reflects the liquid occupancy state of different pore cluster units within the exhaust film.

[0041] Specifically, the steps for training the multimodal computational model include: continuously collecting microscopic image data, sensor data, and control command data corresponding to different operating stages during multiple actual operations of the microfluidic system; uniformly organizing the microscopic image data, sensor data, and control command data collected within the same operating stage to form data samples that reflect the overall state of that operating stage; simultaneously, obtaining liquid occupancy state data of the pore cluster units inside the exhaust membrane corresponding to each data sample by recording the operating results of the microfluidic system or by acquiring the corresponding state of the exhaust membrane; and associating the pore cluster unit liquid occupancy state data with the data samples as supervisory data; and based on this, dividing the multiple sets of data samples and their corresponding pore cluster unit liquid occupancy state data into datasets for model learning, ensuring that each set of data samples includes microscopic image data, sensor data, control command data, and the corresponding pore cluster unit liquid occupancy state data. Subsequently, the dataset was input into the multimodal computational model. By repeatedly calculating the differences between the model output and the corresponding liquid occupancy state data of the pore cluster units, the internal parameters of the multimodal computational model were continuously adjusted, gradually establishing a correspondence between the multimodal input data and the liquid occupancy state of the pore cluster units. During model training, the model parameters were repeatedly updated using different data samples until the multimodal computational model could stably output data results consistent with the liquid occupancy status of the pore cluster units inside the exhaust membrane when inputting microscopic image data, sensor data, and control command data, thus completing the training of the multimodal computational model.

[0042] S204: Calculate the liquid filament connectivity based on the pore cluster occupancy vector and adjacency matrix.

[0043] It should be noted that liquid filament connectivity refers to the data used to describe the degree to which multiple liquid-occupied pore cluster units inside the exhaust membrane are interconnected in the pore cluster network to form a continuous liquid passage. Its magnitude reflects whether the liquid inside the exhaust membrane has developed from local occupation to the formation of a through connection between multiple pore cluster units, thus characterizing the change of the liquid from a dispersed state to a continuous connected state inside the exhaust membrane.

[0044] Specifically, the pore cluster occupancy vector corresponding to each pore cluster unit in the exhaust membrane is obtained, where each component of the vector represents the degree to which the corresponding pore cluster unit is occupied by liquid at the current moment. Simultaneously, an adjacency matrix constructed from the structural connectivity relationships between pore cluster units within the exhaust membrane is obtained, where each element of the adjacency matrix represents whether a connectivity relationship exists between two corresponding pore cluster units. Then, all pore cluster units in the exhaust membrane are paired. For any pair of pore cluster units, two components of its occupancy vector and the corresponding element in the adjacency matrix are taken, and the occupancy values ​​of both are multiplied by the matrix element to obtain an intermediate value representing the degree of connectivity between the pair of pore cluster units in the current state through liquid. After calculating the intermediate values ​​for all pore cluster unit pairs, all intermediate values ​​are summed to obtain the liquid filament connectivity, which represents the overall degree to which the liquid inside the exhaust membrane forms a connectivity path between multiple pore cluster units, allowing the liquid filament connectivity to reflect the changes in the liquid connectivity state within the exhaust membrane.

[0045] In this embodiment of the invention, by calculating the liquid filament connectivity, the liquid occupancy state and connectivity relationships among multiple pore cluster units within the exhaust membrane can be integrated into a single quantitative result. This allows for a numerical description of whether the liquid within the exhaust membrane forms continuous connectivity paths between different pore cluster units, thus avoiding the problem of relying solely on local pore cluster occupancy and failing to reflect the overall connectivity state. This liquid filament connectivity reflects the degree to which the liquid within the exhaust membrane evolves from dispersed occupancy to continuous connectivity. This facilitates continuous tracking of changes in liquid connectivity within the exhaust membrane even with only external multimodal observation data, providing a unified data foundation for subsequent construction of operational state sequences based on liquid filament connectivity, memory organization, and support for control decisions.

[0046] S3: Based on the liquid filament connectivity, obtain the topological memory unit corresponding to the running stage.

[0047] In one possible implementation, S3 specifically includes sub-steps S301 to S307: S301: Determine the operating phase based on multiple control cycles of the microfluidic system.

[0048] It should be noted that the control cycle refers to the time interval during which the microfluidic system performs one complete adjustment and response according to the control command during operation. Each control cycle corresponds to one stable operation process of the pump and valve devices. The operation phase refers to the time interval composed of multiple consecutive control cycles, used to describe the overall state of the microfluidic system during a continuous operation process.

[0049] Specifically, during operation, continuously executed control commands are recorded, and the time interval between each control command completion and the pump and valve devices reaching a stable operating state is marked as a control cycle. Based on this, multiple control cycles that are sequentially arranged in time and maintain consistent operating states are aggregated, and the time intervals corresponding to these consecutive control cycles are merged to determine the same operating phase. In this determination process, by analyzing the changes in control commands and the continuity of execution between control cycles, it is ensured that the operating phase covers a continuous, uninterrupted operating process, thus enabling the determined operating phase to fully reflect the overall changes in the liquid connectivity state and control behavior within the exhaust film during that period.

[0050] S302: Sort the liquid filament connectivity according to the order of the control cycles to obtain the liquid filament connectivity sequence.

[0051] It should be noted that the liquid filament connectivity sequence refers to a set of liquid filament connectivity data arranged in the order of the control cycle, used to describe the evolution of the liquid connectivity state inside the exhaust membrane during the operation phase.

[0052] Specifically, within the defined operating phase, the filament connectivity data corresponding to each control cycle are read sequentially, and the data are arranged according to the chronological order of the control cycles. During the arrangement process, it is ensured that each filament connectivity data point corresponds one-to-one with its corresponding control cycle, so that the arranged data order is consistent with the actual sequence of control cycles during operation. By completing the above sorting operation, a set of filament connectivity data arranged continuously in chronological order is formed, thus obtaining a filament connectivity sequence, which is used to describe the continuous evolution of the liquid connectivity state inside the exhaust membrane during the operating phase with changes in the control cycle.

[0053] S303: Calculate the change in connectivity between two adjacent control cycles based on the liquid filament connectivity sequence.

[0054] It should be noted that the change in connectivity refers to the difference in the connectivity between two adjacent control cycles, which is used to reflect the change in the liquid connectivity state inside the exhaust membrane between adjacent operating cycles.

[0055] Specifically, in the obtained filament connectivity sequence, the filament connectivity data corresponding to each control cycle is read sequentially according to the order of the control cycles. Two adjacent filament connectivity data points in the sequence are grouped together, and the filament connectivity data corresponding to the later control cycle is subtracted from the filament connectivity data corresponding to the earlier control cycle to obtain the difference data reflecting the magnitude of the change in liquid connectivity state between these two adjacent control cycles. By performing the above difference calculation sequentially on the data at all adjacent positions in the filament connectivity sequence, the connectivity change corresponding to each adjacent control cycle is obtained.

[0056] S304: Sort the changes in connectivity sequentially to obtain a sequence of connectivity changes.

[0057] It should be noted that the connectivity change sequence refers to the set of connectivity change data arranged in the order of the control cycle, which is used to describe the continuity and trend of the changes in the connectivity state of the liquid filament.

[0058] Specifically, after calculating the connectivity changes in adjacent control cycles, all connectivity change data are arranged according to the chronological order of the control cycles corresponding to the connectivity changes. During the arrangement process, the temporal order of each connectivity change data is maintained with its corresponding adjacent control cycle, thus forming a sequence of connectivity changes arranged continuously according to the control cycle order.

[0059] S305: During the operation phase, acquire the control action sequence corresponding to each control cycle.

[0060] It should be noted that the control action sequence refers to the set of control command data corresponding to each control cycle during the operation phase, which is used to reflect the adjustment behavior taken by the microfluidic system during operation.

[0061] Specifically, at the start of the operation phase, each subsequent control cycle is marked, and within each control cycle, the control commands used to adjust the fluid operating state are recorded. During the recording process, a correspondence is established between each control command and its corresponding control cycle, thus forming a set of control commands arranged in the order of the control cycles. After recording all control cycles within the operation phase, the control commands corresponding to each control cycle are arranged sequentially to form a control action sequence, ensuring that the control action sequence fully reflects the adjustment behavior taken as the control cycles change within the operation phase.

[0062] S306: After the operation phase is completed, obtain the detection quality indicators of the operation phase.

[0063] It should be noted that the testing quality indicators refer to the data obtained by the testing module after the operation phase ends, which are used to reflect the overall quality level of the testing results during that operation phase.

[0064] Specifically, after the operation phase is completed, the test results generated during that phase are summarized and processed, and data for evaluating the overall testing effectiveness of that phase are extracted from the test results. Testing quality indicators are obtained by comprehensively calculating the completeness, consistency, and stability of the test results within the operation phase, and are used to reflect the overall quality level of the testing process within that phase. By establishing a correlation between the testing quality indicators and the corresponding operation phase, the testing quality indicators can serve as a unified evaluation result of the operation effectiveness of that phase, providing a basis for subsequent analysis of the relationship between changes in liquid connectivity and control behavior within the operation phase.

[0065] S307: Correlate the liquid filament connectivity sequence, connectivity change sequence, control action sequence, and detection quality index to obtain the topological memory unit corresponding to the operation stage.

[0066] It should be noted that a topological memory unit refers to a data unit formed by associating the liquid filament connectivity sequence, connectivity change sequence, control action sequence, and detection quality index. It is used to describe the correspondence between the liquid connectivity state change process inside the exhaust membrane and the control behavior and detection results.

[0067] Specifically, after the operation phase ends, the obtained liquid filament connectivity sequence, connectivity change sequence, and control action sequence are uniformly organized to ensure consistency in the time sequence of these data types. Subsequently, the liquid filament connectivity sequence and connectivity change sequence are used to describe the changes in the liquid connectivity state within the exhaust membrane, the control action sequence is used to describe the regulatory behavior during the operation phase, and the detection quality indicators are used to evaluate the overall detection effect of the operation phase. These different types of data are bound together according to the same operation phase to form combined data units containing information on liquid connectivity state changes, regulatory behavior, and detection result evaluation. This results in a topological memory unit that corresponds one-to-one with that operation phase, enabling the topological memory unit to fully reflect the relationship between the evolution of the liquid connectivity state, regulatory behavior, and detection results during that operation phase, for subsequent memory retrieval and decision analysis.

[0068] In this embodiment of the invention, by obtaining the topological memory unit corresponding to the operation stage, the changes in the liquid connectivity state inside the exhaust membrane during a continuous operation process, the corresponding adjustment behavior, and the final detection results can be uniformly encapsulated into a complete data unit. This allows the operation experience of each operation stage to be stored in a structured form, which is beneficial for recalling and comparing historical operation experience based on the liquid connectivity state change process rather than a single moment's state during subsequent operation. This avoids ignoring the potential risks brought about by the continuous evolution of the liquid connectivity state based solely on short-term stable performance, thereby providing a reliable experience basis for the subsequent establishment of a perception memory and autonomous decision-making mechanism, and improving the overall stability of the operation process and the reliability of the detection results.

[0069] S4: Sort all topological memory units to obtain sorted topological memory units.

[0070] In one possible implementation, S4 specifically includes sub-steps S401 to S403: S401: Get the current filament connectivity sequence of the current running phase.

[0071] It should be noted that the current liquid filament connectivity sequence refers to a set of liquid filament connectivity data arranged in the order of the control cycle within the current operating phase, used to reflect the change process of the liquid connectivity state inside the exhaust membrane within the current operating phase.

[0072] Specifically, within the current operating phase, the filament connectivity data corresponding to each control cycle is read sequentially according to the predetermined control cycle order, and the filament connectivity data is mapped one-to-one with the corresponding control cycle. During the reading process, it is ensured that the acquired filament connectivity data all originate from continuous control cycles within the current operating phase, so that the data can fully reflect the process of changes in the liquid connectivity state inside the exhaust membrane with the control cycle within the current operating phase. After acquiring the filament connectivity data corresponding to all control cycles within the current operating phase, the filament connectivity data is arranged according to the chronological order of the control cycles, thereby forming the current filament connectivity sequence.

[0073] S402: Calculate the absolute difference distance between the current liquid filament connectivity sequence and the liquid filament connectivity sequence in the topological memory cell.

[0074] It should be noted that the absolute difference distance refers to the data obtained by calculating the difference between the current filament connectivity sequence and the filament connectivity sequence in the topological memory unit at the corresponding positions, and then summing the values. It is used to represent the degree of similarity between the current operating stage and the historical operating stage in the process of changing the liquid connectivity state.

[0075] Specifically, after acquiring the current filament connectivity sequence for the current operating phase, the corresponding filament connectivity sequence is read from each topological memory unit. The current filament connectivity sequence and the filament connectivity sequences in the topological memory units are aligned according to the control cycle order, ensuring that filament connectivity data in the same order positions correspond to each other. After sequence alignment, for each pair of corresponding filament connectivity data, the difference between them is calculated and its magnitude is recorded to obtain the difference data reflecting the degree of difference in the liquid connectivity state at that position. The difference data obtained from all corresponding positions are accumulated to obtain the absolute difference distance, which represents the overall difference in the liquid connectivity state change process between the current operating phase and the operating phase corresponding to that topological memory unit, thereby achieving a quantitative description of the similarity of the filament connectivity evolution process across different operating phases.

[0076] S403: Sort all topological memory cells according to the absolute difference distance to obtain the sorted topological memory cells.

[0077] It should be noted that the sorted topological memory units refer to the sequence of topological memory units obtained by arranging all topological memory units according to the size of their absolute difference distances. The topological memory units that are closer to the current operating stage in terms of the liquid filament connectivity state change process are placed at the beginning of the sequence for subsequent memory retrieval and decision analysis.

[0078] Specifically, a one-to-one correspondence is established between each topological memory unit and its corresponding absolute difference distance. All topological memory units are arranged according to their absolute difference distances, with units having smaller absolute differences placed at the beginning and those with larger absolute differences placed at the end. During the sorting process, the data content within each topological memory unit remains unchanged; only the order of the units is adjusted. By completing this sorting operation, a sequence of topological memory units is formed, arranged from highest to lowest similarity to the current liquid connectivity state change process, thus obtaining the sorted topological memory units.

[0079] In this embodiment of the invention, by sorting all topological memory units according to their absolute difference distance, topological memory units whose liquid connectivity state changes in historical operating stages are more closely related to the current operating stage can be identified first, thus avoiding interference caused by indiscriminate processing of all historical memories. This sorting method arranges the topological memory units in an ordered manner according to the similarity of their liquid filament connectivity evolution processes. This facilitates the rapid location of historical operating experiences that match the current liquid connectivity state evolution trend in subsequent processing, providing a clear reference order for analysis and decision-making based on historical experience, and improving the targeting and effectiveness of memory retrieval.

[0080] S5: Calculate the average neighbor detection quality based on the sorted topological memory units; In one possible implementation, S5 specifically includes sub-steps S501 and S502: S501: Based on the absolute difference distance, select L topological memory units from the sorted topological memory units as the nearest neighbor set.

[0081] It should be noted that the nearest neighbor memory set refers to a set of topological memory units selected from the sorted topological memory units. The topological memory units in this set have a high similarity to the current operating stage in terms of the process of changes in the liquid connectivity state, and are used as a historical reference for the current operating state.

[0082] Specifically, after sorting all topological memory units, they are read sequentially according to their order in the sorting results. During the reading process, topological memory units are selected one by one, starting from those at the top of the sorting order, and added to the same set. This selection process continues, following the sorting order, until a predetermined number of topological memory units are contained in the set, thus forming a proximity memory set. The topological memory units in the proximity memory set correspond to historical operating stages where the liquid connectivity state change process is relatively close to the current operating stage, for subsequent analysis and processing based on historical similar operating experience.

[0083] S502: Calculate the average neighbor detection quality based on all detection quality metrics in the neighbor memory set.

[0084] It should be noted that the average quality of neighbor detection refers to the data obtained by summarizing and calculating the detection quality indicators corresponding to each topological memory unit in the neighbor memory set. It is used to reflect the overall level of detection performance in historical operating stages similar to the current operating state.

[0085] Specifically, the detection quality index data corresponding to each topological memory unit in the neighbor memory set is read sequentially, and the detection quality index data is uniformly organized into the same data set. The detection quality index data in this data set are then numerically accumulated one by one to obtain the cumulative result of the detection quality index. After accumulating all the detection quality index data, the cumulative result is correlated with the number of topological memory units contained in the neighbor memory set to calculate the average neighbor detection quality, which reflects the overall detection performance level of the neighbor memory set. This allows the average neighbor detection quality to comprehensively represent the overall performance in terms of detection result quality in historical operating stages similar to the current operating state.

[0086] In this embodiment of the invention, by calculating the average quality of neighboring detections, the detection results of multiple historical operating stages that are similar to the current operating stage's liquid connectivity state change process can be comprehensively summarized, thereby avoiding the random influence caused by judging based solely on the detection results of a single historical operating stage. This average quality of neighboring detections can reflect the overall level of detection performance under similar liquid connectivity state evolution conditions, which is beneficial for providing a stable reference for subsequent construction of operating state assessment and decision-making basis, and improving the reliability and consistency of analysis based on historical similarity experience.

[0087] S6: Generate the liquid filament topology state vector based on the average quality of neighbor detections.

[0088] In one possible implementation, S6 specifically includes sub-steps S601 to S603: S601: Calculate the average filament connectivity based on the filament connectivity sequence.

[0089] It should be noted that the average liquid filament connectivity refers to the result obtained by summarizing and calculating the liquid filament connectivity data in the liquid filament connectivity sequence, and is used to represent the overall level of liquid connectivity within the exhaust membrane during this operating phase.

[0090] Specifically, within a defined operational phase, a sequence of filament connectivity values ​​arranged according to the control cycle is acquired, and each filament connectivity value in the sequence is read sequentially. During the reading process, all filament connectivity values ​​are accumulated to form a cumulative result. After accumulating all data in the filament connectivity sequence, the cumulative result is correlated with the number of data points contained in the sequence to obtain the average filament connectivity value, which represents the overall level of liquid connectivity within the exhaust membrane during this operational phase.

[0091] S602: Calculate the average connectivity change based on the connectivity change sequence.

[0092] It should be noted that the average connectivity change refers to the result obtained by summarizing and calculating the data of each change in the connectivity change sequence, and is used to reflect the strength of the overall change in the liquid connectivity state inside the exhaust membrane during this operating phase.

[0093] Specifically, each connectivity change data point in the connectivity change sequence is read sequentially. Then, these connectivity change data points are accumulated one by one to obtain the cumulative connectivity change result. After accumulating all connectivity change data, the cumulative result is compared with the number of data points in the connectivity change sequence to obtain the average connectivity change, which reflects the overall strength of the change in the liquid connectivity state inside the exhaust membrane during this operating phase. This average connectivity change allows the average connectivity change to describe the overall magnitude of the change in the liquid connectivity state during the operating phase.

[0094] S603: Combine the average filament connectivity, the average connectivity change, and the average neighbor detection quality into a filament topological state vector.

[0095] It should be noted that the liquid filament topology state vector refers to the data formed by combining the average liquid filament connectivity, the average connectivity change, and the average quality of neighboring detections. It is used to comprehensively describe the overall state of the liquid connectivity state inside the exhaust membrane, its changes, and the quality of the corresponding detection results during the current operating phase.

[0096] It should be noted that liquid filaments refer to elongated liquid pathways formed by the continuous occupation and interconnection of liquid in micropores or channels within porous venting membranes or microporous structures. These liquid pathways typically extend along the connecting paths between micropores, and their lateral dimensions are limited by the micropore scale. In the longitudinal direction, they can span multiple pore cluster regions. When multiple micropore regions that are partially occupied by liquid are interconnected, a liquid filament is formed. This term is used to describe the existence of liquid in porous structures, where it gradually develops from partial occupation to a continuous interconnected state.

[0097] Specifically, the average liquid filament connectivity is used to describe the overall level of liquid connectivity during the operation phase; the change in average connectivity is used to describe the strength of changes in liquid connectivity; and the average quality of neighboring detections is used to describe the detection performance level of similar operation phases. Based on this, the above three types of data are combined in a fixed order to form a unified dataset, thereby obtaining the liquid filament topology state vector. This vector comprehensively reflects the liquid connectivity state inside the exhaust membrane during the current operation phase, its changes, and the corresponding detection result quality level, for use in subsequent decision processing.

[0098] In this embodiment of the invention, by obtaining the liquid filament topological state vector, the degree of liquid connectivity within the exhaust membrane, the overall variation range of the liquid connectivity state, and the quality of detection results under similar operating conditions during the operation phase can be uniformly summarized into a structured data expression, thereby avoiding the limitations of relying solely on a single connectivity or instantaneous change information to judge the operating state. This liquid filament topological state vector can comprehensively reflect the liquid connectivity state within the exhaust membrane and its evolution trend. Combined with the detection results of similar historical operating phases, it provides a stable and comprehensive state description basis for subsequent analysis and decision-making based on the liquid connectivity topological state, improving the reliability of the operating state assessment and decision-making process.

[0099] S7: Based on the liquid filament topological state vector, the updated action value data is obtained.

[0100] In one possible implementation, S7 specifically includes sub-steps S701 to S704: S701: Select the first control action from the control action sequence for reinforcement learning update.

[0101] It should be noted that the first control action refers to the adjustment operation selected from the control action sequence that corresponds to the starting position of the operation phase or the initial change phase of the liquid communication state, and is used to represent the earliest control behavior applied within that operation phase.

[0102] It should be noted that the data or operation objects used for reinforcement learning updates refer to the data or operation objects selected during the operation and participating in subsequent learning adjustments. These data or operation objects are associated with the corresponding operating states and their results, and are used to correct and improve the existing action value data in subsequent processing. This enables the adjustment behavior to be improved based on the obtained result information under the same or similar operating states, thereby gradually forming a decision-making basis based on continuous self-adjustment based on operating feedback.

[0103] Specifically, after acquiring the control action sequence within the operational phase, the sequence is read in chronological order according to the control cycle. During the reading process, the control action corresponding to the initial control cycle of the operational phase is identified as the first control action, reflecting the earliest regulatory behavior applied to the fluid operation within that operational phase. By establishing a correspondence between the first control action and the liquid filament topology state vector formed during the operational phase, as well as the detection quality index obtained after the phase ends, the first control action can serve as a control action for value assessment and adjustment in subsequent reinforcement learning updates. This provides a clear operational object for learning and updating regulatory behavior based on operational results.

[0104] S702: The detection quality index is used as the reward value for the liquid filament topology state vector and the first control action.

[0105] It should be noted that the return value refers to the data formed by establishing the association between the detection quality index and the corresponding liquid filament topology state vector and the first control action, and is used to represent the result effect of taking the control action in the liquid communication state.

[0106] Specifically, the detection quality index is used as the return value of the liquid filament topology state vector and the first control action. This means that the detection result quality evaluation data obtained after the end of the operation phase is established with the corresponding liquid filament topology state vector and the earliest applied control action within the operation phase. This allows the detection quality index to reflect the overall effect of the control action taken in the liquid communication state, thereby providing a clear result feedback basis for the subsequent learning and adjustment of the effect relationship between the liquid filament topology state vector and the control action.

[0107] Specifically, the detection quality index is used as the reward value for the liquid filament topology state vector and the first control action because the detection quality index can comprehensively reflect the evolution of the liquid connectivity state inside the exhaust film during an operation phase and the overall impact of the control actions taken on the final detection result. Using it as the reward value allows the liquid connectivity state described by the liquid filament topology state vector and the adjustment behavior represented by the first control action to be directly correlated with the actual operating result. This enables the subsequent reinforcement learning process to no longer adjust based solely on intermediate state changes, but to use the final detection effect as the evaluation basis. This helps guide the control action to gradually optimize in a direction that is conducive to improving detection quality under similar liquid connectivity states, avoiding the formation of learning results that are unrelated to detection reliability.

[0108] S703: Generates action value data based on the liquid filament topology state vector and target control action.

[0109] It should be noted that action value data refers to data used to represent the level of effect of a certain control action under a specific liquid filament topology state vector condition, and is used to reflect the relative advantages and disadvantages of the control action in the liquid connectivity state.

[0110] Specifically, after obtaining the filament topology state vector describing the overall state of the liquid connectivity in the current operating phase, the target control action corresponding to this operating phase is read, and a one-to-one correspondence is established between the filament topology state vector and the target control action. The filament topology state vector is used as input data to characterize the current liquid connectivity state, and the target control action is used as descriptive data for the regulatory behavior taken in this liquid connectivity state. These two are combined to form a data record representing the correspondence between state and action. After establishing the correspondence, the data record is stored as an action value data point, enabling this action value data to represent the basis for evaluating the effect of the target control action taken under the current liquid connectivity state conditions.

[0111] S704: Based on the reward value, the action value data is updated using reinforcement learning to obtain the updated action value data.

[0112] It should be noted that the updated action value data refers to the data obtained by adjusting the original action value data after introducing the reward value. It is used to reflect the changes in the effect of control actions in the corresponding liquid connectivity state after combining historical operation results.

[0113] Specifically, the steps to obtain the updated action value data include: for each action value data to be updated, reading its corresponding reward value, which is determined by the detection quality index of that operation phase and reflects the actual operating effect under the corresponding liquid filament topology state vector and target control action conditions. Simultaneously, reading the action value data corresponding to each candidate control action subsequently related to the liquid filament topology state vector, and taking the one with the largest value as the maximum action value data for the subsequent state, and reading the pre-set learning rate parameter α and discount factor parameter γ. The learning rate parameter α controls the proportion of the current correction in the old action value data, and the discount factor parameter γ controls the influence weight of the maximum action value data for the subsequent state in the expected return. Then, adding the reward value to the product of the discount factor parameter γ and the maximum action value data for the subsequent state yields the expected return. Finally, subtracting the original action value data from the expected return yields the deviation data reflecting the degree of difference between the two. Based on this, the deviation data is multiplied by the learning rate parameter α to obtain the correction amount used to correct the original action value data. This correction amount is then added to the original action value data to form new action value data, which replaces the original action value data for storage. This process of reading the return value, calculating the expected return, obtaining the deviation, and correcting it according to the learning rate and discount factor parameters is repeated multiple times in different operating phases. This allows the action value data corresponding to each liquid filament topology state vector and control action combination to gradually accumulate and reflect long-term operating results, thereby obtaining updated action value data. This provides a more reliable basis for subsequent control decisions based on action value data.

[0114] Specifically, the learning rate parameter α and the discount factor parameter γ were set through an experimental calibration process based on historical operating data. This involved: first, collecting multiple sets of operating phase data representing different working conditions; then, repeatedly executing the action value update process in an offline environment using this data; initially, selecting intermediate values ​​between zero and one for both the learning rate parameter α and the discount factor parameter γ, and recording the convergence speed and numerical fluctuations of the action value data under this parameter combination. Subsequently, while keeping other conditions constant, the learning rate parameter α was adjusted multiple times. When α increased, it was observed whether significant oscillations occurred in the action value data update; when α decreased, it was observed whether convergence became too slow. By comparing the combined performance of convergence speed and stability under different α values, a learning rate parameter α that could converge within an acceptable time without producing large oscillations was determined. After obtaining a suitable learning rate parameter α, the discount factor parameter γ is adjusted in the same way. By changing the value of γ between zero and one, the sensitivity of action value data to long-term and short-term returns is compared. The discount factor parameter γ that can reflect the long-term running effect without causing the update process to be too slow is selected. Finally, the learning rate parameter α and the discount factor parameter γ obtained from the experiment are used as the specific values ​​within the parameter range used in the action value data update process.

[0115] It's important to note that reinforcement learning is manifested in the continuous updating of action value data. This is achieved by incorporating the detection quality index as a reward into the adjustment of the data corresponding to the liquid filament topology state vector and the control action. This ensures that the action value data is not a one-time, predetermined result, but rather repeatedly corrected and accumulated based on actual operational performance at different stages of operation. In this process, the liquid filament topology state vector corresponds to the state in reinforcement learning, the control action corresponds to the action in reinforcement learning, and the detection quality index corresponds to the reward in reinforcement learning. By continuously feeding the reward back into the action value data, subsequent actions under the same or similar liquid filament topology states can select a better control action based on the updated action value data. This demonstrates the reinforcement learning mechanism of continuously learning and improving decision-making behavior based on operational feedback.

[0116] In this embodiment of the invention, by obtaining updated action value data, the effect relationship between the liquid filament topology state vector and control actions can be continuously corrected and improved during operation, thereby gradually aligning the evaluation results of control actions with changes in actual detection effects. This updated action value data comprehensively reflects the historical operational results generated by corresponding control actions under different liquid connectivity conditions. This facilitates the selection of more appropriate adjustment behaviors based on accumulated operational feedback during subsequent operations, avoiding reliance on fixed rules or single-episode experience for control decisions, and improving the overall stability of the operational process and the reliability of the detection results.

[0117] S8: Perform control operations based on the updated action value data.

[0118] In one possible implementation, S8 specifically includes sub-steps S801 to S804: S801: Use the updated action value data with the largest value as the target action value data.

[0119] Specifically, after completing the reinforcement learning update of the action value data, all updated action value data corresponding to the current liquid filament topology state vector are read one by one, and the values ​​of each action value data are compared. During the comparison process, the action value data with the highest value under the current liquid filament topology state is determined and marked as the target action value data. The target action value data represents the evaluation result of the best regulation behavior performance under the current liquid connectivity state based on the comprehensive historical operating results.

[0120] Specifically, the updated action value data with the largest value is selected as the target action value data because action value data reflects the overall effect of corresponding control actions taken under specific liquid filament topology conditions. A larger value indicates better detection quality and operational performance of the control action in historical operations. By selecting the updated action value data with the largest value, subsequent control decisions can prioritize adjustment behaviors that have been validated as more advantageous in similar liquid connectivity conditions through multiple runs. This avoids random or inexperienced action selection and gradually guides adjustment behaviors towards improving detection quality and operational stability.

[0121] S802: Use the control action corresponding to the target action value data as the target control action.

[0122] Specifically, after determining the target action value data, control action information corresponding to that target action value data is read. This control action information originates from the state-action correspondence established during the aforementioned action value data generation and update process. By matching the target action value data with its corresponding control action, this control action is determined as the target control action, making it represent the optimal regulatory behavior evaluated under the current liquid filament topology.

[0123] S803: Generates target control instructions based on target control actions.

[0124] Specifically, the target control action is converted into a corresponding control command expression, and target control commands for execution are generated according to the adjustment content contained in the target control action. During the generation process, the target control commands are designed to fully reflect the adjustment intent and execution requirements of the target control action. By completing the above generation process, target control commands are obtained, which are used to adjust the fluid operating state accordingly during subsequent operation.

[0125] S804: Sends the target control command to the actuator of the microfluidic system to perform the control operation corresponding to the target control action.

[0126] Specifically, after generating the target control command, the command format is rectified to conform to a control command form recognizable by the actuator. The target control command is then transmitted to the actuator related to fluid regulation via a control interface, allowing the actuator to receive and parse the regulation content of the target control command. After parsing the target control command, the actuator adjusts the operating status of the pump and valve devices accordingly based on the regulation requirements described in the target control command. This alters the flow state of the fluid in the channel, completing the control operation corresponding to the target control action, and achieving actual regulation of the operation process based on the action value assessment results.

[0127] It should be noted that in this scheme, the multimodal agent's perception, memory, and autonomous decision-making are manifested in a complete closed-loop process formed by the same decision-making entity during operation. This decision-making entity first acquires microscopic image data, sensor data, and control command data simultaneously to perceive the fluid's operating state and the internal state of the exhaust membrane through multi-source information, and inputs various types of data into a multimodal computational model to form a unified internal state representation. Subsequently, a topological memory unit is constructed based on the changes in the liquid connectivity state during the operation phase. By comparing the current liquid connectivity state with historical operating experience, neighboring memories are retrieved to achieve memory recall. Based on this, a reinforcement learning mechanism is used, with detection quality indicators as rewards, to update the action value relationship between the liquid filament topological state vector and control actions. Based on the updated action value data, the entity autonomously selects target control actions, and finally converts the selected control actions into control commands and executes them. This allows the decision-making entity to form a closed-loop operation between multimodal perception, memory accumulation, and autonomous decision-making, demonstrating the multimodal agent's perception, memory, and autonomous decision-making capabilities.

[0128] Reference manual attached Figure 2 The diagram shows a schematic representation of the structure of a multimodal agent perception memory and autonomous decision-making system provided in an embodiment of the present invention.

[0129] This invention provides a multimodal agent perception memory and autonomous decision-making system 20, including: a processor 201 and a memory 202; The memory 202 stores programs or instructions that can run on the processor 201. When the program or instructions are executed by the processor 201, they implement the steps of the multimodal agent perception memory and autonomous decision-making method described above and achieve the same technical effect. To avoid repetition, the present invention will not elaborate further.

[0130] It should be understood that the processor 201 in this embodiment of the invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0131] It should also be understood that the memory 202 in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDRSDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DRRAM).

[0132] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0133] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0134] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0135] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0136] In the embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0137] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0138] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0139] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. A multimodal agent perception, memory, and autonomous decision-making method, characterized in that, include: S1: Divide the exhaust membrane of the microfluidic system into regions to obtain multiple pore cluster units; S2: Calculate the liquid filament connectivity based on the structural connectivity between each of the pore cluster units; S3: Based on the fluid filament connectivity, obtain the topological memory unit corresponding to the operation stage; S4: Sort all topological memory units to obtain sorted topological memory units; S5: Calculate the average neighbor detection quality based on the sorted topological memory units; S6: Generate a liquid filament topology state vector based on the average quality of the neighbor detection; S7: Based on the liquid filament topological state vector, obtain the updated action value data; S8: Perform control operations based on the updated action value data.

2. The multimodal agent perception, memory, and autonomous decision-making method according to claim 1, characterized in that, S1 specifically includes: S101: Acquire multimodal observation data of the microfluidic system, wherein the multimodal observation data includes microscopic image data, sensor data, and control command data; S102: Determine the venting membrane disposed on the fluid channel of the microfluidic system; S103: Based on the structural parameters of the exhaust film on the fluid channel, the exhaust film on the fluid channel is divided into regions to obtain multiple pore cluster units.

3. The multimodal agent perception, memory, and autonomous decision-making method according to claim 2, characterized in that, S2 specifically includes: S201: Determine the structural connectivity between each of the said pore cluster units; S202: Construct an adjacency matrix based on the aforementioned structural connectivity. S203: The microscopic image data, the sensor data, and the control command data are used as input data and input into the trained multimodal computing model to obtain the pore cluster occupancy vector; S204: Calculate the liquid filament connectivity based on the pore cluster occupancy vector and adjacency matrix.

4. The multimodal agent perception, memory, and autonomous decision-making method according to claim 1, characterized in that, S3 specifically includes: S301: Determine the operating phase based on the multiple control cycles of the microfluidic system; S302: Sort the liquid filament connectivity according to the order of the control cycles to obtain a liquid filament connectivity sequence; S303: Based on the liquid filament connectivity sequence, calculate the connectivity change between two adjacent control cycles; S304: Sort the connectivity changes sequentially to obtain a connectivity change sequence; S305: During the operation phase, acquire the control action sequence corresponding to each control cycle; S306: After the operation phase ends, obtain the detection quality indicators of the operation phase; S307: Associate the liquid filament connectivity sequence, the connectivity change sequence, the control action sequence, and the detection quality index to obtain the topological memory unit corresponding to the operation stage.

5. The multimodal agent perception, memory, and autonomous decision-making method according to claim 4, characterized in that, S4 specifically includes: S401: Obtain the current filament connectivity sequence for the current running phase; S402: Calculate the absolute difference distance between the current liquid filament connectivity sequence and the liquid filament connectivity sequence in the topological memory unit; S403: Sort all topological memory units according to the absolute difference distance to obtain sorted topological memory units.

6. The multimodal agent perception, memory, and autonomous decision-making method according to claim 5, characterized in that, S5 specifically includes: S501: Based on the absolute difference distance, select L topological memory units from the sorted topological memory units as the nearest neighbor set; S502: Calculate the average neighbor detection quality based on all detection quality indicators in the neighbor memory set.

7. The multimodal agent perception, memory, and autonomous decision-making method according to claim 4, characterized in that, S6 specifically includes: S601: Calculate the average filament connectivity based on the filament connectivity sequence; S602: Calculate the average connectivity change based on the connectivity change sequence; S603: Combine the average filament connectivity, the average connectivity change, and the average neighbor detection quality to form the filament topology state vector.

8. The multimodal agent perception, memory, and autonomous decision-making method according to claim 4, characterized in that, Specifically, S7 includes: S701: Select a first control action from the control action sequence for reinforcement learning update; S702: Use the detection quality index as the reward value of the liquid filament topology state vector and the first control action; S703: Generate action value data based on the liquid filament topology state vector and the target control action; S704: Based on the reward value, the action value data is updated using reinforcement learning to obtain the updated action value data.

9. The multimodal agent perception, memory, and autonomous decision-making method according to claim 1, characterized in that, S8 specifically includes: S801: Use the updated action value data with the largest value as the target action value data; S802: Take the control action corresponding to the target action value data as the target control action; S803: Generate target control instructions based on the target control action; S804: The target control command is sent to the actuator of the microfluidic system to execute the control operation corresponding to the target control action.

10. A multimodal agent-based perception, memory, and autonomous decision-making system, characterized in that, include: Processor and memory; The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the multimodal agent perception memory and autonomous decision-making method as described in any one of claims 1 to 9.

Images