A skin state prediction method and system based on time series data

Abstract

The application discloses a skin state prediction method and system based on time series data, and relates to the technical field of skin state monitoring. The method collects image, sensing, environmental exposure and nursing event data according to skin regions, and obtains consistent directional features from a probe or an existing image model. Personal baseline is established according to day cycle and seasonal cycle phase, and skin residual error is obtained. The residual error is divided into bins, change direction and change acceleration to form trajectory pattern code, and statistical feature evolution evidence is obtained. Environmental exposure memory and nursing intervention memory are recursively generated to form unoffset exposure-nursing stress. Delayed response evidence is generated according to the time lag between the peak value of environmental exposure and the peak value of skin residual error. State deviation index is obtained by fusing the above data, similar trajectories in history are searched to predict future skin state, and confidence and supplementary collection prompts are output according to the number of similar windows, collection quality and change dispersion.

Description

Technical Field

[0001] This invention relates to the field of skin condition monitoring technology, and in particular to an electromechanical intelligent linkage control system and method based on edge computing. Background Technology

[0002] With the development of mobile terminal imaging, portable skin analyzers, wearable environmental sensors, and personalized care recording technologies, users can continuously obtain skin condition data. Existing skin condition analysis technologies are mostly based on a single image or a single sensor reading to score the skin condition in real time. Although such solutions are convenient and quick to use, their evaluation results are easily affected by shooting angle, lighting conditions, collection time, seasonal changes, and individual user differences. For the same user, morning and evening oil secretion, seasonal moisture changes, short-term sun exposure reactions, and recent care behaviors will all cause natural fluctuations in skin characteristics. If the skin condition is judged only by a fixed threshold or a single static reference value, it is easy to mistake normal rhythm fluctuations for a worsening condition, and it may also miss delayed changes in condition that are about to occur.

[0003] Existing solutions attempt to input multiple data collections into time-series models for prediction. However, skin condition changes exhibit significant time lags and cumulative characteristics. For example, changes in redness or pigmentation after UV exposure may appear with a lag, and moisture loss due to low humidity environments may accumulate over multiple time windows. The effects of moisturizing, repairing, and sun protection treatments also gradually diminish over time. More importantly, skin condition is not determined solely by a single current feature value but is related to its trajectory over a recent period. The same moisture residual, if in a state of continuous and accelerated decline, carries a significantly different future risk compared to a state just recovering from a low point. Existing models, without distinguishing between trajectory patterns and exposure stress that has been mitigated by treatments versus that that has not, struggle to produce interpretable early predictions.

[0004] Therefore, there is a need for a technical solution that can process continuous time-series data on a skin region basis, establish phase-locked personal baselines, express recent changes in morphology using trajectory morphology codes, express the offsetting relationship between external exposure and nursing intervention using unoffset exposure and nursing stress, and predict future skin conditions using evidence of delayed response. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, this invention provides a skin condition prediction method and system based on time-series data to solve the problems that existing single static skin scores are greatly affected by collection conditions and individual rhythms, ordinary time-series models lack trajectory morphology interpretation, the offsetting relationship between environmental exposure and nursing intervention is unclear, and delayed skin changes are difficult to predict in a timely manner.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A skin condition prediction method based on time-series data includes: By continuously collecting skin images, skin sensor data, environmental exposure data, and nursing event data for skin regions, regional-level time-series skin data is formed. The quality of the acquired images is screened based on their blurriness, pose deviation, and illumination intensity, and daily and seasonal phases are generated based on the acquisition time. Establish a phase-locked personal baseline in historical observations of the same skin region that are of acceptable quality and whose phase distance is no greater than the phase threshold, and convert the current skin features into residuals relative to the phase-locked personal baseline; Trajectory morphology codes are formed based on residual binning, residual change direction, and residual change acceleration, and feature evolution evidence is generated from historical variation windows and stability windows. Environmental exposures such as UV radiation, low humidity, excessive heat, and particulate matter are converted into environmental exposure memories; nursing events such as sun protection, moisturizing, repair, or cleaning are converted into nursing intervention memories; and uncompensated exposure-nursing stress is calculated. Evidence of delayed response is generated based on the response lag between peak environmental exposure and peak skin residual. The Current Skin State Deviation Index was derived by integrating evidence of uncompensated exposure-care stress, characteristic evolution, and delayed response. Based on the current skin condition deviation index, trajectory morphology code, uncompensated exposure-care stress, and phase distance, historical similar trajectory windows are retrieved, and future skin condition is predicted based on future changes of the historical similar trajectory windows; The prediction confidence and supplementary sampling prompts are output based on the number of historical similar trajectory windows, acquisition quality, and future change dispersion.

[0008] Preferably, the phase-locked personal baseline is not a fixed population mean, but a historical valid observation of the same user and the same skin region under similar diurnal and seasonal cyclic phases, used to eliminate skin natural rhythms and individual differences.

[0009] Preferably, the trajectory morphology code is not a single current value, but a combination of residual binning, residual direction and residual acceleration from multiple recent acquisition times, used to distinguish between continuous deterioration, rapid recovery, stable fluctuation and short-term disturbance.

[0010] Preferably, the unoffset exposure and nursing stress are not simply environmental values, but rather truncated results of subtracting the memory of nursing intervention from the memory of the same type of environmental exposure, used to characterize whether external stress has been partially offset by nursing behavior.

[0011] Preferably, the delayed response evidence is not an instantaneous correlation, but is determined based on the response lag between the environmental peak and the skin residual peak, the uncompensated pressure bin, the residual bin, and the residual direction.

[0012] Preferably, the future skin condition prediction is not directly output by a black box model, but is based on the current state, trajectory morphology, uncompensated stress, and time phase to retrieve historical similar trajectory windows, and uses robust statistical values ​​of future changes in historical similar trajectory windows to form the prediction result.

[0013] Secondly, the present invention provides a skin condition prediction system based on time-series data, comprising: The time-series data acquisition module is used to collect skin images, skin sensor data, environmental exposure data, and care event data; The quality phase processing module is used to filter the quality of the acquired data and generate daily and seasonal phases. The phase-locked baseline module is used to establish a phase-locked personal baseline and calculate skin residuals; The trajectory morphology evidence module is used to generate trajectory morphology codes and feature evolution evidence; The exposure-nursing stress module is used to generate environmental exposure memories, nursing intervention memories, and unmitigated exposure-nursing stress. The delayed response evidence module is used to generate delayed response evidence based on the time lag between the peak environmental exposure and the peak skin residual. The similar trajectory prediction module is used to generate the deviation index of the current skin condition and predict the future skin condition; The confidence-based supplementary sampling module is used to calculate the prediction confidence level and output supplementary sampling prompts.

[0014] The beneficial effects of this invention are: 1. This invention eliminates daily cycles, seasonal cycles, and long-term individual differences by phase-locking personal baselines, transforming skin condition prediction from "single static scoring" to "trajectory deviation prediction relative to personal rhythms," thereby reducing misjudgments caused by collection time and individual differences.

[0015] 2. This invention uses trajectory morphology codes to express the changes over a recent period, which can distinguish between continuous variation, deceleration recovery, and stable fluctuations under the same current residual, so that prediction no longer depends solely on the current value; 3. This invention cancels out environmental exposure memories and nursing intervention memories by category to obtain uncancelled exposure-nursing stress, which can express the difference between "exposure has occurred but has been partially offset by nursing care" and "exposure continues to accumulate and nursing care is insufficient"; 4. This invention generates delayed response evidence through response lag, which can identify delayed changes in skin condition such as sunburn redness, dryness due to low humidity, and particulate matter irritation; 5. This invention uses future changes of historical similar trajectory windows for prediction, so that the prediction results can be traced back to historical segments similar to the current state, which is more conducive to interpretation and verification than pure black box time series models.

[0016] 6. This invention sets prediction confidence and supplementary sampling prompts. When there are insufficient similar trajectories, low sampling quality, or excessive dispersion of future changes, it can prompt to re-acquire images, supplement skin sensor readings, or supplement nursing events, thereby improving prediction reliability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0018] Figure 1 This is a flowchart of the overall process for a skin condition prediction method based on time-series data in Example 1; Figure 2 This is an overall flowchart of a skin condition prediction system based on time-series data in Example 1; Figure 3 This is a flowchart of the delayed response evidence in Example 1. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the terms "an embodiment" and "embodiment" as used herein refer to specific features, structures, and characteristics that may be included in at least one implementation of the present invention. The phrase "in an embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single, selective embodiment that excludes other embodiments.

[0022] Example 1, referring to Figure 1 , Figure 2 and Figure 3 This is the first embodiment of the present invention, which provides a skin condition prediction method based on time-series data, comprising the following steps: S1. Collect regional skin time-series data.

[0023] This step takes skin images, skin sensor data, environmental exposure data, care event data, and collection timestamps as input. The target skin region is selected by the user on the acquisition interface or located by an existing DeepLabV3+ image segmentation model. When using this model, the input is a skin image, and the output is a pixel-wise class probability map of the forehead, cheek, nose, jaw, periorbital area, neck, back of hand, and background; the non-background region corresponding to the highest class probability is used as a mask for the skin region.

[0024] For the Each skin area, at any time Observation records were generated: (1) in, This refers to skin images captured by mobile devices or skin detection equipment. This indicates skin sensor data input via Bluetooth, USB, or a device data interface; This represents environmental exposure data input from environmental sensors or weather service interfaces; This indicates nursing event data entered by the user interface or nursing device; This indicates the timestamp of the data acquisition given by the device's system clock. This indicates at least two feature numbers selected from eight features during device initialization and written as a configuration field to each observation record. The system aligns each timestamp to a fixed acquisition grid, and takes the median of multiple readings within the same grid. Moisture and oil are obtained using a moisture and oil probe, elasticity is obtained using a negative pressure adsorption or rebound elastic probe, transdermal moisture loss is obtained using an open-cavity or closed-cavity probe, and skin temperature is obtained using a contact or infrared thermometer.

[0025] Redness, pigment, and texture roughness are obtained from an imaging sensor and an existing ResNet-50 multi-task regression model. This model takes a skin region image segmented by DeepLabV3+ as input, replaces the ResNet-50 end classification layer with three regression output nodes, and fine-tunes it using the labeled redness, pigment, and texture roughness values ​​of the same skin region as supervision labels. After deployment, it outputs the aforementioned three feature values.

[0026] Formula (2) is derived from Formula (1) Read from and with The corresponding image or probe data. letters in Indicates measurement, It is determined by the feature number The operator that uniquely determines the original value is used, rather than the parameters to be trained, when When indicating moisture, oil, elasticity, transepidermal water loss, and skin temperature, Call the aforementioned corresponding probe and output the probe's raw reading; when When indicating redness, pigmentation, or texture roughness, Call the aforementioned ResNet-50 multi-task regression model and output the corresponding regression values.

[0027] letters in Indicates calibration. It is the first The directional calibration operator corresponds to each characteristic. Its source is the equipment's factory calibration table, or a calibration table formed by pairing readings from the same model of equipment with readings from a standard reference instrument. Each calibration table records at least two columns: the original value and the zero-to-one direction consistent value. These are fixed before prediction; during prediction, only table lookup and linear interpolation between adjacent calibration points are performed. The calibration output increases when redness, pigmentation, texture roughness, and transdermal moisture loss increase; the calibration output also increases when moisture and elasticity decrease; and the calibration output increases when oil and skin temperature deviate from the equipment's calibrated suitable range. The output is directly used as Input.

[0028] (2) in, It is the first output of formula (1) Each skin area at any time Observational records, It is a skin feature number; It is for the first during device initialization At least two feature numbers selected for each skin region and their fixed arrangement order; The first one obtained from the observation record Original eigenvalues, Is the original eigenvalue processed by the first The direction consistency value obtained after Zhang's calibration table conversion is, i.e. ; According to The eigenvectors are formed by arranging all selected direction values ​​in a fixed order. Indicates the first Each skin area at any time skin feature vectors, This represents the skin feature vector after orientation unification. Indicates the skin region number, such as cheek, forehead, nose. Indicates the acquisition time, where each component is denoted as... , indicating the first Calibration values ​​for various skin characteristics. For example, selecting three characteristics—moisture, redness, and texture—for the cheek area can be understood as: Each component is calibrated to 0~1, and the larger the value, the worse the skin condition. Therefore, the output of formula (1) first enters the system. , The output then enters , The final output composition There is no unexplained feature extraction black box. The historical components of formula (2) are entered into formula (7) to establish the personal baseline, and the current components are entered into formula (8) to calculate the skin residual.

[0029] S2. Generate time phase and filter acquisition quality.

[0030] This step uses the complete observation record of formula (1) As input, read the acquisition timestamp and skin image, and output the time phase and acquisition quality flag respectively.

[0031] Generate time phase based on the acquisition time: (3) in, Indicates from Read timestamp Deterministic phase transition, superscript Indicates phase, It is a built-in time field reading and trigonometric function conversion operator in the system, with a daily cycle phase of... The seasonal cycle phase is The output of formula (3) is . Indicates the number of hours in a day. Indicates the date sequence number of that year. This represents the total number of days in the year. The output of formula (3) is fed into formula (5) and formula (21).

[0032] from Image blur after reading skin image The blur score is obtained from the normalized blur score output by the existing Laplacian variance sharpness algorithm. The smaller the Laplacian variance, the larger the blur score; shooting posture deviation. Keypoints were obtained by aligning the keypoints output from the existing HRNet keypoint model with standard camera keypoints; illumination intensity Obtained from the device's light sensor field; if no light sensor is configured, the average brightness of the target skin area is taken.

[0033] (4) in, , and These represent the ambiguity threshold, attitude deviation threshold, and illumination threshold, respectively. These three are not arbitrarily input on-site, but rather selected and fixed before deployment using pre-labeled samples of the same model of equipment (those deemed acceptable or unacceptable for data acquisition) based on the receiver's operational characteristic curve. The multiplication of the three indicator functions indicates that all three conditions must be met simultaneously. If any condition is not met The output of formula (4) is fed into formula (6), the effective evaluation starting point set, the effective residual peak set, formula (21), and formula (23).

[0034] S3. Establish phase-locked personal baselines and generate residuals.

[0035] This step takes the direction consistency feature of formula (2), the time phase of formula (3), and the acquisition quality indicator of formula (4) as inputs, and outputs the individual baseline and residual.

[0036] Calculate the phase distance between the historical acquisition time and the current acquisition time: (5) (6) in, Representing the historical phase of formula (3) and current phase Directly calculated phase distance This represents the phase threshold. Since the diurnal and seasonal phases are located on the unit circle, ; The candidate values ​​are selected from this range through rolling validation using only data prior to the time to be predicted, and then fixed before the formal prediction. It only includes historical observations from the same user, the same skin region, earlier than the current time, of acceptable quality, and with a phase distance not exceeding the threshold. The output of formula (5) The output of formula (4) Both are entered into formula (6), and the output of formula (6) is entered into formula (7).

[0037] Calculate the phase-locked individual baseline in the historical phase neighborhood: (7) in, This indicates that the phase is locked to the individual baseline, and the system is in Formal prediction begins when the result is not empty; when the result is empty, data collection continues but no formal prediction results are generated, thus eliminating the need for intermediate parameters such as the population default baseline, minimum scale, and sample sufficiency. The output of formula (7) is fed into formula (8).

[0038] Transform the current directional consistency feature into a skin damage residual relative to the individual baseline: (8) in, This represents the residual of skin damage. The characteristics of formula (2) and the median baseline of formula (7) are both between zero and one, therefore... The unfavorable directions of each feature have been unified, and the difference can be directly compared without introducing the fluctuation scale and the zero constant. The output of formula (8) is fed into formula (9), formula (10), formula (11), the effective residual peak set and formula (17).

[0039] S4. Generate trajectory morphology codes and evidence of feature evolution.

[0040] This step only retains two necessary processing steps: trajectory formation and historical frequency difference. First-order difference is defined. Define the second-order difference and formed the latest Trajectory morphology code at each moment: (9) In formula (9), the three elements are, in order, the residual binning, the direction of residual change, and the direction of residual change acceleration. The upper and lower third boundaries are derived from the historical calibration set before prediction; The zero-value tolerance originates from the device resolution of the corresponding probe or the minimum resolvable variation in the image model validation set; trajectory length It is a positive integer, selected and fixed by rolling verification using only data prior to the time to be predicted. Indicates an increase, Indicates a decrease or recovery; and Indicates an accelerated rise; and This indicates a decrease in recovery speed; two differences being zero within the stated resolution and the binning remaining unchanged indicates stable fluctuations. This is due to the first term of the trajectory. The second-order difference also needs to be read and Only when to common Formula (9) is generated when all continuous time grids have obtained residuals of acceptable quality; otherwise, let And continue collecting. Only when all All meet Only then can a regional trajectory morphology code be formed. And perform formal prediction, the output of formula (9) enters the following set of valid evaluation starting points, formula (12) and formula (20).

[0041] Write down the historical moments in which the trajectory code has been generated by formula (9), the evaluation results have been observed, and both the starting point and the evaluation endpoint are of acceptable quality as the set of valid evaluation starting points. Formula (9) And formula (4) This becomes the direct input for formulas (10) and (11). It forms the historical state variation window and the historical state stability window, respectively: (10) (11) in, The evaluation span is derived from the candidate time span preset by the prediction task; Indicates the first Variation threshold of the feature Indicates the first The stability thresholds for these two features are derived from the historical residual variation distribution. , and All were selected through rolling validation using only data prior to the time to be predicted, and met the following requirements. This is to avoid the same historical change falling into both the variation window and the stability window simultaneously. The effective evaluation starting point set includes... Ensure that only results that have occurred up to the current moment are used, and do not read future data; This indicates that both the historical starting point and the evaluation endpoint are selected through quality screening using formula (4). The outputs of formulas (10) and (11) are simultaneously fed into formulas (12) and (18).

[0042] To ensure a direct connection between formulas (9) and (12), the trajectory occurrence rate is defined. The occurrence rate compares the current trajectory morphology code in formula (9) with the trajectory morphology codes at each moment within a given historical window, without designing additional variables. Evidence for feature evolution is as follows: (12) Formula (12) outputs in the range of -1 to 1; the larger the value, the higher the proportion of the current trajectory appearing in the historical variation window. This output is obtained through step S7. Enter formula (19).

[0043] S5. Generate environmental exposure memories, nursing intervention memories, and uncompensated exposure-nursing stress.

[0044] This step still follows formula (1). For data source. From Read environment fields The device SDK converts the sensor's calibrated range to the first... Exposure intensity of zero to one, denoted as The categories include ultraviolet radiation, low humidity, overheating, and particulate matter. From the same... Read nursing field When there is no matching nursing event, the nursing intensity is taken as zero; when there is a matching event, the ratio of the actual dosage to the product's recommended dosage is taken and truncated to zero or one, denoted as . Sunscreen, moisturizing, cooling and soothing, and cleansing and repair should be matched in that order with the four types of exposure mentioned above.

[0045] First-order recursive memory is used instead of repeated time window summation. Environmental exposure memory is: (13) No. The memory of nursing interventions is as follows: (14) in, and They represent the first The retention coefficients for individual time grids of both environmental exposures and nursing interventions are greater than or equal to zero and less than one. These coefficients are derived from the conversion from a preset half-life to a fixed acquisition grid: if the fixed grid duration is... The environmental and nursing half-lives are respectively and ,but , The candidate half-life value is selected based on the prediction error using a historical validation window and fixed before deployment. The initial memory satisfies... .because and The recursive result always satisfies and The newer the current input, the higher the weight; the influence of history gradually decreases with recursion. The outputs of formulas (13) and (14) are combined into formula (15).

[0046] Subtracting nursing memories from similar environmental memories yields the uncompensated exposure-nursing stress: (15) in, If it is in the range of 0 to 1, it means that the first... External pressures that have not yet been offset by care; its compartments are composed of uniform... Obtain. Formula (15) enters formula (17), formula (19) and formula (20).

[0047] S6. Generate evidence of delayed response.

[0048] Please see Figure 3 In this step, Indicates the skin region number. Indicates the skin feature number. Indicates the environmental exposure category number, Indicates the current evaluation time grid. This represents the candidate time grid that is traversed sequentially within the historical observation window. express The previous time grid. During the search process. , , and Keep it unchanged, only make Changes within the window.

[0049] No. Environmental exposure intensity is denoted as Its source is the historical observation records of formula (1). The environment field is converted to 0~1 according to the device calibration range in step S5. Indicates the intensity of environmental exposure; These represent the time grid and exposure category, respectively. Environmental exposure is shared data for the same user at the same time, therefore... No longer setting skin area subscripts and feature subscript Formula (8) Indicates the first The first region Features in time grid The skin damage residual; Equation (4) This indicates that the area is in A quality indicator of the data acquisition at any given moment.

[0050] symbol and Representing time grids respectively and The corresponding actual timestamp, This indicates the actual elapsed time between the current time and the previous historical candidate time. Indicates the maximum response observation time. This only indicates that the parameter is used for time delay calculations, not for exponentiation. Observation window Indicates going back from the current moment. All candidate time grids within the specified duration. Represents a collection of windows, with subscripts. The window indicates the current time as the endpoint; curly braces represent a set; vertical bars indicate... This indicates that the subsequent conditions are "met". It indicates less than or equal to.

[0051] Peak environmental exposure time is .in, Indicates the first The time grid where the class is exposed to the maximum intensity in the current observation window. This is the English term for environmental exposure, and does not indicate exponentiation. Return to Time grid for obtaining the maximum value Instead of returning the exposure intensity itself, it returns the value. When multiple identical maximum values ​​occur, Return the earliest time grid.

[0052] The effective residual peak candidate set is Superscript This indicates that the set is used to select the peak value of the skin residual, and does not represent the exponentiation operation; This indicates that only the peak value of skin residuals is allowed to be selected at and after the peak of environmental exposure. Because and Both sides take only zero or one, and the product of the two sides of the multiplication sign equals one, indicating that... and The acquisition quality of both adjacent time grids is acceptable.

[0053] exist At that time, the peak time of skin residual was Indicates the first The first region The first feature is for the first The residual peak time grid corresponding to the class exposure. The residual change at this moment is Since the candidate set requires that the quality of two adjacent time grids be qualified, the difference can be calculated directly. Represents the empty set; if This indicates that there is no residual peak value within the current window that simultaneously meets the time sequence and acquisition quality conditions. In this case, let... . Indicates a delayed status code. It indicates a delayed category label, not an exponentiation operation.

[0054] The response time lag is the actual timestamp difference between the two peaks: (16) In formula (16), Indicates the current time grid within, no. Environmental exposure to type 1 The first skin area The response time delay corresponding to each characteristic. Represents the peak time grid of skin residuals The corresponding actual timestamp; Representing the peak time grid of environmental exposure The corresponding actual timestamp. Subtracting the two gives the actual elapsed time, with the unit being the same as the unit of the timestamp difference, such as hours or days.

[0055] Maximum response observation time This represents the longest actual time allowed for pairing an environmental peak with its subsequent skin residual peak; its candidate values ​​are derived from a preset response duration range for the environmental category to be evaluated and selected through rolling validation using only data prior to the time to be predicted. Because... Both peak times came from Therefore, formula (16) satisfies . Indicates the response time delay binning operator, subscript This indicates that its processing object is response time delay; the operator uses the lower and upper thirds of the previous historical time delay calibration set as boundaries, and... The mapping is to the first box, the second box, or the third box. The output of formula (16) Proceed directly to formula (17).

[0056] The uncompensated pressure at the peak of environmental exposure is denoted as The peak value of the skin residual is denoted as The direction of residual change at the peak time is denoted as... Write these four types of information into the delayed status code: (17) Formula (17) is only used in The calculation is performed on time; the four elements represent the response time delay bin, the uncompensated pressure bin at the peak of environmental exposure, the residual bin at the peak of skin residual, and the direction of residual change at that peak, respectively, with the element order being fixed. The output of formula (17) is fed into formula (18).

[0057] Define the occurrence rate of the delayed status code when the current delayed status code is not empty. The occurrence rate is directly compared with the current delayed status code in formula (17) and the delayed status code in the historical window; the empty code at a historical moment does not meet the equality condition, and no intermediate counting symbol in formula (12) is used. The result is: (18) Formula (18) outputs in the range of negative one to one. The larger the value, the more common the current time delay combination is in the historical variation window. When the delay status code of formula (17) is not generated, the corresponding... This indicates that there is currently no available evidence of delay, and the output of formula (18) is obtained through step S7. Enter formula (19) and no longer repeatedly construct the smooth probability and the total number of combinations.

[0058] S7. Calculate the current skin condition deviation index.

[0059] This step only aggregates the data already output in the previous step. The fixed-order set of environment categories is denoted as... And define the mean of uncompensated pressure. Mean of evidence for feature evolution and the mean of evidence for delayed response The three equations respectively follow formulas (15), (12), and (18), and the negative evidence, after being truncated, is all in the range of zero to one; the average value is not a parameter to be trained. The current skin state deviation index is: (19) Formula (19) adopts an equal-weighted average, no longer introducing three fusion weights and two evidence normalization scales, and is derived from... ensure . The larger the value, the greater the overall deviation of the current state from the individual baseline; its state binning is composed of... Obtain. Formula (19) is used to calculate the future variation dispersion before entering formulas (20), (22), and (23).

[0060] S8. Predict future skin condition based on historical similarity trajectory windows.

[0061] This step compresses the current state, trajectory, and pressure into a short signature used only for retrieval: (20) in, The fixed-order set of environment categories defined in step S7 is used. The three parts of formula (20) correspond to the current skin state deviation index bin, trajectory morphology code and various uncompensated pressure bins in turn, without adding any new continuous parameters; the output of formula (20) is entered into formula (21).

[0062] The historical similar trajectory window is: (twenty one) in, Represents the number of positive integer time grids specified for the prediction task and satisfies ;therefore At the same time ensure Furthermore, historical results have been observed. Two quality conditions ensure that the historical starting point and the result time are both valid; equal signatures ensure that the state bin, trajectory morphology code, and pressure bin are the same; the last condition ensures that the phase distance is not greater than the phase threshold. Formula (21) therefore covers the retrieval conditions item by item, and its output enters formula (22) and formula (23).

[0063] when When not empty, predict the future state using the median state change of each historical segment within the prediction span: (twenty two) when When empty, the system command In step S9, the prediction confidence is set to zero to avoid calculating the median of an empty set. The output of formula (22) is fed into formula (24).

[0064] S9. Calculate the prediction confidence level and output a supplementary sampling prompt.

[0065] This step retains only the three confidence sources described in the claims. The number of historical similar windows is... The acquisition quality is as shown in formula (4). The future variation dispersion is directly obtained from the historical state deviation index of formula (19) and the historical similar trajectory window of formula (21), denoted as The dispersion is zero when the set has only one element, and the prediction confidence is zero when the set is empty.

[0066] When the historical similarity window is not empty, the prediction confidence is: (twenty three) in, The number of similar windows with a value greater than zero is represented by the candidate value derived from the historical distribution of similar window numbers. This value is selected based on the validation window prediction error and the supplementary sampling rate, and is fixed before formal prediction. All three multiplicative terms in formula (23) are between zero and one. The three multiplication terms correspond to acquisition quality, sufficiency of similar window quantity, and consistency of future changes, respectively. The average quality window, baseline sufficiency, exponential function, or dispersion normalization scale are no longer used. The output of formula (23) is entered into formula (24).

[0067] The supplementary sampling prompt is as follows: (twenty four) in, This represents the minimum confidence threshold. Candidate values ​​are derived from the confidence distribution of the validation set and selected based on both the supplementary sampling rate and the error prediction rate. The threshold for indicating future states is represented by candidate values ​​derived from historical states that have been manually confirmed to require prompting, deviating from an exponential distribution and selected based on prompt sensitivity. Both are between zero and one and are fixed before deployment. The multiplication of the two indicator functions accurately expresses the "and" in step S9 of claim 1, and makes... A supplementary sampling prompt is only output when the confidence level is insufficient and the predicted value reaches the prompt threshold. At that time, if If the number of similar windows is insufficient or the variation is too large, the system will prompt you to retake the photo; if the number of similar windows is insufficient or the variation is too large, the system will prompt you to supplement the selected skin features, environmental exposures or care event records.

[0068] Please see Figure 2 The present invention also provides a skin condition prediction system based on time-series data, including a time-series data acquisition module, a quality phase processing module, a phase-locked baseline module, a trajectory morphology evidence module, an exposure care stress module, a delayed response evidence module, a similar trajectory prediction module, and a confidence re-acquisition module.

[0069] The time-series data acquisition module is used to acquire skin images, skin sensor data, environmental exposure data, nursing event data, and acquisition timestamps, and executes formulas (1) and (2). The quality phase processing module is used to execute formulas (3) and (4) to generate time phase and acquisition quality indicators. The phase-locked baseline module is used to execute formulas (5) to (8) to output historical phase neighborhood, phase-locked personal baseline, and skin damage residuals.

[0070] The trajectory morphology evidence module executes formulas (9) to (12) and outputs trajectory morphology codes, historical state variation windows, historical state stability windows, and feature evolution evidence. The exposure to nursing stress module executes formulas (13) to (15) and outputs environmental exposure memories, nursing intervention memories, and unmitigated exposure-nursing stress. The delayed response evidence module executes formulas (16) to (18) and outputs response lag, delayed status codes, and delayed response evidence. The similar trajectory prediction module executes formulas (19) to (22) and outputs the current skin state deviation index, historical similar trajectory windows, and future skin state prediction values. The confidence supplementation module executes formulas (23) and (24) and outputs prediction confidence and supplementation prompts.

[0071] Glossary Regional-level skin time-series data refers to the observation records of formula (1). Selected feature set As a configuration field for this record, it limits the first... At least two skin features were actually collected from each skin region. The current skin feature vector refers to the complete observation record of formula (2). The ordered vector obtained from the input Among them, the first Each component is denoted as An increase in the numerical value generally indicates that the feature is changing in an unfavorable direction.

[0072] The time phase refers to the four-dimensional vector in formula (3). The quality indicator is the one defined in formula (4). The historical phase neighborhood is the one defined by formula (6). The phase-locked personal baseline is the formula (7). The skin damage residual is the result of formula (8). .

[0073] Trajectory morphology code refers to the formula (9) that combines the residual binning, residual change direction, and residual change acceleration direction of the most recent multiple time points in chronological order. The historical skin condition deterioration window and the historical skin condition stability window are respectively the values ​​of formula (10) and formula (11). and The evidence for feature evolution is formula (12). .

[0074] Environmental exposure memory and nursing intervention memory are respectively the values ​​of formula (13) and formula (14). and Uncompensated exposure-nursing stress is the result of formula (15). The response delay is the time delay of formula (16). The delayed status code is the one in formula (17). The evidence for the delayed response is the formula (18). .

[0075] The current skin condition deviation index is the formula (19). The state-trajectory-pressure signature is that of formula (20). The historical similarity trajectory window is the formula (21). The predicted value of future skin condition is given by formula (22). The prediction confidence level is the value of formula (23). The supplementary sampling prompt is formula (24). .

[0076] This embodiment also provides a skin condition prediction system based on time-series data, including: The time-series data acquisition module is used to collect skin images, skin sensor data, environmental exposure data, and care event data; The quality phase processing module is used to filter the quality of the acquired data and generate daily and seasonal phases. The phase-locked baseline module is used to establish a phase-locked personal baseline and calculate skin residuals; The trajectory morphology evidence module is used to generate trajectory morphology codes and feature evolution evidence; The exposure-nursing stress module is used to generate environmental exposure memories, nursing intervention memories, and unmitigated exposure-nursing stress. The delayed response evidence module is used to generate delayed response evidence based on the time lag between the peak environmental exposure and the peak skin residual. The similar trajectory prediction module is used to generate the deviation index of the current skin condition and predict the future skin condition; The confidence-based supplementary sampling module is used to calculate the prediction confidence level and output supplementary sampling prompts.

[0077] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications and equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A skin condition prediction method based on time-series data, characterized in that, Includes the following steps: S1. Continuously acquire skin images, skin sensor data, environmental exposure data, and nursing event data according to skin regions to form regional skin time-series data; S2. The regional skin time-series data is subjected to quality screening and time phase generation, wherein the time phase includes daily cycle phase and seasonal cycle phase; S3. Establish a phase-locked personal baseline in the same skin region, in historical observations with qualified acquisition quality and a phase distance from the current acquisition time not greater than the phase threshold, and convert the current skin features into residuals relative to the phase-locked personal baseline. S4. Construct a trajectory morphology code based on the binning, direction of change, and acceleration of change of the residual, and generate feature evolution evidence based on the historical skin condition deterioration window and the historical skin condition stability window. S5. Calculate environmental exposure memory based on environmental exposure data, calculate nursing intervention memory based on nursing event data, and form uncompensated exposure-nursing stress from the environmental exposure memory and nursing intervention memory; S6. Generate evidence of delayed response based on the response lag between the peak environmental exposure and the peak skin residual, the uncompensated exposure-care stress binning, the residual binning, and the direction of residual change. S7. By integrating the uncompensated exposure-care stress, evidence of characteristic evolution, and evidence of delayed response, the current skin condition deviation index is obtained; S8. Based on the current skin condition deviation index, trajectory morphology code, uncompensated exposure-care stress, and phase distance, retrieve historical similar trajectory windows, and predict future skin condition based on future changes of historical similar trajectory windows within the prediction span; S9. Calculate the prediction confidence based on the number of historical similar trajectory windows, acquisition quality, and future change dispersion; when the prediction confidence is lower than the confidence threshold and the future skin condition prediction value reaches the prompt threshold, output a supplementary acquisition prompt.

2. The skin condition prediction method based on time-series data according to claim 1, characterized in that: In step S1, the skin area includes at least one of the following: forehead, cheek, nose, chin, periorbital area, neck, back of hand, or a user-specified skin area. The skin sensing data includes at least two of the following: moisture, oil, redness, pigmentation, texture roughness, elasticity, transepidermal water loss, and skin temperature.

3. The skin condition prediction method based on time-series data as described in claim 1, characterized in that: In S2, the acquisition quality screening is determined based on image blur, shooting posture deviation, and light intensity. An observation is considered valid only if the image blur is not greater than the blur threshold, the shooting posture deviation is not greater than the posture deviation threshold, and the illumination intensity is not less than the illumination threshold.

4. The skin condition prediction method based on time-series data as described in claim 1, characterized in that: In step S3, the phase-locked personal baseline is calculated from historical observations of the same user, the same skin region, with qualified acquisition quality and a phase distance not greater than the phase threshold; The phase distance is determined by the vector difference between the historically observed diurnal and seasonal phases and the currently observed diurnal and seasonal phases.

5. The skin condition prediction method based on time-series data as described in claim 1, characterized in that: In step S4, the trajectory morphology code includes residual binning, residual change direction, and residual change acceleration at multiple recent acquisition times, which is used to characterize the trajectory morphology of skin features as they are rising, falling, accelerating upward, decelerating and recovering, and stabilizing.

6. The skin condition prediction method based on time-series data as described in claim 1, characterized in that: In step S5, the uncountered exposure-care stress is obtained by subtracting similar nursing intervention memories from environmental exposure memories and then truncating it, and is used to characterize external stress that has not yet been countered by nursing events.

7. The skin condition prediction method based on time-series data as described in claim 1, characterized in that: In step S6, the evidence of delayed response is determined by the number of occurrences of the same response time delay bin, the same uncompensated pressure bin, the same residual bin, and the same residual change direction in the historical state variation window and the historical state stability window.

8. The skin condition prediction method based on time-series data as described in claim 1, characterized in that: In step S8, the historical similar trajectory window simultaneously satisfies the following conditions: the current skin state deviation index bin is the same, the trajectory morphology code is the same, the uncompensated exposure-care stress bin is the same, and the phase distance is not greater than the phase threshold.

9. A skin condition prediction system based on time-series data, based on the skin condition prediction method based on time-series data according to any one of claims 1 to 8, characterized in that: The time-series data acquisition module is used to collect skin images, skin sensor data, environmental exposure data, and care event data; The quality phase processing module is used to filter the quality of the acquired data and generate daily and seasonal phases. The phase-locked baseline module is used to establish a phase-locked personal baseline and calculate skin residuals; The trajectory morphology evidence module is used to generate trajectory morphology codes and feature evolution evidence; The exposure-nursing stress module is used to generate environmental exposure memories, nursing intervention memories, and unmitigated exposure-nursing stress. The delayed response evidence module is used to generate delayed response evidence based on the time lag between the peak environmental exposure and the peak skin residual. The similar trajectory prediction module is used to generate the deviation index of the current skin condition and predict the future skin condition; The confidence-based supplementary sampling module is used to calculate the prediction confidence level and output supplementary sampling prompts.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the skin condition prediction method based on time-series data as described in any one of claims 1 to 8.

Images