Gynecological patient real-time monitoring method and system based on edge computing

Abstract

This invention discloses a real-time monitoring method for gynecological patients based on edge computing. To address the difficulties in real-time quantification of vaginal bleeding and lochia, and the increased cumulative error caused by interference from factors such as weighing vibrations, humidity saturation, light shifts, and occlusion, this invention collects the weight, humidity, and surface color data of the absorbent carrier within a preset update cycle, collects environmental parameter data, and reads the absorbent carrier material parameter data to generate multimodal features. Based on these multimodal features, quality indicators are calculated, and a quality assessment model generates modal weight coefficients. In a lightweight temporal Transformer network, the cross-modal attention weights are gated and scaled according to these weight coefficients to obtain an aligned multimodal temporal representation. This representation is input into a physical constraint regression model, which outputs and updates the incremental volume of liquid entering the absorbent carrier, the incremental volume of evaporation, and the volume retained within the absorbent carrier. The cumulative bleeding volume is calculated and accumulated under mass conservation constraints, non-negativity constraints, and upper limit constraints on absorption capacity. Furthermore, online calibration of fast-loop calibration and slow-loop update is performed in conjunction with the monitoring results, and parameters are reset during pad replacement events. This achieves real-time, stable, and robust quantitative monitoring of bleeding volume at the edge, while reducing long-term cumulative errors.

Description

Technical Field

[0001] This invention relates to the field of physiological parameter monitoring, and in particular to a method and system for real-time monitoring of gynecological patients based on edge computing. Background Technology

[0002] For gynecological patients during postoperative recovery, postpartum recovery, or abnormal bleeding observation periods, changes in vaginal discharge are crucial for assessing the risk of blood loss, infection, and recovery status. Current clinical practice often relies on nurses visually observing the soaking area and color changes of dressings or sanitary napkins, or estimating based on dressing frequency and subjective experience. Some institutions also use a weighing method to calculate fluid volume by weighing used absorbent pads and deducting dry weight. Furthermore, with the development of sensor and IoT technologies, solutions integrating weighing, humidity detection, and image acquisition modules into nursing pads or bedside devices have emerged, combining mobile terminals or cloud platforms for data recording, alarm notifications, and trend analysis. Other solutions attempt to introduce machine learning models to fit and regress multi-source data to improve estimation accuracy and achieve continuous monitoring.

[0003] Current technologies still have the following shortcomings, which limit the effectiveness of real-time quantification and long-term stable monitoring of bleeding volume:

[0004] 1. Insufficient anti-interference capability: Single-mode or simple fusion methods are easily affected by weighing jitter and zero drift, humidity sensor saturation, light changes and shading, etc., leading to short-term misjudgment and long-term cumulative deviation.

[0005] 2. Lack of material and environment adaptation: The absorption capacity, diffusion characteristics and initial dry weight of different absorbent carriers vary significantly, and evaporation is affected by temperature, relative humidity and airflow. Existing solutions often do not explicitly incorporate material and environmental parameters into the model, resulting in insufficient generalization ability across consumables and scenarios.

[0006] 3. Lack of physical constraints and online calibration mechanisms: Some data-driven regression methods do not incorporate physical constraints such as mass conservation, non-negativity, and capacity limits, which can easily lead to unreasonable estimates. At the same time, they lack online calibration and pad replacement event handling mechanisms for long-term operation at the edge, making it difficult to maintain stable accuracy without interrupting monitoring.

[0007] Therefore, a method and system for real-time monitoring of gynecological patients that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a real-time monitoring method for gynecological patients based on edge computing. Addressing the problems in existing technologies where vaginal bleeding and lochia bleeding are difficult to quantify continuously and in real-time, and are easily affected by factors such as weighing jitter and zero-point drift, humidity saturation, and light shift and occlusion, leading to unstable estimations and large cumulative errors, this invention proposes a monitoring technology solution that combines multimodal "weight, humidity, and color" fusion with physical constraint regression. Within a preset update cycle, it collects the weight, humidity, and surface color data of the absorber, and collects environmental parameter data and reads material parameter data to generate a specific monitoring system. The invention employs a multimodal temporal representation model. It calculates quality indicators such as weighing jitter, humidity saturation, illumination shift, and occlusion, and generates modal weight coefficients based on the quality assessment model. In a lightweight temporal Transformer network, the cross-modal attention weights are gated and scaled according to the weight coefficients to obtain an aligned multimodal temporal representation. This representation is then input into a physical constraint regression model that satisfies mass conservation, non-negativity, and upper limit constraints on absorption capacity. The model outputs and updates the input volume increment, evaporation volume increment, and retention volume, and calculates the cumulative bleeding. Finally, it performs online calibration and reset of pad replacement events, combining monitoring results with fast-loop calibration and slow-loop updates. This invention offers advantages such as strong real-time performance at the edge, robustness to artifacts and changes in operating conditions, good adaptability across consumables and environments, and low long-term cumulative error.

[0009] This invention provides a method for real-time monitoring of gynecological patients based on edge computing, comprising:

[0010] S1. Within a preset update cycle, collect weight, humidity, and surface color data of the absorbent carrier used to collect vaginal discharge, collect environmental parameter data, read absorbent carrier material parameter data, and generate weight feature sequences, humidity feature sequences, color feature sequences, environmental parameter feature vectors, and absorbent carrier material parameter feature vectors; S2. Calculate quality indicators based on the weight feature sequences, humidity feature sequences, and color feature sequences; S3. Input the quality indicators into the quality assessment model to generate weight coefficients corresponding to the weight feature sequences, humidity feature sequences, and color feature sequences, respectively; S4. Construct a time-series input sequence from the weight feature sequences, humidity feature sequences, color feature sequences, environmental parameter feature vectors, and absorbent carrier material parameter feature vectors, and input it into a lightweight time-series Transf dataset. The ORMER network encoding scales the cross-modal attention weights according to the weight coefficients to generate an aligned multimodal temporal representation; S5, the aligned multimodal temporal representation is input into the physical constraint regression model, which outputs and updates the liquid volume increment entering the absorber, the evaporation volume increment, and the retained volume within the absorber. The physical constraint regression model satisfies mass conservation constraints, non-negativity constraints, and upper limit constraints on absorption capacity during the regression process. The upper limit of absorption capacity is characterized by the eigenvector of the absorber material parameters; S6, the bleeding volume increment is generated based on the liquid volume increment entering the absorber, and the bleeding volume increment is accumulated to obtain the cumulative bleeding volume, which is the bleeding volume monitoring result; S7, online calibration is performed based on the bleeding volume monitoring result and combined with weight data, humidity data, and surface color data.

[0011] Optionally, S1 includes:

[0012] Within the preset update cycle, the weight data of the absorbent carrier used to receive vaginal fluid is collected by a weighing sensor, the humidity data of the absorbent carrier is collected by a humidity sensor, the color data of the surface of the absorbent carrier is collected by an image acquisition module, the environmental parameter data is collected by an environmental sensor, and the absorbent carrier material parameter data is read from local storage or identification information.

[0013] The weight data, humidity data, color data, and environmental parameter data all carry a collection timestamp.

[0014] The weight data, humidity data, and color data are time-aligned based on the acquisition timestamp so that the time-aligned weight data, humidity data, and color data correspond to the same time axis.

[0015] Feature extraction is performed on the time-aligned weight data to generate a weight feature sequence, which includes a weight value sequence and a weight change sequence calculated from the weight value sequence.

[0016] A humidity feature sequence is generated by extracting features from the time-aligned humidity data. The humidity feature sequence includes a humidity value sequence and a humidity change rate sequence calculated from the humidity value sequence.

[0017] A color feature sequence is generated by extracting features from the time-aligned color data. The color feature sequence includes a luminance feature sequence and a chromaticity feature sequence calculated from the color data.

[0018] The environmental parameter data is vectorized to generate environmental parameter feature vectors. The vectorization process includes normalizing the numerical environmental parameters and concatenating them according to a preset field order.

[0019] The absorber material parameter data is vectorized to generate absorber material parameter feature vectors. The vectorization process includes normalizing the numerical material parameters and concatenating them according to a preset field order.

[0020] Output weight data, humidity data, color data, weight feature sequence, humidity feature sequence, color feature sequence, environmental parameter feature vector, and absorber material parameter feature vector.

[0021] Optionally, S2 includes:

[0022] The short-term fluctuation of the weight change sequence is calculated within a preset sliding time window based on the weight feature sequence, and the short-term fluctuation is used as the weighing jitter index.

[0023] The degree of closeness of the humidity value sequence to the preset saturation threshold is determined based on the humidity feature sequence, and the degree of slowing down of humidity change is determined by combining the absolute value of the humidity change rate sequence, so as to generate a humidity saturation index.

[0024] The degree of deviation of the brightness feature sequence from the preset reference brightness is calculated based on the color feature sequence to generate an illumination offset index;

[0025] The effective pixel ratio of color data or the confidence level of color feature extraction is determined based on the color feature sequence, and an occlusion index is generated based on the effective pixel ratio or the confidence level of color feature extraction.

[0026] The quality index is generated by combining the weighing vibration index, the humidity saturation index, the illumination offset index, and the shading index.

[0027] Optionally, S3 includes:

[0028] The quality indicators are constructed into a quality indicator vector and input into the quality assessment model. The quality assessment model outputs the reliability scores corresponding to the weight feature sequence, humidity feature sequence and color feature sequence, respectively.

[0029] Based on the reliability score, weight coefficients are generated corresponding to the weight feature sequence, humidity feature sequence, and color feature sequence, respectively. The weight coefficients are obtained by normalizing the reliability score and satisfy the value range of [0,1] and the sum of the three is 1.

[0030] When any reliability score is lower than the preset reliability threshold, the weight coefficient corresponding to that reliability score is reduced to the preset lower limit or set to zero, and the remaining weight coefficients are renormalized to obtain the final weight coefficient.

[0031] Optionally, S4 includes:

[0032] The weight feature sequence, humidity feature sequence, and color feature sequence are respectively converted into feature embedding sequences with the same input dimension as the lightweight temporal Transformer network through linear mapping. At each time step, the corresponding weight feature embedding, humidity feature embedding, and color feature embedding are concatenated to form the multimodal feature representation of that time step.

[0033] The environmental parameter feature vector and the absorber material parameter feature vector are mapped to environmental parameter embedding and absorber material parameter embedding, respectively, and combined with the multimodal feature representation as additional input to form a time-series input sequence;

[0034] The time-series input sequence is fed into the lightweight temporal Transformer network after adding position encoding.

[0035] In the lightweight temporal Transformer network, cross-modal attention calculation is performed on the temporal input sequence to obtain cross-modal attention weights, and the cross-modal attention weights are scaled according to the weight coefficients corresponding to the weight feature sequence, humidity feature sequence and color feature sequence, respectively.

[0036] Subsequently, causal self-attention encoding is performed on the scaled attention results. The causal self-attention encoding restricts any time step to establish attention associations only with the inputs of that time step and the time steps before it through a causal mask.

[0037] Output the aligned multimodal temporal representation obtained by the causal self-attention encoding.

[0038] Optionally, S5 includes:

[0039] Within the preset update cycle, the aligned multimodal time series representation is input into the physical constraint regression model, and the liquid volume increment, evaporation volume increment, and retention volume in the absorbent are output. The retention volume in the absorbent in the current update cycle is updated based on the retention volume in the absorbent in the previous update cycle.

[0040] The physical constraint regression model satisfies at least the following constraints during the regression process: the weight change is determined based on the weight data, and a mass conservation constraint is established based on the weight change. The mass conservation constraint is used to ensure that the mass change corresponding to the liquid volume increment entering the absorbent carrier, the evaporation volume increment, and the volume retained in the absorbent carrier is consistent with the weight change.

[0041] The volume increment of the liquid entering the absorption carrier and the volume increment of evaporation are both defined as non-negative.

[0042] The volume of the absorbent carrier remaining within the absorbent carrier is limited to not exceeding the upper limit of the absorbent capacity, as characterized by the eigenvector of the absorbent carrier material parameters. The non-negative constraint and the upper limit of the absorbent capacity constraint are achieved by constraining the output of the physical constraint regression model and / or by setting constraint terms in the physical constraint regression model.

[0043] Optionally, S6 includes:

[0044] The increase in the volume of liquid entering the absorbent carrier is taken as the increase in bleeding volume, or the increase in bleeding volume is calculated based on the increase in the volume of liquid entering the absorbent carrier, combined with the increase in the volume of evaporation and the volume of liquid retained in the absorbent carrier.

[0045] Within each preset update cycle, the incremental bleeding amount is accumulated to generate a cumulative bleeding amount, and the cumulative bleeding amount is used as the bleeding amount monitoring result.

[0046] Optionally, the S7 includes:

[0047] Based on the blood loss monitoring results, combined with weight data, humidity data, and color data, online calibration is performed, which includes fast loop calibration and slow loop update.

[0048] The fast-loop calibration is performed in each preset update cycle. It estimates and updates the zero-point correction parameters of the weight data by estimating the zero-point drift of the weight data and updates the baseline correction parameters of the humidity data by estimating the baseline drift of the humidity data. Based on the updated zero-point correction parameters of the weight data, it performs zero-point correction on the weight data and baseline correction on the humidity data based on the updated baseline correction parameters of the humidity data, so as to obtain the corrected weight data and corrected humidity data used to update the bleeding monitoring results.

[0049] The slow loop update is performed at a lower update frequency than the fast loop calibration. It estimates the white balance deviation of the color data and updates the white balance parameters, and then performs white balance correction on the color data based on the updated white balance parameters.

[0050] Furthermore, the eigenvector of the absorbent carrier material parameters is updated by estimating the long-term deviation of the bleeding volume monitoring results, so that the upper limit of the absorption capacity characterized by the eigenvector of the absorbent carrier material parameters matches the actual absorption state.

[0051] The pad replacement event is detected based on changes in weight data and humidity data. The pad replacement event detection includes: when the change in weight data meets a first change condition and the change in humidity data meets a second change condition, a pad replacement event is determined to be detected.

[0052] When a pad replacement event is detected, the weight zero-point correction parameter and the humidity baseline correction parameter are reset;

[0053] Outputs online calibrated bleeding monitoring results, updated weight zero-point correction parameters, updated humidity baseline correction parameters, updated white balance parameters, and updated absorbent carrier material parameter feature vectors.

[0054] Optionally, the evaporation volume increment is constrained by the evaporation prior value calculated by the evaporation prior model, which is determined based on the temperature, relative humidity and / or airflow parameters in the environmental parameter feature vector, and uses the evaporation prior value as the input feature of the physical constraint regression model and / or as the regularization constraint term in the regression loss function.

[0055] On the other hand, the present invention also provides a real-time monitoring system for gynecological patients based on edge computing, comprising:

[0056] The data acquisition module is used to collect weight data, humidity data, and surface color data of the absorbent carrier used to receive vaginal discharge within a preset update cycle, collect environmental parameter data, and read absorbent carrier material parameter data.

[0057] The feature generation module is used to generate weight feature sequences, humidity feature sequences, color feature sequences, as well as environmental parameter feature vectors and absorber material parameter feature vectors.

[0058] The quality assessment module is used to calculate quality indicators based on the weight feature sequence, humidity feature sequence and color feature sequence, and generate weight coefficients corresponding to the weight feature sequence, humidity feature sequence and color feature sequence respectively based on the quality indicators;

[0059] The Transformer encoding module includes a lightweight temporal Transformer network, which is used to construct and encode the temporal input sequence from the weight feature sequence, humidity feature sequence, color feature sequence, environmental parameter feature vector, and absorber material parameter feature vector, and scale the cross-modal attention weights according to the weight coefficients to generate an aligned multimodal temporal representation.

[0060] The physical constraint regression module includes a physical constraint regression model, which is used to input the aligned multimodal time series representation into the physical constraint regression model, output and update the liquid volume increment, evaporation volume increment and retention volume in the absorbent carrier, wherein the regression process satisfies the mass conservation constraint, the non-negativity constraint and the upper limit constraint of the absorption capacity characterized by the eigenvector of the absorbent carrier material parameters.

[0061] The bleeding volume calculation module is used to generate a bleeding volume increment based on the increase in the volume of liquid entering the absorbent carrier, and to accumulate the bleeding volume increment to obtain a cumulative bleeding volume, which is the bleeding volume monitoring result;

[0062] The online calibration module is used to perform online calibration based on the bleeding monitoring results and in combination with the weight data, humidity data and surface color data;

[0063] An edge computing terminal includes a processor and a memory, wherein the memory stores instructions to cause the processor to perform the functions corresponding to the above-mentioned modules.

[0064] The beneficial effects of this invention are:

[0065] 1. By generating modal weight coefficients based on quality indicators such as weighing jitter, humidity saturation, illumination shift and occlusion, and gating and scaling the cross-modal attention weights in a lightweight temporal Transformer network, dynamic suppression of low-reliability modes and adaptive enhancement of high-reliability modes are achieved, thereby improving the robustness of bleeding estimation under complex working conditions and reducing short-term misjudgment and abnormal fluctuations.

[0066] 2. The feature vectors of the absorber material parameters and the feature vectors of the environmental parameters are used as additional inputs in the temporal coding. In the physical constraint regression, the material parameters are used to represent the upper limit of the absorption capacity. At the same time, the evaporation volume increment constraint is introduced so that the model can adapt to the absorption characteristics of different pad materials and the evaporation differences under different temperature and humidity environments, thereby improving the generalization ability across consumables and scenarios and reducing the cumulative bias.

[0067] 3. The physical constraint regression model outputs and updates the liquid volume increment, evaporation volume increment, and retention volume entering the absorbent carrier, while satisfying mass conservation, non-negativity, and capacity upper limit constraints. Combined with online calibration of fast-loop calibration and slow-loop update, as well as the pad replacement event reset mechanism, it realizes parameter drift suppression and stability improvement during long-term continuous monitoring at the edge, thereby reducing long-term cumulative errors and improving the reliability of monitoring results. Attached Figure Description

[0068] 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:

[0069] Figure 1 This is a flowchart of a real-time monitoring method and system for gynecological patients based on edge computing, as proposed in this invention. Detailed Implementation

[0070] 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.

[0071] refer to Figure 1 A real-time monitoring method for gynecological patients based on edge computing, comprising:

[0072] S1. Within a preset update cycle, collect weight, humidity, and surface color data of the absorbent carrier used to collect vaginal discharge, collect environmental parameter data, read absorbent carrier material parameter data, and generate weight feature sequences, humidity feature sequences, color feature sequences, environmental parameter feature vectors, and absorbent carrier material parameter feature vectors; S2. Calculate quality indicators based on the weight feature sequences, humidity feature sequences, and color feature sequences; S3. Input the quality indicators into the quality assessment model to generate weight coefficients corresponding to the weight feature sequences, humidity feature sequences, and color feature sequences, respectively; S4. Construct a time-series input sequence from the weight feature sequences, humidity feature sequences, color feature sequences, environmental parameter feature vectors, and absorbent carrier material parameter feature vectors, and input it into a lightweight time-series Transf dataset. The ORMER network encoding scales the cross-modal attention weights according to the weight coefficients to generate an aligned multimodal temporal representation; S5, the aligned multimodal temporal representation is input into the physical constraint regression model, which outputs and updates the liquid volume increment entering the absorber, the evaporation volume increment, and the retained volume within the absorber. The physical constraint regression model satisfies mass conservation constraints, non-negativity constraints, and upper limit constraints on absorption capacity during the regression process. The upper limit of absorption capacity is characterized by the eigenvector of the absorber material parameters; S6, the bleeding volume increment is generated based on the liquid volume increment entering the absorber, and the bleeding volume increment is accumulated to obtain the cumulative bleeding volume, which is the bleeding volume monitoring result; S7, online calibration is performed based on the bleeding volume monitoring result and combined with weight data, humidity data, and surface color data.

[0073] In this specific embodiment, S1 includes:

[0074] On edge computing terminals with update cycles Execute step S1 continuously and set The edge computing terminal includes a processor, a memory, and a data bus for communication with the data acquisition module. The data acquisition module includes a weighing sensor, a humidity sensor, an image acquisition module, and an environmental sensor. The weighing sensor... The humidity sensor outputs the weight data of the absorber and appends a timestamp to each output. The image acquisition module outputs humidity data of the absorbing carrier and appends a timestamp to each output. The environmental sensor outputs color data from the surface of the absorbing carrier and appends a timestamp to each frame of image. Output environmental parameter data and append a collection timestamp to each output, wherein the collection timestamp is a monotonically increasing millisecond count and denoted as . At the beginning of each update cycle, the edge computing terminal reads the raw data from the local cache queue of each sensor within the most recent period and performs time alignment so that the time-aligned weight data, time-aligned humidity data, and time-aligned color data correspond to the same time axis. The same time axis uses the sampling time based on the end time of the current update cycle. and This indicates the update cycle number, and in each Linear interpolation resampling is performed on the weight and humidity data respectively to obtain aligned weight and humidity values. The same interpolation rule is used for any single-channel sensor sequence to be aligned.

[0075] ;

[0076] in This indicates the weight or humidity data to be aligned. Indicates the moment of target alignment. and Indicates satisfaction The timestamps of two consecutive original data collections and They represent in and The raw sensor readings collected at the location;

[0077] The color data is time-aligned with distance The most recent and satisfied The image frame is used as the aligned color data and the pixel matrix of the frame is used as the input for subsequent color feature extraction. If there is no image frame within the time tolerance, the aligned color data of the previous update cycle is used as the aligned color data of the current update cycle to ensure the continuity of the timing input.

[0078] After obtaining the time-aligned weight data, the edge computing terminal generates a weight feature sequence, which consists of a weight value sequence and a weight change sequence, with the weight value sequence denoted as... ,in Indicates in The aligned weight values ​​in grams, the sequence of weight changes is denoted as... and And in When Set to 0;

[0079] After obtaining the time-aligned humidity data, the edge computing terminal generates a humidity feature sequence, which consists of a humidity value sequence and a humidity change rate sequence, and the humidity value sequence is denoted as... ,in Indicates in The aligned humidity values ​​are expressed as a percentage of relative humidity, and the humidity change rate sequence is denoted as... and And in When Set to 0;

[0080] After obtaining the aligned color data, the edge computing terminal performs a fixed-process color feature extraction on the image frame to generate a color feature sequence. This fixed process includes converting the image from the RGB color space to the HSV color space and calculating luminance and chromaticity features within a preset region of interest (ROI) on the absorption carrier. The ROI is defined by the mounting structure as a rectangular region at the center of the image, with a width 0.6 times the image width and a height 0.6 times the image height. The luminance feature sequence is denoted as... and Defined as HSV in the region of interest The arithmetic mean of the channel pixel values, the chromaticity feature sequence is denoted as and Defined as HSV in the region of interest The arithmetic mean of the channel pixel values, the and A new sample is generated in each update cycle and... Binding;

[0081] The edge computing terminal simultaneously collects environmental parameter data once in each update cycle and forms an environmental parameter feature vector. The environmental parameter data includes temperature, relative humidity, and airflow velocity, which are denoted as follows: and The environmental parameter feature vector is denoted as And the fields are concatenated in the order of temperature, relative humidity, and airflow velocity. By By interval The value is obtained by performing minimum-maximum normalization on the Celsius value. By By interval The percentage is obtained by minimum-maximum normalization. By By interval The value is obtained by performing minimum-maximum normalization on meters per second, and the normalization interval is stored as a constant in the memory to ensure consistency across cycles.

[0082] The edge computing terminal reads the absorber carrier material parameter data and forms an absorber carrier material parameter feature vector when the absorber carrier is placed in position. This reading is accomplished through identification information on the absorber carrier, which is encoded using a QR code and decoded by an image acquisition module during placement confirmation. The absorber carrier material parameter data includes the absorber carrier dry weight, upper limit of absorption capacity, and thickness, respectively denoted as... and The characteristic vector of the absorber material parameters is denoted as... And the data is concatenated in the order of dry weight, maximum absorption capacity, and thickness. By By interval The min-max normalization is obtained by performing min-max normalization on the gram. By By interval The milliliters were obtained by min-max normalization. By By interval The millimeter is obtained by minimum-maximum normalization, and the above normalization interval is stored as a constant in the memory;

[0083] To form a time-series data structure that can be directly input into subsequent steps, the edge computing terminal maintains a length of [length missing] for each of the weight feature sequence, humidity feature sequence, and color feature sequence. A circular buffer and in each update cycle, the newly generated one... and Append to the corresponding buffer; if the buffer is not yet full, fill it again with the earliest obtained sample until the length is reached. To ensure that the sequence length remains constant;

[0084] After completing the above acquisition, alignment, feature extraction and vectorization, the output includes weight data, humidity data, color data, and time-aligned weight feature sequence, humidity feature sequence, color feature sequence, environmental parameter feature vector and absorber material parameter feature vector.

[0085] In this specific embodiment, S2 includes:

[0086] On the edge computing terminal, during each update cycle The end of the process executes step S2 and uses a weight feature sequence. Humidity characteristic sequence and color feature sequences As input, where Indicates the update cycle number. This represents the aligned weight value in grams. Indicates the change in weight and This represents the aligned humidity value in percentage of relative humidity. This indicates the rate of change in humidity, expressed as a percentage per second. This indicates the HSV color space within the preset area of ​​interest. The average value of the channel and its range is This indicates the HSV color space within the preset area of ​​interest. The average value of the channel and its range is ;

[0087] The weighing jitter index is characterized by the short-term volatility of the weight change sequence within a sliding time window, with the sliding time window length set to [value missing]. Each update cycle and each time... Incrementing forward by one cycle, through the set The standard deviation is calculated to obtain the volatility, which is then normalized to a threshold and mapped to a weighing jitter index. The formula for its calculation is:

[0088] ;

[0089] in The range of values ​​is Furthermore, a larger value indicates more severe jitter. This indicates that the input will be truncated to... Interval cutoff function, This represents the number of samples within the sliding time window. Indicates the first Weight change per update cycle Indicates the sliding time window The arithmetic mean, This indicates that the jitter is normalized and scaled and taken as a fixed value. ;

[0090] The humidity saturation index is determined by combining the degree to which the humidity value is close to the saturation threshold with the degree of slowness in humidity change, and is denoted as [insert index here]. The saturation threshold is fixed at 1. And the threshold for slowing down the change is fixed at 1. ,when When Set to 0 to indicate that the humidity is far from the saturation range, when When Set to 0.4 to characterize near saturation while still possessing a certain dynamic range, when and When Set to 1 to characterize that humidity has reached saturation and its change has slowed down, making it insensitive to new liquid additions. and When Set to 0.7 to indicate that although it is close to saturation, there is still usable information about changes;

[0091] The illumination offset index is determined and denoted by the degree of deviation of the brightness feature sequence from a preset reference brightness. The reference brightness is denoted as And before the detection of the pad replacement event Within each update cycle, by The median is determined and fixed until the next pad replacement event, thus ensuring that the reference brightness is consistent with the current absorber and the current installation position. When Set to 0 to characterize stable illumination. When Set to 0.5 to characterize the presence of a moderate illumination shift, when When Set to 1 to characterize the reduced reliability of color features caused by significant illumination shift;

[0092] The occlusion index is determined and denoted by the effective pixel percentage of the color data. The effective pixels are defined as those that simultaneously satisfy the following conditions within a preset region of interest: and The pixels are selected to exclude unusable pixels caused by being too dark, overexposed, or undersaturated. and These are the luminance and saturation channels in the HSV color space, respectively, and the effective pixel percentage is denoted as... It is obtained by dividing the number of effective pixels by the total number of pixels in the region of interest. When Set to 0 to indicate no occlusion or negligible occlusion effect. When Set to 0.5 to represent partial occlusion, when When Set to 1 to indicate that severe occlusion leads to distortion in color feature extraction;

[0093] Finally, the weighing vibration index will be... Humidity saturation index Illumination shift index and occlusion index The quality indicator vector is composed according to the field order. .

[0094] In this specific embodiment, S3 includes:

[0095] On the edge computing terminal, during each update cycle The final step S3 is executed and the quality index vector is... Input the quality assessment model to generate weight coefficients corresponding to the weight feature sequence, humidity feature sequence, and color feature sequence, respectively. Indicates the update cycle number. This represents a quality index vector, and its four components are the weighing jitter indexes. Humidity saturation index Illumination shift index With occlusion index The quality assessment model operates with fixed parameters at the edge and employs a two-layer feedforward neural network structure to map quality indicators to modal reliability. The two-layer feedforward neural network has an input dimension of 4, a fixed hidden layer dimension of 8, and an output dimension of 3, outputting three reliability scores in the order of "weight, humidity, and color." Specifically, The three outputs are calculated sequentially through the first and second fully connected layers, and the values ​​are compressed to their maximum values ​​using the Sigmoid function at the output. Intervals are used to form a reliability score vector:

[0096] ;

[0097] in The reliability score represents the weight feature sequence. The reliability score represents the humidity feature sequence. The reliability score represents the color feature sequence, and the parameters of the first fully connected layer are the weight matrix. With bias vector And the activation function uses The parameters of the second fully connected layer are the weight matrix. With bias vector And the output activation function adopts The and The data is obtained through offline training and permanently stored in the non-volatile memory of the edge computing terminal, and is not updated during the online inference phase;

[0098] In the Set a preset reliability threshold when generating weighting coefficients. The weight lower bound rule is set to zero, and threshold suppression and renormalization are performed to obtain the final weight coefficients. Specifically, a threshold comparison is performed on each reliability score. The score corresponding to the weight after suppression is set to 0. ,when The score corresponding to the humidity suppression is set to 0. ,when The suppressed score corresponding to the color is then set to 0. When the corresponding score is not less than Keep and in The weight coefficient vector is obtained by normalizing it according to the following formula:

[0099] ;

[0100] in This represents the reliability score vector after threshold suppression. This represents the final weight coefficient vector, corresponding to the weight feature sequence, humidity feature sequence, and color feature sequence, respectively. Represents a vector consisting entirely of 1s. Indicates to The scalar normalized denominator obtained by summing the three components, the fraction represents the expression for... Perform element-wise division by the denominator to ensure... The range of values ​​is And the sum of the three is 1;

[0101] When it appears In the case of Set to the weight coefficient vector of the previous update cycle. To maintain cross-cycle gating stability and in the first update cycle after system startup Fixed initialization ;

[0102] Output the and .

[0103] In this specific embodiment, S4 includes:

[0104] On the edge computing terminal, during each update cycle The final step S4 is executed, and the weight feature sequence, humidity feature sequence, color feature sequence, environmental parameter feature vector and absorber material parameter feature vector are constructed into a time-series input sequence and then input into a lightweight temporal Transformer network to complete the encoding. The cross-modal attention weights are scaled according to the weight coefficients to generate an aligned multimodal temporal representation.

[0105] The timing input sequence adopts a length The sliding sequence is arranged from oldest to newest as a time index. The weight characteristic is in the first The input vector at each time step is denoted as... Humidity characteristics in the first The input vector at each time step is denoted as... Color features in the first The input vector at each time step is denoted as... The environmental parameter feature vector is denoted as Furthermore, the eigenvectors of the absorber material parameters are shared across all time steps within this update cycle, and are denoted as follows: Furthermore, it is shared across all update cycles and all time steps within the same pad replacement cycle. This represents the aligned weight value in grams. It represents the change in weight, and the unit is grams. This represents the aligned humidity value in percentage of relative humidity. This indicates the rate of change in humidity, expressed as a percentage per second. Represents brightness characteristics and has a value range of . Represents chromaticity characteristics and has a value range of . and All are normalized values ​​obtained in step S;

[0106] The model dimension of the lightweight temporal Transformer network is set to 1. It includes a cross-modal attention alignment layer and two causal self-attention coding layers. The cross-modal attention alignment layer performs cross-modal attention on the three embeddings of weight, humidity and color at each time step and outputs the three-way aligned modal representation. The two causal self-attention coding layers perform self-attention with causal mask on the time dimension to ensure that any time step only establishes attention association with that time step and the time steps before it, thereby satisfying real-time inference at the edge.

[0107] During the embedding construction phase, respectively for Performing a linear mapping yields modal embeddings. The linear mapping parameters are fixed as follows: as well as as well as And it is solidified and stored at the edge, and meets the requirements. , ;

[0108] At the same time and Conditional embedding is obtained through linear mapping. and The linear mapping parameters are fixed as follows: as well as and satisfy and ;

[0109] At each time step, the conditional embeddings are combined with the three-way modal embeddings to form the input for cross-modal attention. The combination method involves performing conditional injection on each modal embedding and consistently using additive injection, i.e. Updated to This imposes consistent conditional constraints on the alignment process of the three modes by both environmental and material parameters;

[0110] The cross-modal attention alignment layer employs multi-head attention with a fixed number of heads. Within the same time step, three modal embeddings are used as three types of tokens, and a cross-modal attention weight matrix is ​​calculated. Then, gating scaling is performed based on the weight coefficients. The weight coefficient vector is denoted as... And respectively correspond to the weight feature sequence, humidity feature sequence and color feature sequence and satisfy the following conditions: The gating scaling applies to any query modality at each time step. The cross-modal attention weights are applied as follows:

[0111] ;

[0112] in Represents the scaled attention weights and represents the first... Query modality in each time step bond mode Attention weights Indicates the first Key modality within each update cycle The weighting coefficients, This represents the unscaled attention weights, calculated from the scaled dot product and Softmax within the cross-modal multi-head attention, and satisfies the condition for the same query modality. have In the denominator This represents the weighted summation and normalization term for the three-way key modes, thus ensuring... The scaled attention weights are used to perform a weighted summation of the value vectors of the three modalities to obtain aligned three-modal representations. At the end of the cross-modal alignment layer, the aligned three-modal representations are concatenated in a fixed order into a single time-step representation. After linear projection matrix With bias Mapped to fused representation ;

[0113] Before entering the causal self-attention encoding layer, a learnable positional encoding is added to each time step to preserve temporal order information. The learnable positional encoding is a parameter table. And the first row vector Directly with Adding them together gives Both causal self-attention coding layers employ Head attention and feedforward network dimensions Furthermore, each layer includes residual connections and layer normalization, and a causal mask matrix is ​​applied in the attention calculation to mask all conditions that satisfy the condition. The future time step key-value pairs thus form a time-series coding result that depends only on history;

[0114] Output aligned multimodal timing representation And each The output is the causal encoding for the corresponding time step.

[0115] In this specific embodiment, S5 includes:

[0116] On the edge computing terminal, during each update cycle The end of the process executes step S5 and then performs the aligned multimodal timing representation. The physical constraint regression model is input to output and update the liquid volume increment entering the absorbent carrier, the evaporation volume increment, and the retention volume within the absorbent carrier. Indicates the update cycle number. Indicates timing length, Indicates the first The encoded vector at each time step and ;

[0117] The physical constraint regression model runs with fixed parameters at the edge and is implemented using a deterministic structure of "temporal representation convergence layer + regression head + constraint mapping + state update". The temporal representation convergence layer takes the latest time step vector as the periodic level representation and denoted as... The regression head is a two-layer feedforward network with parameters for its first fully connected layer as follows: and And adopt As the activation function, its second fully connected layer parameters are and And output scalar The The non-negative liquid volume residual entering the absorber is obtained through constrained mapping and denoted as . And the mapping method is fixed as follows ,in This indicates the upper limit of the allowed entry volume residual within a single update cycle. Representing the Sigmoid function to ensure ;

[0118] The mass conservation constraint determines the weight change based on weight data and introduces a liquid density constant between the volume domain and the mass domain. The weight change is calculated from the weight value sequence and denoted as... And the unit is grams, of which For the first The weight value after each update cycle alignment, wherein the liquid density constant is fixed as... And it is stored as a fixed parameter in the edge computing terminal;

[0119] To simultaneously satisfy both the non-negativity constraint and the upper limit constraint of absorption capacity, the retention volume within the absorber in the previous update cycle is first determined. With upper limit of absorption capacity Perform a physically feasible trimming on the weight change to obtain the effective weight change for establishing mass conservation. ,in The unit is milliliters and is reset to [value] when a pad replacement event is detected. The material parameter readings, determined by the feature vector of the absorbent carrier material parameters and in milliliters, remain constant within the same pad replacement cycle. The physically feasible cutting will... The constraint is to ensure that the updated retention volume falls within the range. The value of is taken and the cropping result is recorded as . ;

[0120] In obtaining The volume increment of the liquid entering the absorbent carrier is then calculated according to the mass conservation constraint. With evaporation volume increment And ensuring that both are non-negative and consistent with the effective weight change, the mass conservation constraint is written as:

[0121] ;

[0122] in This represents the effective change in weight used for conservation constraints, expressed in grams. This indicates the density of a liquid as a constant, expressed in grams per milliliter. This indicates the volume increase of liquid entering the absorbent carrier, expressed in milliliters. Indicates the volume increase during evaporation, with the unit being milliliters;

[0123] To ensure that the equation holds strictly on a cycle-by-cycle basis during the edge-end inference phase and simultaneously satisfies the non-negativity constraint, the volumetric basis directly given by the weight change is first calculated. And order Then solve using the mass conservation equation. Thus guarantee and ;

[0124] In obtaining and Then, the state of the retained volume within the absorbent carrier is updated, and an upper limit constraint on the absorption capacity is applied. The state update uses the retained volume from the previous update cycle. Perform recursion and denote it as and will By limiting the saturation mapping The interval is defined as the retention volume within the absorber that forms the final output, wherein the saturation mapping is to first set values ​​less than 0 to 0 and then set values ​​greater than 0 to 0. Set the value to ;

[0125] The volume increment of liquid entering the absorbent carrier is output in each update cycle. Evaporation volume increment and the retention volume within the updated absorbent carrier .

[0126] In this specific embodiment, S6 includes:

[0127] On the edge computing terminal, during each update cycle The final step S6 is performed based on the liquid volume increment entering the absorbent carrier. The incremental bleeding volume is generated and accumulated to obtain the cumulative bleeding volume as the bleeding volume monitoring result. Indicates the update cycle number. The unit is milliliters and it indicates the number of milliliters in the first milliliter. The increase in the volume of liquid entering and being absorbed by the absorbent carrier within each update cycle, wherein the increase in bleeding volume is fixedly defined in this embodiment as... ,in Indicates the first The bleeding volume increment is measured in milliliters for each update cycle, thus ensuring that the bleeding volume calculation is consistent with the mass conservation constraint and avoiding duplicate measurements caused by secondary derivation of the evaporation volume increment and the retention volume.

[0128] The cumulative blood loss is recorded as Furthermore, the main output, which serves as the real-time monitoring output at the edge, is continuously accumulated and updated within the same pad replacement cycle. This accumulation is implemented using a recursive register and is performed according to the update cycle. Execution, in which This represents the cumulative bleeding amount in the previous update cycle, expressed in milliliters. This indicates the increase in bleeding volume during the current update cycle, expressed in milliliters. This indicates the updated cumulative blood loss, expressed in milliliters.

[0129] When the system starts up The initial value is fixed at 0 mL. When a pad replacement event is detected and a reset is performed in step S7, the value will be... The volume is simultaneously reset to 0 mL to ensure that the cumulative blood loss strictly corresponds to the current usage cycle of the absorption carrier;

[0130] In When the output is the blood loss monitoring result, the output for the current update period will also be output. As a short-term bleeding trend quantity and and Write records with timestamps to local storage.

[0131] In this specific embodiment, S7 includes:

[0132] On the edge computing terminal, during each update cycle The final step S7 will be executed and the cumulative bleeding volume will be recorded. and weight data Humidity data With color data As an input for online calibration, Indicates the update cycle number. The unit is milliliters and represents the number up to the specified number. The cumulative bleeding amount at the end of each update cycle. The unit is gram and represents the first. Aligned weight value at the end of each update cycle The unit is relative humidity percentage and represents the first... Aligned humidity value at the end of each update cycle Indicates the first RGB image frames aligned with the timeline at the end of each update cycle are used for white balance estimation and correction.

[0133] The online calibration includes fast-loop calibration and slow-loop updates, both of which are performed with defined parameters and defined judgment conditions. The fast-loop calibration is performed and maintains the zero-point weight correction parameters in each update cycle. Humidity baseline correction parameters ,in The unit is grams and represents the zero-point drift of the weighing channel. The unit is relative humidity percentage and represents the baseline drift of the humidity channel. The slow loop update is performed at a fixed period. Perform and maintain white balance parameters With the eigenvector of the absorber material parameters The upper limit of the absorption capacity corresponds to the parameter ,in It is a three-channel gain vector and Corresponding to The channel gain coefficient and used to... The color data is corrected to white balance. The unit is milliliters and is used for the upper limit constraint of absorption capacity in step S5, and is recalculated synchronously after the update. The corresponding normalized component ;

[0134] The drift estimation of the fast-loop calibration is implemented using a calibrator structure of "stability interval determination + exponential sliding update", and the stability interval determination is fixed in this embodiment as simultaneously satisfying and The update cycle, of which This indicates the change in weight data. This indicates the amount of change in humidity data. This indicates an increase in bleeding volume;

[0135] The instantaneous estimate of the zero-point drift of the weight is obtained within the update period that satisfies the stability interval criterion. Set as To ensure that the corrected weight remains constant between adjacent periods and to provide an instantaneous estimate of the humidity baseline drift. Set as To ensure that the corrected humidity remains constant in adjacent cycles, in update cycles that do not meet the stability interval criteria, respectively set... and Estimated by frozen fast-loop drift;

[0136] The drift parameters are updated using an exponential sliding update rule:

[0137] ;

[0138] in This indicates the calibration parameters to be updated and are taken during fast loop calibration. This indicates the calibration parameter with the same name from the previous update cycle. This represents the instantaneous estimate of the same name obtained from the stability interval determination. This represents the exponential sliding update coefficient, and in this embodiment, the weight zero-point correction parameter is taken as... And take the humidity baseline correction parameters The update rule is executed once in each update cycle and implemented at the edge using floating-point numbers, and the updated... and Write to non-volatile memory to support power-off resume operation;

[0139] After completing the fast ring parameter update, the weight data will be... Subtract weight zero-point correction parameters Obtain the corrected weight data and humidity data Subtract humidity baseline correction parameters Obtain the corrected humidity data The and This serves as the input cache value for the next update cycle step S1, enabling subsequent steps S2 to S6 to calculate the monitoring results based on the corrected weight and humidity.

[0140] The slow loop update satisfies The update cycle is executed, and white balance deviation estimation and white balance parameter updates are performed on the color data. The white balance deviation estimation uses the same region of interest as in step S1, and pixels with a channel pixel value less than 10 or greater than 245 are removed from the region before calculation. The average of the three channels is used, with the arithmetic mean of the three channel averages as the target average. The instantaneous white balance gain of each channel is then calculated. Set as a scaling factor to scale the channel mean to the target mean and clip each gain component to... The interval is used to limit the amplified noise under extreme lighting conditions, and then... Update element by element according to the exponential sliding update rule, with the slow loop update coefficient fixed. And will update The original image frame is applied to the next update cycle to generate white balance corrected color data for step S1 to re-extract the luminance feature sequence and chrominance feature sequence, thereby suppressing the systematic deviation of white balance drift on color features;

[0141] The slow-loop update simultaneously performs a long-term deviation-driven update on the absorber material parameter feature vector M, and the long-term deviation is calculated in this embodiment when a pad replacement event occurs. The pad replacement event detection is based on the simultaneous satisfaction of two conditions: a sudden change in weight data and a sudden change in humidity data, with the first sudden change condition fixed as follows: And the second mutation condition is fixed as follows: When the condition is met, the replacement pad flag is set to 1, and the replacement pad flag is paused during the period when the condition is set to 1. The new additions are accumulated and reset and restored after the stable conditions for the new absorbent carrier to be in place are detected. The stable conditions for the new absorbent carrier to be in place are fixed as continuous. Each update cycle satisfies and ;

[0142] Latch the calibrated weight of the previous cycle before the pad replacement event is detected and the system reaches a stable condition. With cumulative blood loss and the dry weight of the absorbent carrier and liquid density constant The reference volume of the liquid in the absorbent carrier before pad replacement is calculated as follows: As the mass of the liquid and divided by The obtained volume value, then with The difference between the volume and the reference volume is taken as the long-term bias and calculated using the learning rate. Weighted summation of the long-term deviation The instantaneous estimate of the upper limit of absorption capacity is formed and then clipped to... Update according to the exponential sliding update rule after the interval. And the slow loop update coefficient is fixed at 1. The updated Write back And update the normalized components simultaneously. To ensure that the upper limit constraint of absorption capacity matches the actual absorption state;

[0143] The weight zero-point correction parameter is adjusted when the stability conditions for the new absorbent carrier in place are met. Reset to make the corrected weight equal to the current dry weight of the absorbent carrier. The values ​​and humidity baseline correction parameters Reset to current humidity value This establishes a baseline for the new pad replacement cycle and resets the pad replacement flag.

[0144] Output online calibrated blood loss monitoring results and updated weight zero-point correction parameters Humidity baseline correction parameters White balance parameters With the eigenvector of the absorber material parameters .

[0145] In this specific embodiment, during each update cycle The internal model constructs an evaporation prior model based on environmental parameter data and calculates evaporation prior values ​​to account for evaporation volume increments. Apply constraints;

[0146] The environmental parameter data includes temperature. relative humidity With airflow speed ,in Indicates the update cycle number. The unit is And by the environmental sensor in the first Data collected in each update cycle, The unit is And the range of values ​​is The unit is And the range of values ​​is ;

[0147] The evaporation prior model is a parameterized model that can be determined at the edge and outputs evaporation prior values ​​in each update cycle. The formula for its calculation is:

[0148] ;

[0149] in Indicates the first The evaporation prior value for each update cycle, and the unit is... This indicates that the input will be truncated to... The truncation function, This is a constant representing the upper limit of the evaporation volume increment in a single update cycle. This represents the constant of the evaporation intensity coefficient. This represents a constant representing the temperature sensitivity coefficient. Indicates the reference temperature constant. Indicates the update cycle duration and the unit is . The proportional value representing relative humidity thus makes the item Indicates the degree of air dryness. This represents the linear enhancement factor of airflow on evaporation. This represents an exponential function, thus increasing the prior value of evaporation as the temperature rises;

[0150] To ensure that the evaporation prior value is robust against transient noise during edge-end operation, the aforementioned The first-order low-pass method is used for smoothing, and the smoothing coefficient is set to 0.2. The final evaporation prior value used for constraint is obtained by updating "0.8 of the smoothing value of the previous period plus 0.2 of the prior value of the current period".

[0151] The evaporation prior value is applied to the physical constraint regression model in step S5 in two ways: firstly, it is used as an input feature to concatenate the evaporation prior value into the periodic representation used for regression in step S5. Forming extended representation Accordingly, the input dimension of the first fully connected layer of the regression head in step S5 is fixedly adjusted from 64 to 65, so that the model can explicitly perceive the environmental evaporation conditions during inference. Secondly, as an output constraint, the evaporation volume increment is obtained in step S5. Then perform hard clipping to meet the requirements. And The amount is fixed as a priori relaxation value to allow for short-term environmental measurement errors and evaporation deviations caused by local differences on the surface of the absorber.

[0152] Simultaneously, during offline training of the physical constraint regression model, the "deviation between the evaporation volume increment and the prior evaporation value" was used as a regularization constraint term with a fixed regularization weight of 0.2. This was to encourage the model to prioritize the evaporation volume increment consistent with environmental conditions, while satisfying the mass conservation constraint, non-negativity constraint, and upper limit constraint of absorption capacity. The final output at the edge... Numerically constrained by evaporation priors and exhibiting interpretable adaptive consistency under changes in ambient temperature, relative humidity, and airflow.

[0153] 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.

[0154] This invention addresses the challenges of continuously and in real-time quantification of vaginal discharge bleeding, which is susceptible to cumulative errors due to operational conditions. It aligns and fuses weight, humidity, and surface color data over time into a fusion model. Furthermore, it incorporates environmental parameters and absorbent carrier material parameters as conditional information in the encoding process. This allows bleeding estimation to go beyond instantaneous readings from a single sensor; instead, it relies on the complementary relationships of multiple information sources to jointly infer the processes of fluid entry, evaporation, and retention. The cumulative bleeding amount is obtained by periodically accumulating the volume increase of the fluid entering the absorbent carrier, enabling continuous monitoring at the edge. This reduces the subjectivity and instability of traditional visual estimation or single-weighing calculations, improving the real-time performance and reliability of bleeding quantification.

[0155] This invention incorporates improvements for both operating conditions and physical processes: First, it generates modal weight coefficients based on quality indicators such as weighing jitter, humidity saturation, illumination shift, and occlusion. These weight coefficients are then applied to the cross-modal attention weight scaling of a lightweight temporal Transformer network, achieving gating suppression of low-reliability modes and adaptive enhancement of high-reliability modes, thus structurally improving robustness under complex operating conditions. Second, it employs decomposed physical constraint regression to output and update the ingress volume increment, evaporation volume increment, and retention volume. Bleeding is obtained under constraints of mass conservation, non-negativity, and upper limit of absorption capacity, avoiding unreasonable estimates arising from relying solely on black-box regression and reducing long-term cumulative bias. Third, it establishes an online calibration mechanism with fast-loop calibration and slow-loop update, resetting key correction parameters upon detecting a pad replacement event to suppress zero-point drift and baseline drift and maintain stability during long-term continuous monitoring. This better achieves the technical effects of real-time quantification of bleeding and low cumulative error.

Claims

1. A real-time monitoring method for gynecological patients based on edge computing, comprising: S1. Within a preset update cycle, collect weight, humidity, and surface color data of the absorbent carrier used to collect vaginal discharge, collect environmental parameter data, read absorbent carrier material parameter data, and generate weight feature sequences, humidity feature sequences, color feature sequences, environmental parameter feature vectors, and absorbent carrier material parameter feature vectors; S2. Calculate quality indicators based on the weight feature sequences, humidity feature sequences, and color feature sequences; S3. Input the quality indicators into the quality assessment model to generate weight coefficients corresponding to the weight feature sequences, humidity feature sequences, and color feature sequences, respectively; S4. Construct a time-series input sequence from the weight feature sequences, humidity feature sequences, color feature sequences, environmental parameter feature vectors, and absorbent carrier material parameter feature vectors, and input it into a lightweight time-series Transf dataset. The ORMER network encoding scales the cross-modal attention weights according to the weight coefficients to generate an aligned multimodal temporal representation; S5, the aligned multimodal temporal representation is input into the physical constraint regression model, which outputs and updates the liquid volume increment entering the absorber, the evaporation volume increment, and the retained volume within the absorber. The physical constraint regression model satisfies mass conservation constraints, non-negativity constraints, and upper limit constraints on absorption capacity during the regression process. The upper limit of absorption capacity is characterized by the eigenvector of the absorber material parameters; S6, the bleeding volume increment is generated based on the liquid volume increment entering the absorber, and the bleeding volume increment is accumulated to obtain the cumulative bleeding volume, which is the bleeding volume monitoring result; S7, online calibration is performed based on the bleeding volume monitoring result and combined with weight data, humidity data, and surface color data.

2. The method for real-time monitoring of gynecological patients based on edge computing according to claim 1, S1 includes: Within the preset update cycle, the weight data of the absorbent carrier used to receive vaginal fluid is collected by a weighing sensor, the humidity data of the absorbent carrier is collected by a humidity sensor, the color data of the surface of the absorbent carrier is collected by an image acquisition module, the environmental parameter data is collected by an environmental sensor, and the absorbent carrier material parameter data is read from local storage or identification information. The weight data, humidity data, color data, and environmental parameter data all carry a collection timestamp. The weight data, humidity data, and color data are time-aligned based on the acquisition timestamp so that the time-aligned weight data, humidity data, and color data correspond to the same time axis. Feature extraction is performed on the time-aligned weight data to generate a weight feature sequence, which includes a weight value sequence and a weight change sequence calculated from the weight value sequence. A humidity feature sequence is generated by extracting features from the time-aligned humidity data. The humidity feature sequence includes a humidity value sequence and a humidity change rate sequence calculated from the humidity value sequence. A color feature sequence is generated by extracting features from the time-aligned color data. The color feature sequence includes a luminance feature sequence and a chromaticity feature sequence calculated from the color data. The environmental parameter data is vectorized to generate environmental parameter feature vectors. The vectorization process includes normalizing the numerical environmental parameters and concatenating them according to a preset field order. The absorber material parameter data is vectorized to generate absorber material parameter feature vectors. The vectorization process includes normalizing the numerical material parameters and concatenating them according to a preset field order. Output weight data, humidity data, color data, weight feature sequence, humidity feature sequence, color feature sequence, environmental parameter feature vector, and absorber material parameter feature vector.

3. The method for real-time monitoring of gynecological patients based on edge computing according to claim 1, S2 includes: The short-term fluctuation of the weight change sequence is calculated within a preset sliding time window based on the weight feature sequence, and the short-term fluctuation is used as the weighing jitter index. The degree of closeness of the humidity value sequence to the preset saturation threshold is determined based on the humidity feature sequence, and the degree of slowing down of humidity change is determined by combining the absolute value of the humidity change rate sequence, so as to generate a humidity saturation index. The degree of deviation of the brightness feature sequence from the preset reference brightness is calculated based on the color feature sequence to generate an illumination offset index; The effective pixel ratio of color data or the confidence level of color feature extraction is determined based on the color feature sequence, and an occlusion index is generated based on the effective pixel ratio or the confidence level of color feature extraction. The quality index is generated by combining the weighing vibration index, the humidity saturation index, the illumination offset index, and the shading index.

4. The method for real-time monitoring of gynecological patients based on edge computing according to claim 1, S3 includes: The quality indicators are constructed into a quality indicator vector and input into the quality assessment model. The quality assessment model outputs the reliability scores corresponding to the weight feature sequence, humidity feature sequence and color feature sequence, respectively. Based on the reliability score, weight coefficients are generated corresponding to the weight feature sequence, humidity feature sequence, and color feature sequence, respectively. The weight coefficients are obtained by normalizing the reliability score and satisfy the value range of [0,1] and the sum of the three is 1. When any reliability score is lower than the preset reliability threshold, the weight coefficient corresponding to that reliability score is reduced to the preset lower limit or set to zero, and the remaining weight coefficients are renormalized to obtain the final weight coefficient.

5. The method for real-time monitoring of gynecological patients based on edge computing according to claim 1, S4 includes: The weight feature sequence, humidity feature sequence, and color feature sequence are respectively converted into feature embedding sequences with the same input dimension as the lightweight temporal Transformer network through linear mapping. At each time step, the corresponding weight feature embedding, humidity feature embedding, and color feature embedding are concatenated to form the multimodal feature representation of that time step. The environmental parameter feature vector and the absorber material parameter feature vector are mapped to environmental parameter embedding and absorber material parameter embedding, respectively, and combined with the multimodal feature representation as additional input to form a time-series input sequence; The time-series input sequence is fed into the lightweight temporal Transformer network after adding position encoding. In the lightweight temporal Transformer network, cross-modal attention calculation is performed on the temporal input sequence to obtain cross-modal attention weights, and the cross-modal attention weights are scaled according to the weight coefficients corresponding to the weight feature sequence, humidity feature sequence and color feature sequence, respectively. Subsequently, causal self-attention encoding is performed on the scaled attention results. The causal self-attention encoding restricts any time step to establish attention associations only with the inputs of that time step and the time steps before it through a causal mask. Output the aligned multimodal temporal representation obtained by the causal self-attention encoding.

6. The method for real-time monitoring of gynecological patients based on edge computing according to claim 1, S5 includes: Within the preset update cycle, the aligned multimodal time series representation is input into the physical constraint regression model, and the liquid volume increment, evaporation volume increment, and retention volume in the absorbent are output. The retention volume in the absorbent in the current update cycle is updated based on the retention volume in the absorbent in the previous update cycle. The physical constraint regression model satisfies at least the following constraints during the regression process: the weight change is determined based on the weight data, and a mass conservation constraint is established based on the weight change. The mass conservation constraint is used to ensure that the mass change corresponding to the liquid volume increment entering the absorbent carrier, the evaporation volume increment, and the volume retained in the absorbent carrier is consistent with the weight change. The volume increment of the liquid entering the absorption carrier and the volume increment of evaporation are both defined as non-negative. The volume of the absorbent carrier remaining within the absorbent carrier is limited to not exceeding the upper limit of the absorbent capacity, as characterized by the eigenvector of the absorbent carrier material parameters. The non-negative constraint and the upper limit of the absorbent capacity constraint are achieved by constraining the output of the physical constraint regression model and / or by setting constraint terms in the physical constraint regression model.

7. The method for real-time monitoring of gynecological patients based on edge computing according to claim 1, S6 includes: The increase in the volume of liquid entering the absorbent carrier is taken as the increase in bleeding volume, or the increase in bleeding volume is calculated based on the increase in the volume of liquid entering the absorbent carrier, combined with the increase in the volume of evaporation and the volume of liquid retained in the absorbent carrier. Within each preset update cycle, the incremental bleeding amount is accumulated to generate a cumulative bleeding amount, and the cumulative bleeding amount is used as the bleeding amount monitoring result.

8. The method for real-time monitoring of gynecological patients based on edge computing according to claim 1, S7 includes: Based on the blood loss monitoring results, combined with weight data, humidity data, and color data, online calibration is performed, which includes fast loop calibration and slow loop update. The fast-loop calibration is performed in each preset update cycle. It estimates and updates the zero-point correction parameters of the weight data by estimating the zero-point drift of the weight data and updates the baseline correction parameters of the humidity data by estimating the baseline drift of the humidity data. Based on the updated zero-point correction parameters of the weight data, it performs zero-point correction on the weight data and baseline correction on the humidity data based on the updated baseline correction parameters of the humidity data, so as to obtain the corrected weight data and corrected humidity data used to update the bleeding monitoring results. The slow loop update is performed at a lower update frequency than the fast loop calibration. It estimates the white balance deviation of the color data and updates the white balance parameters, and then performs white balance correction on the color data based on the updated white balance parameters. Furthermore, the eigenvector of the absorbent carrier material parameters is updated by estimating the long-term deviation of the bleeding volume monitoring results, so that the upper limit of the absorption capacity characterized by the eigenvector of the absorbent carrier material parameters matches the actual absorption state.

9. Detecting pad replacement events based on sudden changes in weight and humidity data, wherein... The pad replacement event detection includes: when the change in weight data meets the first abrupt change condition and the change in humidity data meets the second abrupt change condition, it is determined that a pad replacement event has been detected. When a pad replacement event is detected, the weight zero-point correction parameter and the humidity baseline correction parameter are reset; Outputs online calibrated bleeding monitoring results, updated weight zero-point correction parameters, updated humidity baseline correction parameters, updated white balance parameters, and updated absorbent carrier material parameter feature vectors.

10. A method for real-time monitoring of gynecological patients based on edge computing according to claim 6, characterized in that, The evaporation volume increment is constrained by the evaporation prior value calculated by the evaporation prior model, which is determined based on the temperature, relative humidity and / or airflow parameters in the environmental parameter feature vector, and uses the evaporation prior value as the input feature of the physical constraint regression model and / or as the regularization constraint term in the regression loss function.

11. A real-time monitoring system for gynecological patients based on edge computing, used to execute the real-time monitoring method for gynecological patients based on edge computing as described in any one of claims 1 to 9, comprising: The data acquisition module is used to collect weight data, humidity data, and surface color data of the absorbent carrier used to receive vaginal discharge within a preset update cycle, collect environmental parameter data, and read absorbent carrier material parameter data. The feature generation module is used to generate weight feature sequences, humidity feature sequences, color feature sequences, as well as environmental parameter feature vectors and absorber material parameter feature vectors. The quality assessment module is used to calculate quality indicators based on the weight feature sequence, humidity feature sequence and color feature sequence, and generate weight coefficients corresponding to the weight feature sequence, humidity feature sequence and color feature sequence respectively based on the quality indicators; The Transformer encoding module includes a lightweight temporal Transformer network, which is used to construct and encode the temporal input sequence from the weight feature sequence, humidity feature sequence, color feature sequence, environmental parameter feature vector, and absorber material parameter feature vector, and scale the cross-modal attention weights according to the weight coefficients to generate an aligned multimodal temporal representation. The physical constraint regression module includes a physical constraint regression model, which is used to input the aligned multimodal time series representation into the physical constraint regression model, output and update the liquid volume increment, evaporation volume increment and retention volume in the absorbent carrier, wherein the regression process satisfies the mass conservation constraint, the non-negativity constraint and the upper limit constraint of the absorption capacity characterized by the eigenvector of the absorbent carrier material parameters. The bleeding volume calculation module is used to generate a bleeding volume increment based on the increase in the volume of liquid entering the absorbent carrier, and to accumulate the bleeding volume increment to obtain a cumulative bleeding volume, which is the bleeding volume monitoring result; The online calibration module is used to perform online calibration based on the bleeding monitoring results and in combination with the weight data, humidity data and surface color data; An edge computing terminal includes a processor and a memory, wherein the memory stores instructions to cause the processor to perform the functions corresponding to the above-mentioned modules.

Images