Infusion dripping speed self-stabilization control method based on deep learning

Abstract

The invention discloses an infusion dripping speed self-stabilization control method based on deep learning. The infusion dripping speed self-stabilization control method comprises the following steps that 1, a multi-source flow state original token set of an infusion dripping bucket is obtained; 2, asynchronous time mark alignment processing is executed to generate a unified time sequence flow state token sequence; 3, inputting the unified time sequence flow state token sequence into the improved Perciver IO model to obtain a latent variable flow state representation matrix; 4, obtaining dripping speed prediction distribution and dripping intensity distribution based on the latent variable flow state representation matrix; 5, constructing a prediction error sample set, dividing the prediction error sample set into a calibration set and an update set, and calculating quantiles corresponding to the target coverage rate; and 6, determining a dripping speed upper bound, a dripping speed lower bound and a pressure change rate upper bound, and generating a dynamic safety constraint set. According to the method, the infusion dripping speed stability constraint control is realized by fusing the token sequence, the improved Perciver IO model and the dynamic safety margin.

Description

Technical Field

[0001] This invention relates to the field of medical infusion control technology, and in particular to a deep learning-based self-stabilizing control method for infusion drip rate. Background Technology

[0002] In clinical intravenous infusion, drip rate is a crucial parameter affecting the dosage and rhythm of drug administration. Current drip rate control typically relies on drip counting or pressure monitoring to achieve closed-loop regulation. One approach uses the count of drip events as feedback, combined with a fixed-parameter adjustment strategy to control the clamp actuator. Another approach estimates the drip rate using pressure or flow-related signals for control. While these approaches can achieve basic drip rate regulation under normal conditions, in practical applications, drip events are often non-uniform and susceptible to optical interference such as drip field obstruction, reflection, and IV drips, leading to unstable drip event recognition and thus drip rate feedback deviations. Furthermore, the state of the infusion tubing fluctuates with changes in patient position, tubing bends, ambient temperature, and infusion bottle height, causing significant time-varying and short-term abrupt changes in drip rate and pressure signals. This makes control methods relying solely on a single signal or fixed sampling method inaccurate in reflecting the actual infusion status.

[0003] Furthermore, existing technologies often process multi-source data using synchronous sampling or single time axis alignment, which makes it difficult to accommodate the asynchronous relationship between the arrival time stamp sequence of the dripping event and continuous sampling signals such as pressure sequences and actuator texture fragments. This results in insufficient information fusion, which in turn affects the ability to predict the future short-term drip rate change trend. At the same time, existing control methods often use fixed thresholds or empirical rules when setting safety boundaries, making it difficult to dynamically adjust the upper and lower limits of the drip rate and the upper limit of the pressure change rate as the prediction error changes. This makes it easy for the drip rate to fluctuate more or for safety constraints to fail when environmental disturbances increase or observation errors increase.

[0004] Therefore, how to provide a deep learning-based self-stabilizing control method for infusion drip rate 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 a deep learning-based self-stabilizing control method for infusion drip rate. This invention calculates the confidence score of dripping events in the infusion drip chamber and fuses multi-source data to form a unified temporal flow state token sequence, reducing interference errors. This token sequence is then input into an improved Perceiver IO model to enhance signal coupling representation and short-term drip rate prediction accuracy. Furthermore, a dynamic safety margin coefficient is constructed based on the predicted distribution to determine dynamic safety constraints, thereby improving drip rate control stability, reducing the risk of misadjustment, and enhancing the effectiveness of the safety boundary.

[0006] A deep learning-based self-stabilizing control method for infusion drip rate according to an embodiment of the present invention includes the following steps: Step 1: Collect the arrival timestamp sequence of dripping events from the infusion drip chamber and calculate the event confidence score. At the same time, collect the pressure sequence of the infusion tubing, the actuator texture fragment, the change in ambient temperature and the relative height of the infusion bottle to obtain the multi-source flow state original token set. Step 2: Perform asynchronous time-stamp alignment processing on the original token set of the multi-source flow state to generate a variable-length unified time-series flow state token sequence; Step 3: Input the unified temporal flow state token sequence into the improved Perceiver IO model, and obtain the latent variable flow state representation matrix through the token type shaping module, the drip rhythm phase encoding module, the token writing module and the representation matrix generation module; Step 4: Based on the latent variable flow regime characterization matrix, obtain the predicted distribution of drip rate and drip intensity for the future short time window; Step 5: Based on the predicted drop rate distribution and the drop intensity distribution, construct a set of prediction error samples in the future short time window, and divide it into a calibration set and an update set. Use the absolute error as the consistency score and calculate the quantile corresponding to the target coverage in the empirical distribution of the consistency score in the calibration set to obtain the dynamic safety margin coefficient of drop rate and the dynamic safety margin coefficient of intensity. Step 6: Determine the upper limit of drip rate, the lower limit of drip rate, and the upper limit of pressure change rate based on the dynamic safety margin coefficient of drip rate and the dynamic safety margin coefficient of strength, and generate a set of dynamic safety constraints.

[0007] Optionally, step one specifically includes: Collect the arrival timestamp sequence of dripping events from the infusion drip chamber, and extract the drop shape feature set, the time interval feature between adjacent dripping events, and the field-of-view optical interference feature for each dripping event in the arrival timestamp sequence. The drop shape consistency index is calculated based on the drop shape feature set; the drop interval rationality index is calculated based on the time interval feature of adjacent drop events; and the occlusion and reflection discrimination index is calculated based on the field of view optical interference feature, wherein the occlusion and reflection discrimination index is the sum of the proportion of occlusion area and the proportion of high-brightness reflection pixels in the field of view of the drop bucket. The event credibility score of the dripping event is obtained by weighting and summing the drop shape consistency index, the drop interval rationality index, and the occlusion and reflection discrimination index according to preset weights. The pressure sequence of the infusion line is acquired, and the pressure sequence is decomposed into low-frequency trend segments and high-frequency texture segments within a sliding window; Collect the pose and clamping force information of the tube clamping actuator, and extract the actuator texture fragments within the sliding window; Collect the changes in ambient temperature and the relative height of the infusion bottle, and perform timestamp alignment on the changes in ambient temperature and the relative height of the infusion bottle; The arrival timestamp of the dripping event, the event credibility score, the low-frequency trend fragment, the high-frequency texture fragment, the actuator texture fragment, the change in ambient temperature and the relative height of the infusion bottle are encapsulated into a multi-source flow state raw token set, where each token contains a corresponding timestamp field, a source identifier field and a numerical payload field.

[0008] Optionally, the asynchronous time-scale alignment process specifically includes: Extract event timestamps from the drip event tokens in the original token set of the multi-source flow regime and sort them in ascending order of event timestamps to obtain the drip event token sequence; Extract the sampling timestamps from the pressure texture tokens and actuator texture tokens in the original token set of the multi-source flow state, and sort them in ascending order of sampling timestamps to obtain the pressure texture token sequence and the actuator texture token sequence; The event interval is determined based on the timestamps of two adjacent drop events in the drop event token sequence. Within each event interval, pressure texture tokens and actuator texture tokens whose sampling timestamps fall into the current event interval are retrieved. The retrieved pressure texture tokens and actuator texture tokens are inserted between the drop event tokens corresponding to the current event interval in ascending order of sampling timestamps, forming a mixed token sequence after event-driven insertion. Sparse rearrangement is performed on consecutively sampled tokens in a mixed token sequence, the sparse rearrangement including: For the pressure texture token, calculate the first-order difference amplitude within the window, and retain pressure mutation fragments and pressure oscillation enhancement fragments whose first-order difference amplitude exceeds a preset threshold; The first-order differential amplitude of the motor current within the window is calculated for the actuator texture token and the actuator current mutation segment exceeding the preset threshold is retained. At the same time, the first-order differential amplitude of the pose information within the window is calculated and the pose rapid change segment exceeding the preset threshold is retained. Continuous sampling tokens that do not meet the retention conditions are deleted. The hybrid token sequence after sparse rearrangement is output as a variable-length unified temporal fluid token sequence, which retains the event order of the drip event tokens and includes the inserted pressure texture tokens and actuator texture tokens.

[0009] Optionally, the improved Perceiver IO model is specifically as follows: The unified temporal streaming token sequence is input into the token type shaping module for field consistency processing. The field consistency processing involves extracting the timestamp field, source identifier field, and numerical payload field of each token, and mapping the numerical payload field to a unified dimension representation according to the source identifier, thereby obtaining a field-consistent input token sequence. Input a consistent sequence of input tokens into the drip rhythm phase encoding module, locate the drip event token in the consistent sequence of input tokens, and construct an event interval sequence based on the timestamp difference between adjacent drip event tokens. The event interval sequence is mapped to the rhythm phase sequence, and then the rhythm phase sequence is written into the extended fields of the corresponding drip event token and the pressure texture token and actuator texture token in its adjacent interval to obtain the rhythm phase enhancement token sequence. The process of mapping the event interval sequence to a rhythmic phase sequence specifically includes: Read the arrival timestamp of each drop from the drop event token and arrange them in chronological order; calculate the time difference between two adjacent drop events in sequence to obtain the event interval sequence, where each event interval corresponds to a time period between adjacent drop events; Using the timestamps of two adjacent dripping events in the event interval sequence as the start and end points of an interval, the time period corresponding to each event interval is defined as a rhythm interval; the start time, end time, and interval number of each rhythm interval are recorded to form a rhythm interval index table; For each token in the unified temporal flow token sequence, read the timestamp, look up the rhythm interval to which the current timestamp falls in the rhythm interval index table, and determine the interval number, interval start time and interval end time corresponding to the current token; After determining the rhythm interval to which the token belongs, calculate the position ratio of the current token within the rhythm interval. The position ratio represents the progress of the current token from the start of the interval to the end of the interval. The progress corresponding to the start of the interval is the start progress, and the progress before the end of the interval is the end progress. The corresponding position ratio is used as the rhythm phase value of the current token to obtain the rhythm phase sequence; Write the rhythm phase sequence into the extended fields of the corresponding drip event token and the pressure texture token and actuator texture token in its adjacent interval to obtain the rhythm phase enhancement token sequence; The rhythm phase enhancement token sequence is input into the token writing module to generate an initial latent variable of a fixed-length latent variable array. The rhythm phase enhancement token sequence and the initial latent variable are cross-written, and the information in the rhythm phase enhancement token sequence is written into the initial latent variable to obtain the written latent variable array. The latent variable array is input into the characterization matrix generation module for dimension shaping and order arrangement, and the latent variable flow state characterization matrix is ​​output.

[0010] Optionally, the step of performing a cross-writing operation between the rhythm phase enhancement token sequence and the initial latent variables specifically involves: For each latent variable cell in the initial latent variables, extract the numerical load field and combine the current numerical load field with the position identifier of the corresponding latent variable cell to form a sequence of cells to be written to the query cell. For each token in the rhythm phase enhancement token sequence, extract its timestamp field, source identifier field, numerical payload field and the field corresponding to the rhythm phase, and combine them in a preset order to obtain the token retrieval unit sequence; Centered on the time anchor point corresponding to each write query unit, the token retrieval units whose timestamps fall into the preset time window are filtered in the token retrieval unit sequence to obtain the time domain candidate token set, wherein the preset time window is determined by the drip event interval; In the temporal candidate token set, the tokens are divided into a dripping event token subset, a pressure texture token subset, and an actuator texture token subset according to the source identifier field. Then, two subsets are selected to participate in the writing together according to the preset writing strategy of the write query unit to obtain a multi-source candidate token set. For each token retrieval unit in the multi-source candidate token set, a matching score is calculated with the currently written query unit. The matching score is determined by the compatibility between the numerical payload field and the rhythm phase extension field. Sort the tokens from highest to lowest according to the matching score, and select a preset number of preceding tokens from the subset corresponding to each source identifier to obtain a fixed-quota selection token group; Statistical summaries of the numerical load fields are calculated for each token group selected for quota. The statistical summaries include a mean summary, a maximum value summary, and a minimum value summary. The statistical summaries corresponding to each source identifier are then concatenated in a fixed order to form a write summary vector. The write summary vector is concatenated with the numerical payload field of the current write query unit and dimension reshaping is performed to obtain the write update payload; Write the update load to the corresponding latent variable cell to update the numerical load field of the initial latent variable, and obtain the write latent variable array.

[0011] Optionally, the dimensional shaping and sequence arrangement specifically refers to: For each latent variable cell written in the latent variable array, read the position identifier and source writing statistics, and generate a latent variable cell sequence number mapping table based on a preset order rule. The preset order rule includes ascending order by position identifier and fixed priority order by source writing statistics. The latent variable array is rearranged according to the latent variable unit number mapping table, and each latent variable unit is arranged in the order of the mapped number to obtain the sequential latent variable sequence. Numerical loading fields are extracted from each latent variable unit in the sequentially arranged latent variable sequence, and the numerical loading fields are converted into a uniform dimension vector according to the zero-filling rule to obtain the dimension-aligned loading sequence. The dimension-aligned load sequences are grouped and spliced ​​according to preset matrix shape parameters, and the unified dimension vector of each latent variable unit is rearranged in the order of row-first and column-later expansion to obtain a matrix load block. The preset matrix shape parameters include row number parameters and column number parameters, and the product of the row number parameters and column number parameters is equal to the vector dimension of the unified dimension of each latent variable unit. The matrix-based load blocks corresponding to each latent variable unit are stacked along a preset stacking dimension to form a two-dimensional latent variable flow regime representation matrix.

[0012] Optionally, step four specifically involves: Determine the start and end times and window length of the future short-term window, and generate a predicted time scale sequence; Based on the predicted time scale sequence, a drop intensity query unit is generated for each time scale to obtain the drop intensity query sequence; The latent variable flow regime characterization matrix and the drop intensity query sequence are input into the drop intensity decoding process. The corresponding drop intensity result is output for each time scale, and the drop intensity results of all time scales are arranged in chronological order to obtain the drop intensity sequence of the future short time window. Based on the future short-term window, the intensity sequence of the droplet is mapped from intensity to droplet count. A candidate set of droplet counts is generated with the short-term window as the statistical boundary, and the droplet count distribution results are generated. The drop count distribution is used as a candidate set of drop rates, and the candidate set of drop rates is divided into intervals to form a set of drop rate intervals. The frequency percentage of the candidate drip rate set falling into each drip rate interval set is counted to generate the drip rate prediction distribution result; The drop rate prediction distribution is checked for consistency with the future short-term drop intensity sequence. Distribution tails that do not meet the short-term window boundary conditions are removed, and the drop rate prediction distribution and the future short-term drop intensity distribution are output after consistency check.

[0013] Optionally, step five specifically includes: The median method was used to analyze the central statistics of the predicted drop rate distribution and the drop intensity distribution within a future short time window, and the statistics of actual drop events and actual drop intensity within the same future short time window were obtained simultaneously. The difference between the central statistic of the predicted drip rate distribution and the actual drip event statistic is used as the drip rate error sample, and the difference between the central statistic of the drip intensity distribution and the actual drip intensity statistic is used as the intensity error sample. The drip rate error sample and the intensity error sample are then merged in chronological order to obtain the prediction error sample set. The prediction error sample set is randomly divided into an equal number of calibration and update sets; For each error sample in the calibration set, a consistency score is calculated. The consistency score is the absolute error value of the current error sample, and drop rate consistency score sequence and intensity consistency score sequence are formed respectively. Empirical distributions are constructed in the drip rate consistency score sequence and the intensity consistency score sequence, respectively. Based on the preset target coverage, the corresponding quantiles are determined in their respective empirical distributions to obtain the drip rate dynamic safety margin coefficient and the intensity dynamic safety margin coefficient. The updated set is used as the source of new samples when updating the sliding time window to refresh the calibration set.

[0014] Optionally, step six specifically includes: The dynamic safety margin coefficient of drip rate is mapped to the drip rate tightening level. The mapping adopts a preset segmentation rule to convert the coefficient range into the drip rate tightening level, and the dynamic safety margin coefficient of strength is mapped to the strength tightening level. Based on the drip rate tightening level and the strength tightening level, the upper limit tightening amount of drip rate, the lower limit tightening amount of drip rate, and the upper limit tightening amount of pressure change rate are determined and encapsulated to obtain a dynamic safety constraint set.

[0015] The beneficial effects of this invention are: This invention calculates event confidence scores from the arrival timestamp sequence of infusion dripping events and simultaneously collects infusion tubing pressure sequences, actuator texture fragments, ambient temperature, and changes in the relative height of the infusion bottle to form a multi-source flow state raw token set. This is then asynchronously time-scaled to generate a variable-length unified temporal flow state token sequence, ensuring consistent representation of dripping event data and continuous sampling data within the same temporal semantics. This reduces error propagation caused by shading, reflection, and drop shape anomalies on drip rate feedback. Based on this, the unified temporal flow state token sequence is input into an improved PerceiverIO model. A token type shaping module achieves field consistency and unified dimensional representation. A dripping rhythm phase encoding module maps event intervals to rhythm phases and writes them into extended fields, giving the model a stronger structural representation of the coupling relationship between dripping rhythms and continuous signal fragments. A token writing module writes variable-length token information into fixed-length latent variables, completes dimensional shaping and sequential arrangement, and outputs a latent variable flow state representation matrix, enabling subsequent prediction processes to... Inferences can be made within a stable representation space to obtain the predicted distribution of drip rate and drip intensity for future short-term windows, thereby improving the ability to characterize short-term fluctuations. Furthermore, based on the predicted distribution, a set of prediction error samples is constructed and divided into a calibration set and an update set. Using the absolute error as a consistency score, the quantile corresponding to the target coverage is calculated in the empirical distribution of the calibration set, forming a dynamic safety margin coefficient for drip rate and a dynamic safety margin coefficient for intensity. This allows the safety margin to automatically adjust with changes in error level, avoiding the failure of fixed thresholds when errors increase or excessive conservatism when errors decrease. Finally, based on the dynamic safety margin coefficients, the upper bound of drip rate, the lower bound of drip rate, and the upper bound of pressure change rate are determined and encapsulated into a dynamic safety constraint set. This enables the control process to maintain drip rate stability and suppress abnormal pressure increases with constraint strength matching the prediction uncertainty when facing changes in ambient temperature, changes in infusion bottle height, and fluctuations in pipeline status. This achieves a comprehensive effect of improving drip rate control stability, reducing the risk of misadjustment caused by observation interference, and enhancing the effectiveness of safety boundaries. Attached Figure Description

[0016] 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 an overall flowchart of a deep learning-based self-stabilizing control method for infusion drip rate proposed in this invention. Figure 2 This is a schematic diagram of the processing steps of the improved Perceiver IO model for a deep learning-based self-stabilizing control method for infusion drip rate proposed in this invention. Detailed Implementation

[0017] 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.

[0018] refer to Figures 1-2 A deep learning-based self-stabilizing control method for infusion drip rate includes the following steps: Step 1: Collect the arrival timestamp sequence of dripping events from the infusion drip chamber and calculate the event confidence score. At the same time, collect the pressure sequence of the infusion tubing, the actuator texture fragment, the change in ambient temperature and the relative height of the infusion bottle to obtain the multi-source flow state original token set. Step 2: Perform asynchronous time-stamp alignment processing on the original token set of the multi-source flow state to generate a variable-length unified time-series flow state token sequence; Step 3: Input the unified temporal flow state token sequence into the improved Perceiver IO model, and obtain the latent variable flow state representation matrix through the token type shaping module, the drip rhythm phase encoding module, the token writing module and the representation matrix generation module; Step 4: Based on the latent variable flow regime characterization matrix, obtain the predicted distribution of drip rate and drip intensity for the future short time window; Step 5: Based on the predicted drop rate distribution and the drop intensity distribution, construct a set of prediction error samples in the future short time window, and divide it into a calibration set and an update set. Use the absolute error as the consistency score and calculate the quantile corresponding to the target coverage in the empirical distribution of the consistency score in the calibration set to obtain the dynamic safety margin coefficient of drop rate and the dynamic safety margin coefficient of intensity. Step 6: Determine the upper limit of drip rate, the lower limit of drip rate, and the upper limit of pressure change rate based on the dynamic safety margin coefficient of drip rate and the dynamic safety margin coefficient of strength, and generate a set of dynamic safety constraints.

[0019] This step, through comprehensive acquisition and event credibility verification of multi-source infusion flow data, effectively filters out invalid information such as optical interference and sampling noise, significantly improving data reliability and solving the one-sidedness problem of traditional single-source data control. Asynchronous timescale alignment and sparse rearrangement processing accurately adapt to the temporal differences of multi-source data, generating a high-quality unified token sequence, providing solid data support for subsequent predictive analysis. The improved Perceiver IO model, combined with rhythmic phase encoding and cross-writing mechanisms, strengthens the capture of the temporal correlation of drip rate, enabling accurate prediction of drip rate and drip intensity within a short time window, significantly improving control foresight. The dynamic safety margin coefficient generated based on real-time error calibration can flexibly adapt to dynamic changes such as ambient temperature, infusion bottle height, and tubing pressure, making the safety boundary more consistent with actual working conditions and effectively reducing infusion risks. The final dynamic safety constraint set clarifies the reasonable boundaries of drip rate and pressure changes, achieving adaptive and stable control of infusion drip rate, reducing the cost of manual monitoring and intervention, and comprehensively improving the safety and intelligence level of clinical infusion.

[0020] In this embodiment, step one specifically includes: Collect the arrival timestamp sequence of dripping events from the infusion drip chamber, and extract the drop shape feature set, the time interval feature between adjacent dripping events, and the field-of-view optical interference feature for each dripping event in the arrival timestamp sequence. The drop-shaped consistency index is calculated based on the drop-shaped feature set, which includes the drop-shaped contour area, drop-shaped contour perimeter, drop-shaped contour roundness, and drop-shaped principal axis direction. The drop-shaped consistency index is determined by the Euclidean distance between the drop-shaped feature vector composed of the drop-shaped feature set in a preset order and the drop-shaped template feature vector corresponding to the preset drop-shaped template feature set. The rationality index of the drip interval is calculated based on the characteristics of the time interval between adjacent drip events. The rationality index of the drip interval is determined by the degree of deviation between the time interval between the current drip event and the previous drip event and the target time interval corresponding to the target drip rate. The target time interval is obtained by converting the reciprocal of the target drip rate. The occlusion and reflection discrimination index is calculated based on the optical interference characteristics of the field of view. The occlusion and reflection discrimination index is the sum of the proportion of occlusion area and the proportion of high-brightness reflection pixels in the field of view of the droplet. The event credibility score of the dripping event is obtained by weighting and summing the drop shape consistency index, the drop interval rationality index, and the occlusion and reflection discrimination index according to preset weights. The pressure sequence of the infusion line is collected and decomposed into low-frequency trend segments and high-frequency texture segments within a sliding window. The low-frequency trend segments are obtained by smoothing the pressure sequence within the sliding window, and the high-frequency texture segments are generated by the difference between the pressure sequence and the low-frequency trend segments. The pose information and clamping force information of the tube clamping actuator are collected, and the actuator texture fragment is extracted within a sliding window. The actuator texture fragment is generated by splicing together the first-order difference amplitude values ​​within the window for calculating the pose information and clamping force information respectively. Collect the changes in ambient temperature and the relative height of the infusion bottle, and perform timestamp alignment on the changes in ambient temperature and the relative height of the infusion bottle; The arrival timestamp of the dripping event, the event credibility score, the low-frequency trend fragment, the high-frequency texture fragment, the actuator texture fragment, the change in ambient temperature and the relative height of the infusion bottle are encapsulated into a multi-source flow state raw token set, where each token contains a corresponding timestamp field, a source identifier field and a numerical payload field.

[0021] This step extracts multi-dimensional features from dripping events and calculates specific confidence scores. Combined with weighted verification using multi-dimensional indicators such as drop shape, time interval, and optical interference, it accurately identifies invalid or interfering dripping events, significantly improving the effectiveness and accuracy of dripping event data and avoiding misjudgments caused by single-feature analysis. Segmenting the pipeline pressure sequence into low-frequency trends and high-frequency textures captures both the overall pressure change trend and instantaneous fluctuations, preserving the core information of the pressure data. Extracting actuator texture fragments focuses on the dynamic changes in pose and clamping force, accurately reflecting the actuator's working state. Simultaneously, timestamp alignment processing of data such as ambient temperature and infusion bottle height achieves standardized encapsulation of multi-source heterogeneous data. The resulting multi-source flow state raw token set possesses both temporal consistency and information integrity, providing a high-quality, high-dimensional data foundation for subsequent asynchronous timestamp alignment and model analysis, effectively improving the data input accuracy of the overall control method.

[0022] In this embodiment, the asynchronous time stamp alignment process specifically includes: Extract event timestamps from the drip event tokens in the original token set of the multi-source flow regime and sort them in ascending order of event timestamps to obtain the drip event token sequence; Extract the sampling timestamps from the pressure texture tokens and actuator texture tokens in the original token set of the multi-source flow state, and sort them in ascending order of sampling timestamps to obtain the pressure texture token sequence and the actuator texture token sequence; The event interval is determined based on the timestamps of two adjacent drop events in the drop event token sequence. Within each event interval, pressure texture tokens and actuator texture tokens whose sampling timestamps fall into the current event interval are retrieved. The retrieved pressure texture tokens and actuator texture tokens are inserted between the drop event tokens corresponding to the current event interval in ascending order of sampling timestamps, forming a mixed token sequence after event-driven insertion. Sparse rearrangement is performed on consecutively sampled tokens in a mixed token sequence, the sparse rearrangement including: For the pressure texture token, calculate the first-order difference amplitude within the window, and retain pressure mutation fragments and pressure oscillation enhancement fragments whose first-order difference amplitude exceeds a preset threshold; The first-order differential amplitude of the motor current within the window is calculated for the actuator texture token and the actuator current mutation segment exceeding the preset threshold is retained. At the same time, the first-order differential amplitude of the pose information within the window is calculated and the pose rapid change segment exceeding the preset threshold is retained. Continuous sampling tokens that do not meet the retention conditions are deleted. The hybrid token sequence after sparse rearrangement is output as a variable-length unified temporal fluid token sequence, which retains the event order of the drip event tokens and includes the inserted pressure texture tokens and actuator texture tokens.

[0023] This step achieves precise temporal alignment of heterogeneous tokens from multiple sources by sorting them by timestamp and inserting pressure and actuator tokens driven by the dripping event. This resolves the issue of inconsistent sampling frequencies and time bases across different data sources, ensuring the correlation and integrity of the time-series data. For the sparse rearrangement strategy of continuously sampled tokens, key feature fragments such as abrupt pressure changes and rapid changes in actuator current / pose are retained, while meaningless redundant sampling data is eliminated. This simplifies the data volume, reduces the computational load of subsequent models, and accurately preserves the core features of flow regime changes, avoiding redundant information from interfering with the analysis results. The resulting unified temporal flow regime token sequence maintains the core temporal logic of the dripping event and condenses the key dynamic information of pressure and actuators, providing a high-quality, lightweight time-series data foundation for accurate analysis by subsequent deep learning models, effectively improving model processing efficiency and analysis accuracy.

[0024] In this embodiment, the improved Perceiver IO model is specifically as follows: The unified temporal streaming token sequence is input into the token type shaping module for field consistency processing. The field consistency processing involves extracting the timestamp field, source identifier field, and numerical payload field of each token, and mapping the numerical payload field to a unified dimension representation according to the source identifier, thereby obtaining a field-consistent input token sequence. Input a consistent sequence of input tokens into the drip rhythm phase encoding module, locate the drip event token in the consistent sequence of input tokens, and construct an event interval sequence based on the timestamp difference between adjacent drip event tokens. The event interval sequence is mapped to the rhythm phase sequence, and then the rhythm phase sequence is written into the extended fields of the corresponding drip event token and the pressure texture token and actuator texture token in its adjacent interval to obtain the rhythm phase enhancement token sequence. The process of mapping the event interval sequence to a rhythmic phase sequence specifically includes: Read the arrival timestamp of each drop from the drop event token and arrange them in chronological order; calculate the time difference between two adjacent drop events in sequence to obtain the event interval sequence, where each event interval corresponds to a time period between adjacent drop events; Using the timestamps of two adjacent dripping events in the event interval sequence as the start and end points of an interval, the time period corresponding to each event interval is defined as a rhythm interval; the start time, end time, and interval number of each rhythm interval are recorded to form a rhythm interval index table; For each token in the unified temporal flow token sequence, read the timestamp, look up the rhythm interval to which the current timestamp falls in the rhythm interval index table, and determine the interval number, interval start time and interval end time corresponding to the current token; After determining the rhythm interval to which the token belongs, calculate the position ratio of the current token within the rhythm interval. The position ratio represents the progress of the current token from the start of the interval to the end of the interval. The progress corresponding to the start of the interval is the start progress, and the progress before the end of the interval is the end progress. The corresponding position ratio is used as the rhythm phase value of the current token to obtain the rhythm phase sequence; Write the rhythm phase sequence into the extended fields of the corresponding drip event token and the pressure texture token and actuator texture token in its adjacent interval to obtain the rhythm phase enhancement token sequence; The rhythm phase enhancement token sequence is input into the token writing module to generate an initial latent variable of a fixed-length latent variable array. The rhythm phase enhancement token sequence and the initial latent variable are cross-written, and the information in the rhythm phase enhancement token sequence is written into the initial latent variable to obtain the written latent variable array. The latent variable array is input into the characterization matrix generation module for dimension shaping and order arrangement, and the latent variable flow state characterization matrix is ​​output.

[0025] This step maps the numerical payloads of tokens from different sources to a unified dimension through field consistency processing, eliminating differences in multi-source data formats and laying the foundation for cross-type data fusion analysis. Drop rhythm phase encoding, by constructing rhythm intervals and calculating token position ratios, integrates the temporal rhythm characteristics of drop events into all relevant tokens, strengthening the correlation between pressure, actuator data, and drop rate changes, allowing the model to more accurately capture flow regime changes. Cross-writing operations efficiently integrate multi-dimensional key information through temporal filtering, multi-source subset matching, and statistical summary extraction, avoiding the limitations of a single data source; dimension shaping and ordering transform the latent variable array into a structured representation matrix, improving data readability and model processing efficiency. The overall process retains the core features of multi-source data while strengthening temporal correlations through rhythm enhancement and intelligent fusion. The generated latent variable flow regime representation matrix is ​​both complete and targeted, providing high-precision feature support for subsequent drop rate prediction and significantly improving the accuracy and reliability of model predictions.

[0026] In this embodiment, the step of performing a cross-writing operation between the rhythm phase enhancement token sequence and the initial latent variable specifically involves: For each latent variable cell in the initial latent variables, extract the numerical load field and combine the current numerical load field with the position identifier of the corresponding latent variable cell to form a sequence of cells to be written to the query cell. For each token in the rhythm phase enhancement token sequence, extract its timestamp field, source identifier field, numerical payload field and the field corresponding to the rhythm phase, and combine them in a preset order to obtain the token retrieval unit sequence; Centered on the time anchor point corresponding to each write query unit, the token retrieval units whose timestamps fall into the preset time window are filtered in the token retrieval unit sequence to obtain the time domain candidate token set, wherein the preset time window is determined by the drip event interval; In the temporal candidate token set, the tokens are divided into a dripping event token subset, a pressure texture token subset, and an actuator texture token subset according to the source identifier field. Two subsets are selected to participate in the writing together according to the preset writing strategy of the write query unit to obtain a multi-source candidate token set. The preset writing strategy is used to ensure that the writing information of each write query unit simultaneously includes event rhythm information and continuous sampling information. For each token retrieval unit in the multi-source candidate token set, a matching score is calculated with the currently written query unit. The matching score is determined by the compatibility between the numerical payload field and the rhythm phase extension field. Sort the tokens from highest to lowest according to the matching score, and select a preset number of preceding tokens from the subset corresponding to each source identifier to obtain a fixed-quota selection token group; Statistical summaries of the numerical load fields are calculated for each token group selected for quota. The statistical summaries include a mean summary, a maximum value summary, and a minimum value summary. The statistical summaries corresponding to each source identifier are then concatenated in a fixed order to form a write summary vector. The write summary vector is concatenated with the numerical payload field of the current write query unit and dimension reshaping is performed to obtain the write update payload; Write the update load to the corresponding latent variable cell to update the numerical load field of the initial latent variable, and obtain the write latent variable array.

[0027] This step constructs a precise matching system between the write query unit and the token retrieval unit, using the drip event interval as a benchmark to define the time domain window, achieving precise screening of multi-source tokens within key time intervals, avoiding irrelevant data interference, and improving the targeting of information extraction. A preset write strategy forces the fusion of two core information types: event rhythm and continuous sampling, ensuring that the latent variable unit simultaneously carries the temporal characteristics of the drip event and the dynamic changes of pressure and actuators, solving the problem of insufficient single-dimensional information representation. Based on a quota selection mechanism using matching scores, core token information with high compatibility with the query unit is prioritized for integration, combined with statistical summary extraction of mean, maximum / minimum values, preserving key data features while simplifying data dimensions and reducing computational complexity. Finally, through dimension shaping and latent variable updates, multi-source heterogeneous information is efficiently fused into a structured write latent variable array, significantly improving the latent variables' ability to represent the infusion flow regime, laying the foundation for subsequent generation of a high-precision flow regime representation matrix, and effectively enhancing the model's ability to capture and predict drip rate changes.

[0028] In this embodiment, the dimensional shaping and sequential arrangement specifically refer to: For each latent variable cell written in the latent variable array, read the position identifier and source writing statistics, and generate a latent variable cell sequence number mapping table based on a preset order rule. The preset order rule includes ascending order by position identifier and fixed priority order by source writing statistics. The latent variable array is rearranged according to the latent variable unit number mapping table, and each latent variable unit is arranged in the order of the mapped number to obtain the sequential latent variable sequence. Numerical loading fields are extracted from each latent variable unit in the sequentially arranged latent variable sequence, and the numerical loading fields are converted into a uniform dimension vector according to the zero-filling rule to obtain the dimension-aligned loading sequence. The dimension-aligned load sequences are grouped and spliced ​​according to preset matrix shape parameters, and the unified dimension vector of each latent variable unit is rearranged in the order of row-first and column-later expansion to obtain a matrix load block. The preset matrix shape parameters include row number parameters and column number parameters, and the product of the row number parameters and column number parameters is equal to the vector dimension of the unified dimension of each latent variable unit. The matrix-based load blocks corresponding to each latent variable unit are stacked along a preset stacking dimension to form a two-dimensional latent variable flow regime representation matrix.

[0029] This step generates a sequence mapping table and rearranges latent variable units using preset order rules. Combining location identifiers and source statistics in the sorting logic, the latent variable sequence possesses both temporal continuity and hierarchical source characteristics, avoiding model analysis biases caused by data disorder. Zero-filling rules achieve a unified dimensional transformation of numerical payloads, eliminating dimensional differences between different latent variable units and providing a standardized data foundation for matrix processing. The construction and stacking of matrix-based payload blocks transforms one-dimensional vector-like payload data into structured two-dimensional matrices. This preserves the complete features of each latent variable unit and adapts to the computational logic of deep learning models through a regular matrix shape, significantly improving model reading and computation efficiency. The resulting latent variable flow regime representation matrix has a clear structure and unified dimensions. It integrates multi-source flow regime information and possesses good computability, enabling models to more efficiently and accurately uncover flow regime change patterns in the data, providing high-quality feature matrix support for subsequent drop rate and intensity distribution predictions. In this embodiment, step four specifically includes: Determine the start and end times and window length of the future short-term window, and generate a predicted time scale sequence; Based on the predicted time scale sequence, a drip intensity query unit is generated for each time scale to obtain the drip intensity query sequence. The drip intensity query unit includes the time identifier of the current time scale and the prediction task identifier. The latent variable flow regime characterization matrix and the drop intensity query sequence are input into the drop intensity decoding process. The corresponding drop intensity result is output for each time scale, and the drop intensity results of all time scales are arranged in chronological order to obtain the drop intensity sequence of the future short time window. Based on the future short-term window, the intensity sequence of the droplet is mapped from intensity to droplet count. A candidate set of droplet counts is generated with the short-term window as the statistical boundary, and the droplet count distribution results are generated. The drop count distribution is used as a candidate set of drop rates, and the candidate set of drop rates is divided into intervals to form a set of drop rate intervals. The frequency percentage of the candidate drip rate set falling into each drip rate interval set is counted to generate the drip rate prediction distribution result; The drop rate prediction distribution is checked for consistency with the future short-term drop intensity sequence. Distribution tails that do not meet the short-term window boundary conditions are removed, and the drop rate prediction distribution and the future short-term drop intensity distribution are output after consistency check.

[0030] This step defines a clear future short-term window and generates a standardized prediction time scale sequence, making the prediction range and granularity more precise and controllable, adapting to the needs of real-time clinical infusion control. The drip intensity decoding process combines a latent variable flow regime representation matrix with a dedicated query unit to achieve accurate intensity prediction on a time scale, providing a reliable basis for drip rate conversion. The mapping from intensity to the number of drops and interval statistics transform continuous intensity data into drip rate distribution results that conform to clinical understanding, improving the practicality of the prediction results. Consistency verification eliminates invalid distribution data that does not conform to the boundaries of the short-term window, further ensuring the accuracy and rationality of the prediction results. The overall process, from intensity prediction to drip rate distribution generation, is progressive, preserving the temporal continuity of the data while strengthening the reliability of the results through statistics and verification. The output drip rate prediction distribution and intensity distribution can accurately reflect the flow regime change trend within the future short-term window, providing a high-precision and high-reliability prediction basis for subsequent dynamic safety margin calculations, significantly improving the foresight and accuracy of drip rate control. In this embodiment, step five specifically includes: The median method was used to analyze the central statistics of the predicted drop rate distribution and the drop intensity distribution within a future short time window, and the statistics of actual drop events and actual drop intensity within the same future short time window were obtained simultaneously. The difference between the central statistic of the predicted drip rate distribution and the actual drip event statistic is used as the drip rate error sample, and the difference between the central statistic of the drip intensity distribution and the actual drip intensity statistic is used as the intensity error sample. The drip rate error sample and the intensity error sample are then merged in chronological order to obtain the prediction error sample set. The prediction error sample set is randomly divided into an equal number of calibration and update sets; For each error sample in the calibration set, a consistency score is calculated. The consistency score is the absolute error value of the current error sample, and drop rate consistency score sequence and intensity consistency score sequence are formed respectively. Empirical distributions are constructed in the drip rate consistency score sequence and the intensity consistency score sequence, respectively. Based on the preset target coverage, the corresponding quantiles are determined in their respective empirical distributions to obtain the drip rate dynamic safety margin coefficient and the intensity dynamic safety margin coefficient. The updated set is used as the source of new samples when updating the sliding time window to refresh the calibration set.

[0031] This step extracts the predicted and actual flow center statistics from the median, robustly suppressing outlier interference and improving the reliability and representativeness of error samples. The predicted errors are used to construct calibration and update sets, enabling online self-calibration of the safety margin and avoiding the shortcomings of fixed parameters failing to adapt to changes in infusion conditions. Using the absolute error as a consistency score and calculating the target coverage quantile based on an empirical distribution, dynamic safety margin coefficients for drip rate and intensity that closely match actual conditions can be dynamically generated, allowing the safety boundary to adaptively adjust in real time with system errors. Simultaneously, a sliding window continuously refreshes the calibration set, ensuring the safety margin maintains high accuracy and adaptability over the long term, significantly improving the rationality and reliability of subsequent safety constraints and providing a dynamic and robust safety benchmark for self-stabilizing infusion drip rate control. In this embodiment, step six specifically includes: The dynamic safety margin coefficient of drip rate is mapped to the drip rate tightening level. The mapping adopts a preset segmentation rule to convert the coefficient range into the drip rate tightening level, and the dynamic safety margin coefficient of strength is mapped to the strength tightening level. Based on the drip rate tightening level and the intensity tightening level, the upper limit tightening amount of drip rate, the lower limit tightening amount of drip rate, and the upper limit tightening amount of pressure change rate are determined and encapsulated to obtain a dynamic safety constraint set. The upper limit tightening amount of pressure change rate is used to limit the control action space when the pressure rises above a preset threshold.

[0032] This step transforms the dynamic safety margin coefficient into tiered tightening amounts, allowing for adaptive adjustment of control constraint strength based on real-time errors. This de-fixes the safety boundary, enabling it to dynamically change with infusion conditions and prediction accuracy. A piecewise mapping rule is used to generate drip rate and intensity tightening levels, resulting in clear logic, efficient execution, and rapid system response. Based on the tightening amounts determined by the levels—including the upper and lower bounds of drip rate and the upper bound of pressure change rate—abnormal pressure rises and drastic drip rate fluctuations are effectively limited, reducing the space for unreasonable control actions and improving system stability and safety. The resulting dynamic safety constraint set is precise, flexible, and robust, providing a reliable safety boundary for self-stabilizing infusion drip rate control, significantly reducing infusion risks, and improving the stability and controllability of clinical infusion procedures.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to the Intensive Care Unit (ICU) of a tertiary-level Class A hospital. This department receives critically ill patients undergoing postoperative recovery or septic shock, requiring long-term infusion of vasoactive drugs, nutritional solutions, and other special medications. These patients have extremely high requirements for the stability of the infusion drip rate; abnormal drip rates can easily lead to heart failure, drug poisoning, or delays in treatment. In clinical practice, existing infusion equipment has significant shortcomings: traditional infusion pumps can only be manually preset to a fixed drip rate, unable to adapt to dynamic changes such as patient position, bottle height, and temperature, resulting in frequent drip rate deviations; ordinary intelligent infusion pumps rely on single-source data acquisition, are susceptible to optical interference, and have poor data reliability; furthermore, both have fixed safety margins and cannot be adjusted in real time based on prediction errors, posing significant infusion risks. To address the aforementioned issues, the ICU introduced the method of this invention, modified existing equipment, and selected 50 critically ill patients (28 males and 22 females, aged 22-78 years) as the application subjects. The preset drip rate was 10-80 drops / minute. The superiority of this invention was verified through clinical application and compared with traditional infusion pumps and ordinary intelligent infusion pumps.

[0034] In this ICU scenario, existing infusion equipment was modified by installing a high-definition industrial camera, pressure sensor, temperature sensor, displacement sensor, and tube clamping actuator on the infusion stand. These sensors collect data on dripping events, tubing pressure, ambient temperature, bottle height changes, and actuator status. All sensors are connected to a backend terminal, which incorporates the control algorithm of this invention, enabling fully automated control of the entire process. Medical staff only need to preset basic parameters such as the target drip rate. In application, the high-definition industrial camera captures dripping events and records timestamps, extracting drop shape features, time interval features, and optical interference features. An event confidence score is calculated using a 3:4:3 weighting, and invalid events with scores below 0.6 are discarded. The pressure sensor collects pressure sequences at 10Hz, decomposing them into low-frequency trends and high-frequency texture fragments within a 500ms sliding window. The tube clamping actuator collects pose and clamping force information at 5Hz, extracting actuator texture fragments. The temperature and displacement sensors collect data at 1Hz and complete timestamp alignment, ultimately encapsulating the data into a multi-source flow state raw token set. The backend terminal performs asynchronous time-stamp alignment on the token set, sorts tokens of each type by timestamp, inserts pressure and actuator tokens with the drip event interval as the boundary, and then removes redundant data through sparse rearrangement to generate a unified time-series flow state token sequence. This sequence is input into the improved Perceiver IO model, and after field consistency, rhythm phase encoding, cross-writing operation, and dimension shaping, a 64×32-dimensional latent variable flow state representation matrix is ​​generated. Based on this matrix, a prediction time scale sequence is generated with a short time window of 10 seconds. The drip intensity sequence is obtained through a decoding process, mapped to the number of drips, and the drip rate prediction distribution is statistically analyzed. After consistency verification, the accurate prediction result is output. The median method is used to extract the predicted and actual statistics, calculate the error samples and divide them into calibration and update sets. The quantiles are determined with 95% target coverage to obtain the dynamic safety margin coefficient, which is mapped to the tightening level to generate a dynamic safety constraint set, which is sent to the clamping actuator to achieve drip rate self-stabilization control. To verify the effectiveness, traditional infusion pumps and ordinary intelligent infusion pumps were selected as the control group, with 50 devices in each group. They were run continuously under the same conditions, and the comparison data of the core indicators were recorded as shown in the table below.

[0035] Table 1 Comparison of Clinical Application Effects As shown in Table 1, the overall effect of this invention is significantly better than the control group, effectively solving the core problems of existing equipment. In terms of drip rate control accuracy, the experimental group's average deviation is only ±0.8 drops / minute, a 78.38% improvement compared to traditional infusion pumps, completely solving the problem of excessive drip rate deviation; the response time to bottle height changes is only 85ms, demonstrating extremely strong adaptability to dynamic operating conditions. In clinical applications, the incidence of adverse events is only 0.8%, and the frequency of medical intervention is reduced to 0.3 times / day / unit, significantly reducing infusion risks and the workload of medical staff. This invention effectively improves the accuracy, stability, and safety of infusion control, is suitable for complex clinical scenarios in the ICU, and has extremely high application value.

[0036] 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 deep learning-based self-stabilizing control method for infusion drip rate, characterized in that, Includes the following steps: Step 1: Collect the arrival timestamp sequence of dripping events from the infusion drip chamber and calculate the event confidence score. At the same time, collect the pressure sequence of the infusion tubing, the actuator texture fragment, the change in ambient temperature and the relative height of the infusion bottle to obtain the multi-source flow state original token set. Step 2: Perform asynchronous time-stamp alignment processing on the original token set of the multi-source flow state to generate a variable-length unified time-series flow state token sequence; Step 3: Input the unified temporal flow state token sequence into the improved Perceiver IO model, and obtain the latent variable flow state representation matrix through the token type shaping module, the drip rhythm phase encoding module, the token writing module and the representation matrix generation module; Step 4: Based on the latent variable flow regime characterization matrix, obtain the predicted distribution of drip rate and drip intensity for the future short time window; Step 5: Based on the predicted drop rate distribution and the drop intensity distribution, construct a set of prediction error samples in the future short time window, and divide it into a calibration set and an update set. Use the absolute error as the consistency score and calculate the quantile corresponding to the target coverage in the empirical distribution of the consistency score in the calibration set to obtain the dynamic safety margin coefficient of drop rate and the dynamic safety margin coefficient of intensity. Step 6: Determine the upper limit of drip rate, the lower limit of drip rate, and the upper limit of pressure change rate based on the dynamic safety margin coefficient of drip rate and the dynamic safety margin coefficient of strength, and generate a set of dynamic safety constraints.

2. The method for self-stabilizing infusion drip rate control based on deep learning according to claim 1, characterized in that, Step one specifically involves: Collect the arrival timestamp sequence of dripping events from the infusion drip chamber, and extract the drop shape feature set, the time interval feature between adjacent dripping events, and the field-of-view optical interference feature for each dripping event in the arrival timestamp sequence. The drop shape consistency index is calculated based on the drop shape feature set; the drop interval rationality index is calculated based on the time interval feature of adjacent drop events; and the occlusion and reflection discrimination index is calculated based on the field of view optical interference feature, wherein the occlusion and reflection discrimination index is the sum of the proportion of occlusion area and the proportion of high-brightness reflection pixels in the field of view of the drop bucket. The event credibility score of the dripping event is obtained by weighting and summing the drop shape consistency index, the drop interval rationality index, and the occlusion and reflection discrimination index according to preset weights. The pressure sequence of the infusion line is acquired, and the pressure sequence is decomposed into low-frequency trend segments and high-frequency texture segments within a sliding window; Collect the pose and clamping force information of the tube clamping actuator, and extract the actuator texture fragments within the sliding window; Collect the changes in ambient temperature and the relative height of the infusion bottle, and perform timestamp alignment on the changes in ambient temperature and the relative height of the infusion bottle; The arrival timestamp of the dripping event, the event credibility score, the low-frequency trend fragment, the high-frequency texture fragment, the actuator texture fragment, the change in ambient temperature and the relative height of the infusion bottle are encapsulated into a multi-source flow state raw token set, where each token contains a corresponding timestamp field, a source identifier field and a numerical payload field.

3. The method for self-stabilizing infusion drip rate control based on deep learning according to claim 1, characterized in that, The asynchronous time stamp alignment process is specifically as follows: Extract event timestamps from the drip event tokens in the original token set of the multi-source flow regime and sort them in ascending order of event timestamps to obtain the drip event token sequence; Extract the sampling timestamps from the pressure texture tokens and actuator texture tokens in the original token set of the multi-source flow state, and sort them in ascending order of sampling timestamps to obtain the pressure texture token sequence and the actuator texture token sequence; The event interval is determined based on the timestamps of two adjacent drop events in the drop event token sequence. Within each event interval, pressure texture tokens and actuator texture tokens whose sampling timestamps fall into the current event interval are retrieved. The retrieved pressure texture tokens and actuator texture tokens are inserted between the drop event tokens corresponding to the current event interval in ascending order of sampling timestamps, forming a mixed token sequence after event-driven insertion. Sparse rearrangement is performed on consecutively sampled tokens in a mixed token sequence, the sparse rearrangement including: For the pressure texture token, calculate the first-order difference amplitude within the window, and retain pressure mutation fragments and pressure oscillation enhancement fragments whose first-order difference amplitude exceeds a preset threshold; The first-order differential amplitude of the motor current within the window is calculated for the actuator texture token and the actuator current mutation segment exceeding the preset threshold is retained. At the same time, the first-order differential amplitude of the pose information within the window is calculated and the pose rapid change segment exceeding the preset threshold is retained. Continuous sampling tokens that do not meet the retention conditions are deleted. The hybrid token sequence after sparse rearrangement is output as a variable-length unified temporal fluid token sequence, which retains the event order of the drip event tokens and includes the inserted pressure texture tokens and actuator texture tokens.

4. The method for self-stabilizing infusion drip rate control based on deep learning according to claim 1, characterized in that, The improved Perceiver IO model is specifically as follows: The unified temporal streaming token sequence is input into the token type shaping module for field consistency processing. The field consistency processing involves extracting the timestamp field, source identifier field, and numerical payload field of each token, and mapping the numerical payload field to a unified dimension representation according to the source identifier, thereby obtaining a field-consistent input token sequence. Input a consistent sequence of input tokens into the drip rhythm phase encoding module, locate the drip event token in the consistent sequence of input tokens, and construct an event interval sequence based on the timestamp difference between adjacent drip event tokens. The event interval sequence is mapped to the rhythm phase sequence, and then the rhythm phase sequence is written into the extended fields of the corresponding drip event token and the pressure texture token and actuator texture token in its adjacent interval to obtain the rhythm phase enhancement token sequence. The process of mapping the event interval sequence to a rhythmic phase sequence specifically includes: Read the arrival timestamp of each drop from the drop event token and arrange them in chronological order; calculate the time difference between two adjacent drop events in sequence to obtain the event interval sequence, where each event interval corresponds to a time period between adjacent drop events; Using the timestamps of two adjacent dripping events in the event interval sequence as the start and end points of an interval, the time period corresponding to each event interval is defined as a rhythm interval; the start time, end time, and interval number of each rhythm interval are recorded to form a rhythm interval index table; For each token in the unified temporal flow token sequence, read the timestamp, find the rhythm interval to which the current timestamp falls in the rhythm interval index table, and determine the interval number, interval start time and interval end time corresponding to the current token. After determining the rhythm interval to which the token belongs, calculate the position ratio of the current token within the rhythm interval. The position ratio represents the progress of the current token from the start of the interval to the end of the interval. The progress corresponding to the start of the interval is the start progress, and the progress before the end of the interval is the end progress. The corresponding position ratio is used as the rhythm phase value of the current token to obtain the rhythm phase sequence; Write the rhythm phase sequence into the extended fields of the corresponding drip event token and the pressure texture token and actuator texture token in its adjacent interval to obtain the rhythm phase enhancement token sequence; The rhythm phase enhancement token sequence is input into the token writing module to generate an initial latent variable of a fixed-length latent variable array. The rhythm phase enhancement token sequence and the initial latent variable are cross-written, and the information in the rhythm phase enhancement token sequence is written into the initial latent variable to obtain the written latent variable array. The latent variable array is input into the characterization matrix generation module for dimension shaping and order arrangement, and the latent variable flow state characterization matrix is ​​output.

5. The method for self-stabilizing infusion drip rate control based on deep learning according to claim 4, characterized in that, The step of performing a cross-writing operation between the rhythm phase enhancement token sequence and the initial latent variable is specifically as follows: For each latent variable cell in the initial latent variables, extract the numerical load field and combine the current numerical load field with the position identifier of the corresponding latent variable cell to form a sequence of cells to be written to the query cell. For each token in the rhythm phase enhancement token sequence, extract its timestamp field, source identifier field, numerical payload field and the field corresponding to the rhythm phase, and combine them in a preset order to obtain the token retrieval unit sequence; Centered on the time anchor point corresponding to each write query unit, the token retrieval units whose timestamps fall into the preset time window are filtered in the token retrieval unit sequence to obtain the time domain candidate token set, wherein the preset time window is determined by the drip event interval; In the temporal candidate token set, the tokens are divided into a dripping event token subset, a pressure texture token subset, and an actuator texture token subset according to the source identifier field. Then, two subsets are selected to participate in the writing together according to the preset writing strategy of the writing query unit to obtain a multi-source candidate token set. For each token retrieval unit in the multi-source candidate token set, a matching score is calculated with the currently written query unit. The matching score is determined by the compatibility between the numerical payload field and the rhythm phase extension field. Sort the tokens from highest to lowest according to the matching score, and select a preset number of preceding tokens from the subset corresponding to each source identifier to obtain a fixed-quota selection token group; Statistical summaries of the numerical load fields are calculated for each token group selected for quota. The statistical summaries include mean summaries, maximum summaries, and minimum summaries. The statistical summaries corresponding to each source identifier are concatenated in a fixed order to form a write summary vector. The write summary vector is concatenated with the numerical payload field of the current write query unit and dimension reshaping is performed to obtain the write update payload; Write the update load to the corresponding latent variable cell to update the numerical load field of the initial latent variable, and obtain the write latent variable array.

6. The method for self-stabilizing infusion drip rate control based on deep learning according to claim 4, characterized in that, The specific steps of performing dimensional shaping and sequential arrangement are as follows: For each latent variable cell written in the latent variable array, read the position identifier and source writing statistics, and generate a latent variable cell sequence number mapping table based on a preset order rule. The preset order rule includes ascending order by position identifier and fixed priority order by source writing statistics. The latent variable array is rearranged according to the latent variable unit number mapping table, and each latent variable unit is arranged in the order of the mapped number to obtain the sequential latent variable sequence. Numerical loading fields are extracted from each latent variable unit in the sequentially arranged latent variable sequence, and the numerical loading fields are converted into a uniform dimension vector according to the zero-filling rule to obtain the dimension-aligned loading sequence. The dimension-aligned load sequences are grouped and spliced ​​according to preset matrix shape parameters, and the unified dimension vector of each latent variable unit is rearranged in the order of row-first and column-later expansion to obtain a matrix load block. The preset matrix shape parameters include row number parameters and column number parameters, and the product of the row number parameters and column number parameters is equal to the vector dimension of the unified dimension of each latent variable unit. The matrix-based load blocks corresponding to each latent variable unit are stacked along a preset stacking dimension to form a two-dimensional latent variable flow regime representation matrix.

7. The method for self-stabilizing infusion drip rate control based on deep learning according to claim 1, characterized in that, Step four specifically involves: Determine the start and end times and window length of the future short-term window, and generate a predicted time scale sequence; Based on the predicted time scale sequence, a drop intensity query unit is generated for each time scale to obtain the drop intensity query sequence; The latent variable flow regime characterization matrix and the drop intensity query sequence are input into the drop intensity decoding process. The corresponding drop intensity result is output for each time scale, and the drop intensity results of all time scales are arranged in chronological order to obtain the drop intensity sequence of the future short time window. Based on the future short-term window, the intensity sequence of the droplet is mapped from intensity to droplet count. A candidate set of droplet counts is generated with the short-term window as the statistical boundary, and the droplet count distribution results are generated. The drop count distribution is used as a candidate set of drop rates, and the candidate set of drop rates is divided into intervals to form a set of drop rate intervals. The frequency percentage of the candidate drip rate set falling into each drip rate interval set is counted to generate the drip rate prediction distribution result; The drop rate prediction distribution is checked for consistency with the future short-term drop intensity sequence. Distribution tails that do not meet the short-term window boundary conditions are removed, and the drop rate prediction distribution and the future short-term drop intensity distribution are output after consistency check.

8. The method for self-stabilizing infusion drip rate control based on deep learning according to claim 1, characterized in that, Step five specifically involves: The median method was used to analyze the central statistics of the predicted drop rate distribution and the drop intensity distribution within a future short time window, and the statistics of actual drop events and actual drop intensity within the same future short time window were obtained simultaneously. The difference between the central statistic of the predicted drip rate distribution and the actual drip event statistic is used as the drip rate error sample, and the difference between the central statistic of the drip intensity distribution and the actual drip intensity statistic is used as the intensity error sample. The drip rate error sample and the intensity error sample are then merged in chronological order to obtain the prediction error sample set. The prediction error sample set is randomly divided into an equal number of calibration and update sets; For each error sample in the calibration set, a consistency score is calculated. The consistency score is the absolute error value of the current error sample, and drop rate consistency score sequence and intensity consistency score sequence are formed respectively. Empirical distributions are constructed in the drip rate consistency score sequence and the intensity consistency score sequence, respectively. Based on the preset target coverage, the corresponding quantiles are determined in their respective empirical distributions to obtain the drip rate dynamic safety margin coefficient and the intensity dynamic safety margin coefficient. The updated set is used as the source of new samples when updating the sliding time window to refresh the calibration set.

9. The method for self-stabilizing infusion drip rate control based on deep learning according to claim 1, characterized in that, Step six specifically involves: The dynamic safety margin coefficient of drip rate is mapped to the drip rate tightening level. The mapping adopts a preset segmentation rule to convert the coefficient range into the drip rate tightening level, and the dynamic safety margin coefficient of strength is mapped to the strength tightening level. Based on the drip rate tightening level and the strength tightening level, the upper limit tightening amount of drip rate, the lower limit tightening amount of drip rate, and the upper limit tightening amount of pressure change rate are determined and encapsulated to obtain a dynamic safety constraint set.

Images