SPR (Surface Plasmon Resonance) microfluidic detection process control system based on reinforcement learning optimization

Abstract

The invention discloses an SPR (Surface Plasmon Resonance) microfluidic detection process control system based on reinforcement learning optimization, which comprises the following steps: starting a baseline fluid and a real-time refractive index change signal in a microfluidic channel, and outputting original SPR signal data; performing preprocessing to generate a standardized state vector sequence; performing decision generation on the standardized state vector, and outputting a candidate operation list; inputting the candidate operation list into a Gumbel-Top-k sequence sampler, and outputting a deterministic control sequence; issuing the deterministic control sequence to a microfluidic execution module, and recording an actually executed operation track; corresponding response curve segments are collected, and an SPR detection curve corresponding to the execution track is output; performing result analysis based on the SPR detection curve to generate a performance evaluation result; and forming an SPR microfluidic detection process control scheme based on the pairing information of the performance evaluation result. According to the invention, SPR microfluidic detection process control is realized.

Description

Technical Field

[0001] This invention relates to the field of biosensing technology, and in particular to an SPR microfluidic detection process control system based on reinforcement learning optimization. Background Technology

[0002] Currently, surface plasmon resonance (SPR) technology, as a highly sensitive, label-free molecular interaction detection method, has been widely applied in fields such as biomedicine, environmental monitoring, and food safety. SPR technology typically relies on microfluidic channels for precise sample injection, rinsing, and replacement. The microfluidic system uses valve switching, flow rate regulation, and pressure control to alternate between the sample and buffer solution, thereby ensuring the integrity and accuracy of the detection curve. In existing technologies, integrated SPR and microfluidic devices are relatively mature, and related systems can achieve automated liquid path switching and signal acquisition, such as publicly available fully automated SPR bioanalyzers capable of performing multi-step sample injection and detection tasks. However, these systems generally rely on preset sequential logic or simple feedback correction mechanisms, resulting in rigid control processes and a lack of adaptive optimization capabilities. They are often inefficient and subject to certain uncertainties when facing complex detection conditions or multi-parameter constraints.

[0003] In terms of process control for detection, existing technologies typically employ a method of first generating an operation sequence and then eliminating non-conforming items through constraint checks. While this ensures basic executability, it generates a large number of invalid operations, reducing overall optimization efficiency. Furthermore, microfluidic detection involves multiple constraints, such as volume budgeting, differential pressure limits, valve switching frequency and sequence. Existing step-by-step processing methods struggle to comprehensively consider all factors during the generation phase, leading to frequent backend corrections and potential delays and performance fluctuations. In addition, traditional methods often process valve control duration and fluid volume budgeting independently in stages, easily causing a disconnect between the operation sequence and duration control, thus affecting the stability and repeatability of experimental results.

[0004] Therefore, how to provide an SPR microfluidic detection process control system based on reinforcement learning optimization is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an SPR microfluidic detection process control system based on reinforcement learning optimization. This invention introduces an optimization mechanism that combines reinforcement learning with Gumbel-Top-k sequence sampling, embedding feasibility constraints and budget gating at the generation stage. This avoids a large number of invalid candidates and back-end corrections, significantly improving the determinism of sequence generation and the success rate of execution. At the same time, it achieves linkage optimization of valve switching, flow rate, and duration, ensuring the stability and repeatability of the detection process, thereby improving the efficiency and reliability of SPR microfluidic detection results.

[0006] The SPR microfluidic detection process control system based on reinforcement learning optimization according to an embodiment of the present invention includes: The signal acquisition module is used to acquire the refractive index change signal during the molecular binding process in real time and output the raw SPR signal data with timestamps. The signal preprocessing module is used to perform noise suppression, baseline correction and amplitude normalization on the raw SPR signal data, and generate a time-aligned normalized state vector sequence. The decision generation module is used to assign a learnable logit and constraint tags to each operation and output a list of candidate operations. The sequence sampling module is used to apply relevant Gumbel perturbations to the logit operation and perform Top-k selection, outputting a deterministic control sequence that satisfies the feasibility conditions; The microfluidic execution module is used to record the actual operation trajectory and generate execution logs in real time; The SPR sensing module is used to collect corresponding response curve segments one by one, splice and verify them, and output a complete SPR detection curve. The performance evaluation and optimization module is used to generate performance evaluation results and bind them to the control sequence. Then, through weighted evaluation and constraint verification, an optimized SPR microfluidic detection process control scheme is formed.

[0007] Optionally, modules can be integrated using the following methods: A baseline fluid is initiated within a microfluidic channel, and the refractive index change signal during the molecular binding process is acquired in real time using an SPR sensing module, outputting raw SPR signal data. The raw SPR signal data is preprocessed, including noise suppression, baseline correction and amplitude normalization, to generate a time-aligned normalized state vector sequence; Decision generation is performed on the standardized state vector, constructing a candidate operation set composed of valve opening and closing sequence, duration, flow rate and pressure, and assigning a learnable logit and constraint label to each decision, outputting a candidate operation list; The candidate operation list is input into the Gumbel-Top-k sequence sampler. Gumbel perturbation is applied to the logit of each operation and Top-k selection is performed. The valve and fluid control sequence that meets the feasibility conditions is generated by combining hard mask and partial order constraints. At the same time, duration and volume budget gating are embedded in the sampling stage to output a deterministic control sequence. The deterministic control sequence is sent to the microfluidic execution module to implement valve switching, flow rate regulation, pressure setting and injection duration control one by one, and the actual operation trajectory is recorded. The operation trajectory is used as a time axis. The corresponding response curve segments are collected by the SPR sensing module, the complete response data is spliced ​​and verified, and the SPR detection curve corresponding to the execution trajectory is output. Based on the SPR detection curve, the results are analyzed, and the peak intensity, rising edge slope, dissociation residue and signal-to-noise ratio are quantitatively calculated to generate performance evaluation results, which are then bound to the control sequence one by one. Based on the pairing information from the performance evaluation results, an optimized SPR microfluidic detection process control scheme is formed, serving as a deterministic result of the performance evaluation.

[0008] Optionally, the output process of the raw SPR signal data specifically includes: The baseline fluid is degassed and filtered, and then injected into a microfluidic channel. A constant temperature and target volumetric flow rate are set, and the target linear velocity is determined according to the effective cross-sectional area of ​​the channel. A stable continuous flow is output through a micro-pump and verified by a pressure sensor and a flow meter. When the fluctuation of the baseline signal within the continuous sampling period does not exceed the preset stability threshold, a baseline stable operating condition parameter set is output. The parameter set includes at least the temperature set value, the volumetric flow rate set value, the channel cross-sectional area, the target linear velocity, the inlet and outlet pressures, and the sampling period. Using the baseline stable operating condition parameter set as input, SPR real-time acquisition is started under the baseline stable condition. The change of reflection intensity over time is recorded with a fixed sampling period and timestamp and channel identifier. According to the pre-calibrated linear conversion relationship between intensity and refractive index, the intensity change is converted into the refractive index change proportionally, the refractive index change data is generated in time sequence, and the corresponding timestamp sequence is output. Based on the refractive index change time series and timestamp sequence, integrity and continuity are verified according to the execution time order. At the same time, the refractive index change data, channel identifier, and baseline stable condition parameter set are encapsulated together to generate the original SPR signal data.

[0009] Optionally, the process of generating the standardized state vector sequence specifically includes: The system receives raw SPR signal data, takes the time series of refractive index change recorded over time and the corresponding timestamp as input, resamples the time axis at equal intervals and discards outliers that exceed the range, selects a continuous data segment as a stable segment to check whether the flow rate and temperature are within the set range, and outputs the resampled and stable raw signal sequence and the corresponding timestamp. Using the stability-checked SPR signal data as input, the sliding window length and endpoint processing method are set. The window slides gradually on the time axis to smooth and denoise the neighborhood of each sampling point. Several sampling points in the initial stage are selected as the baseline segment. The average value of the current baseline segment is calculated and subtracted point by point from the whole sequence to eliminate baseline offset and slow drift. The output is the SPR signal data with denoising and baseline correction completed. The amplitude range of the SPR signal data after denoising and baseline correction is calculated and normalized. When the amplitude range is zero, the entire sequence is set as a zero vector. The sliding window length and step size are set according to the preset state dimension. The continuous normalized samples are assembled into a state vector sequence of equal length in chronological order. At the same time, a relative time index corresponding to each state vector is generated. The time-aligned normalized state vector sequence and the corresponding relative time index are output.

[0010] Optionally, the process for outputting the candidate operation list includes: Based on the standardized state vector sequence, the valve identifier set is called according to the experimental configuration to determine all executable valve opening and closing commands, and an optional duration range is set for each valve. At the same time, the discrete values ​​of the target flow rate and the discrete values ​​of the target pressure are combined item by item to form atomic operations. Experimental constraint parameters are loaded when forming operations, and the constraint parameters are bound to each atomic operation. The bound atomic operation set is output. Based on the set of atomic operations, candidate operation sequences are generated step by step according to the preset maximum sequence length. Constraint indicators are calculated for each candidate operation sequence. The total injection volume is obtained by summing the products of the flow rate and duration of each atomic operation. The number of valve switching is obtained by comparing the valve states of adjacent operations. The maximum pressure difference is obtained by calculating the difference between the highest and lowest pressure values ​​in the sequence. The order of operations is checked. If all conditions are within the constraints, the sequence is marked as feasible; otherwise, it is marked as infeasible. A set of candidate operation sequences with feasibility marks is output. For each atomic operation in the candidate operation sequence set, a scoring function with learnable parameters is called under the current state vector to calculate the score. The scores of all atomic operations in the same sequence are accumulated sequentially to form a basic score. Then, the constraint cost is calculated according to three types of constraint indicators. The portions exceeding the volume limit, the pressure difference limit, and the switching number limit are accumulated according to preset weights to obtain the final constraint cost. The final constraint cost is then subtracted from the basic score to obtain the aggregate score. Finally, the candidate operation list is sorted by aggregate score from high to low based on the candidate operation sequence set, feasibility marker, and aggregate score.

[0011] Optionally, the output process of the deterministic control sequence specifically includes: Read the learnable logit score, valve identifier, duration, target flow rate and target pressure of each atomic operation in the candidate operation list one by one. Apply a feasibility hard mask to shield operations that violate the volume budget, differential pressure limit and switching number limit. Load partial order constraints to limit the allowed order relationship. Perform budget and safety constraints to remove items that are definitely not executable under the current budget. Form a candidate score table with mask mark and constraint label. Based on the candidate scoring table, a noise vector with a covariance structure is generated for each operation using relevant Gumbel perturbation. The noise is superimposed on the logit score to generate the perturbed score. Temperature parameters and random seeds are recorded to ensure that the sampling is reproducible. The perturbed score vector is checked for order consistency and stabilized according to partial order constraints. The perturbed score and the corresponding candidate index set that pass the check are output. Based on the verified perturbation scores, a cost matrix is ​​constructed, with rows corresponding to Top-k positions and columns corresponding to candidate operations. A rectangular assignment sparse selection model with entropy regularization is established, and the sparse selection matrix is ​​obtained by solving it. The candidate operation index corresponding to each Top-k position is determined according to the sparse selection matrix, and the candidates are sorted from high to low according to the perturbation scores. In the case of the same score, the position order is stabilized to obtain an ordered candidate set. Taking an ordered candidate set as input, duration and volume budget gating are sequentially embedded to simultaneously determine dwell time and volume consumption during the generation stage. When the remaining budget is insufficient to support the next item, the dwell time is reduced and the budget is updated according to the gating rules. For adjacent candidates with similar scores, a one-time Sinkhorn order refinement is performed to stabilize the sorting. At the same time, CVaR risk reweighting is applied to historical failed samples to suppress unstable sequences, and a deterministic control sequence is output.

[0012] Optionally, the recording of the actual executed operation trajectory specifically includes: The valve identifier, gating duration, target flow rate, and target pressure contained in the deterministic control sequence are parsed one by one, and a corresponding execution instruction queue is generated during the parsing process. Each instruction is bound to an index order and a preset timestamp. The execution command queue is sent to the microfluidic execution module one by one, driving the micropump to maintain the target flow rate within the duration specified by the command, controlling the valve to complete the opening and closing switching at the corresponding start and end time, and monitoring whether the pressure changes at the inlet and outlet meet the set target pressure. The flow rate, pressure and valve status collected during the execution process are recorded as the original execution log. The original execution log is synchronized with time and its duration is checked. For each entry, the actual duration is compared with the gating setting. When a deviation occurs, the pump speed is automatically adjusted or the valve holding time is extended to compensate for the difference. At the same time, the switching delay with the previous target is calculated and written into the log. The actual execution operation trajectory is recorded, including valve switching sequence, actual duration, real-time flow rate, real-time pressure and switching delay.

[0013] Optionally, the output process of the SPR detection curve specifically includes: The complete execution trajectory is used as an event list. The valve switching sequence, start and end time, actual duration, real-time flow rate, real-time pressure and switching delay in the event list are read one by one to generate the corresponding collection time window. The valve sequence and channel identifier are bound to each time window to form a collection schedule and index sorted by item. Using the collection schedule and index as input, the SPR sensing module is driven to collect the reflection intensity sequence within the time window corresponding to each entry, and the entry number, channel identifier, and actual duration are added. At the same time, the sampling validity is checked using the flow rate and pressure fields, and the valid signals are converted into a refractive index change time series point by point. The response segments with quality labels are generated by combining the start and end times. Based on the response fragment set, the response fragments are spliced ​​in the order of the execution event list items. The continuity of the response fragments of adjacent items is checked at the end and start times. If there is a gap due to switching delay, it is filled with the baseline segment. If there is an overlap due to drift, the tail of the previous item is trimmed. After all boundary processing is completed, the SPR detection curve is output.

[0014] Optionally, the process of generating the performance evaluation results specifically includes: Read the segments corresponding to the SPR detection curve and execution trajectory entries one by one, and perform signal baseline subtraction and amplitude normalization on each segment to obtain a standardized response sequence that can be directly used for feature analysis; For each segment of the standardized response sequence, the detection performance parameters are calculated, including the peak response amplitude, the slope of the rising phase, the residual value of the dissociation phase, and the signal-to-noise ratio. The parameters are then organized by entry number, and the performance evaluation results with time index and entry identifier are output. Based on the performance evaluation results and combined with the deterministic control sequence, the valve switching sequence, duration, target flow rate and target pressure in each control sequence are bound to the corresponding sections in the performance parameter table one by one.

[0015] Optionally, the formation of the optimized SPR microfluidic detection process control scheme specifically includes: Based on the performance evaluation results, set the weights and pass thresholds for the detection performance parameters, calculate the comprehensive score and attach compliance marks for volume budget, pressure limit and number of switching times, and generate a sequence evaluation table containing serial number, comprehensive score, compliance status and key descriptions, arranged in descending order of comprehensive score; The sequence evaluation table is subjected to constraint review and executability verification from high to low. Entries that do not meet the volume budget, pressure limit or switching number limit are skipped. The first control sequence that meets both constraints and quality requirements is selected as the optimal solution and output together with the corresponding performance parameters as the optimized SPR microfluidic detection process control scheme.

[0016] The beneficial effects of this invention are: 1. This invention combines feasibility hard masking and partial order constraints in the decision generation stage, and moves the volume budget, pressure limit and valve switching number to the sampling stage in advance, thereby avoiding the generation of invalid sequences from the source and greatly improving the determinism and execution success rate of the control scheme.

[0017] 2. This invention embeds a gating mechanism of duration and volume budget in the sequence sampling process, which enables the "operation selection" and "duration allocation" to be optimized synchronously, avoiding the timing drift problem caused by traditional step-by-step processing, thereby improving the stability and repeatability of the detection process.

[0018] 3. This invention achieves ordered Top-k selection through sparse assignment and entropy regularization, and combines Sinkhorn order refinement and CVaR risk reweighting to effectively reduce the uncertainty caused by the proximity scores between candidate sequences, thus ensuring the robustness and anti-interference ability of the generated results. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the SPR microfluidic detection process control system based on reinforcement learning optimization proposed in this invention. Figure 2 This is a schematic diagram of the SPR microfluidic detection process control system based on reinforcement learning optimization proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1-2 The SPR microfluidic detection process control system based on reinforcement learning optimization includes the following steps: The signal acquisition module is used to acquire the refractive index change signal during the molecular binding process in real time and output the raw SPR signal data with timestamps. The signal preprocessing module is used to perform noise suppression, baseline correction and amplitude normalization on the raw SPR signal data, and generate a time-aligned normalized state vector sequence. The decision generation module is used to assign a learnable logit and constraint tags to each operation and output a list of candidate operations. The sequence sampling module is used to apply relevant Gumbel perturbations to the logit operation and perform Top-k selection, outputting a deterministic control sequence that satisfies the feasibility conditions; The microfluidic execution module is used to record the actual operation trajectory and generate execution logs in real time; The SPR sensing module is used to collect corresponding response curve segments one by one, splice and verify them, and output a complete SPR detection curve. The performance evaluation and optimization module is used to generate performance evaluation results and bind them to the control sequence. Then, through weighted evaluation and constraint verification, an optimized SPR microfluidic detection process control scheme is formed.

[0022] This invention achieves end-to-end optimization of the SPR microfluidic detection process through modular design, maintaining efficient connection and consistency from signal acquisition to result evaluation. Signal preprocessing improves the stability of input data, decision generation and sequence sampling ensure the feasibility and determinism of control sequences under multiple constraints, microfluidic execution and SPR sensing achieve precise correspondence between the operation trajectory and the detection curve, and the performance evaluation and optimization module completes result-oriented improvements based on feedback. The overall solution significantly improves the efficiency, robustness, and reliability of detection, providing high-quality support for complex biological detection.

[0023] In this embodiment, the modules are interconnected using the following method: A baseline fluid is initiated within a microfluidic channel, and the refractive index change signal during the molecular binding process is acquired in real time using an SPR sensing module, outputting raw SPR signal data. The raw SPR signal data is preprocessed, including noise suppression, baseline correction and amplitude normalization, to generate a time-aligned normalized state vector sequence; Decision generation is performed on the standardized state vector, constructing a candidate operation set composed of valve opening and closing sequence, duration, flow rate and pressure, and assigning a learnable logit and constraint label to each decision, outputting a candidate operation list; The candidate operation list is input into the Gumbel-Top-k sequence sampler. Gumbel perturbation is applied to the logit of each operation and Top-k selection is performed. The valve and fluid control sequence that meets the feasibility conditions is generated by combining hard mask and partial order constraints. At the same time, duration and volume budget gating are embedded in the sampling stage to output a deterministic control sequence. The deterministic control sequence is sent to the microfluidic execution module to implement valve switching, flow rate regulation, pressure setting and injection duration control one by one, and the actual operation trajectory is recorded. The operation trajectory is used as a time axis. The corresponding response curve segments are collected by the SPR sensing module, the complete response data is spliced ​​and verified, and the SPR detection curve corresponding to the execution trajectory is output. Based on the SPR detection curve, the results are analyzed, and the peak intensity, rising edge slope, dissociation residue and signal-to-noise ratio are quantitatively calculated to generate performance evaluation results, which are then bound to the control sequence one by one. Based on the pairing information from the performance evaluation results, an optimized SPR microfluidic detection process control scheme is formed, serving as a deterministic result of the performance evaluation.

[0024] This invention introduces reinforcement learning optimization and Gumbel-Top-k sequence sampling mechanisms into the SPR microfluidic detection process, achieving a deep integration of control sequence generation and constraint conditions, thus avoiding the problems of invalid instructions and repeated corrections. Signal preprocessing and standardized vector generation improve the stability of data input, and the strict correspondence between the execution trajectory and the detection curve enhances the traceability of the experiment.

[0025] In this embodiment, the output process of the original SPR signal data specifically includes: The baseline fluid is degassed and filtered, and then injected into a microfluidic channel. A constant temperature and target volumetric flow rate are set, and the target linear velocity is determined according to the effective cross-sectional area of ​​the channel. A stable continuous flow is output through a micro-pump and verified by a pressure sensor and a flow meter. When the fluctuation of the baseline signal within the continuous sampling period does not exceed the preset stability threshold, a baseline stable operating condition parameter set is output. The parameter set includes at least the temperature set value, the volumetric flow rate set value, the channel cross-sectional area, the target linear velocity, the inlet and outlet pressures, and the sampling period. Using the baseline stable operating condition parameter set as input, SPR real-time acquisition is started under the baseline stable condition. The change of reflection intensity over time is recorded with a fixed sampling period and timestamp and channel identifier. According to the pre-calibrated linear conversion relationship between intensity and refractive index, the intensity change is converted into the refractive index change proportionally, the refractive index change data is generated in time sequence, and the corresponding timestamp sequence is output. Based on the refractive index change time series and timestamp sequence, integrity and continuity are verified according to the execution time order. At the same time, the refractive index change data, channel identifier, and baseline stable condition parameter set are encapsulated together to generate the original SPR signal data.

[0026] This invention significantly reduces the interference of background noise and baseline drift on the detection results by degassing and filtering the baseline fluid and maintaining a stable fluid environment under constant temperature and flow conditions. The dual verification mechanism combining a pressure sensor and a flow meter effectively ensures the stability and controllability of fluid conditions. Precise signal quantification is achieved through intensity and refractive index conversion, and operating parameters and timestamps are introduced into the data encapsulation, ensuring the integrity and traceability of the original signal. This provides a highly stable and accurate data input foundation for SPR microfluidic detection.

[0027] In this embodiment, the process of generating the standardized state vector sequence specifically includes: The system receives raw SPR signal data, takes the time series of refractive index change recorded over time and the corresponding timestamp as input, resamples the time axis at equal intervals and discards outliers that exceed the range, selects a continuous data segment as a stable segment to check whether the flow rate and temperature are within the set range, and outputs the resampled and stable raw signal sequence and the corresponding timestamp. Using the stability-checked SPR signal data as input, the sliding window length and endpoint processing method are set. The window slides gradually on the time axis to smooth and denoise the neighborhood of each sampling point. Several sampling points in the initial stage are selected as the baseline segment. The average value of the current baseline segment is calculated and subtracted point by point from the whole sequence to eliminate baseline offset and slow drift. The output is the SPR signal data with denoising and baseline correction completed. The amplitude range of the SPR signal data after denoising and baseline correction is calculated and normalized. When the amplitude range is zero, the entire sequence is set as a zero vector. The sliding window length and step size are set according to the preset state dimension. The continuous normalized samples are assembled into a state vector sequence of equal length in chronological order. At the same time, a relative time index corresponding to each state vector is generated. The time-aligned normalized state vector sequence and the corresponding relative time index are output.

[0028] This invention significantly improves the stability and accuracy of raw SPR signal data by resampling, smoothing and denoising, baseline correction, and amplitude normalization. Stability checks eliminate outliers and correct slow drift, effectively eliminating interference from environmental fluctuations. Amplitude normalization and state vector assembly achieve temporal alignment and unified feature representation of the data.

[0029] In this embodiment, the process of outputting the candidate operation list includes: Based on the standardized state vector sequence, the valve identifier set is called according to the experimental configuration to determine all executable valve opening and closing commands, and an optional duration range is set for each valve. At the same time, the discrete values ​​of the target flow rate and the discrete values ​​of the target pressure are combined item by item to form atomic operations. Experimental constraint parameters are loaded when forming operations, and the constraint parameters are bound to each atomic operation. The bound atomic operation set is output. Based on the set of atomic operations, candidate operation sequences are generated step by step according to the preset maximum sequence length. Constraint indicators are calculated for each candidate operation sequence. The total injection volume is obtained by summing the products of the flow rate and duration of each atomic operation. The number of valve switching is obtained by comparing the valve states of adjacent operations. The maximum pressure difference is obtained by calculating the difference between the highest and lowest pressure values ​​in the sequence. The order of operations is checked. If all conditions are within the constraints, the sequence is marked as feasible; otherwise, it is marked as infeasible. A set of candidate operation sequences with feasibility marks is output. For each atomic operation in the candidate operation sequence set, a scoring function with learnable parameters is called under the current state vector to calculate the score. The scores of all atomic operations in the same sequence are accumulated sequentially to form a basic score. Then, the constraint cost is calculated according to three types of constraint indicators. The portions exceeding the volume limit, the pressure difference limit, and the switching number limit are accumulated according to preset weights to obtain the final constraint cost. The final constraint cost is then subtracted from the basic score to obtain the aggregate score. Finally, the candidate operation list is sorted by aggregate score from high to low based on the candidate operation sequence set, feasibility marker, and aggregate score.

[0030] This invention achieves unified optimization of multi-dimensional parameters such as valve switching, flow rate, pressure, and duration by introducing constraint binding, aggregation scoring, and feasibility marking mechanisms during the candidate operation generation and screening process. This effectively avoids the problems of operational redundancy and constraint conflicts in traditional methods. By calculating the constraint cost of each candidate sequence and introducing weighted penalties, the feasibility and priority ranking of the generated sequences under conditions such as volume, pressure, and number of switching operations are guaranteed, thereby improving the rationality and efficiency of sequence generation and providing a more stable and reliable control scheme for SPR microfluidic detection.

[0031] In this embodiment, the output process of the deterministic control sequence specifically includes: Read the learnable logit score, valve identifier, duration, target flow rate and target pressure of each atomic operation in the candidate operation list one by one. Apply a feasibility hard mask to shield operations that violate the volume budget, differential pressure limit and switching number limit. Load partial order constraints to limit the allowed order relationship. Perform budget and safety constraints to remove items that are definitely not executable under the current budget. Form a candidate score table with mask mark and constraint label. Based on the candidate scoring table, a noise vector with a covariance structure is generated for each operation using relevant Gumbel perturbation. The noise is superimposed on the logit score to generate the perturbed score. Temperature parameters and random seeds are recorded to ensure that the sampling is reproducible. The perturbed score vector is checked for order consistency and stabilized according to partial order constraints. The perturbed score and the corresponding candidate index set that pass the check are output. Based on the verified perturbation scores, a cost matrix is ​​constructed, with rows corresponding to the Top-k positions and columns corresponding to candidate operations. A rectangular assignment sparse selection model with entropy regularization is established, and the sparse selection matrix is ​​obtained by solving it. ; in, This represents the sparse choice matrix obtained from the final solution. Represents the cost matrix, Represents the selection matrix. This represents the inner product of the cost matrix and the selection matrix. This represents the number of Top-k choices, i.e., the final number of operations that need to be selected from the candidate operation set. This represents the total number of candidate operations. Indicates candidate operation Assigned to the Top-k The degree of each position, Represents the entropy regularity coefficient. This represents the entropy regularization term. Let represent a logarithmic function with base 2; determine the candidate operation index corresponding to each Top-k position based on the sparse selection matrix, sort them from high to low according to the perturbation score, stabilize them according to position order in the case of the same score, and obtain an ordered candidate set; Taking an ordered candidate set as input, duration and volume budget gating are sequentially embedded to simultaneously determine dwell time and volume consumption during the generation stage. When the remaining budget is insufficient to support the next item, the dwell time is reduced and the budget is updated according to the gating rules. For adjacent candidates with similar scores, a one-time Sinkhorn order refinement is performed to stabilize the sorting. At the same time, CVaR risk reweighting is applied to historical failed samples to suppress unstable sequences, and a deterministic control sequence is output.

[0032] This invention achieves efficient sequence optimization under multiple constraints by introducing Gumbel perturbation and Top-k sparse selection mechanisms into the candidate operation screening process, avoiding the inefficiency problems of invalid candidates and repeated corrections in traditional methods. By combining hard masking, partial order constraints, and budget gating, the feasibility of the control sequence in terms of volume, pressure, and timing is ensured. At the same time, the stability and robustness of the sequence are improved by utilizing Sinkhorn order refinement and CVaR risk reweighting, significantly enhancing the determinism, reliability, and experimental success rate of the SPR microfluidic detection process.

[0033] In this embodiment, the recording of the actual executed operation trajectory specifically includes: The valve identifier, gating duration, target flow rate, and target pressure contained in the deterministic control sequence are parsed one by one, and a corresponding execution instruction queue is generated during the parsing process. Each instruction is bound to an index order and a preset timestamp. The execution command queue is sent to the microfluidic execution module one by one, driving the micropump to maintain the target flow rate within the duration specified by the command, controlling the valve to complete the opening and closing switching at the corresponding start and end time, and monitoring whether the pressure changes at the inlet and outlet meet the set target pressure. The flow rate, pressure and valve status collected during the execution process are recorded as the original execution log. The original execution log is synchronized with time and its duration is checked. For each entry, the actual duration is compared with the gating setting. When a deviation occurs, the pump speed is automatically adjusted or the valve holding time is extended to compensate for the difference. At the same time, the switching delay with the previous target is calculated and written into the log. The actual execution operation trajectory is recorded, including valve switching sequence, actual duration, real-time flow rate, real-time pressure and switching delay.

[0034] This invention achieves a high degree of consistency in valve switching, flow rate maintenance, and pressure control through precise analysis and execution verification of the control sequence. It also records and corrects operational deviations in real time during execution, ensuring the stability and controllability of the experimental process. This method significantly reduces experimental errors caused by valve delays, flow rate fluctuations, or pressure deviations, improves the consistency between the operational trajectory and the preset control, thereby enhancing the repeatability and reliability of SPR microfluidic detection and providing accurate data support for subsequent signal analysis.

[0035] In this embodiment, the output process of the SPR detection curve specifically includes: The complete execution trajectory is used as an event list. The valve switching sequence, start and end time, duration, real-time flow rate, real-time pressure and switching delay in the event list are read one by one to generate the corresponding collection time window. The valve sequence and channel identifier are bound to each time window to form a collection schedule and index sorted by item. Using the collection schedule and index as input, the SPR sensing module is driven to collect the reflection intensity sequence within the time window corresponding to each entry, and the entry number, channel identifier, and actual duration are added. At the same time, the sampling validity is checked using the flow rate and pressure fields, and the valid signals are converted into a refractive index change time series point by point. The response segments with quality labels are generated by combining the start and end times. Based on the response fragment set, the response fragments are spliced ​​in the order of the execution event list items. The continuity of the response fragments of adjacent items is checked at the end and start times. If there is a gap due to switching delay, it is filled with the baseline segment. If there is an overlap due to drift, the tail of the previous item is trimmed. After all boundary processing is completed, the SPR detection curve is output.

[0036] This invention establishes a precise correspondence between the execution trajectory and the acquisition process, enabling segmented acquisition, continuity verification, and high-quality stitching of SPR signals. This effectively avoids curve gaps and overlaps caused by valve switching delays or signal drift. This method ensures the integrity and consistency of the detection curves in both the time and parameter dimensions, making the experimental data more stable and repeatable. It improves the signal reliability and analytical accuracy of SPR microfluidic detection, providing a solid data foundation for subsequent performance evaluation and process optimization.

[0037] In this embodiment, the process of generating the performance evaluation results specifically includes: Read the segments corresponding to the SPR detection curve and execution trajectory entries one by one, and perform signal baseline subtraction and amplitude normalization on each segment to obtain a standardized response sequence that can be directly used for feature analysis; For each segment of the standardized response sequence, the detection performance parameters are calculated, including the peak response amplitude, the slope of the rising phase, the residual value of the dissociation phase, and the signal-to-noise ratio. The parameters are then organized by entry number, and the performance evaluation results with time index and entry identifier are output. Based on the performance evaluation results and combined with the deterministic control sequence, the valve switching sequence, duration, target flow rate and target pressure in each control sequence are bound to the corresponding sections in the performance parameter table one by one.

[0038] This invention achieves accurate evaluation of multi-dimensional indicators such as peak response amplitude, rise slope, dissociation residue, and signal-to-noise ratio by standardizing and quantifying the SPR detection curve, and establishes a correspondence between these indicators and specific control sequences.

[0039] In this embodiment, the formation of the optimized SPR microfluidic detection process control scheme specifically includes: Based on the performance evaluation results, set the weights and pass thresholds for the detection performance parameters, calculate the comprehensive score and attach compliance marks for volume budget, pressure limit and number of switching times, and generate a sequence evaluation table containing serial number, comprehensive score, compliance status and key descriptions, arranged in descending order of comprehensive score; The sequence evaluation table is subjected to constraint review and executability verification from high to low. Entries that do not meet the volume budget, pressure limit or switching number limit are skipped. The first control sequence that meets both constraints and quality requirements is selected as the optimal solution and output together with the corresponding performance parameters as the optimized SPR microfluidic detection process control scheme.

[0040] This invention introduces a weighted calculation and threshold setting mechanism based on the performance evaluation results, and integrates multiple constraints such as detection performance parameters, volume budget, pressure limit, and valve switching times into a comprehensive scoring system, thereby realizing a quantitative evaluation of candidate control sequences.

[0041] Example 1: To verify the feasibility of this invention in practice, it was applied to an SPR microfluidic detection experiment on antibody-antigen binding kinetics. In this experiment, multiple sets of antigen solutions of different concentrations are continuously injected into the microfluidic channel to obtain complete binding and dissociation curves, thereby evaluating the binding rate constant and dissociation rate constant. Traditional SPR microfluidic detection systems typically rely on preset valve switching sequences and constant flow rate schemes, failing to dynamically adjust the control sequence based on real-time signals. This leads to numerous invalid attempts under complex constraints, poor experimental repeatability, and frequent issues of volume budget exceeding limits or valve switching delays during multi-channel parallel testing.

[0042] After applying the detection system of this invention, thanks to the introduction of reinforcement learning and the Gumbel-Top-k sampling method, feasibility constraints, volume budget, and duration are considered together in the sequence generation stage, resulting in output control sequences with high determinism and execution success rate. Experiments show that under the same detection conditions, this invention can generate control sequences that meet the requirements in one go, and no back-end correction is needed in actual execution, significantly improving the success rate. Simultaneously, by recording the execution trajectory in real time and combining it with sensor feedback, the collected response curves maintain a strict correspondence with the operation sequence, ensuring the integrity and continuity of the data. To verify the effectiveness of this invention, a comparative experiment was conducted between this invention and a traditional system; the experimental data are shown in Table 1. Table 1. Performance Comparison of Different Detection Systems in SPR Microfluidic Experiments Based on the comparison results in Table 1, it can be seen that the present invention outperforms the traditional SPR microfluidic detection system in several key performance indicators. The total experimental time was reduced from 145 minutes to 102 minutes, improving detection efficiency; sample volume consumption was reduced from 580 μL to 420 μL, significantly reducing resource waste; the proportion of invalid or corrective instructions was reduced from 30% to less than 5%, ensuring the determinism and stability of the execution process; the maximum valve switching delay was reduced from 200 milliseconds to 50 milliseconds, making system operation more precise; the signal-to-noise ratio of the detection curve was improved from 11.4 to 15.9, significantly improving data quality; the curve overlap rate in 10 repeated experiments increased from 76% to 92%, demonstrating better experimental repeatability; and the overall experimental success rate increased from 78% to 96%, indicating that the present invention has achieved substantial improvements in efficiency, reliability, and robustness.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A SPR microfluidic detection process control system based on reinforcement learning optimization, characterized in that, include: The signal acquisition module is used to acquire the refractive index change signal during the molecular binding process in real time and output the raw SPR signal data with timestamps. The signal preprocessing module is used to perform noise suppression, baseline correction and amplitude normalization on the raw SPR signal data, and generate a time-aligned normalized state vector sequence. The decision generation module is used to assign a learnable logit and constraint tags to each operation and output a list of candidate operations. The sequence sampling module is used to apply relevant Gumbel perturbations to the logit operation and perform Top-k selection, outputting a deterministic control sequence that satisfies the feasibility conditions; The microfluidic execution module is used to record the actual operation trajectory and generate execution logs in real time; The SPR sensing module is used to collect corresponding response curve segments one by one, splice and verify them, and output a complete SPR detection curve. The performance evaluation and optimization module is used to generate performance evaluation results and bind them to the control sequence. Then, through weighted evaluation and constraint verification, an optimized SPR microfluidic detection process control scheme is formed.

2. The SPR microfluidic detection process control system based on reinforcement learning optimization according to claim 1, characterized in that, The modules are connected in the following way: A baseline fluid is initiated within a microfluidic channel, and the refractive index change signal during the molecular binding process is acquired in real time using an SPR sensing module, outputting raw SPR signal data. The raw SPR signal data is preprocessed to generate a time-aligned normalized state vector sequence; Decision generation is performed on the standardized state vector, and a learnable logit and constraint label are assigned to each decision, outputting a list of candidate operations; The candidate operation list is input into the Gumbel-Top-k sequence sampler. Gumbel perturbation is applied to the logit of each operation and Top-k selection is performed. The valve and fluid control sequence that meets the feasibility conditions is generated by combining hard mask and partial order constraints, and the deterministic control sequence is output. The deterministic control sequence is sent to the microfluidic execution module, and the actual execution trajectory is recorded; The operation trajectory is used as a time axis. The corresponding response curve segments are collected by the SPR sensing module, the complete response data is spliced ​​and verified, and the SPR detection curve corresponding to the execution trajectory is output. The results are analyzed based on the SPR detection curve, performance evaluation results are generated, and they are bound one by one with the control sequences. Based on the pairing information from the performance evaluation results, an optimized SPR microfluidic detection process control scheme is formed.

3. The SPR microfluidic detection process control system based on reinforcement learning optimization according to claim 2, characterized in that, The output process of the raw SPR signal data specifically includes: The baseline fluid is degassed and filtered, and the treated baseline fluid is injected into the microfluidic channel. A constant temperature and target volumetric flow rate are set, and the baseline stable operating condition parameter set is output. Using the baseline stable operating condition parameter set as input, SPR real-time acquisition is initiated under baseline stable conditions, with timestamps and channel identifiers attached, refractive index change data is generated in chronological order, and the corresponding timestamp sequence is output. Based on the refractive index change time series and timestamp series, integrity and continuity are verified to generate the original SPR signal data.

4. The SPR microfluidic detection process control system based on reinforcement learning optimization according to claim 2, characterized in that, The process of generating the standardized state vector sequence specifically includes: The system receives raw SPR signal data, takes the time series of refractive index changes recorded over time and the corresponding timestamps as input, performs equal-interval resampling on the time axis, and outputs the raw signal sequence and corresponding timestamps for stability verification. Using the stability-checked SPR signal data as input, the sliding window length and endpoint processing method are set, and the window is gradually slid along the time axis to smooth and denoise the neighborhood of each sampling point, eliminating baseline offset and slow drift, and outputting the denoised and baseline-corrected SPR signal data. The amplitude range of the SPR signal data after denoising and baseline correction is calculated and the amplitude is normalized. The continuous normalized samples are assembled into a state vector sequence of equal length in chronological order. At the same time, a relative time index corresponding to each state vector is generated, and a time-aligned normalized state vector sequence is output.

5. The SPR microfluidic detection process control system based on reinforcement learning optimization according to claim 2, characterized in that, The process of outputting the candidate operation list includes: Based on the standardized state vector sequence, all valve opening and closing commands are determined, and an optional duration range is set for each valve to form atomic operations. Experimental constraint parameters are loaded when forming operations, and the constraint parameters are bound to each atomic operation. The set of bound atomic operations is output. Based on the set of atomic operations, candidate operation sequences are generated step by step according to the preset maximum sequence length. Constraint indicators are calculated for each candidate operation sequence, and a set of candidate operation sequences with feasibility labels is output. For each atomic operation in the candidate operation sequence set, a scoring function with learnable parameters is called under the current state vector to calculate the score. The scores of all atomic operations in the same sequence are accumulated sequentially to form the basic score, and the final constraint cost is obtained. The aggregate score is obtained by subtracting the final constraint cost from the basic score. The candidate operation list is output in descending order of aggregate score.

6. The SPR microfluidic detection process control system based on reinforcement learning optimization according to claim 2, characterized in that, The output process of the deterministic control sequence specifically includes: Read the learnable logit score of each atomic operation in the candidate operation list one by one, apply a feasibility hard mask to block operations that violate the volume budget, pressure difference limit and switching number limit, load partial order constraints to limit the allowed order relationship, and perform front-end pruning of the execution budget and safety constraints to remove items that are definitely not executable under the current budget, forming a candidate score table with mask mark and constraint label; Based on the candidate scoring table, relevant Gumbel perturbation is used to generate noise vectors with covariance structure for each operation. The noise is superimposed on the logit score to generate the perturbed score. The perturbed score vector is checked for order consistency and stabilized according to partial order constraints. The perturbed score and the corresponding candidate index set that pass the check are output. Based on the verified perturbation scores, a cost matrix is ​​constructed, with rows corresponding to Top-k positions and columns corresponding to candidate operations. A rectangular assignment sparse selection model with entropy regularization is established, and the sparse selection matrix is ​​obtained by solving it. The candidate operation index corresponding to each Top-k position is determined according to the sparse selection matrix, and the candidates are sorted from high to low according to the perturbation scores. In the case of the same score, the position order is stabilized to obtain an ordered candidate set. Taking an ordered candidate set as input, duration and volume budget gating are sequentially embedded to simultaneously determine dwell time and volume consumption during the generation stage. When the remaining budget is insufficient to support the next item, the dwell time is reduced and the budget is updated according to the gating rules. For adjacent candidates with similar scores, a one-time Sinkhorn order refinement is performed to stabilize the sorting. At the same time, CVaR risk reweighting is applied to historical failed samples to suppress unstable sequences, and a deterministic control sequence is output.

7. The SPR microfluidic detection process control system based on reinforcement learning optimization according to claim 2, characterized in that, The recording of the actual operation trajectory includes: The deterministic control sequence is parsed one by one to generate a corresponding execution instruction queue, and each instruction is bound to an index order and a preset timestamp; The execution command queue is sent to the microfluidic execution module one by one, driving the micropump to maintain the target flow rate within the duration specified by the command, controlling the valve to complete the opening and closing switching at the corresponding start and end time, and recording the flow rate, pressure and valve status collected during the execution process as the original execution log; The original execution logs are synchronized with time and their duration is verified to record the actual execution trajectory.

8. The SPR microfluidic detection process control system based on reinforcement learning optimization according to claim 2, characterized in that, The output process of the SPR detection curve specifically includes: The complete execution trajectory is used as an event list. The event list is read one by one to form a collection schedule and index. Using the collection schedule and index as input, the SPR sensing module is driven to collect the reflection intensity sequence within the time window corresponding to each entry, convert the effective signal point by point into the refractive index change time sequence, and generate a response segment with quality label by combining the start and end times. Based on the response fragment set, the fragments are assembled in the order of the execution event list entries, and the SPR detection curve is output after all boundary processing is completed.

9. The SPR microfluidic detection process control system based on reinforcement learning optimization according to claim 2, characterized in that, The process of generating the performance evaluation results specifically includes: Read the segments corresponding to the SPR detection curve and execution trajectory entries one by one, and perform signal baseline subtraction and amplitude normalization on each segment to obtain a standardized response sequence; Calculate the detection performance parameters for each segment of the standardized response sequence, organize them by item number, and output the performance evaluation results; Based on the performance evaluation results and combined with the deterministic control sequence, each segment is bound to the corresponding section in the performance parameter table.

10. The SPR microfluidic detection process control system based on reinforcement learning optimization according to claim 2, characterized in that, The optimized SPR microfluidic detection process control scheme is specifically formed by: Based on the performance evaluation results, set the weights and pass thresholds for the detection performance parameters, calculate the comprehensive score, and generate a sequence evaluation table; Constraint verification and executability check are performed on the sequence evaluation table from high to low. The first control sequence that simultaneously meets the constraints and quality requirements is selected as the optimal solution, and the corresponding performance parameters are output together as the optimized SPR microfluidic detection process control scheme.

Images