A high-precision liquid chip sample positioning and identification system

Abstract

The present invention discloses a high-precision liquid chip sample positioning and identification system, which belongs to the field of high-precision liquid chip sample positioning and identification. The system comprises an active liquid sample positioning control module, which is used to collect liquid flow state signals and adjust the speed and position of the input sample in real time through an adjustable microfluidic structure, thereby realizing fixed-point residence and directional transportation of the sample in the chip channel; an image recognition and feature position fine-tuning module, which is used to perform multimodal image acquisition of the area where the sample is located, extract spatial feature information of the target sample based on a deep learning recognition model, and perform fine-tuning correction on the sample position according to the feedback signal; by constructing an active liquid sample positioning and identification system, high-precision directional control and spatial residence of the liquid sample in the chip channel are realized, effectively avoiding the sample drift and identification instability problems under the traditional passive transportation method, and improving the timeliness and accuracy of sample identification.

Description

Technical Field

[0001] The present invention relates to the technical field of high-precision liquid chip sample positioning and identification, and more particularly, to a high-precision liquid chip sample positioning and identification system. Background Art

[0002] With the widespread application of microfluidic liquid-phase chip technology in the field of biological sample detection, liquid-phase chips have gradually become a key platform for high-throughput detection tasks such as cell analysis, protein detection, and nucleic acid identification. In the actual detection process, the target sample is usually injected into the chip channel in liquid form, and precise positioning, residence control, and signal acquisition are completed within the set recognition area. Then, the composition information and structural characteristics of the sample are obtained through various detection methods such as optical imaging and electrical impedance response.

[0003] However, existing liquid-phase chip recognition systems mostly use fixed channel structures and passive sample transport methods, lacking effective control over the sample flow path and recognition behavior. This can lead to problems such as sample position drift, unstable residence, or recognition errors within the recognition area, which in turn affects the accuracy and repeatability of test results. To improve recognition performance, some studies have attempted to introduce various methods such as image recognition, electrode arrays, and electrical impedance analysis to assist in sample recognition tasks. However, the following technical limitations still exist:

[0004] First, existing technologies lack an active control mechanism for liquid states, making it impossible to achieve real-time adjustment and precise control of sample flow rate, positioning trajectory, and behavior pattern, making it difficult to adapt to detection needs in complex liquid environments.

[0005] Secondly, in terms of multimodal recognition information fusion, such as between image features and electrical signal responses, an effective dynamic fusion mechanism has not yet been established. This makes it impossible to optimize the judgment path in a timely manner when there are conflicts in recognition results or modal inconsistencies, which can easily lead to unstable system output.

[0006] Finally, traditional systems mostly rely on static parameter configurations and fixed recognition processes, lacking adaptive adjustment capabilities. This makes it difficult to adjust recognition strategies and path structures based on real-time changes in sample status or channel environment, resulting in insufficient recognition stability under complex working conditions.

[0007] Based on this, we proposed a high-precision liquid chip sample recognition system that integrates active liquid regulation, multimodal fusion recognition and behavior prediction feedback mechanism. It can achieve accurate positioning and stable recognition of target samples within the recognition area, thereby effectively improving the detection sensitivity and recognition accuracy of complex biological samples, as well as the stability and adaptability of the system operation. Summary of the Invention

[0008] In order to solve the problems raised in the background technology, the present invention provides the following technical solutions:

[0009] A high-precision liquid chip sample positioning and identification system, comprising:

[0010] Active liquid sample positioning control module, used to collect liquid flow state signals and adjust the speed and position of the input sample in real time through adjustable microfluidic structure, so as to achieve fixed-point residence and directional transportation of samples in the chip channel;

[0011] The image recognition and feature position fine-tuning module is used to collect multimodal images of the sample area, extract the spatial feature information of the target sample based on the deep learning recognition model, and perform fine-tuning correction on the sample position according to the feedback signal;

[0012] The multi-scale nanoelectrode array recognition module is deployed in the sample recognition area to collect the electrical impedance response information of the sample in the microchannel, construct a spatial electrical signal map, and fuse it with the image recognition results to enhance recognition accuracy.

[0013] A digital twin prediction and feedback control module is used to establish a dynamic simulation model of sample recognition behavior based on chip structural parameters, historical recognition data, and current recognition status, predict sample behavior deviations in real time, and output correction parameters to jointly adjust the control strategies of the positioning control module and the image recognition module;

[0014] The modules form a closed-loop feedback mechanism by interconnecting recognition results, prediction data, and control signals. The system has online learning and adaptive control capabilities, and can dynamically optimize the recognition path under conditions of sample recognition anomalies, variable liquid states, or recognition conflicts, achieving high-confidence, continuous, and high-stability recognition and control of target samples in the liquid chip.

[0015] Furthermore, the active liquid sample positioning control module includes:

[0016] A multi-dimensional liquid state sensing unit is used to collect the flow rate, pressure, and particle concentration of the input sample in the chip channel, as well as the position information fed back by the image recognition module;

[0017] An adjustable microfluidic drive actuator, including an oscillating micropump, a phase-controllable microvalve, and a variable cavity pressure regulator, is used to dynamically adjust the sample's propulsion direction and speed based on the data output by the sensing unit to control the sample's flow trajectory within the chip.

[0018] The local disturbance and channel limiting structure is set at the entrance of the recognition area to generate controllable flow field disturbance or regional throttling to guide the sample to achieve refined position correction and spatial orientation before recognition;

[0019] Positioning behavior control logic unit, which is used to integrate liquid state parameters, image recognition error information, and sample behavior trends provided by the digital twin prediction module to generate dynamic control instructions to optimize sample delivery paths and positioning actions;

[0020] Among them, the active liquid sample positioning control module, the image recognition module and the digital twin prediction and feedback control module construct a feedback closed-loop mechanism through the interconnection of recognition results, prediction outputs and control instructions to achieve sample path optimization and posture adjustment under abnormal recognition situations, thereby cooperating with the system to achieve continuity, high confidence and adaptive control capabilities of the recognition path.

[0021] Furthermore, the image recognition and feature position fine-tuning module includes:

[0022] A multimodal image acquisition unit is used to obtain composite image data of the target sample in the identification area under the conditions of fluorescence imaging, bright field imaging and scattered light;

[0023] The deep feature recognition unit extracts spatial features, locates center coordinates, and outputs recognition confidence of target samples in the sampled images based on a multi-layer convolutional neural network;

[0024] The coordinate regression and error fine-tuning unit is used to combine the previous frame trajectory information, the digital twin prediction path and the feedback results of the electrical impedance recognition module to perform fine regression and coordinate adjustment on the current image recognition position information;

[0025] Heterogeneous modality collaborative fusion unit, used to perform heterogeneous feature matching and conflict determination between image recognition results and electrical impedance response information, triggering a multimodal weight reconstruction mechanism when recognition results are inconsistent;

[0026] The model parameter adaptive update unit is used to monitor the trend of recognition confidence over time and trigger the model structure update or feature extraction strategy adjustment when the recognition accuracy falls below a preset threshold to improve the adaptability of the image model;

[0027] Among them, a two-way collaborative path is established between the image recognition and feature position fine-tuning module and the digital twin prediction module and the electrode array recognition module, which supports fault-tolerant recognition processing, feedback-driven fine-tuning correction and model parameter evolution update under abnormal conditions, and is used to improve the continuous stability of sample recognition in complex environments and the overall response performance of the system.

[0028] Furthermore, the digital twin prediction and feedback control module includes:

[0029] Structural parameter digital modeling unit, used to establish digital geometric models and simulation parameter models based on the microchannel structure, sample flow characteristics and multi-module interconnection topology within the liquid chip;

[0030] The historical behavior sample library construction unit is used to record the flow trajectory of past samples in the recognition channel, recognition deviation, image confidence, and multi-dimensional behavior state information of electrical impedance response, and to build a sample behavior library based on time series;

[0031] The recognition behavior dynamic prediction unit uses a dynamic Bayesian network or long short-term memory model to make real-time predictions on the spatial behavior trend, trajectory deviation, and recognition confidence interval of the current sample based on structural modeling parameters and historical sample libraries;

[0032] The behavioral anomaly classification judgment unit is used to determine whether the current sample has entered an abnormal state type based on the dynamic evolution trend of the recognition confidence and the behavioral deviation indicator, and trigger the preset protection strategy or activate the correction path generation mechanism;

[0033] The control strategy generation and output unit is used to integrate the prediction output and behavior discrimination results, generate microfluidic drive adjustment instructions, image recognition weight reconstruction instructions and model self-correction parameters, and send them to the positioning control module and image recognition module for feedback adjustment;

[0034] The twin model self-evolution optimization unit is used to adaptively update the twin model structure or control strategy based on incremental learning or reinforcement learning mechanisms when the recognition success rate decreases or continuous deviations accumulate;

[0035] Among them, the units build multi-dimensional interactive links through identification data, behavioral trends and feedback control instructions, supporting adaptive adjustment and dynamic reconstruction of the identification path in the event of identification anomalies, drastic fluctuations in liquid state or sudden changes in sample behavior, and realizing high-stability and high-confidence target recognition control of the liquid chip system under complex conditions.

[0036] Furthermore, the multi-scale nanoelectrode array recognition module includes:

[0037] The electrode array structure layout unit is used to layout multiple electrode arrays at different spatial scales in the liquid chip identification channel area to adapt to different sample particle sizes and identification accuracy requirements;

[0038] A dynamic electrical impedance acquisition unit is used to continuously collect the complex impedance response data between different electrode pairs while the target sample flows through the electrode array, and to construct a multi-dimensional spatial electrical signal map;

[0039] Asynchronous sampling and multi-frequency drive modulation unit, used to apply high-frequency drive signals of different proportions to each level of electrodes, and extract weak impedance characteristics in an asynchronous sampling manner to improve the response resolution in multi-dimensional scenarios;

[0040] The electrical signal feature extraction and matching judgment unit is used to convert the collected electrical impedance features into a standardized vector space expression, and jointly judge them with the spatial coordinate features output by the image recognition module to output the sample recognition consistency confidence value;

[0041] The conflict identification and feature reconstruction mechanism is used to trigger the reconstruction of the electrode array sampling strategy when there is a discrepancy between the electrical signal recognition and image recognition results. This includes adjusting the sampling density, switching feature channels, and reconfiguring the modal fusion weights to improve the fusion robustness under extreme working conditions.

[0042] The signal-to-noise ratio self-optimizing filter unit is used to monitor the background noise level in the identification channel and adjust the filter parameter group in real time based on the dynamic changes of noise, thereby enhancing the weak signal recognition capability and ensuring the accuracy of low-concentration sample recognition.

[0043] The electrode temperature drift compensation control unit is used to correct the impedance baseline offset or drift based on the chip temperature change curve and historical sampling offset trend, thereby improving the measurement stability under long-term operation;

[0044] The multimodal feature weight learning unit is used to automatically learn the recognition contribution of electrical signal features and image signals in multiple sample categories through deep fusion model training, and dynamically allocate modal weights in real-time recognition to improve the overall judgment accuracy of the system.

[0045] Furthermore, the system also includes a multimodal collaborative fusion and recognition fault-tolerant control mechanism, which includes:

[0046] The conflict recognition unit is used to compare the recognition results of the image recognition module, the electrode array recognition module and the digital twin prediction module, and extract the difference and confidence range deviation between the outputs of each modality;

[0047] A fusion strategy selection unit, configured to select a fusion path from a preset fusion strategy set based on the conflict identification result, including image-dominant, electrode-dominant, or prediction-first paths;

[0048] The modality weighting adjustment unit is used to set the fusion weight of each modality output according to the currently selected strategy path, adjust the feature fusion order and complete the alignment process;

[0049] The model update trigger unit is used to call the local update process of the image model, electrode parameters or twin model when the continuous recognition deviation or conflict anomaly reaches the set conditions;

[0050] Identification consistency check unit, used to perform consistency comparison on the final results of the three types of modalities after fusion output, and generate output confirmation mark or feedback control instruction;

[0051] Among them, the multimodal collaborative fusion and recognition fault-tolerant control mechanism forms an identification process adjustment path through the linkage of feature output and fusion strategy with the image recognition module, electrode array recognition module and digital twin prediction module, which is used to support structural response adjustment in the event of recognition deviation, data conflict or reduced system stability.

[0052] Furthermore, the system also includes a modality fusion evolution and redundant path determination mechanism, which includes:

[0053] The modal evolution structure modeling unit constructs the modal evolution map between the image recognition module, the electrode array recognition module, and the digital twin prediction module;

[0054] A multi-path fusion configuration unit configures redundant fusion paths based on modal output results, including weight distribution relationships between modalities and path priority sequences;

[0055] Modal stability monitoring unit, which monitors the temporal fluctuation of each modal recognition result and generates recognition consistency feedback;

[0056] The trust region adjustment unit downgrades or suppresses unstable modes based on the modal stability output results;

[0057] Feedback consistency verification module, establishes the modal comparison structure after fusion output, and outputs structural parameters and identification adjustment instructions for subsequent identification path optimization;

[0058] Among them, the modal fusion dynamic control mechanism is interconnected with each recognition module through a structural path to form a closed-loop regulation for recognition optimization, which is used to improve the stability, adaptability and fusion accuracy of the system under multimodal recognition deviation conditions.

[0059] Furthermore, the system also includes a recognition system linkage control module for self-evolution of cross-modal recognition strategy and self-optimization of recognition path structure, and the module includes:

[0060] Modal adaptability adjustment unit, used to dynamically analyze the stability and response delay of each modal recognition result under different sample types and channel states, and adjust the modal call priority and structural configuration;

[0061] The strategy migration guidance unit guides the system to migrate between strategy sets based on the evolution trend of behavioral deviations during the recognition process and the historical fusion path records, executing the path switching from image-driven and electrode-led to twin prediction joint decision-making;

[0062] The linkage architecture reconstruction unit triggers the adjustment of the modal path topology structure according to the current modal performance indicators, including modal insertion, switching or bypass clipping operations;

[0063] Redundant recognition result confirmation mechanism, which is used to construct a redundant judgment set through multi-path parallel output results, and implement minority obeys majority, multimodal consistency voting or confidence weighted average strategy to form the final recognition output;

[0064] The adaptive strategy evolution update unit combines the system recognition success rate, conflict frequency and error distribution to perform incremental training and structural update of strategy parameters.

[0065] In summary, the present invention has the following beneficial effects:

[0066] The active positioning control module, equipped with a multi-dimensional liquid state sensing unit and an adjustable microfluidic drive structure, enables fine adjustment of the sample's flow rate, direction, and residence position within the microchannel. This significantly improves the positioning accuracy and path stability of the liquid sample within the identification area, effectively avoiding the drift and identification errors that are common with traditional passive transport methods.

[0067] By building a fusion judgment mechanism among three heterogeneous modalities: image recognition, electrical impedance response, and digital twin prediction, the system can automatically perform fusion strategy selection, modality weighting adjustment, and recognition fault tolerance judgment under conditions of conflicting recognition results or recognition signal interference. This enhances the system's recognition stability and result consistency in complex liquid states, and improves the system's error suppression capability.

[0068] By introducing a digital twin prediction mechanism based on time-series sample behavior modeling, it is possible to make real-time predictions on the behavior trends, recognition trajectories, and deviation development of target samples, and jointly with the feedback control module to dynamically correct the image model, electrical signal acquisition strategy, and sample path, thereby achieving self-optimization of the recognition path and intelligent evolution of the strategy structure, ensuring that the system still has high-confidence output capabilities even under long-term operation and non-ideal working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0070] Figure 1 This is a simplified system architecture diagram of the present invention;

[0071] Figure 2 Schematic diagram of the multi-scale electrode array structure of the present invention;

[0072] Figure 3 This is the digital twin prediction flow chart of the present invention;

[0073] Figure 4 Schematic diagram of multimodal fusion control of the present invention;

[0074] Figure 5 This is a sample positioning control flow chart of the present invention;

[0075] Figure 6 Schematic diagram of the closed-loop control system of the present invention. DETAILED DESCRIPTION

[0076] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0077] Example:

[0078] The following is combined with Figures 1-6 The present invention is described in further detail.

[0079] See also Figures 1-6 The present invention provides a technical solution: a high-precision liquid chip sample positioning and identification system, comprising: an active liquid sample positioning control module for collecting liquid flow state signals and adjusting the speed and position of the input sample in real time through an adjustable microfluidic structure to achieve fixed-point residence and directional transportation of the sample in the chip channel;

[0080] The image recognition and feature position fine-tuning module is used to collect multimodal images of the sample area, extract the spatial feature information of the target sample based on the deep learning recognition model, and perform fine-tuning correction on the sample position according to the feedback signal;

[0081] The multi-scale nanoelectrode array recognition module is deployed in the sample recognition area to collect the electrical impedance response information of the sample in the microchannel, construct a spatial electrical signal map, and fuse it with the image recognition results to enhance recognition accuracy.

[0082] A digital twin prediction and feedback control module is used to establish a dynamic simulation model of sample recognition behavior based on chip structural parameters, historical recognition data, and current recognition status, predict sample behavior deviations in real time, and output correction parameters to jointly adjust the control strategies of the positioning control module and the image recognition module;

[0083] The modules form a closed-loop feedback mechanism by interconnecting recognition results, prediction data, and control signals. The system has online learning and adaptive control capabilities, and can dynamically optimize the recognition path under conditions of sample recognition anomalies, variable liquid states, or recognition conflicts, achieving high-confidence, continuous, and high-stability recognition and control of target samples in the liquid chip.

[0084] In this embodiment, the above-mentioned content is applicable to a variety of liquid-phase biological sample identification scenarios, such as single-cell analysis, DNA droplet capture, protein chip labeling, etc., all of which have strict requirements for high spatial accuracy and dynamic recognition stability. The system can achieve adaptive adjustment of the recognition path through feedback linkage between modules.

[0085] The active liquid-phase sample positioning control module uses an integrated micropump, microvalve, and microchannel perturbator structure to detect sample flow rate and pressure and adjust its propulsion direction and retention position in the identification channel. To avoid ambiguous functional terms, the control parameter range is clearly defined: the micropump oscillation frequency adjustment range is 0.1–2 Hz, and the flow rate control range under microvalve regulation is 0.1–5 mm / s, which can adapt to the retention and positioning requirements of different samples in the identification area.

[0086] The image recognition module combines convolutional neural networks to perform real-time analysis of sample images, extracting the spatial features, center coordinates, and confidence values ​​of target samples. It then combines the previous frame trajectory with digital twin prediction results to implement positioning fine-tuning and error regression.

[0087] The multi-scale nanoelectrode array recognition module collects complex impedance response feature information from an electrode array deployed in the chip recognition area and integrates it with image recognition results to enhance recognition accuracy, further supporting modal consistency verification and anomaly identification processing.

[0088] The digital twin prediction and feedback control module is used to establish a dynamic simulation model of sample recognition behavior based on chip structural parameters, historical recognition data, and current recognition status. It predicts sample behavior deviations in real time and outputs correction parameters to jointly adjust the control strategies of the above modules to maintain the stability and reliability of the recognition path.

[0089] The system's modules establish a closed-loop feedback mechanism through identification results, prediction data, and control signals. The system possesses online learning and adaptive control capabilities, and can dynamically optimize the identification path when identifying abnormalities, variable liquid states, or conflicting conditions, achieving high-confidence, continuity, and high-stability identification and control of target samples in liquid chips.

[0090] Those skilled in the art can combine the existing microfluidic chip platform with the image recognition module, and use the existing deep learning model and electrical impedance acquisition equipment to achieve collaborative work between the modules in the present invention and complete the structural integration and control optimization of the sample recognition process.

[0091] like Figures 1-6 As shown, the active liquid sample positioning control module includes:

[0092] A multi-dimensional liquid state sensing unit is used to collect the flow rate, pressure, and particle concentration of the input sample in the chip channel, as well as the position information fed back by the image recognition module;

[0093] An adjustable microfluidic drive actuator, including an oscillating micropump, a phase-controllable microvalve, and a variable cavity pressure regulator, is used to dynamically adjust the sample's propulsion direction and speed based on the data output by the sensing unit to control the sample's flow trajectory within the chip.

[0094] The local disturbance and channel limiting structure is set at the entrance of the recognition area to generate controllable flow field disturbance or regional throttling to guide the sample to achieve refined position correction and spatial orientation before recognition;

[0095] Positioning behavior control logic unit, which is used to integrate liquid state parameters, image recognition error information, and sample behavior trends provided by the digital twin prediction module to generate dynamic control instructions to optimize sample delivery paths and positioning actions;

[0096] Among them, the active liquid sample positioning control module, the image recognition module and the digital twin prediction and feedback control module construct a feedback closed-loop mechanism through the interconnection of recognition results, prediction outputs and control instructions to achieve sample path optimization and posture adjustment under abnormal recognition situations, thereby cooperating with the system to achieve continuity, high confidence and adaptive control capabilities of the recognition path.

[0097] In this embodiment, in order to verify the actual performance of the active liquid-phase sample positioning control module in a high-precision liquid-phase chip identification system, a comparative experimental platform was constructed, including three types of sample channel control schemes. We used the traditional constant current pump control as prior art A, the micropump system with a simple feedback regulation function as prior art B, and the "active liquid-phase sample positioning control module" proposed in the present invention as schemes 1 and 2. The difference between schemes 1 and 2 lies in the control strategy and the disturbance frequency setting.

[0098] Specifically, 30 cell nucleus-labeled droplet samples were selected and injected into the microchannel chip, and the four control methods mentioned above were used to conduct a complete experiment on the sample flow positioning and identification process;

[0099] The experimental system was equipped with a high-precision image recognition platform, a microfluidic pump control platform, and an electrical impedance recognition device. The flow paths of the samples collected under each scheme were recorded and key parameters were extracted: average positioning error, recognition response delay, consistency score of trimodal recognition results, recognition stability index, and false recognition rate. To ensure no statistical bias between samples, the experiment was repeated three times under constant temperature conditions, and the average value was used for evaluation.

[0100] Solution 1 of the present invention adopts multi-dimensional liquid state sensing and microfluidic perturbation mechanism. Solution 2 adds a phase-controlled microvalve structure and carries a digital twin prediction correction feedback module to achieve dynamic correction of the sample path.

[0101] In the experiment, Scheme 2 achieved real-time flow rate adjustment within 0.1 seconds based on image error feedback. It also stabilized the sample's residence in the center of the recognition area for more than 500ms through a coordinated micropump and valve mechanism, improving recognition quality. Specific experimental data are as follows:

[0102] Liquid chip recognition system comparison experiment data table (Table 1)

[0103] Test parameter name (unit) Prior Art A Prior Art B Solution 1 of the present invention Solution 2 of the present invention Average positioning error (μm) 12.4 10.8 6.3 5.8 Identification time delay (ms) 185.0 162.0 109.0 97.0 Identification consistency score (0-1) 0.82 0.86 0.93 0.95 Identification stability index (%) 87.6 89.1 95.2 96.7 Misidentification rate (%) 6.2 5.7 2.1 1.8

[0104] The experimental data show that the active liquid sample positioning control module proposed in this invention significantly outperforms the existing technology in terms of average positioning error. The positioning errors of existing technologies A and B were 12.4μm and 10.8μm, respectively, while those of solutions 1 and 2 of this invention were reduced to 6.3μm and 5.8μm, respectively. This indicates that the multi-dimensional liquid state sensing and disturbance guidance mechanism significantly improves the sample positioning accuracy.

[0105] In terms of recognition time delay, existing solutions have a response time of over 160ms due to the lack of a dynamic flow rate response mechanism. However, the proposed solution can control the delay to 109ms or even 97ms, indicating that the active control mechanism shortens the feedback link between system decision-making and action execution.

[0106] In terms of recognition consistency, Schemes 1 and 2 achieved high confidence scores of 0.93 and 0.95, respectively, significantly outperforming the traditional scheme (only 0.82–0.86). This is closely related to its real-time fusion of electrical impedance signals and image recognition. Furthermore, in terms of recognition stability, Scheme 2 achieved a stability of 96.7%, reflecting its superior ability to regulate multi-source modal coupling paths.

[0107] In terms of false recognition rate, the existing technology B still has an identification error of 5.7% at a higher flow rate, while the solution of the present invention reduces the false recognition rate to 2.1% and 1.8% through real-time feedback control, proving that the modal collaboration mechanism of the present invention is more robust under the complex changing conditions of the liquid environment.

[0108] like Figures 1-6 As shown, the image recognition and feature position fine-tuning module includes:

[0109] A multimodal image acquisition unit is used to obtain composite image data of the target sample in the identification area under the conditions of fluorescence imaging, bright field imaging and scattered light;

[0110] The deep feature recognition unit extracts spatial features, locates center coordinates, and outputs recognition confidence of target samples in the sampled images based on a multi-layer convolutional neural network;

[0111] The coordinate regression and error fine-tuning unit is used to combine the previous frame trajectory information, the digital twin prediction path and the feedback results of the electrical impedance recognition module to perform fine regression and coordinate adjustment on the current image recognition position information;

[0112] Heterogeneous modality collaborative fusion unit, used to perform heterogeneous feature matching and conflict determination between image recognition results and electrical impedance response information, triggering a multimodal weight reconstruction mechanism when recognition results are inconsistent;

[0113] The model parameter adaptive update unit is used to monitor the trend of recognition confidence over time and trigger the model structure update or feature extraction strategy adjustment when the recognition accuracy falls below a preset threshold to improve the adaptability of the image model;

[0114] Among them, the image recognition and feature position fine-tuning module establishes a bidirectional collaborative path with the digital twin prediction module and the electrode array recognition module, supporting recognition fault-tolerant processing under abnormal conditions, feedback-driven fine-tuning correction and model parameter evolution update, which is used to improve the continuous stability of sample recognition in complex environments and the overall response performance of the system;

[0115] In this example, in order to verify the performance advantage of the "image recognition and feature position fine-tuning module" in the present invention in the complex liquid chip sample positioning and recognition scenario, two categories of four level models were selected for comparative testing, including:

[0116] The old baseline CNN recognition system;

[0117] synthesized Bayesian networks with limited input;

[0118] The present invention introduces a multimodal synthesis model of feature fine-tuning and matching writing, which is compared with 1;

[0119] The present invention introduces the third modality association and digital twin prediction to fine-tune and expand the integrated model, which is comparison 2;

[0120] The selected samples for the test comparison included giving liquid samples of different corresponding integrated components under the same chip channel conditions, creating the same light environment through a projection device and light radiation, and performing 10 repeated recognition tests for each type of sample to obtain the average data;

[0121] Output indicators include recognition accuracy, mean position deviation, mean confidence level given by the recognition system, single recognition time, and average number of times the fine-tuning function is triggered;

[0122] In order to express the introduction of composite mathematical output algorithm in this model to expand the performance and versatility of the recognition program, the following mathematical equation is designed:

[0123] ;

[0124] Among them, Ω is the sample multimodal consistency confidence score after the fusion of image recognition and electrical signal, which is an important metric to measure the current recognition accuracy and modality consistency; n is the number of different image modalities under the image channel; Φ i represents the pixel response function under the i-th image modality, which represents the response change of the target sample to the spatial position x in the image modality i under unit intensity illumination; α i is the modal attenuation factor, which indicates that the more the image modality deviates from the center in the sample space, the weaker the response; m is the number of all samples in the current recognition frame; ψ j Predict the offset value for the center coordinate of the jth sample in the image recognition model; ∈ j Indicates the spatial response error value of the sample in the electrical impedance recognition module; λ j It represents the inconsistency index of the modal response between the image channel and the electrical signal channel of the sample; θ is the modal fusion sensitivity parameter, which is more sensitive to smaller modal deviations when set to 2, and linear fusion response when set to 1; к j The normalized value of the frequency at which the jth sample triggers parameter update or image error correction in the model feedback path. This formula outputs a single fusion evaluation index Ω by weighted integral image modal response and fusion error signal, which is used by the digital twin feedback module to determine whether the current recognition path needs to be adjusted or whether the model triggers a self-update strategy.

[0125] The experimental data is: Image recognition experimental data table (Table 2)

[0126] Identification indicators Base-CNN (basic convolutional neural network) Bayes-CNN (Bayesian Convolutional Neural Network) Comparison 1 Comparison 2 Image recognition accuracy (%) 88.4 90.7 94.6 96.2 Position deviation mean (μm) 7.8 6.7 4.4 3.2 Identification confidence mean 0.82 0.88 0.93 0.97 Single recognition time (ms) 42.0 48.0 56.0 63.0 Trigger fine-tuning times 2.1 1.1 0.8 0.4

[0127] It can be seen from the test data table that the image recognition and feature position fine-tuning model designed by the present invention outperforms the existing technology in all indicators;

[0128] Among them, the recognition accuracy increased from 88.4% of Base-CNN to 96.2%, an improvement of approximately 29%; the position deviation was reduced from 7.8μm to 3.2μm, approaching the zero deviation of clear recognition;

[0129] The mean recognition confidence is also significantly higher than that of other models, indicating that the present invention has obvious advantages in the stability of recognition results.

[0130] It is particularly noteworthy that the mathematical equation designed in this invention summarizes the characteristic description, recognition bias and confidence of each model through a composite function structure, providing a unified measurement standard for the synthetic recognition output;

[0131] According to the various interaction indicators in the full equation, it can be observed that the numerical value of the result synthesized by the present invention is significantly higher than that of Base-CNN and Bayes-CNN, proving that the algorithm combination program itself has strong matching and support capabilities, and realizes extreme micro-recognition reasoning for complex level samples.

[0132] like Figures 1-6 As shown, the digital twin prediction and feedback control module includes:

[0133] Structural parameter digital modeling unit, used to establish digital geometric models and simulation parameter models based on the microchannel structure, sample flow characteristics and multi-module interconnection topology within the liquid chip;

[0134] The historical behavior sample library construction unit is used to record the flow trajectory of past samples in the recognition channel, recognition deviation, image confidence, and multi-dimensional behavior state information of electrical impedance response, and to build a sample behavior library based on time series;

[0135] The recognition behavior dynamic prediction unit uses a dynamic Bayesian network or long short-term memory model to make real-time predictions on the spatial behavior trend, trajectory deviation, and recognition confidence interval of the current sample based on structural modeling parameters and historical sample libraries;

[0136] The behavioral anomaly classification judgment unit is used to determine whether the current sample has entered an abnormal state type based on the dynamic evolution trend of the recognition confidence and the behavioral deviation indicator, and trigger the preset protection strategy or activate the correction path generation mechanism;

[0137] The control strategy generation and output unit is used to integrate the prediction output and behavior discrimination results, generate microfluidic drive adjustment instructions, image recognition weight reconstruction instructions and model self-correction parameters, and send them to the positioning control module and image recognition module for feedback adjustment;

[0138] The twin model self-evolution optimization unit is used to adaptively update the twin model structure or control strategy based on incremental learning or reinforcement learning mechanisms when the recognition success rate decreases or continuous deviations accumulate;

[0139] The units build multi-dimensional interactive links between each other through identification data, behavior trends, and feedback control instructions. This supports adaptive adjustment and dynamic reconstruction of the identification path in the event of anomalies, drastic fluctuations in liquid state, or sudden changes in sample behavior, thus achieving high-stability and high-confidence target recognition control of the liquid chip system under complex conditions.

[0140] In this embodiment, the system first establishes a digital simulation model based on the geometric structure and channel topology of the liquid phase chip through the structural parameter digital modeling unit. Taking a typical multi-branch microfluidic chip as the basis, the system extracts its core structural parameters such as channel width, bifurcation angle, liquid viscosity, and sample particle radius to form an initial digital geometric model, which is then loaded into the twin modeling engine.

[0141] Subsequently, the historical behavior sample library construction unit calls the real-time recorded data from the past 300 sample recognition tasks to collect multi-dimensional features such as the image center offset, electrical impedance response curve, recognition confidence evolution trend, and microfluidic flow velocity change trajectory generated by each sample during the flow process;

[0142] The library is stored in the form of time series and provides training support for subsequent behavior prediction;

[0143] When the system detects a new sample flowing into the chip's recognition area, the recognition behavior dynamic prediction unit immediately calls the historical sample library and current structural parameters, and uses a pre-trained long-short-term memory neural network model to synchronously predict the sample's position evolution, confidence fluctuation trend, and electrical impedance response behavior. This module outputs the predicted trajectory and recognition confidence interval for the current sample in the next five-frame time window. For example, for a sample with an initial position confidence of only 0.72, the system predicts that it will reach 0.91 in the third frame and stabilize until the fifth frame.

[0144] If the system detects that the sample's behavior prediction deviation is significantly higher than the set threshold, such as a trajectory drift exceeding 12μm and a confidence level falling below 0.6, the behavior anomaly classification unit will classify it as a "stability mutation anomaly" and activate the control strategy generation and output unit to trigger feedback actions. The feedback instructions include:

[0145] Adjust the weight of the image recognition module to 0.4 and the weight of the electrode recognition module to 0.6;

[0146] Sending perturbation commands to the active positioning module to make the sample regress to the predicted trajectory;

[0147] Synchronously start the twin model self-evolution optimization unit to perform local parameter fine-tuning;

[0148] The twin model self-evolution optimization unit performs incremental training on historical processing data from the past 10 similar abnormal situations, adjusts the weight coefficient of the control feedback logic in the twin model, and records the improvement in the sample recognition success rate after adjustment, thereby continuously optimizing future control strategies;

[0149] This module forms a multi-dimensional interactive link with the positioning control module, image recognition module, and electrode recognition module to achieve dynamic reconstruction, error tolerance, and recognition accuracy improvement control under abnormal conditions of the recognition path. This implementation method significantly enhances the adaptability of the system in complex liquid environments through the linkage mechanism of real-time prediction, feedback adjustment, and adaptive evolution.

[0150] like Figures 1-6 As shown, the multi-scale nanoelectrode array recognition module includes:

[0151] The electrode array structure layout unit is used to layout multiple electrode arrays at different spatial scales in the liquid chip identification channel area to adapt to different sample particle sizes and identification accuracy requirements;

[0152] A dynamic electrical impedance acquisition unit is used to continuously collect the complex impedance response data between different electrode pairs while the target sample flows through the electrode array, and to construct a multi-dimensional spatial electrical signal map;

[0153] Asynchronous sampling and multi-frequency drive modulation unit, used to apply high-frequency drive signals of different proportions to each level of electrodes, and extract weak impedance characteristics in an asynchronous sampling manner to improve the response resolution in multi-dimensional scenarios;

[0154] The electrical signal feature extraction and matching judgment unit is used to convert the collected electrical impedance features into a standardized vector space expression, and jointly judge them with the spatial coordinate features output by the image recognition module to output the sample recognition consistency confidence value;

[0155] The conflict identification and feature reconstruction mechanism is used to trigger the reconstruction of the electrode array sampling strategy when there is a discrepancy between the electrical signal recognition and image recognition results. This includes adjusting the sampling density, switching feature channels, and reconfiguring the modal fusion weights to improve the fusion robustness under extreme working conditions.

[0156] The signal-to-noise ratio self-optimizing filter unit is used to monitor the background noise level in the identification channel and adjust the filter parameter group in real time based on the dynamic changes of noise, thereby enhancing the weak signal recognition capability and ensuring the accuracy of low-concentration sample recognition.

[0157] The electrode temperature drift compensation control unit is used to correct the impedance baseline offset or drift based on the chip temperature change curve and historical sampling offset trend, thereby improving the measurement stability under long-term operation;

[0158] A multimodal feature weight learning unit, which automatically learns the recognition contribution of electrical signal features and image signals in multiple sample categories through deep fusion model training, and dynamically allocates modal weights in real-time recognition to improve the overall judgment accuracy of the system;

[0159] In this embodiment, the multi-scale nanoelectrode array recognition module is constructed on the inner walls of the main recognition channel area of ​​the liquid phase chip. Three groups of nanoelectrode arrays of different scales are arranged along the longitudinal direction of the chip fluid channel, specifically including:

[0160] The first group is a high-density array with a spacing of approximately 5μm, used to identify the electrical impedance change signals of submicron target samples; the second group is a medium-density array with a spacing of approximately 10μm, suitable for detecting most conventional cell-level particles; the third group is a low-density array with a spacing of approximately 20μm, mainly used for detecting larger particles or droplet structures. The electrode material is platinum-coated microelectrodes to improve conductive stability and corrosion resistance under high-frequency drive;

[0161] During chip operation, the dynamic electrical impedance acquisition unit performs real-time impedance sampling of samples flowing through the identification zone based on embedded control logic. The sampling method adopts a step-by-step trigger mechanism. That is, when the first set of electrodes captures a weak signal but does not meet the identification threshold, the system automatically activates the second or even third set of electrodes for supplementary sampling, thereby forming a multi-scale redundant acquisition mechanism to improve the overall recognition coverage and accuracy of the system. Each set of electrode pairs is connected to the main processing chip through an independent signal acquisition channel, which can process the real and imaginary parameters of the complex impedance in parallel to construct a three-dimensional electrical signal response spectrum.

[0162] The asynchronous sampling and multi-frequency drive modulation unit uses a high-frequency signal source to load alternating voltage signals of different frequencies to each electrode group, such as 10MHz and 50MHz alternating drive, so that the electrode response produces differentiated frequency penetration characteristics for samples of different particle sizes, thereby improving the recognition resolution.

[0163] The sampler adopts an asynchronous sampling strategy to avoid frequency interference and supports a signal oversampling mechanism to extract the instantaneous electrical impedance fluctuations during the short-term residence of the sample in the channel.

[0164] In addition, the system integrates a signal-to-noise ratio self-optimizing filter unit, which dynamically selects the filter window function and bandwidth based on the background noise level in the identification channel, such as liquid flow disturbances or electronic device thermal noise, to improve the effective signal-to-noise ratio of weak target signals. When the temperature change in the identification area exceeds a set threshold, such as ±5, the temperature drift compensation control unit automatically intervenes and dynamically corrects the reference voltage of each electrode to avoid systematic shifts in the impedance baseline caused by temperature drift.

[0165] During the actual recognition process, if the image recognition module matches the electrical impedance signal, the system outputs a fusion recognition result with a higher confidence level. If there is a deviation between the two, the conflict recognition mechanism will trigger the recognition path reconstruction program, such as adjusting the electrode array sampling density, selecting the main modal channel, or activating the modal fusion weight adjustment mechanism to ensure the stability and accuracy of the system output.

[0166] Through the above-mentioned structural deployment and recognition process control, the multi-scale nanoelectrode array recognition module described in the present invention not only has the adaptive recognition capability for samples of different particle sizes, but also has complex functions such as noise adaptation, temperature drift compensation, modal fusion and redundant sampling, which effectively enhances the recognition accuracy and stability of the system in complex fluid environments, and ensures that the recognition mechanism has sufficient response closed-loop capability in dynamic scenarios.

[0167] like Figures 1-6 As shown, the system also includes a multimodal collaborative fusion and recognition fault-tolerant control mechanism, which includes:

[0168] The conflict recognition unit is used to compare the recognition results of the image recognition module, the electrode array recognition module and the digital twin prediction module, and extract the difference and confidence range deviation between the outputs of each modality;

[0169] A fusion strategy selection unit, configured to select a fusion path from a preset fusion strategy set based on the conflict identification result, including image-dominant, electrode-dominant, or prediction-first paths;

[0170] The modality weighting adjustment unit is used to set the fusion weight of each modality output according to the currently selected strategy path, adjust the feature fusion order and complete the alignment process;

[0171] The model update trigger unit is used to call the local update process of the image model, electrode parameters or twin model when the continuous recognition deviation or conflict anomaly reaches the set conditions;

[0172] Identification consistency check unit, used to perform consistency comparison on the final results of the three types of modalities after fusion output, and generate output confirmation mark or feedback control instruction;

[0173] The multimodal collaborative fusion and recognition fault-tolerant control mechanism forms an identification process adjustment path through the linkage of feature output and fusion strategy with the image recognition module, electrode array recognition module and digital twin prediction module, which is used to support structural response adjustment in the event of recognition deviation, data conflict or reduced system stability.

[0174] In this embodiment, the multimodal collaborative fusion and recognition fault-tolerant control mechanism is deployed in the central coordination control module of the liquid chip system. As the three main recognition sub-modules, it serves as the central linkage layer between the image recognition module, the electrode array recognition module, and the digital twin prediction module. It is responsible for implementing dynamic fusion strategy decision-making, modality weighting adjustment, and abnormal behavior response in the multimodal recognition path. To achieve high-reliability fusion output, the mechanism integrates multiple functional sub-modules, and each sub-module is related and complementary.

[0175] First, the conflict recognition unit periodically calls the output data of the three recognition submodules, including spatial position coordinates, recognition confidence, timestamp, and prediction deviation, to construct a recognition result comparison matrix and calculate the difference between the three pairs of modes and the degree of confidence interval deviation.

[0176] This unit does not rely on fixed threshold judgments, but instead combines historical recognition consistency indicators with context fluctuations to dynamically adjust the conflict judgment threshold to ensure that the system has the sensitivity and stability to adapt to different sample states and flow speeds;

[0177] When the conflict recognition unit confirms that there is a modal output deviation, the fusion strategy selection unit triggers the strategy path evaluation mechanism. Based on the current recognition environment status, such as image blur, weak electrical impedance signal, and twin prediction lag, it automatically selects the optimal fusion path from the preset fusion strategy set.

[0178] For example, when the confidence level of image data decreases but the electrical impedance response is clear, the "electrode-dominated" path will be selected; vice versa. In addition, when the sample enters the recognition blind spot or exhibits rapid jumping behavior, the "prediction-first" path will be preferred to maintain recognition continuity.

[0179] Subsequently, the modal weighting adjustment unit adjusts the feature weights and priorities in the modal fusion process according to the selected strategic path. Specifically, this unit controls the order of modal feature vector alignment processing and redistributes modal contributions based on the weight matrix to improve the consistency of the fusion results. This process is completed by calling the image-to-electrical signal feature matching mapping table and the historical deviation adjustment factor. The fusion result is used as the candidate output of the system's final recognition path.

[0180] To address continuous recognition anomalies or accumulated conflict offsets, the model update trigger unit, upon determining that a preset update threshold has been reached, invokes the local parameter adjustment function in the image model, the real-time calibration logic for electrode parameters, or the incremental training module of the digital twin model to perform a lightweight online model update operation, thereby ensuring that the system maintains recognition stability under various changes in input conditions.

[0181] Finally, the recognition consistency check unit performs a three-modal consistency check on the fusion result, including modal output coordinate center of gravity deviation detection, confidence interval cross-comparison and time consistency judgment. If the fusion output meets the consistency conditions, it is packaged as the final recognition result of the system and an execution path confirmation signal is sent; if there is an inconsistency risk, an identification reconstruction request is output, returned to the conflict recognition unit, and the fusion strategy is reselected and the cycle is repeated.

[0182] like Figures 1-6 As shown, the system also includes a modality fusion evolution and redundant path determination mechanism, which includes:

[0183] The modal evolution structure modeling unit constructs the modal evolution map between the image recognition module, the electrode array recognition module, and the digital twin prediction module;

[0184] A multi-path fusion configuration unit configures redundant fusion paths based on modal output results, including weight distribution relationships between modalities and path priority sequences;

[0185] Modal stability monitoring unit, which monitors the temporal fluctuation of each modal recognition result and generates recognition consistency feedback;

[0186] The trust region adjustment unit downgrades or suppresses unstable modes based on the modal stability output results;

[0187] Feedback consistency verification module, establishes the modal comparison structure after fusion output, and outputs structural parameters and identification adjustment instructions for subsequent identification path optimization;

[0188] The modal fusion dynamic control mechanism is interconnected with each recognition module through a structural path to form a closed-loop regulation for recognition optimization, which is used to improve the stability, adaptability and fusion accuracy of the system under multimodal recognition deviation conditions.

[0189] In this embodiment, the modal fusion evolution and redundant path determination mechanism is embedded in the fusion control logic layer of the system main controller, and a closed-loop information path is established through a multi-threaded interconnected data channel with the image recognition module, the electrode array recognition module, and the digital twin prediction module. The function of this mechanism is to dynamically adjust the fusion path structure and modal participation weights in scenarios where there are fluctuations, deviations, or inconsistent confidence levels in the multimodal recognition results, so as to improve the stability, continuity, and accuracy of the fusion results.

[0190] In this embodiment, the modal fusion evolution and redundant path determination mechanism is embedded in the fusion control logic layer of the system main controller, and forms a closed-loop information path with the image recognition module, electrode array recognition module, and digital twin prediction module through multi-threaded interconnected data channels. The function of this mechanism is to dynamically adjust the fusion path structure and modal participation weights in scenarios where there are fluctuations, deviations, or inconsistent confidence levels in multimodal recognition results, thereby improving the stability, continuity, and accuracy of the fusion results.

[0191] First, the modal evolution structure modeling unit constructs an evolutionary graph between the three types of recognition modes based on the real-time output data during the operation of the recognition module, including recognition coordinates, feature distribution, confidence curves, and response delays. This graph uses the recognition time window as the horizontal axis and the modal state index as the vertical axis to establish the dynamic influence function of each mode on the fusion result. The graph has a 1-second update cycle and can be dynamically scaled according to the sample flow rate, forming a real-time modeling capability for the evolution trend of modal behavior.

[0192] Based on this atlas information, the multi-path fusion configuration unit establishes multiple candidate fusion paths by calling a preset modality fusion path template set. Each path includes modality weight settings, fusion order priority, and conditional trigger thresholds. For example, if the confidence level is less than 80% or the number of conflicts is greater than 3 times, for example, when image recognition is stable but electrode recognition fluctuates frequently, the system automatically selects the image priority and prediction auxiliary paths, and sets the lowest fusion weight for the electrode modality to reduce its interference;

[0193] Subsequently, the modal stability monitoring unit quantitatively analyzes the output fluctuation amplitude, identification offset, and continuity consistency of each mode within the current identification cycle to generate a modal stability evaluation value. When the MSV is continuously lower than the set threshold, such as 0.75, the system considers the mode to be unstable.

[0194] To deal with unstable modes, the trust region adjustment unit performs dynamic weight reduction or output suppression operations. It calculates the modal trust factor based on the MSV value and adjusts the weighted proportion of the modal participation in the fusion process accordingly. If necessary, the modal output is completely blocked to ensure that the fusion process is not interfered with by abnormal modes. The weight reduction process is accompanied by a modal state recovery mechanism, which can automatically restore its weight after stability improves.

[0195] Finally, before outputting the fusion results, the feedback consistency verification module compares the final outputs of the three identification modules, calculates the difference between the fusion results and the results of each modal, and forms the consistency feedback structure parameters. These parameters are fed into the modal evolution graph to drive the update of the modal path and the optimization of the fusion structure in the next cycle, forming a closed-loop regulation logic of prediction-fusion-verification-optimization.

[0196] Through the above mechanism, the system can dynamically allocate fusion paths and modal roles in the case of imbalance or conflict in multiple modal recognitions, and realize fusion output under multi-source heterogeneous data conditions. This mechanism effectively solves the problem of fusion instability of traditional single-path fusion mechanism through structural modeling and weight reconstruction, and significantly improves the fault tolerance and adaptive control capabilities of the liquid chip recognition system.

[0197] like Figures 1-6 As shown, the system also includes a recognition system linkage control module for self-evolution of cross-modal recognition strategy and self-optimization of recognition path structure, and the module includes:

[0198] Modal adaptability adjustment unit, used to dynamically analyze the stability and response delay of each modal recognition result under different sample types and channel states, and adjust the modal call priority and structural configuration;

[0199] The strategy migration guidance unit guides the system to migrate between strategy sets based on the evolution trend of behavioral deviations during the recognition process and the historical fusion path records, executing the path switching from image-driven and electrode-led to twin prediction joint decision-making;

[0200] The linkage architecture reconstruction unit triggers the adjustment of the modal path topology structure according to the current modal performance indicators, including modal insertion, switching or bypass clipping operations;

[0201] Redundant recognition result confirmation mechanism, which is used to construct a redundant judgment set through multi-path parallel output results, and implement minority obeys majority, multimodal consistency voting or confidence weighted average strategy to form the final recognition output;

[0202] The adaptive strategy evolution update unit performs incremental training and structural updates of strategy parameters based on the system's recognition success rate, conflict frequency, and error distribution.

[0203] In this embodiment, the recognition system linkage control module is integrated into the high-level decision-making layer of the liquid chip main control system, and a multi-path decision-making feedback mechanism is established between the image recognition module, the electrode array recognition module, and the digital twin prediction module. This is used to dynamically adjust the recognition strategy and reconstruct the fusion path structure according to the modal output state during the sample recognition process, thereby improving the system's multimodal adaptability and redundant fault tolerance.

[0204] First, the modal adaptability adjustment unit receives performance parameters output by each modal module during different recognition cycles in real time, including recognition confidence, response delay, coordinate drift amplitude, and path stability coefficient. The system uses an exponentially weighted moving average mechanism to model the short-term trends of these parameters, thereby evaluating the stability and real-time performance of each modal recognition capability under the current sample type (such as low-concentration cells and high-viscosity liquids) and channel conditions (such as temperature fluctuations and unstable electrode voltage).

[0205] For example, when the image recognition delay exceeds 200ms and the confidence level is less than 70% for three consecutive frames, the system lowers the priority of image recognition in path fusion;

[0206] Next, the strategy migration guidance unit uses the fusion path data and conflict response records recorded in the recognition behavior history library to identify whether the current fusion strategy has a performance degradation trend;

[0207] If the system's recognition behavior shows a decrease in fusion consistency, an increase in conflict frequency, or a surge in error rate over multiple cycles, the unit will guide the system to perform an orderly migration of recognition strategies based on the strategy migration matrix.

[0208] For example, the strategy of migrating from "image-led-prediction-assisted" to "electrode-priority-image-weighted" is accompanied by confidence interval correction and error recalibration to ensure a smooth transition of the recognition path.

[0209] Subsequently, the linkage architecture reconstruction unit triggers structural-level linkage operations based on the modal performance evaluation indicators. The supported operations include: inserting new modal paths, including but not limited to temporarily introducing a temperature control feedback recognition module as an auxiliary modality, replacing existing modal channels, such as changing the image recognition path to an image-prediction joint fusion path, and trimming abnormal modalities, such as suspending its output participation in the event of continuous image recognition failures;

[0210] The above structural adjustments are performed through a soft switching mechanism, which enables dynamic reconstruction of the control link and data synchronization path;

[0211] To further improve the stability of recognition results, the redundant recognition result confirmation mechanism uses a multi-path parallel fusion judgment structure. The system performs recognition calculations in three modes and outputs preliminary results to build a candidate recognition set.

[0212] If there is a situation where the modal outputs are inconsistent, the fusion decision is made through one of three mechanisms: the first is the "majority voting method", which selects the modality with the most consistent outputs; the second is the "confidence-weighted averaging method", which calculates the fusion value by weighting the results of each modality according to its confidence; the third is the "preferred modal arbitration method", which gives priority to the output of the dominant modality in the current strategy in the event of a conflict;

[0213] The fusion results eventually form the system confirmation output and are written into the recognition record library to participate in subsequent path optimization;

[0214] Finally, the adaptive strategy evolution update unit uses the recognition success rate index, conflict trigger frequency and error distribution sequence recorded in the main control system to dynamically update the strategy selection parameters and path structure template through an incremental learning mechanism;

[0215] The system adopts a reinforcement learning structure based on policy feedback loss to perform real-time fine-tuning of the policy switching threshold, modal priority dynamic table, fusion policy library content, etc., thereby improving the system's policy evolution response capability when sample complexity increases or recognition performance degrades.

[0216] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0217] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may be subject to various changes and improvements, and these changes and improvements shall fall within the scope of the invention claimed for protection.

Claims

1. A high-precision liquid chip sample positioning and identification system, characterized in that: include: Active liquid sample positioning control module, used to collect liquid flow state signals and adjust the speed and position of the input sample in real time through adjustable microfluidic structure, so as to achieve fixed-point residence and directional transportation of samples in the chip channel; The image recognition and feature position fine-tuning module is used to collect multimodal images of the sample area, extract the spatial feature information of the target sample based on the deep learning recognition model, and perform fine-tuning correction on the sample position according to the feedback signal; The multi-scale nanoelectrode array recognition module is deployed in the sample recognition area to collect the electrical impedance response information of the sample in the microchannel, construct a spatial electrical signal map, and fuse it with the image recognition results to enhance recognition accuracy. The digital twin prediction and feedback control module is used to establish a dynamic simulation model of sample recognition behavior based on chip structure parameters, historical recognition data and current recognition status, predict sample behavior deviations in real time, and output correction parameters to jointly adjust the control strategies of the positioning control module and the image recognition module.

2. A high-precision liquid chip sample positioning and identification system according to claim 1, characterized in that: The active liquid sample positioning control module includes: A multi-dimensional liquid state sensing unit is used to collect the flow rate, pressure, and particle concentration of the input sample in the chip channel, as well as the position information fed back by the image recognition module; An adjustable microfluidic drive actuator, including an oscillating micropump, a phase-controllable microvalve, and a variable cavity pressure regulator, is used to dynamically adjust the sample's propulsion direction and speed based on the data output by the sensing unit to control the sample's flow trajectory within the chip. The local disturbance and channel limiting structure is set at the entrance of the recognition area to generate controllable flow field disturbance or regional throttling to guide the sample to achieve refined position correction and spatial orientation before recognition; The positioning behavior control logic unit is used to integrate liquid state parameters, image recognition error information and sample behavior trends provided by the digital twin prediction module to generate dynamic control instructions to optimize the sample transport path and positioning action.

3. The high-precision liquid chip sample positioning and identification system according to claim 1, characterized in that: The image recognition and feature position fine-tuning module includes: A multimodal image acquisition unit is used to obtain composite image data of the target sample in the identification area under the conditions of fluorescence imaging, bright field imaging and scattered light; The deep feature recognition unit extracts spatial features, locates center coordinates, and outputs recognition confidence of target samples in the sampled images based on a multi-layer convolutional neural network; The coordinate regression and error fine-tuning unit is used to combine the previous frame trajectory information, the digital twin prediction path and the feedback results of the electrical impedance recognition module to perform fine regression and coordinate adjustment on the current image recognition position information; Heterogeneous modality collaborative fusion unit, used to perform heterogeneous feature matching and conflict determination between image recognition results and electrical impedance response information, triggering a multimodal weight reconstruction mechanism when recognition results are inconsistent; The model parameter adaptive update unit is used to monitor the trend of recognition confidence over time and trigger the model structure update or feature extraction strategy adjustment when the recognition accuracy is lower than the preset threshold to improve the adaptability of the image model.

4. The high-precision liquid chip sample positioning and identification system according to claim 1, characterized in that: The digital twin prediction and feedback control module includes: Structural parameter digital modeling unit, used to establish digital geometric models and simulation parameter models based on the microchannel structure, sample flow characteristics and multi-module interconnection topology within the liquid chip; The historical behavior sample library construction unit is used to record the flow trajectory of past samples in the recognition channel, recognition deviation, image confidence, and multi-dimensional behavior state information of electrical impedance response, and to build a sample behavior library based on time series; The recognition behavior dynamic prediction unit uses a dynamic Bayesian network or long short-term memory model to make real-time predictions on the spatial behavior trend, trajectory deviation, and recognition confidence interval of the current sample based on structural modeling parameters and historical sample libraries; The behavioral anomaly classification judgment unit is used to determine whether the current sample has entered an abnormal state type based on the dynamic evolution trend of the recognition confidence and the behavioral deviation indicator, and trigger the preset protection strategy or activate the correction path generation mechanism; The control strategy generation and output unit is used to integrate the prediction output and behavior discrimination results, generate microfluidic drive adjustment instructions, image recognition weight reconstruction instructions and model self-correction parameters, and send them to the positioning control module and image recognition module for feedback adjustment; The twin model self-evolution optimization unit is used to adaptively update the twin model structure or control strategy based on incremental learning or reinforcement learning mechanism when the recognition success rate decreases or continuous deviations accumulate.

5. The high-precision liquid chip sample positioning and identification system according to claim 1, characterized in that: The multi-scale nanoelectrode array recognition module includes: The electrode array structure layout unit is used to layout multiple electrode arrays at different spatial scales in the liquid chip identification channel area to adapt to different sample particle sizes and identification accuracy requirements; A dynamic electrical impedance acquisition unit is used to continuously collect the complex impedance response data between different electrode pairs while the target sample flows through the electrode array, and to construct a multi-dimensional spatial electrical signal map; Asynchronous sampling and multi-frequency drive modulation unit, used to apply high-frequency drive signals of different proportions to each level of electrodes, and extract weak impedance characteristics in an asynchronous sampling manner to improve the response resolution in multi-dimensional scenarios; The electrical signal feature extraction and matching judgment unit is used to convert the collected electrical impedance features into a standardized vector space expression, and jointly judge them with the spatial coordinate features output by the image recognition module to output the sample recognition consistency confidence value; The conflict identification and feature reconstruction mechanism is used to trigger the reconstruction of the electrode array sampling strategy when there is a discrepancy between the electrical signal recognition and image recognition results. This includes adjusting the sampling density, switching feature channels, and reconfiguring the modal fusion weights to improve the fusion robustness under extreme working conditions. The signal-to-noise ratio self-optimizing filter unit is used to monitor the background noise level in the identification channel and adjust the filter parameter group in real time based on the dynamic changes of noise, thereby enhancing the weak signal recognition capability and ensuring the accuracy of low-concentration sample recognition. The electrode temperature drift compensation control unit is used to correct the impedance baseline offset or drift based on the chip temperature change curve and historical sampling offset trend, thereby improving the measurement stability under long-term operation; The multimodal feature weight learning unit is used to automatically learn the recognition contribution of electrical signal features and image signals in multiple sample categories through deep fusion model training, and dynamically allocate modal weights in real-time recognition to improve the overall judgment accuracy of the system.

6. The high-precision liquid chip sample positioning and identification system according to claim 1, characterized in that: The system also includes a multimodal collaborative fusion and recognition fault-tolerant control mechanism, which includes: The conflict recognition unit is used to compare the recognition results of the image recognition module, the electrode array recognition module and the digital twin prediction module, and extract the difference and confidence range deviation between the outputs of each modality; A fusion strategy selection unit, configured to select a fusion path from a preset fusion strategy set based on the conflict identification result, including image-dominant, electrode-dominant, or prediction-first paths; The modality weighting adjustment unit is used to set the fusion weight of each modality output according to the currently selected strategy path, adjust the feature fusion order and complete the alignment process; The model update trigger unit is used to call the local update process of the image model, electrode parameters or twin model when the continuous recognition deviation or conflict anomaly reaches the set conditions; The identification consistency check unit is used to perform consistency comparison on the final results of the three types of modalities after the fusion output, and generate an output confirmation mark or feedback control instruction.

7. The high-precision liquid chip sample positioning and identification system according to claim 1, characterized in that: The system also includes a modality fusion evolution and redundant path determination mechanism, which includes: The modal evolution structure modeling unit constructs the modal evolution map between the image recognition module, the electrode array recognition module, and the digital twin prediction module; A multi-path fusion configuration unit configures redundant fusion paths based on modal output results, including weight distribution relationships between modalities and path priority sequences; Modal stability monitoring unit, which monitors the temporal fluctuation of each modal recognition result and generates recognition consistency feedback; The trust region adjustment unit downgrades or suppresses unstable modes based on the modal stability output results; Feedback consistency verification module establishes the modal comparison structure after fusion output, and outputs structural parameters and identification adjustment instructions for subsequent identification path optimization.

8. The high-precision liquid chip sample positioning and identification system according to claim 1, characterized in that: The system also includes a recognition system linkage control module for self-evolution of cross-modal recognition strategies and self-optimization of recognition path structures, which includes: Modal adaptability adjustment unit, used to dynamically analyze the stability and response delay of each modal recognition result under different sample types and channel states, and adjust the modal call priority and structural configuration; The strategy migration guidance unit guides the system to migrate between strategy sets based on the evolution trend of behavioral deviations during the recognition process and the historical fusion path records, executing the path switching from image-driven and electrode-led to twin prediction joint decision-making; The linkage architecture reconstruction unit triggers the adjustment of the modal path topology structure according to the current modal performance indicators, including modal insertion, switching or bypass clipping operations; Redundant recognition result confirmation mechanism, which is used to construct a redundant judgment set through multi-path parallel output results, and implement minority obeys majority, multimodal consistency voting or confidence weighted average strategy to form the final recognition output; The adaptive strategy evolution update unit combines the system recognition success rate, conflict frequency and error distribution to perform incremental training and structural update of strategy parameters.

Images