A personalized skincare product dynamic formulation system and method based on real-time monitoring of the skin microbiome

Abstract

This application relates to a personalized skincare product dynamic formulation system and method based on real-time monitoring of the skin microbiome, belonging to the fields of intelligent skincare and biosensing technology. The personalized skincare product dynamic formulation system includes: a microbiome detection module, an edge computing processing module, a state calibration mapping module, and a cloud-based intelligent decision-making module, used to receive microbiome state estimates and generate an initial skincare formula by combining environmental factors and user profile data; a dynamic formula calculation module, a micro-precision formulation module, and a user terminal interaction module; by combining an impedance spectroscopy sensing array with a calibration mapping model, the skin microbiome state can be rapidly estimated in a home setting, significantly shortening the detection cycle compared to traditional gene sequencing methods, establishing a systematic decision pipeline from sensor data to skincare formula, which helps reduce subjectivity and inconsistency in human interpretation; and multi-layered ingredient safety constraints help improve the safety of the formula.

Description

Technical Field

[0001] This application relates to the fields of intelligent skincare and biosensing technology, and in particular to a personalized skincare product dynamic formulation system and method based on real-time monitoring of the skin microbiome. Background Technology

[0002] The surface of human skin is home to a rich and diverse microbial community, including bacteria, fungi, viruses, and archaea, collectively known as the skin microbiome. Studies have shown that the skin microbiome plays a crucial role in maintaining skin barrier function, immune regulation, and combating pathogen invasion. Imbalances in the skin microbiome are associated with various skin problems, including acne, atopic dermatitis, rosacea, psoriasis, and seborrheic dermatitis.

[0003] Current skincare product formulation design is primarily based on coarse-grained classification of skin type, supplemented by demographic factors such as age and gender for product recommendations. Most personalized skincare customization services on the market rely on static formula recommendations based on user-completed questionnaires or one-time skin test results, making it difficult to achieve continuous dynamic optimization and closed-loop feedback adjustment of the formula.

[0004] Existing patents disclose an AI-screened exosome composition comprising the following components: exosomes of a specific type, proportion, and concentration screened by AI; a pharmaceutically or cosmetically acceptable carrier; and an exosome stabilizer. This invention, through multimodal fusion, links the type, proportion, and concentration of exosomes to the molecular mechanisms of skin problems, user lifestyle habits, and environmental factors, eliminating blind trials and potential risks. Personalized solutions combined with exosome stabilizers ensure maximum exosome efficacy, improve user experience and satisfaction, guarantee the safety of the skincare composition, and extend product shelf life. The established feedback optimization loop allows the skincare plan to evolve with skin condition, and combined with skin condition change reports, enables dynamic adjustment and quantitative management of the skincare process, ensuring long-term, sustainable skin-beautifying effects.

[0005] The existing technical solutions mentioned above have the following drawbacks: 1. Existing skincare solutions do not fully consider the important biological factor of the skin microbiome. The composition of the skin microbiome varies significantly among different individuals, and even for the same individual, the composition of the microbiome changes dynamically at different times, in different locations, and under different environmental conditions; 2. Existing skin microbiome testing methods mainly rely on swab sampling and subsequent gene sequencing. These methods require specialized laboratory equipment, and the time from sampling to obtaining results is usually several days to several weeks, making it difficult to support the dynamic monitoring needs in real-time, home settings. Even if microbiome analysis results are obtained, there is currently a lack of systematic and intelligent decision-making methods to automatically transform microbiome data into skincare formulations. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this application is to provide a personalized skincare product dynamic formulation system and method based on real-time monitoring of the skin microbiome. By combining an impedance spectroscopy sensing array with a calibration mapping model, the state of the skin microbiome can be rapidly estimated in a home setting. Compared with traditional gene sequencing methods, this significantly shortens the detection cycle and establishes a systematic decision pipeline from sensor data to skincare formulation, which helps to reduce subjectivity and inconsistency in human interpretation.

[0007] This was achieved using the following technical solutions: In a first aspect, this application provides a personalized skincare product dynamic formulation system based on real-time monitoring of the skin microbiome, comprising: The microbiome detection module is used to monitor the user's skin surface and collect electrochemical impedance spectroscopy parameters and auxiliary physicochemical parameters; The edge computing processing module is used to clean and feature electrochemical impedance spectroscopy parameters and auxiliary physicochemical parameters, and to generate and transmit impedance spectroscopy feature vectors. The state calibration and mapping module is used to register and transform the impedance spectrum feature vector according to the impedance spectrum-microbiome state mapping mechanism to obtain the estimated value of the microbiome state. The cloud-based intelligent decision-making module receives microbiome state estimates and combines them with environmental factors and user profile data to generate an initial skincare formula. The dynamic formula calculation module is used to examine and optimize the component ratio parameters of the initial skin care formula based on the ingredient safety constraint mechanism, and generate a personalized skin care formula. The micro-precision formulation module is used to extract and mix the corresponding active ingredient raw materials according to the personalized skin care formula to form personalized skin care products. The user terminal interaction module is used by users to control the type and proportion of active ingredient raw materials, store personalized skin care formulas, and generate alarm notifications during the formulation process of personalized skin care products.

[0008] By adopting the above technical solution, skin electrochemical impedance spectroscopy and physicochemical parameter acquisition are carried out, feature vectors are extracted through edge computing, skin condition is estimated using an impedance spectroscopy-microbiome mapping model, and combined with cloud environment and user profile data, a constrained optimization algorithm is used to generate a personalized skin care formula with safe ingredients and automatically adjust active raw materials, thus realizing a precise, real-time and safe personalized skin care closed loop.

[0009] Furthermore, the microbiome detection module is in the form of a flexible patch, including: A microfluidic sampling layer is used to automatically collect surface sweat and sebum secretions from the user's skin using capillary force through an embedded array of microchannels. Impedance spectroscopy sensing array is used to detect the electrochemical properties of surface sweat and sebum secretions by scanning different frequency bands, and obtain electrochemical impedance signals; The signal conditioning and conversion layer is used for synchronous acquisition and digitization of electrochemical impedance signals to obtain electrochemical impedance spectroscopy parameters and auxiliary physicochemical parameters.

[0010] By adopting the above technical solution, skin secretions are automatically collected through microfluidic capillary force. Combined with multi-band impedance spectroscopy sensing and synchronous digital conditioning technology, non-invasive, real-time, and multi-channel accurate detection of electrochemical parameters is achieved, improving the convenience and data reliability of skin microbiome monitoring.

[0011] Furthermore, the edge computing processing module includes: The filtering and noise reduction layer is used to filter and reduce noise in the electrochemical impedance spectroscopy parameters and auxiliary physicochemical parameters to obtain noise-free impedance spectroscopy parameters and noise-free physicochemical parameters. The parameter correction layer is used to perform baseline correction and correlation of the noiseless impedance spectrum parameters and noiseless physicochemical parameters based on the acquisition timestamp, so as to obtain the impedance spectrum physicochemical parameters. The feature extraction layer is used to extract features from the physicochemical parameters of the impedance spectrum to obtain the impedance spectrum feature vector; The wireless communication layer is used to monitor and transmit impedance spectrum eigenvectors to the cloud.

[0012] By adopting the above technical solution, data quality is improved through filtering and noise reduction and baseline correction. Impedance spectrum feature vectors are generated using feature extraction algorithms and wirelessly transmitted to the cloud, which significantly reduces the amount of data transmission and ensures real-time performance.

[0013] Furthermore, the impedance spectroscopy-microbiome state mapping mechanism includes: Impedance spectroscopy sensing and swab sampling were performed on the target skin area of ​​the subjects to obtain impedance spectral feature vectors and microbial group sequences; Based on the collection timestamp, impedance spectrum feature vectors and microbial group sequences are paired to construct a training dataset and an independent validation set; Based on the pre-defined improved MobileNet network, the training dataset is iterated several times to generate an initial state mapping model and determine the microbiome state indicators. The initial state mapping model was evaluated and calibrated based on the independent validation set. The microbiome state values ​​were calculated, and the model accuracy was verified by combining the microbiome state indicators. Based on specific paired data and a paired sampling calibration mechanism, the initial state mapping model is corrected to obtain a customized state mapping model, which outputs the target microbial group sequence.

[0014] By adopting the above technical solution, the impedance spectrum features and microbial sequences are paired and learned based on the improved MobileNet network. Through independent validation set evaluation and calibration and special paired data correction, a high-precision personalized mapping from electrochemical signals to microbiome states is achieved, which significantly improves the accuracy of non-invasive detection of skin microbiome and the model's generalization ability.

[0015] Furthermore, based on a pre-defined improved MobileNet network, the training dataset is iteratively learned several times to generate an initial state mapping model and determine microbiome state indicators, including: Microbial group sequences in the training dataset are used as target labels, and impedance spectrum feature vectors in the training dataset are labeled as training data. Based on the impedance spectrum parameters in the training data of the scanning frequency band, resampling and linear interpolation are performed to calculate the frequency band impedance magnitude and frequency band phase angle. The auxiliary parameters in the training data are globally scalarized and then concatenated with the frequency band impedance magnitude and frequency band phase angle to construct the feature vector to be trained. Based on the data source, frequency drift simulation and noise generalization are performed on the feature vectors to be trained to obtain enhanced feature vectors; Based on the one-dimensional convolutional input layer in the improved MobileNet network and the batch data volume, the dimension of the enhanced feature vector is unified to obtain a feature vector of the same dimension. Based on the preset learning rate decay function and the iterative training layer, the same-dimensional feature vector is iterated several times to obtain the impedance spectrum feature weights. Based on the frequency band attention mechanism and target label, the impedance spectrum feature weights are temporally correlated and frequency band weighted to obtain the microbial mapping weight matrix. Based on the microbial mapping weight matrix and the parallel output layer, the enhanced feature vector is identified and predicted, and the metabolic activity value and the number of bacterial genera are output. Clustering and normalization of microbial taxonomic sequences, and calculation of microbiome state indices; The metabolic activity values ​​and the number of bacterial genera were compared based on the microbiome status indicators; If both are within the corresponding target tolerance range, the current microbial mapping weight matrix is ​​deemed compliant and used as the initial state mapping model.

[0016] By adopting the above technical solution, based on the improved MobileNet and frequency band attention mechanism, and through data augmentation, one-dimensional convolutional input and parallel output layers, the impedance spectrum is accurately mapped to microbial metabolic activity and genus type, which significantly improves the model's lightweightness and recognition accuracy.

[0017] Furthermore, the cloud-based intelligent decision-making module includes: The data input device is used to receive multimodal data including impedance spectrum eigenvectors, auxiliary physicochemical parameters, microbiome state estimates, historical microbiome state sequences, environmental factors, allergen data, and user profile data. The feature encoder is used to perform feature mapping on the microbiome state estimate based on the microbiome embedding layer of the preset multimodal attention fusion network, capture the temporal state sequence of the historical microbiome state by combining the temporal feature extraction layer, and cross-fuse multimodal data by combining the cross-modal fusion layer to obtain multidimensional feature encoding. The graph inference engine is used to infer the multidimensional feature codes based on the knowledge graph formed by the interaction between microorganisms and active ingredients, and obtain the initial skin care formula.

[0018] By adopting the above technical solution, based on multimodal attention fusion network and temporal feature extraction, cross-integrating microbiome, environmental and user data, and combining microbial-active ingredient knowledge graph for reasoning, the intelligent generation of personalized skin care formulas is realized, which significantly improves the scientific nature and accurate adaptation of the formulas.

[0019] Furthermore, the dynamic formula calculation module includes: The formula optimizer is used to modify the microbiome diversity recovery index, target microbial abundance adjustment precision, and total number of ingredients among the components in the initial skin care formula according to a preset multi-objective optimization mechanism, thereby generating a modified skin care formula. Safety calibrators are used to screen modified skincare formulas for allergens, verify concentration safety thresholds, and validate the stability of multiple coexisting ingredients, resulting in safe skincare formulas. The formula output device is used to decompose safe skin care formulas according to the skin care product formulation process, output formula component vectors, and generate a series of formulation instructions. The feedback learner is used to incrementally update the initial state mapping model based on changes in the skin microbiome state before and after skincare and the user's subjective satisfaction rating, and to optimize safe skincare formulas to form personalized skincare formulas.

[0020] By adopting the above technical solutions, a multi-objective optimization algorithm (such as NSGA-II) is used to balance microbial diversity, microbial regulation precision and total number of ingredients. A safe formula is generated through safety constraint verification (allergen screening, concentration threshold, coexistence stability). The mapping model is incrementally updated by using changes in the user's skin condition before and after skin care and satisfaction scores. This achieves a dynamic self-optimization closed loop for the formula, which significantly improves the safety and accuracy of personalized skin care.

[0021] Furthermore, the micro-precision dispensing module includes: The raw material storage area is used to store several active ingredient raw materials, identify the type and expiration date of the raw materials, and load gel matrix or emulsion matrix. Micropump arrays are used to extract active ingredient raw materials by calling up a corresponding number of micropump channels according to the ingredient ratio of personalized skin care formulas, so as to obtain a quantitative amount of ingredient raw materials. The control and detection unit is used to decode the dispensing instruction sequence, generate channel timing drive waveforms, control and monitor the actual output of the micro-pump channel, and verify the active ingredient raw materials. An online mixing chamber is used to mix active ingredient raw materials to obtain an initial skincare product; The online quality inspection unit is used to test the viscosity, pH and uniformity of the initial skin care product and make a judgment based on the preset target threshold range; If all three are within the corresponding target threshold range, then the current initial skincare product is determined to be a personalized skincare product. The finished product outlet is used for the preservation and bottling of personalized skincare products.

[0022] By adopting the above technical solution, based on micro-pump array and closed-loop control algorithm, and decoding and generating time-driven waveforms according to personalized formula, quantitative active ingredients are accurately extracted and mixed. Viscosity, pH and uniformity are detected online and compared with thresholds, realizing the precise micro-scale formulation and real-time quality verification of active ingredients, which significantly improves the accuracy, safety and consistency of the formula and the finished product.

[0023] Secondly, this application also provides a method for dynamic formulation of personalized skincare products based on real-time monitoring of the skin microbiome, which adopts the following technical solution; A method for dynamic formulation of personalized skincare products based on real-time monitoring of the skin microbiome includes: The user attaches a microfluidic sampling layer to the target skin to collect surface sweat and sebum secretions. An impedance spectroscopy sensor array scans and detects its electrochemical properties, while simultaneously collecting skin temperature and pH. A signal conditioning and conversion layer digitizes the electrochemical impedance signal to obtain impedance spectroscopy parameters and auxiliary physicochemical parameters. A filtering and denoising layer filters and reduces noise on the above parameters. A parameter correction layer performs baseline correction and correlation based on timestamps to obtain impedance spectroscopy physicochemical parameters. A feature extraction layer extracts impedance spectroscopy feature vectors. A wireless communication layer transmits the feature vectors to the cloud. A customized state mapping model transforms the estimated microbiome state into a value, which is then input into the data input device in combination with environmental factors, user history, and profile data. The feature encoder performs feature mapping on the estimated value through a multi-head self-attention mechanism, capturing the temporal state and fusing multimodal data to obtain a multidimensional feature code. The graph inference engine infers the initial skincare formula based on the knowledge graph of microorganisms and active ingredients. The formula optimizer modifies the initial formula through multi-objective optimization to generate a modified skincare formula; the safety verifier screens allergens, verifies concentration thresholds and coexistence stability to obtain a safe skincare formula; the formula outputter decomposes the formula according to the blending process and outputs the component vector and blending instruction sequence. The control and detection unit decodes instructions, generates channel timing drive waveforms, controls and monitors the output of the micro-pump channel, and verifies the active ingredients; the online mixing chamber mixes raw materials to obtain the initial finished product; the online quality inspection unit detects the viscosity, pH, and uniformity of the finished product, and if all meet the standards, it is determined to be a personalized skin care product; The feedback learner incrementally updates the state mapping model based on changes in the user's skin microbiome and subjective satisfaction before and after use, and optimizes safe skincare formulas to form personalized skincare formulas.

[0024] By adopting the above technical solution, skin secretions are collected through flexible patches, and feature vectors are extracted using impedance spectroscopy sensing and edge computing. These vectors are then transformed into microbiome states through a customized state mapping model (improved MobileNet and frequency band attention mechanism). An initial formula is generated in the cloud based on multi-head self-attention and knowledge graph reasoning. A personalized formula is then obtained through multi-objective optimization (such as NSGA-II) and safety constraint verification. The formula is precisely dispensed through a micro-pump array and quality inspected online. Incremental model updates are achieved by combining feedback learning. This forms a personalized skin care system that integrates non-invasive detection, intelligent decision-making, precise dispensing, and closed-loop optimization, significantly improving the scientific validity, safety, and user suitability of the formula.

[0025] Thirdly, this application also provides a storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to realize the personalized skin care product dynamic formulation method based on real-time monitoring of the skin microbiome as described above.

[0026] In summary, the beneficial technical effects of this application are as follows: By combining impedance spectroscopy sensing arrays with calibration mapping models, the skin microbiome status can be rapidly estimated in home settings, significantly shortening the detection cycle compared to traditional gene sequencing methods. The micro-precision dispensing device enables instant customized dispensing of skincare products, avoiding the limitations of traditional pre-made product formulas. Furthermore, through multi-layered ingredient safety constraints (allergen screening, concentration safety threshold verification, and multi-ingredient coexistence stability verification), the safety of skincare formulas is improved. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the personalized skincare product dynamic formulation system in this application; Figure 2This is a cross-sectional view of the skin microbiome real-time sensing module structure in this application; Figure 3 This is a diagram of the cloud-based AI formula generation module architecture in this application; Figure 4 This is a schematic diagram of the micro-precision dispensing module structure in this application; Figure 5 This is a flowchart of the personalized skincare product dynamic formulation method in this application. Detailed Implementation

[0028] The present application will be further described in detail below with reference to the accompanying drawings.

[0029] Reference Figure 1 This application discloses a personalized skincare product dynamic formulation system based on real-time monitoring of the skin microbiome, comprising a real-time skin microbiome sensing module, a cloud-based artificial intelligence formula generation module, and a micro-volume precision formulation interactive module: wherein, The real-time skin microbiome sensing module includes a microbiome detection module and an edge computing processing module; The microbiome detection module is used to monitor the user's skin surface and collect electrochemical impedance spectroscopy parameters and auxiliary physicochemical parameters; The edge computing processing module is used to clean and feature electrochemical impedance spectroscopy parameters and auxiliary physicochemical parameters, and to generate and transmit impedance spectroscopy feature vectors. The cloud-based AI-powered recipe generation module includes a state calibration and mapping module, a cloud-based intelligent decision-making module, and a dynamic recipe calculation module. The state calibration and mapping module is used to register and transform the impedance spectrum feature vector according to the impedance spectrum-microbiome state mapping mechanism to obtain the estimated value of the microbiome state. The cloud-based intelligent decision-making module receives microbiome state estimates and combines them with environmental factors and user profile data to generate an initial skincare formula. The dynamic formula calculation module is used to examine and optimize the component ratio parameters of the initial skin care formula based on the ingredient safety constraint mechanism, and generate a personalized skin care formula. The micro-volume precision dispensing interaction module includes a micro-volume precision dispensing module and a user terminal interaction module; The micro-precision formulation module is used to extract and mix the corresponding active ingredient raw materials according to the personalized skin care formula to form personalized skin care products. The user terminal interaction module is used by users to control the type and proportion of active ingredient raw materials, store personalized skin care formulas, and generate alarm notifications during the formulation process of personalized skin care products.

[0030] In this embodiment, the skin microbiome detection module (101) is responsible for collecting electrochemical impedance spectroscopy data and auxiliary physicochemical parameters of the skin surface; the edge computing processing module (102) completes real-time preprocessing and feature extraction of the signal to generate impedance spectrum feature vectors; the calibration mapping model (108) converts the impedance spectrum feature vectors into microbiome state estimates; the cloud-based artificial intelligence decision engine (103) generates personalized formulas based on microbiome state estimates; the dynamic formula calculation module (104) performs multi-objective constraint optimization; the micro-precision dispensing module (105) performs physical dispensing operations; the user terminal interaction module (106) provides a human-computer interaction interface; and the standardized raw material consumables module (107), i.e., the raw material storage area, provides the raw materials required for the formula. A closed loop of "perception-calibration-analysis-decision-execution-feedback" is formed between the various components.

[0031] Reference Figure 2 Preferably, the real-time skin microbiome sensing module includes: The microbiome detection module is in the form of a flexible patch, including: A microfluidic sampling layer is used to automatically collect surface sweat and sebum secretions from the user's skin using capillary force through an embedded array of microchannels. Impedance spectroscopy sensing array is used to detect the electrochemical properties of surface sweat and sebum secretions by scanning different frequency bands, and obtain electrochemical impedance signals; The signal conditioning and conversion layer is used to synchronously acquire and digitize electrochemical impedance signals to obtain electrochemical impedance spectroscopy parameters and auxiliary physicochemical parameters. The edge computing processing module includes: The filtering and noise reduction layer is used to filter and reduce noise in the electrochemical impedance spectroscopy parameters and auxiliary physicochemical parameters to obtain noise-free impedance spectroscopy parameters and noise-free physicochemical parameters. The parameter correction layer is used to perform baseline correction and correlation of the noiseless impedance spectrum parameters and noiseless physicochemical parameters based on the acquisition timestamp, so as to obtain the impedance spectrum physicochemical parameters. The feature extraction layer is used to extract features from the physicochemical parameters of the impedance spectrum to obtain the impedance spectrum feature vector; The wireless communication layer is used to monitor and transmit impedance spectrum eigenvectors to the cloud.

[0032] In this embodiment, the overall thickness of the impedance spectrum sensing array body (201) is no more than 3.5 mm, the effective detection area is 20 mm × 20 mm, and a flexible patch shape is adopted to ensure good conformal contact with the skin.

[0033] The microfluidic sampling layer (202) is fabricated on a flexible polydimethylsiloxane (PDMS) substrate and has an embedded array of microchannels (in one embodiment, the channel width is about 50 μm and the depth is about 30 μm) to automatically collect sweat and sebum secretions from the skin surface using capillary force without the need for an external driving device.

[0034] In one embodiment, the impedance spectroscopy sensing array (203) consists of 64×64 interdigitated electrodes with an electrode spacing of approximately 10 μm and a finger width of approximately 5 μm, fabricated using a gold / chromium bilayer thin film via photolithography-stripping. The sensing array covers a frequency sweep range of 100 Hz to 10 MHz, and the impedance spectroscopy data in different frequency bands can reflect different levels of electrochemical information: the low-frequency band (100 Hz to 10 kHz range) mainly reflects changes in extracellular matrix conductivity; the mid-frequency band (10 kHz to 1 MHz range) mainly reflects cell membrane capacitance characteristics; and the high-frequency band (1 MHz to 10 MHz range) mainly reflects changes in intracellular impedance.

[0035] It should be noted that the impedance spectroscopy data mentioned above are not directly equivalent to the taxonomic composition or quantitative abundance data of the microbiome. The impedance spectroscopy data must be transformed through a calibration mapping model (see Section 5.3) before an estimate of the microbiome state can be obtained.

[0036] The signal conditioning and analog-to-digital conversion layer (204) integrates a multi-channel analog-to-digital converter for synchronous acquisition and digitization of impedance spectrum signals. In one implementation, a 16-channel 16-bit analog-to-digital converter is used.

[0037] In one implementation, the edge processing microcontroller (205) employs an ARM Cortex-M7 core processor, runs a lightweight inference model, performs signal denoising (e.g., adaptive Wiener filtering), baseline correction and preliminary feature extraction on the edge, and outputs an impedance spectrum feature vector.

[0038] The wireless communication module (206) supports dual-mode communication of Bluetooth Low Energy (BLE) and Near Field Communication (NFC). BLE is used for continuous transmission of monitoring data, and NFC is used for device pairing and firmware upgrades.

[0039] The impedance spectroscopy sensing array also integrates a temperature sensor (207) and a pH sensor (208) based on ion-sensitive field-effect transistor (ISFET) technology to acquire auxiliary parameters of skin surface temperature and pH.

[0040] Preferably, the impedance spectroscopy-microbiome state mapping mechanism includes: Impedance spectroscopy sensing and swab sampling were performed on the target skin area of ​​the subjects to obtain impedance spectral feature vectors and microbial group sequences; Based on the collection timestamp, impedance spectrum feature vectors and microbial group sequences are paired to construct a training dataset and an independent validation set; Based on a pre-defined improved MobileNet network, the training dataset is iteratively learned several times to generate an initial state mapping model and determine microbiome state indicators, including: Microbial group sequences in the training dataset are used as target labels, and impedance spectrum feature vectors in the training dataset are labeled as training data. Based on the impedance spectrum parameters in the training data of the scanning frequency band, resampling and linear interpolation are performed to calculate the frequency band impedance magnitude and frequency band phase angle. The auxiliary parameters in the training data are globally scalarized and then concatenated with the frequency band impedance magnitude and frequency band phase angle to construct the feature vector to be trained. Based on the data source, frequency drift simulation and noise generalization are performed on the feature vectors to be trained to obtain enhanced feature vectors; Based on the one-dimensional convolutional input layer in the improved MobileNet network and the batch data volume, the dimension of the enhanced feature vector is unified to obtain a feature vector of the same dimension. Based on the preset learning rate decay function and the iterative training layer, the same-dimensional feature vector is iterated several times to obtain the impedance spectrum feature weights. Based on the frequency band attention mechanism and target label, the impedance spectrum feature weights are temporally correlated and frequency band weighted to obtain the microbial mapping weight matrix. Based on the microbial mapping weight matrix and the parallel output layer, the enhanced feature vector is identified and predicted, and the metabolic activity value and the number of bacterial genera are output. Clustering and normalization of microbial taxonomic sequences, and calculation of microbiome state indices; The metabolic activity values ​​and the number of bacterial genera were compared based on the microbiome status indicators; If both are within the corresponding target tolerance range, the current microbial mapping weight matrix is ​​deemed compliant and used as the initial state mapping model. The initial state mapping model was evaluated and calibrated based on the independent validation set. The microbiome state values ​​were calculated, and the model accuracy was verified by combining the microbiome state indicators. Based on specific paired data and a paired sampling calibration mechanism, the initial state mapping model is corrected to obtain a customized state mapping model, which outputs the target microbial group sequence.

[0041] In this embodiment, (a) paired sampling: impedance spectroscopy sensing and swab sampling are performed simultaneously on the target skin area of ​​the subject. The swab samples are sent to the laboratory for 16S rRNA gene amplicon sequencing or metagenomic sequencing to obtain quantitative data such as the taxonomic composition of the microbiome and the relative abundance of various groups as labels.

[0042] (b) Data Alignment: The feature vectors acquired from each impedance spectroscopy acquisition are paired one-to-one with the sequencing results at the same time point to construct a training dataset. The training set should contain paired samples that meet the preset size requirements and cover different individuals, different sites, and different time points.

[0043] (c) Model Training: Using impedance spectrum feature vectors as input and microbiome state indicators represented by sequencing results as target labels, a regression or classification model is trained. Target labels may include: estimated values ​​of microbial diversity indices (such as the Shannon index), estimated values ​​of the relative abundance of major microbial communities, and estimated values ​​of overall metabolic activity indicators. In one implementation, a deep neural network is used as the mapping model.

[0044] (d) Calibration and verification: The predictive performance of the calibration mapping model is evaluated using an independent validation set. The prediction accuracy on the validation set is required to meet the preset requirements (e.g., using the coefficient of determination R² or mean absolute error as the evaluation index).

[0045] (e) Individualized fine-tuning: When a user uses the system for the first time, the system can choose to perform a paired sampling calibration process to fine-tune the pre-trained calibration mapping model using a small amount of individualized paired data, so as to improve the prediction accuracy for that user.

[0046] In this embodiment, the impedance spectrum (magnitude |Z| and phase angle φ) of the continuously swept frequency is divided into three frequency bands according to physiological meaning: Low frequency band (100Hz–10kHz): Reflects the extracellular matrix, with Nlow points sampled at logarithmic intervals.

[0047] Mid-frequency band (10kHz–1MHz): reflects cell membrane capacitance, taken as Nmid point.

[0048] High frequency band (1MHz–10MHz): reflects intracellular impedance, taking the Nhigh point.

[0049] Linear interpolation is performed on the data for each frequency band to ensure that each sample has a vector of the same dimension at a fixed frequency index.

[0050] Main feature: The impedance magnitude and phase angle of the three frequency bands are concatenated into a dual-channel one-dimensional vector (Channel0: magnitude, Channel1: phase).

[0051] Auxiliary parameters: Skin surface temperature and pH are used as global scalars and copied and expanded into bias vectors of the same length as the main features (or used as independent input branches).

[0052] The final input dimension is: [Batch,2,Sequence_Length], where Sequence_Length=N_{low}+N_{mid}+N_{high}.

[0053] The Shannon diversity index and metabolic activity indicators (such as ATP concentration) were standardized using Z-scores.

[0054] The relative abundance of major phyla / genus (e.g., the proportion of Staphylococcus genus, the proportion of Propionibacterium genus) is normalized using Softmax or retained as a continuous value in the 0-1 range.

[0055] Remove the original MobileNet 224×224×3 input layer and replace it with a 1D convolutional input layer (in_channels=2, out_channels=32).

[0056] Replace the 2D depthwise convolution Conv2d(3x3) with the 1D depthwise convolution Conv1d(kernel_size=3).

[0057] The channel-preserving pointwise convolution Conv1d(kernel_size=1) is used for cross-channel information fusion.

[0058] The linear bottleneck layer of the inverted residual structure remains unchanged, and only the stride is adjusted to control the time-length downsampling rate.

[0059] Embed a frequency-wise Squeeze-and-Excitation Block before the global average pooling layer: Global average pooling is performed on the one-dimensional feature map, and frequency band weights are generated through two fully connected layers, which are then weighted back to the original feature map.

[0060] This move aims to enhance the model's sensitivity to specific frequency bands, such as the mid-frequency band that reflects membrane capacitance, in predicting microbiome metabolism.

[0061] Remove the original classification head and construct a parallel output branch: Branch A (Regression): Fully connected layer → Output:1 (Shannon index) or Output:1 (metabolic activity value).

[0062] Branch B (Abundance Prediction): Fully connected layer → Softmax → Output: N_classes (Number of major bacterial genera).

[0063] The loss function used is joint loss: Loss = λ1 * MSE (regression) + λ2 * CrossEntropy (abundance classification).

[0064] A small frequency drift simulation is applied to the impedance spectrum sequence (after adding a random perturbation to the original frequency index and then resampling).

[0065] Gaussian noise was added to the auxiliary parameters (temperature, pH) to simulate fluctuations in the real measurement environment.

[0066] Optimizer: AdamW (decouples weight decay to prevent overfitting); Batch size: 32 or 64 depending on the sample size. Linear warmup for the first 5 epochs. Cosine Annealing is then used to decay the learning rate to 1% of the initial learning rate.

[0067] Every 10 iterations (Epochs) are completed, the mean absolute error (MAE) (for diversity indices) and F1 score (for dominant microbial community predictions) are calculated on the validation set.

[0068] If the loss does not decrease after 20 consecutive iterations, early stopping is triggered, and the optimal weights are saved as the initial state mapping model.

[0069] Input the newly acquired impedance spectrum data into the trained model.

[0070] Output 1: Read the regression branch values, and obtain the Shannon diversity estimate and metabolic activity estimate after destandardization.

[0071] Output 2: Read the index of the highest probability of the classification branch, map it back to the name of the genus, and obtain the label of the dominant bacterial community on the skin surface and its relative abundance estimate.

[0072] Enable Monte Carlo Dropout (keep the Dropout layer enabled during inference and perform multiple forward propagations).

[0073] Calculate the variance of multiple prediction results. If the variance is too large (exceeding the set threshold), determine that the current input feature is in the model blind zone, mark it as a low-confidence result, and suggest supplementary sequencing verification.

[0074] Reference Figure 3 Preferably, the cloud-based AI recipe generation module includes: The data input device is used to receive multimodal data including impedance spectrum eigenvectors, auxiliary physicochemical parameters, microbiome state estimates, historical microbiome state sequences, environmental factors, allergen data, and user profile data. The feature encoder is used to perform feature mapping on the microbiome state estimate based on the microbiome embedding layer of the preset multimodal attention fusion network, capture the temporal state sequence of the historical microbiome state by combining the temporal feature extraction layer, and cross-fuse multimodal data by combining the cross-modal fusion layer to obtain multidimensional feature encoding. The graph inference engine is used to infer the multidimensional feature codes based on the knowledge graph formed by the interaction between microorganisms and active ingredients, and obtain the initial skin care formula.

[0075] The formula optimizer is used to modify the microbiome diversity recovery index, target microbial abundance adjustment precision, and total number of ingredients among the components in the initial skin care formula according to a preset multi-objective optimization mechanism, thereby generating a modified skin care formula. Safety calibrators are used to screen modified skincare formulas for allergens, verify concentration safety thresholds, and validate the stability of multiple coexisting ingredients, resulting in safe skincare formulas. The formula output device is used to decompose safe skin care formulas according to the skin care product formulation process, output formula component vectors, and generate a series of formulation instructions. The feedback learner is used to incrementally update the initial state mapping model based on changes in the skin microbiome state before and after skincare and the user's subjective satisfaction rating, and to optimize safe skincare formulas to form personalized skincare formulas.

[0076] In this embodiment, the input layer (401) receives multiple types of data: (a) impedance spectrum feature vectors; (b) measured values ​​of temperature and pH; (c) microbiome state estimates output by the calibration mapping model; (d) time-series data of the user's historical microbiome state; (e) environmental factors (UV index, temperature, humidity, etc., obtained through an external interface); (f) user profile (age, gender, skin type classification, past skincare product usage records); and (g) allergen database matching results. In one embodiment, the calibration mapping model (108) can be deployed in the input preprocessing stage to generate microbiome state estimates based on the impedance spectrum feature vectors.

[0077] The feature encoder (402) uses a multimodal attention fusion network to process input features from different modalities. In one implementation, a multi-head self-attention mechanism based on the Transformer architecture is used. The microbiome embedding layer maps discrete microbial community features to a continuous vector space, the temporal feature extraction module (e.g., a temporal convolutional network, TCN) captures the changing trend of microbiome state over time, and the cross-modal fusion layer fuses information from different data sources through cross-attention.

[0078] The knowledge graph (403) encodes known interactions (promoting, inhibiting, neutral) between skin-associated microbes and active ingredients, as well as compatibility constraints between ingredients. A graph inference engine performs inference on this knowledge graph to discover potential ingredient-microbe synergies. In one implementation, a graph neural network (GNN) is used as the inference engine.

[0079] The formulation optimizer (404) employs a multi-objective optimization algorithm, simultaneously optimizing the following objective functions: microbiome diversity recovery index, target community abundance adjustment accuracy, total number of components (with a tendency to simplify), and cost. In one implementation, the NSGA-III multi-objective genetic algorithm is used to perform Pareto front search, generating a set of non-dominated solutions, which are then filtered by a security verification module and user preferences.

[0080] The safety verification module (405) performs triple component safety constraint verification: allergen screening (comparing to the user's known allergen database), concentration safety threshold verification, and multi-component coexistence stability verification.

[0081] The output layer (406) generates a formula component vector (mass percentage of each component) and a dispensing instruction sequence (timing control parameters of each micropump).

[0082] The feedback learning module (407) converts changes in the skin microbiome state and user subjective satisfaction ratings over a certain period after user use into feedback signals, and incrementally updates the model parameters. In one implementation, a reinforcement learning algorithm is used to achieve the feedback update.

[0083] Reference Figure 4 Preferably, the micro-volume precision dispensing module includes: The raw material storage area is used to store several active ingredient raw materials, identify the type and expiration date of the raw materials, and load gel matrix or emulsion matrix. Micropump arrays are used to extract active ingredient raw materials by calling up a corresponding number of micropump channels according to the ingredient ratio of personalized skin care formulas, so as to obtain a quantitative amount of ingredient raw materials. The control and detection unit is used to decode the dispensing instruction sequence, generate channel timing drive waveforms, control and monitor the actual output of the micro-pump channel, and verify the active ingredient raw materials. An online mixing chamber is used to mix active ingredient raw materials to obtain an initial skincare product; The online quality inspection unit is used to test the viscosity, pH and uniformity of the initial skin care product and make a judgment based on the preset target threshold range; If all three are within the corresponding target threshold range, then the current initial skincare product is determined to be a personalized skincare product. The finished product outlet is used for the preservation and bottling of personalized skincare products.

[0084] In this embodiment, the raw material storage area (501) adopts a replaceable capsule design and is equipped with a radio frequency identification (RFID) chip for automatic identification of raw material type and expiration date. The system can support the simultaneous installation of multiple active ingredient capsules. The substrate carrier storage area (502) can be loaded with gel matrix or emulsion matrix, selected according to formulation requirements.

[0085] Each channel of the micropump array (503) consists of an independent piezoelectric ceramic actuator and a precision one-way valve. In one embodiment, the single-step resolution is 0.1 μL. Compared with traditional stepper motor pumps, the piezoelectric drive scheme has advantages such as low pulsation, high precision, fast response speed, and low power consumption.

[0086] The online mixing chamber (504) adopts a spiral micromixer design to achieve rapid mixing of multiple components through the Dean vortex effect.

[0087] The discharge port (505) adopts a micro-jet design to reduce the introduction of air bubbles.

[0088] The main microcontroller in the control and detection unit (506) receives formula instructions from the cloud and decodes them into timing drive waveforms for each channel's micropumps. A flow sensor array monitors the actual output of each channel in real time, forming a closed-loop flow control. A near-infrared spectroscopy detection module can be used to verify the correctness of the output composition.

[0089] The online quality inspection module (507) includes a viscosity sensor (based on the principle of a quartz crystal microbalance in one embodiment), an online pH electrode, and a uniformity scattering detector. Example

[0090] User A, female, 28 years old, with combination skin, is known to be allergic to parabens. One evening, the user applied a sensor patch to her right cheek, and the system collected impedance spectroscopy data.

[0091] After feature extraction at the edge, the impedance spectrum feature vector is input into the calibration mapping model. The calibration mapping model outputs the following microbiome state estimates: the estimated microbial diversity index is 2.1 (the user's healthy baseline value is approximately 3.2); the relative abundance of Propionibacterium acnes in the target microbial community associated with acne is estimated to be high (approximately 35%, while the user's healthy range is approximately 10% to 20%); and the diversity has shown a decreasing trend over the past 7 days.

[0092] It should be noted that the above estimates are all outputs of the calibration mapping model, and their accuracy is affected by the calibration dataset and the degree of individualized calibration.

[0093] The cloud-based decision engine initiates the formula generation process based on the above estimates. The knowledge graph reasoning stage outputs the following associations: Niacinamide is positively correlated with inhibiting the excessive proliferation of Propionibacterium acnes within a specific concentration range, and has little effect on commensal bacteria such as Staphylococcus epidermidis; prebiotic ingredients (such as fructooligosaccharides) help promote the restoration of microbial diversity; ceramides help strengthen the skin barrier function.

[0094] The formulation optimizer solves the problem under the following ingredient safety constraints: exclusion of known allergens (parabens); concentrations of each ingredient not exceeding the upper limits specified in the "Cosmetic Safety Technical Specifications"; and successful verification of multi-ingredient coexistence stability. The optimizer primarily optimizes the microbial diversity recovery index and the accuracy of target microbial abundance adjustment, with minimizing the number of ingredients as a secondary objective. After multi-objective optimization search, it outputs a recommended formulation. In this embodiment, the recommended formulation includes: approximately 2% niacinamide, approximately 1.5% fructooligosaccharides, and approximately 0.5% ceramides, with a gel matrix as the base.

[0095] The blending device completes the blending according to the formula instructions and discharges the material after passing online quality inspection. After user use, the system observes a rebound in the estimated value of the microbial diversity index and a decrease in the estimated value of the relative abundance of the target microbial community in subsequent monitoring periods. After the user submits a satisfaction rating, the feedback learning module updates the model parameters accordingly.

[0096] Reference Figure 5 This application discloses a method for dynamically formulating personalized skincare products based on real-time monitoring of the skin microbiome, comprising: The user attaches a microfluidic sampling layer to the target skin to collect surface sweat and sebum secretions. An impedance spectroscopy sensor array scans and detects its electrochemical properties, while simultaneously collecting skin temperature and pH. A signal conditioning and conversion layer digitizes the electrochemical impedance signal to obtain impedance spectroscopy parameters and auxiliary physicochemical parameters. A filtering and denoising layer filters and reduces noise on the above parameters. A parameter correction layer performs baseline correction and correlation based on timestamps to obtain impedance spectroscopy physicochemical parameters. A feature extraction layer extracts impedance spectroscopy feature vectors. A wireless communication layer transmits the feature vectors to the cloud. A customized state mapping model transforms the estimated microbiome state into a value, which is then input into the data input device in combination with environmental factors, user history, and profile data. The feature encoder performs feature mapping on the estimated value through a multi-head self-attention mechanism, capturing the temporal state and fusing multimodal data to obtain a multidimensional feature code. The graph inference engine infers the initial skincare formula based on the knowledge graph of microorganisms and active ingredients. The formula optimizer modifies the initial formula through multi-objective optimization to generate a modified skincare formula; the safety verifier screens allergens, verifies concentration thresholds and coexistence stability to obtain a safe skincare formula; the formula outputter decomposes the formula according to the blending process and outputs the component vector and blending instruction sequence. The control and detection unit decodes instructions, generates channel timing drive waveforms, controls and monitors the output of the micro-pump channel, and verifies the active ingredients; the online mixing chamber mixes raw materials to obtain the initial finished product; the online quality inspection unit detects the viscosity, pH, and uniformity of the finished product, and if all meet the standards, it is determined to be a personalized skin care product; The feedback learner incrementally updates the state mapping model based on changes in the user's skin microbiome and subjective satisfaction before and after use, and optimizes safe skincare formulas to form personalized skincare formulas.

[0097] In this embodiment: Step S301: The user attaches the skin microbiome detection module to the target skin area (such as the T-zone of the face, cheeks or forearms) to ensure that the sensor array is fully attached to the skin.

[0098] Step S302: The impedance spectrum sensing array performs an impedance spectrum scan within a set frequency range, simultaneously acquiring skin surface temperature and pH value. In one embodiment, the sweep frequency range is 100Hz to 10MHz, and a single complete acquisition cycle is approximately 5 seconds.

[0099] Step S303: The edge computing module performs preprocessing operations on the raw impedance spectroscopy data, including noise reduction, baseline correction, and feature extraction, to generate impedance spectroscopy feature vectors. These feature vectors are then transformed into microbiome state estimates using a calibration mapping model (see Section 5.3). In one implementation, distributed relaxation time (DRT) analysis is used to extract characteristic peak parameters, followed by principal component analysis (PCA) for dimensionality reduction.

[0100] Step S304: The estimated microbiome state value, along with environmental factors, user historical data, and user profile, is wirelessly transmitted to the cloud-based artificial intelligence decision engine. A personalized formulation solution is generated through a cascade pipeline of feature encoding → knowledge graph reasoning → formulation optimization.

[0101] Step S305: Perform component safety constraint verification on the generated formulation, including allergen screening, concentration safety threshold verification, and multi-component coexistence stability verification. If the verification fails, return to step S304 to modify the constraints and regenerate.

[0102] Step S306: The validated formulation instruction is sent to the micro-precision dispensing device. The multi-channel micropump array precisely extracts each active ingredient and substrate carrier sequentially according to the instruction, injecting them into the online mixing chamber for thorough mixing. In one embodiment, the single dispensing volume is 1 to 5 mL, and the dispensing time does not exceed 30 seconds.

[0103] Step S307: The online quality inspection module tests the viscosity, pH value, and uniformity of the prepared product to confirm that all indicators are within the target range. If any indicator fails to meet the requirements, the product is re-prepared.

[0104] Step S308: The user uses the prepared product. After use, the system continuously monitors changes in the skin microbiome, and the user submits a subjective evaluation of their experience through the terminal interface.

[0105] Step S309: The system decides whether to continue the next round of monitoring-allocation cycle based on the monitoring frequency set by the user.

[0106] Step S310: The feedback learning module updates the model parameters of the decision engine based on the microbiome state change data and user satisfaction evaluations after use. At the end of the monitoring period, the system generates a comprehensive report containing the microbiome state change trend and the history of formula adjustments.

[0107] This application discloses a storage medium storing at least one instruction, at least one program, code set, or instruction set. The at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to realize the personalized skincare product dynamic formulation method based on real-time monitoring of the skin microbiome as described above.

[0108] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A personalized skincare product dynamic formulation system based on real-time monitoring of the skin microbiome, characterized in that, include: The microbiome detection module is used to monitor the user's skin surface and collect electrochemical impedance spectroscopy parameters and auxiliary physicochemical parameters; An edge computing processing module is used to clean and characterize the electrochemical impedance spectroscopy parameters and the auxiliary physicochemical parameters, and to generate and transmit impedance spectroscopy feature vectors. The state calibration and mapping module is used to register and transform the impedance spectrum feature vector according to the impedance spectrum-microbiome state mapping mechanism to obtain the estimated value of the microbiome state. The cloud-based intelligent decision-making module is used to receive the estimated microbiome state value, combine it with environmental factors and user profile data, and generate an initial skincare formula. The dynamic formula calculation module is used to examine and optimize the component ratio parameters of the initial skin care formula according to the ingredient safety constraint mechanism, and generate a personalized skin care formula. The micro-precision formulation module is used to extract and mix the corresponding active ingredient raw materials according to the personalized skin care formula to form personalized skin care products.

2. The personalized skincare product dynamic formulation system based on real-time monitoring of the skin microbiome according to claim 1, characterized in that, The microbiome detection module is in the form of a flexible patch, including: A microfluidic sampling layer is used to automatically collect surface sweat and sebum secretions from the user's skin using capillary force through an embedded array of microchannels. Impedance spectroscopy sensing array is used to detect the electrochemical properties of surface sweat and sebum secretions by scanning different frequency bands, and obtain electrochemical impedance signals; The signal conditioning and conversion layer is used to synchronously acquire and digitize the electrochemical impedance signal to obtain electrochemical impedance spectral parameters and auxiliary physicochemical parameters.

3. The personalized skincare product dynamic formulation system based on real-time monitoring of the skin microbiome according to claim 1, characterized in that, The edge computing processing module includes: The filtering and noise reduction layer is used to filter and reduce noise in the electrochemical impedance spectroscopy parameters and auxiliary physicochemical parameters to obtain noise-free impedance spectroscopy parameters and noise-free physicochemical parameters. The parameter correction layer is used to perform baseline correction and correlation on the noiseless impedance spectrum parameters and the noiseless physicochemical parameters according to the acquisition timestamp, so as to obtain the impedance spectrum physicochemical parameters. The feature extraction layer is used to extract features from the impedance spectrum physicochemical parameters to obtain the impedance spectrum feature vector; A wireless communication layer is used to monitor and transmit the impedance spectrum feature vector to the cloud.

4. The personalized skincare product dynamic formulation system based on real-time monitoring of the skin microbiome according to claim 1, characterized in that, The impedance spectrum-microbiome state mapping mechanism includes: Impedance spectroscopy sensing and swab sampling were performed on the target skin area of ​​the subjects to obtain impedance spectral feature vectors and microbial group sequences; The impedance spectrum feature vector and the microbial group sequence are paired according to the collection timestamp to construct a training dataset and an independent validation set; Based on a pre-defined improved MobileNet network, the training dataset is iterated several times to generate an initial state mapping model and determine microbiome state indicators. The initial state mapping model is evaluated and calibrated based on the independent validation set, the microbiome state value is calculated, and the model accuracy is verified by combining the microbiome state index. Based on specific paired data and a paired sampling calibration mechanism, the initial state mapping model is corrected to obtain a customized state mapping model, which outputs the target microbial group sequence.

5. The personalized skincare product dynamic formulation system based on real-time monitoring of the skin microbiome according to claim 4, characterized in that, The step of performing several iterations of learning on the training dataset based on a preset improved MobileNet network to generate an initial state mapping model and determine microbiome state indicators includes: Microbial group sequences in the training dataset are used as target labels, and impedance spectrum feature vectors in the training dataset are labeled as training data. Based on the scanning frequency band, the impedance spectrum parameters in the training data are resampled and linearly interpolated to calculate the frequency band impedance magnitude and frequency band phase angle. The auxiliary parameters in the training data are globally scalarized and then concatenated with the frequency band impedance magnitude and the frequency band phase angle to construct the training feature vector. Based on the data source, the feature vector to be trained is subjected to frequency drift simulation and noise generalization to obtain an enhanced feature vector; Based on the one-dimensional convolutional input layer in the improved MobileNet network and the batch data volume, the enhanced feature vector is dimension-unified to obtain a feature vector of the same dimension. Based on the preset learning rate decay function and combined with the iterative training layer, the same-dimensional feature vector is iteratively learned several times to obtain the impedance spectrum feature weights. Based on the frequency band attention mechanism and target label, the impedance spectrum feature weights are temporally correlated and frequency band weighted to obtain the microbial mapping weight matrix. Based on the microbial mapping weight matrix and the parallel output layer, the enhanced feature vector is identified and predicted, and the metabolic activity value and the number of bacterial genera are output. Clustering and normalization of the microbial group sequences were performed to calculate microbiome state indices; The metabolic activity value and the number of bacterial genera were compared based on the microbiome status indicators; If both are within the corresponding target tolerance range, the current microbial mapping weight matrix is ​​deemed compliant and used as the initial state mapping model.

6. The personalized skincare product dynamic formulation system based on real-time monitoring of the skin microbiome according to claim 1, characterized in that, The cloud-based intelligent decision-making module includes: The data input device is used to receive multimodal data including impedance spectrum eigenvectors, auxiliary physicochemical parameters, microbiome state estimates, historical microbiome state sequences, environmental factors, allergen data, and user profile data. The feature encoder is used to perform feature mapping on the microbiome state estimate based on the microbiome embedding layer of the preset multimodal attention fusion network, capture the temporal state of the historical microbiome state sequence by combining the temporal feature extraction layer, and cross-fuse the multimodal data by combining the cross-modal fusion layer to obtain multidimensional feature encoding. The graph inference engine is used to infer the multidimensional feature encoding based on the knowledge graph formed by the interaction between microorganisms and active ingredients to obtain an initial skin care formula.

7. The personalized skincare product dynamic formulation system based on real-time monitoring of the skin microbiome according to claim 1, characterized in that, The dynamic formula calculation module includes: The formula optimizer is used to modify the microbiome diversity recovery index, target microbial abundance adjustment precision, and total number of ingredients among the components in the initial skin care formula according to a preset multi-objective optimization mechanism, thereby generating a modified skin care formula. A safety calibrator is used to screen the modified skincare formula for allergens, verify the concentration safety threshold, and validate the stability of multiple coexisting ingredients to obtain a safe skincare formula. A formula output device is used to decompose the safe skin care formula according to the skin care product formulation process, output the formula component vector, and generate a formulation instruction sequence. The feedback learner is used to incrementally update the initial state mapping model based on changes in the skin microbiome state before and after skincare and the user's subjective satisfaction rating, and to optimize safe skincare formulas to form personalized skincare formulas.

8. The personalized skincare product dynamic formulation system based on real-time monitoring of the skin microbiome according to claim 1, characterized in that, The micro-volume precision dispensing module includes: The raw material storage area is used to store several active ingredient raw materials, identify the type and expiration date of the raw materials, and load gel matrix or emulsion matrix. The micropump array is used to extract the active ingredient raw materials by calling the corresponding number of micropump channels according to the ingredient ratio of the personalized skin care formula, so as to obtain a quantitative amount of ingredient raw materials. The control and detection unit is used to decode the dispensing instruction sequence, generate channel timing drive waveforms, control and monitor the actual output of the micropump channel, and verify the active ingredient raw materials. An online mixing chamber is used to mix the active ingredient raw materials to obtain an initial skincare product; An online quality inspection unit is used to detect the viscosity, pH, and uniformity of the initial skincare product and make a judgment based on a preset target threshold range. If all three are within the corresponding target threshold range, then the current initial skincare product is determined to be a personalized skincare product. The finished product outlet is used for preserving and bottling the personalized skincare products.

9. The personalized skincare product dynamic formulation system based on real-time monitoring of the skin microbiome according to claim 1, characterized in that, The personalized skincare product dynamic blending system also includes: The user terminal interaction module is used by users to control the type and proportion of active ingredient raw materials, store personalized skin care formulas, and generate alarm notifications during the formulation process of personalized skin care products.

10. A method for dynamically adjusting personalized skincare products based on real-time monitoring of the skin microbiome, applied to the system described in any one of claims 1-9, characterized in that, include: Users attach the microfluidic sampling layer to the target skin to collect surface sweat and sebum secretions. The impedance spectroscopy sensor array scans and detects its electrochemical properties, while simultaneously collecting skin temperature and pH. The signal conditioning and conversion layer digitizes the electrochemical impedance signal to obtain impedance spectrum parameters and auxiliary physicochemical parameters; the filtering and denoising layer filters and reduces noise on the above parameters; the parameter correction layer performs baseline correction and correlation based on the timestamp to obtain the impedance spectrum physicochemical parameters; the feature extraction layer extracts the impedance spectrum feature vector; and the wireless communication layer transmits the feature vector to the cloud. A customized state mapping model is used to transform the estimated values ​​of the microbiome state. These values ​​are then input into the data input device in combination with environmental factors, user history, and profile data. The feature encoder performs feature mapping on the estimated values ​​through a multi-head self-attention mechanism, capturing the temporal state and fusing multimodal data to obtain multidimensional feature encoding. The graph inference engine uses a knowledge graph of microorganisms and active ingredients to deduce an initial skincare formula. The formula optimizer refines the initial formula through multi-objective optimization, generating a modified skincare formula. The safety calibrator screens for allergens, verifies concentration thresholds and coexistence stability, and obtains safe skincare formulas. The formula output device breaks down the formula according to the blending process and outputs the component vector and blending instruction sequence; The control and detection unit decodes instructions, generates channel timing drive waveforms, controls and monitors the output of the micro-pump channel, and verifies the active ingredients; the online mixing chamber mixes raw materials to obtain the initial finished product; The online quality inspection unit tests the viscosity, pH, and uniformity of the finished product. If all of these meet the standards, it is determined to be a personalized skincare product. The feedback learner incrementally updates the state mapping model based on changes in the user's skin microbiome and subjective satisfaction before and after use, and optimizes safe skincare formulas to form personalized skincare formulas.

Images