Skin care product whole-process detection equipment and production management and control method

Abstract

The invention discloses skin care product whole-process detection equipment and a production management and control method, and relates to the technical field of cosmetic detection and production control. The method comprises the following steps: step 1, acquiring multi-source detection data of the whole production process; 2, adopting a data cleaning algorithm to remove abnormal values, and extracting feature vectors representing process states through a feature engineering method to form a standardized feature data set; and 3, based on the standardized feature data set, establishing a nonlinear mapping relationship between the process parameters and the quality indexes by using a machine learning algorithm, and outputting a process state score and a deviation degree in real time. According to the method, deep cleaning, feature extraction and fusion analysis are performed on multi-source data, a dynamic process state evaluation mechanism is constructed by using a machine learning model, and a complex nonlinear relationship between process parameters and final quality indexes is accurately described, so that early recognition of fine process fluctuation and potential quality risks is realized.

Description

Technical Field

[0001] This invention relates to the field of cosmetic testing and production control technology, and in particular to a whole-process testing device and production control method for skin care products. Background Technology

[0002] As the skincare industry rapidly develops towards intelligent, standardized, and traceable processes, higher demands are placed on the full-process testing and control of product quality during production. As daily chemical products that come into direct contact with human skin, skincare products require strict monitoring of key quality parameters such as ingredient stability, microbiological indicators, physicochemical properties, and packaging sealing throughout the entire chain, from raw material input, mixing and emulsification, filling and sealing to finished product delivery. Traditional quality testing relies heavily on offline sampling and manual interpretation, with testing processes scattered across different workstations, resulting in isolated and delayed data, making it difficult to achieve real-time interception and root cause tracing of abnormal batches. Furthermore, existing automated testing equipment typically only monitors a single process (such as filling volume or appearance defects), lacking the ability to dynamically analyze cross-process quality parameters. This makes it impossible to identify downstream quality problems caused by fluctuations in upstream processes (such as emulsification temperature deviations or insufficient stirring time), leading to defective products entering the end market, increasing recall risks and brand reputation damage.

[0003] The production control method for skincare product end-to-end testing equipment aims to build a closed-loop system for quality perception and decision-making covering the entire lifecycle from raw materials to finished products. This method requires the integration of multi-source sensor data (such as near-infrared spectroscopy, visual images, temperature, humidity, and pressure signals), and through a unified data acquisition architecture and real-time analysis engine, to achieve synchronous evaluation and coordinated control of the quality status of each process node.

[0004] In existing technologies, on the one hand, most production lines still adopt a segmented testing mode, with each testing unit operating independently. This results in inconsistent data formats and asynchronous timestamps, making it difficult to support quality traceability across processes. On the other hand, even if some companies have introduced online testing devices, their control logic is still based on static threshold alarms, unable to dynamically adjust judgment criteria based on historical batch data and current operating conditions. This leads to a severe deficiency in identifying gradual process drift or complex defects (such as minor contamination combined with packaging micro-leakage). Furthermore, in highly flexible production scenarios (such as small-batch, multi-formula switching), existing systems lack an adaptive learning mechanism for new formula process windows, resulting in frequent reliance on manual resetting of testing rules, which is inefficient and prone to errors. Therefore, there is an urgent need for a production control method for skincare product end-to-end testing equipment capable of multi-process data fusion, dynamic threshold optimization, and anomaly propagation path tracking to improve the real-time performance, accuracy, and intelligence of quality control. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a skincare product end-to-end testing device and production control method. By introducing cross-process collaborative analysis and Bayesian networks, it proactively simulates the propagation path of quality risks between processes, significantly improving the success rate of real-time interception of defective products and fundamentally enhancing the ability to trace the root causes of quality problems.

[0006] In a first aspect, the present invention provides a production control method for a skincare product end-to-end testing equipment, comprising: Step 1: Acquire multi-source detection data throughout the entire production process. By deploying a sensor network at the raw material feeding, mixing and reaction, and filling and packaging stages, detection data from each process is collected and multi-source detection data is generated. Step 2: Preprocess and extract features from the multi-source detection data, use data cleaning algorithms to remove outliers, and use feature engineering methods to extract feature vectors that characterize the process state to form a standardized feature dataset; Step 3: Construct a dynamic process status assessment model. Based on the standardized feature dataset, use machine learning algorithms to establish a nonlinear mapping relationship between process parameters and quality indicators, and output the process status score and deviation in real time. Step 4: Perform cross-process collaborative analysis and risk warning, integrate the process status assessment results of each process, identify potential quality risk propagation paths through multi-source information fusion algorithms, and generate graded warning signals; Step 5: Implement adaptive production control and quality traceability. Dynamically adjust relevant process parameters based on the risk warning signals, record the data correlation of the entire process, and construct a traceable quality archive.

[0007] Preferably, the sensor network includes: The sensor network includes a high-precision flow sensor, an online spectral analyzer, a machine vision inspection unit, and a temperature and pressure sensor array. The high-precision flow sensor is deployed at the inlet of the raw material feeding pipeline to monitor the instantaneous flow rate and cumulative feeding amount of various liquid or paste raw materials in real time. The online spectrometer is installed in the observation window or bypass flow cell of the mixing reactor and uses near-infrared or Raman spectroscopy to perform non-contact online analysis of the component concentration, moisture content and active ingredient ratio of the materials in the reactor. The machine vision inspection unit includes a high-resolution industrial camera, a ring LED light source, and an image processing industrial control computer, which are respectively deployed at the raw material barrel label recognition station, the mixing kettle liquid level observation point, below the filling head, and the finished product sealing inspection area. It is used to capture static or dynamic high-definition images and to identify label information, liquid level height, filling volume, and sealing integrity. The temperature and pressure sensor array is distributed and embedded in the wall of the mixing reactor, the pipe connections and the filling cavity to monitor temperature and pressure changes in real time. The data collected by the sensor network is uploaded to the central processing unit in real time via the industrial Ethernet protocol and constructed as multi-source detection data.

[0008] Preferably, the data cleaning algorithm includes: For text data, the data cleaning algorithm adopts a statistical process control method based on sliding windows; an independent sliding window is established for each sensor channel, and the window size is not a fixed value, but is adaptively adjusted according to the current production line's process cycle time; For image data, a method combining morphological filtering and background subtraction is used to remove noise and irrelevant background interference.

[0009] Preferably, the feature engineering method includes: Feature engineering methods include principal component analysis and time-domain / frequency-domain feature extraction; wherein, the principal component analysis is used for dimensionality reduction, and the time-domain / frequency-domain feature extraction is used to capture dynamic characteristics; For spectral data or vectors composed of physical parameters, principal component analysis is applied to retain principal components with a cumulative contribution rate of not less than 95%, and the original data is projected into a low-dimensional feature space to eliminate redundant information and reduce computational complexity. For time-series data, time-domain feature extraction and frequency-domain feature extraction are performed. By performing a fast Fourier transform on the data, the energy proportion and dominant frequency in a specific frequency band are obtained.

[0010] Preferably, the dynamic process status evaluation model includes: The dynamic process status evaluation model adopts the gradient boosting decision tree algorithm, takes the standardized feature dataset generated in step 2 as input, and outputs a process stability index and a component uniformity score. The process stability index is used to reflect the smoothness of the current process operation, and the component uniformity score is used to quantify the uniformity level of the mixture. The dynamic process status assessment model establishes a mapping between features and quality indicators through supervised learning. In addition, the dynamic process status assessment model introduces an online learning mechanism. When new production batch data is generated, the new production batch data is added to the training set, and the incremental update of the dynamic process status assessment model is triggered. The update process adopts the mini-batch gradient descent method, and only performs a limited number of iterations on the new data to avoid overfitting of the dynamic process status assessment model.

[0011] Preferably, step 4 includes: The multi-source information fusion algorithm adopts a Bayesian network structure; the network nodes of the Bayesian network correspond to the state variables of the raw material feeding, mixing reaction, filling and packaging processes, and the state of the node corresponds to the process stability index and component uniformity score; the edges of the Bayesian network correspond to the influence relationship between processes. During real-time operation, the latest status assessment results of each process are input into the Bayesian network, and the posterior probability of each node is calculated using Bayes' theorem. When the posterior probability of a node is lower than the preset safety threshold, a quality risk is determined, and a graded warning signal is generated based on the degree to which the posterior probability deviates from the threshold. The warning levels are usually divided into three levels: Level 1 warning indicates that there is a potential risk and requires operator attention; Level 2 warning indicates that the risk is relatively high and process fine-tuning is recommended; Level 3 warning indicates that the risk is extremely high and the machine must be stopped immediately for investigation.

[0012] Preferably, step 5 includes: The adaptive production control adopts a fuzzy logic controller; the input of the fuzzy logic controller is the process status score and the early warning level, and the output is the adjustment amount sent to the actuator. The fuzzy logic controller includes fuzzy rules to convert numerical inputs into linguistic variables; after fuzzy inference, it generates control outputs through defuzzification to ensure that process parameters are maintained within the optimal operating range. In addition, a globally unique identifier is assigned to each production batch; the identifier is associated with all data of the corresponding batch and stored in the distributed database of the central processing unit; A data association map is constructed based on the identification code. The data association map supports forward tracing and backward tracing, providing a data foundation for quality accident analysis and continuous process improvement.

[0013] Preferably, the process knowledge base includes: The process knowledge base includes historical production data, empirical rules, and process optimization cases. During the production process, the current production situation is matched with the process optimization cases in the process knowledge base using case reasoning technology. The most similar case is retrieved, and its corresponding optimization strategy or risk avoidance measures are used as auxiliary decision-making suggestions for the fuzzy logic controller to refer to.

[0014] Preferably, the central processing unit includes: The central processing unit adopts a distributed computing architecture, including a data acquisition layer, a business logic layer, and a decision application layer. The data acquisition layer is responsible for interfacing with the underlying sensor network to receive, cache, and perform preliminary verification of data; the business logic layer carries the core data processing, model inference, and risk analysis tasks; and the decision application layer is responsible for knowledge base management, human-computer interaction, issuing control commands, and constructing quality archives.

[0015] Secondly, the present invention also provides a skincare product end-to-end testing device, including a central processing unit and a computer program. When the central processing unit executes the computer program, it implements the production control method of the skincare product end-to-end testing device as described above.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects: This invention integrates a multi-source sensor network deployed across key processes, enabling synchronous, continuous, and high-precision acquisition of critical parameters in raw material feeding, mixing, and filling / packaging. This effectively solves the data isolation and information lag problems caused by traditional offline sampling and testing methods. Through deep cleaning, feature extraction, and fusion analysis of multi-source data, and by utilizing machine learning models to construct a dynamic process status assessment mechanism, the complex nonlinear relationship between process parameters and final quality indicators is accurately characterized, thereby achieving early identification of subtle process fluctuations and potential quality risks. Crucially, by introducing cross-process collaborative analysis and Bayesian networks, this invention proactively simulates the propagation path of quality risks between processes and generates tiered early warning signals. This not only significantly improves the success rate of real-time interception of defective products but also fundamentally enhances the ability to trace the root causes of quality problems. Furthermore, the system's embedded adaptive control mechanism and continuous online learning function enable it to dynamically optimize process parameters based on real-time operating conditions and adapt to frequent formula changes, significantly improving the flexibility and intelligence of the production line. Simultaneously, the traceable quality archive built based on full-process data association provides solid data support for quality incident analysis, continuous process optimization, and production decision-making.

[0017] In summary, this invention not only significantly improves the real-time nature, accuracy, and reliability of quality control in skincare product production, but also provides an effective solution for the digital and intelligent upgrading of the process manufacturing industry. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a production control method for a skincare product end-to-end testing equipment.

[0019] Figure 2 This is a general framework diagram of a production control method for a full-process testing equipment for skincare products.

[0020] Figure 3 This is a schematic diagram of the central processing unit in a production control method for a skincare product end-to-end testing equipment. Detailed Implementation

[0021] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. It should be noted that relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0022] Example 1 Please see Figures 1-3 This embodiment provides a production control method for a skincare product end-to-end testing equipment, including: Step 1: Acquire multi-source detection data throughout the entire production process. By deploying a sensor network at each key process, including raw material feeding, mixing and reaction, filling and packaging, detection data for each key process is collected and multi-source detection data is generated. Step 2: Preprocess and extract features from the multi-source detection data, use data cleaning algorithms to remove outliers, and use feature engineering methods to extract key feature vectors that characterize the process state to form a standardized feature dataset; Step 3: Construct a dynamic process status assessment model. Based on the standardized feature dataset, use machine learning algorithms to establish a nonlinear mapping relationship between process parameters and quality indicators, and output the process status score and deviation in real time. Step 4: Perform cross-process collaborative analysis and risk warning, integrate the process status assessment results of each process, identify potential quality risk propagation paths through multi-source information fusion algorithms, and generate graded warning signals; Step 5: Implement adaptive production control and quality traceability. Dynamically adjust relevant process parameters based on the risk warning signals, record the data correlation of the entire process, and construct a traceable quality archive.

[0023] Specifically, step 1 includes: The sensor network includes a high-precision flow sensor, an online spectrometer, a machine vision inspection unit, and a temperature and pressure sensor array. The high-precision flow sensor is deployed at the inlet of the raw material feeding pipeline to monitor the instantaneous flow rate and cumulative feeding amount of various liquid or paste raw materials in real time. Its measurement accuracy is no less than 0.5%, and its range covers 0.1 liters per minute to 50 liters per minute to meet the precise metering requirements of materials with different viscosities. The online spectrometer is installed in the observation window or bypass flow cell of the mixing reactor. It uses near-infrared or Raman spectroscopy to perform non-contact online analysis of chemical parameters such as component concentration, moisture content, and active ingredient ratio of the materials in the reactor. The spectral scanning frequency is no less than 1 Hz, and the wavelength resolution is better than 2 nm. The machine vision inspection unit consists of a high-resolution industrial camera, a ring LED light source, and an image processing industrial control computer. It is deployed at the raw material barrel label recognition station, the mixing reactor liquid level observation point, below the filling head, and the finished product sealing inspection area to capture static or dynamic high-definition images to meet the requirements for accurate identification of visual features such as label information, liquid level, filling volume, and sealing integrity. The temperature and pressure sensor arrays are distributed and embedded in the walls of the mixing reactor, pipe connections, and filling chamber, monitoring temperature and pressure changes at key points in real time. The temperature sensors have a measurement range of -20 to 150°C and an accuracy of ±0.1°C, while the pressure sensors have a range of 0 to 2 MPa and an accuracy of 0.25% of full scale. All data collected by the sensors is uploaded to the central processing unit in real time via the industrial Ethernet protocol. The data acquisition frequency is uniformly set to no less than 10Hz to ensure complete capture of rapidly changing processes, thereby constructing a multi-source detection data system with strictly aligned timestamps, clear spatial locations, and rich data types.

[0024] Specifically, step 2 includes: For text data, the data cleaning algorithm employs a sliding window-based statistical process control method. This method first establishes an independent sliding window for each sensor channel. The window size is not fixed but adaptively adjusted according to the current production line's cycle time. For example, when the production line speed increases and the cycle time shortens to 30 seconds, the window size automatically adjusts to 300 sampling points (corresponding to 30 seconds multiplied by 10Hz); when the cycle time slows to 60 seconds, the window size expands to 600 sampling points. Within each sliding window, the mean μ and standard deviation σ of the data are calculated, and any values ​​exceeding σ are excluded. Data points within a specified range are identified as outliers and removed. The value of k is dynamically optimized based on historical data distribution characteristics, initially set to 3. For image data, a method combining morphological filtering and background subtraction is used to remove noise and irrelevant background interference. After data cleaning, the feature engineering stage begins. Feature engineering methods include principal component analysis (PCA) and time-domain / frequency-domain feature extraction; PCA is used for dimensionality reduction, and time-domain / frequency-domain feature extraction is used to capture dynamic characteristics. For high-dimensional spectral data or vectors composed of multiple physical parameters, principal component analysis is applied to retain principal components with a cumulative contribution rate of no less than 95%, and the original high-dimensional data is projected into a low-dimensional feature space, effectively eliminating redundant information and reducing the computational complexity of subsequent models. For time-series flow, temperature, and pressure data, time-domain feature extraction is performed, including calculating statistics such as mean, variance, skewness, kurtosis, maximum value, minimum value, and rate of change within the sliding window. Simultaneously, frequency-domain feature extraction is performed by performing a Fast Fourier Transform on the data to obtain its energy proportion and dominant frequency in specific frequency bands, used to characterize the stability and oscillation characteristics of the process. Finally, all extracted feature vectors are Z-score standardized to form a standardized feature dataset with fixed dimensions and uniform format, which serves as the input for subsequent models.

[0025] Specifically, step 3 includes: The dynamic process status assessment model employs a gradient boosting decision tree algorithm. This model takes the standardized feature dataset generated in step 2 as input and outputs two core indicators: a process stability index and a component uniformity score. The process stability index comprehensively reflects the smoothness of the current process operation, ranging from 0 to 100, with higher values ​​indicating greater stability. The component uniformity score quantifies the homogeneity level of the mixture, also ranging from 0 to 100. The training data for the dynamic process status assessment model comes from qualified and unqualified samples from historical production batches that have undergone offline laboratory verification. A mapping between features and quality indicators is established through supervised learning. To address frequent formula changes in actual production, the dynamic process status assessment model incorporates an online learning mechanism. Whenever a new, verified production batch of data is generated, this data is added to the training set, triggering an incremental update of the dynamic process status assessment model. The update process uses a mini-batch gradient descent method, iterating only a limited number of times on the new data. This prevents the dynamic process status assessment model from forgetting old knowledge due to overfitting to new data, thus ensuring that the dynamic process status assessment model can continuously adapt to new formula systems. The dynamic process status assessment model is deployed in the business logic layer of the central processing unit. It receives real-time feature data streams from the data acquisition layer and outputs process status scores and deviations at a frequency of no less than 1Hz, providing a basis for decision-making for subsequent risk warnings.

[0026] Specifically, step 4 includes: The multi-source information fusion algorithm employs a Bayesian network structure. A Bayesian network is a directed acyclic graph (DAG), where each node corresponds to the state variables of the three core processes: raw material input, mixing reaction, and filling / packaging. The state of each node is a comprehensive representation of the process stability index and component uniformity score output in step 3. The network edges represent the influence relationships between processes. For example, deviations in raw material input flow rate directly affect the mixing accuracy, thus impacting the component consistency of the final packaged product. These influence relationships are quantified using conditional probability tables. The values ​​in the conditional probability tables are derived from a large amount of historical production data, reflecting the probabilistic influence of upstream process states on downstream process states. During real-time operation, the system inputs the latest state evaluation results of each process into the Bayesian network and uses Bayes' theorem to calculate the posterior probability of each node. When the posterior probability of a key quality node (such as the final product qualification rate) falls below a preset safety threshold, the system determines that a quality risk exists and generates a graded early warning signal based on the degree to which the posterior probability deviates from the threshold. Warning levels are typically divided into three levels: Level 1 (yellow) indicates a potential risk requiring operator attention; Level 2 (orange) indicates a relatively high risk, suggesting minor process adjustments; and Level 3 (red) indicates an extremely high risk, requiring immediate shutdown for investigation. The warning signals and their associated risk propagation path analysis results are pushed in real-time to the human-machine interface and the decision-making application layer of the central processing unit.

[0027] Specifically, step 5 includes: The adaptive production control employs a fuzzy logic controller. The controller's inputs are the process status score output in step 3 and the warning level generated in step 4, with the output being the adjustment amount sent to the actuator. The fuzzy logic controller internally defines multiple fuzzy rules, such as "if the process stability index is low and the warning level is level two, increase the stirring motor speed" or "if the component uniformity score is low and the warning level is level one, fine-tune the raw material flow rate setpoint." These rules convert precise numerical inputs into linguistic variables (such as "low," "medium," and "high"), and after fuzzy inference, a defuzzification process (such as the center-of-gravity method) yields precise control outputs, ensuring that process parameters are always dynamically adjusted and maintained within the optimal operating range. Simultaneously, the system synchronously executes a quality traceability function. The construction of a traceable quality archive specifically includes: assigning a globally unique 128-bit identifier to each initiated production batch; the identifier is associated with all data in that batch and stored in the distributed database of the central processing unit; based on this, a time-series-based data association graph is established, where nodes represent various data entities, and edges represent the temporal or causal relationships between them. This map supports forward tracing (from raw materials to all affected finished products) and backward tracing (from problematic finished products to specific raw material batches or abnormal process parameters), providing a solid data foundation for quality incident analysis and continuous process improvement.

[0028] Furthermore, the method also includes establishing a process knowledge base. This knowledge base is stored in the decision application layer of the central processing unit, and its content covers massive amounts of historical production data, empirical rules summarized by domain experts, and past successful process optimization cases. During production, the system uses case-based reasoning technology to perform similarity matching between the current production situation (including formula, environment, equipment status, etc.) and historical cases in the knowledge base, retrieving the most similar cases and using their corresponding optimization strategies or risk avoidance measures as auxiliary decision-making suggestions. These suggestions are then pushed to operators or directly provided to the fuzzy logic controller for reference, thereby improving the system's intelligence level and decision-making quality.

[0029] When the method is applied to a continuous production line, redundant detection nodes are set up in the system to ensure absolute continuity of detection. For example, at the critical filling volume detection station, in addition to the machine vision detection unit, a backup laser rangefinder sensor array is deployed in parallel to monitor the health status of the main detection channel in real time. Once a fault such as data loss, image blurring, or communication interruption is detected in the main channel, the built-in fault diagnosis module will complete the fault confirmation within 10ms and complete the control switch within the following 90ms, automatically and seamlessly migrating the data acquisition and analysis tasks to the backup channel. The entire switching response time is strictly controlled within 100ms to ensure that the production line does not need to stop and the detection data flow is not interrupted.

[0030] The central processing unit adopts a distributed computing architecture, comprising a data acquisition layer, a business logic layer, and a decision application layer. The data acquisition layer interfaces with the underlying sensor network to receive, cache, and perform preliminary verification of data. The business logic layer handles core data processing, model inference, and risk analysis tasks. The decision application layer is responsible for knowledge base management, human-computer interaction, issuing control commands, and building quality archives. Data exchange between layers occurs through standard interface protocols. This loosely coupled architecture allows for modular expansion of system functionality. For example, when a new detection algorithm needs to be introduced, only a new microservice module needs to be developed in the business logic layer and registered to the service bus, without modifying the code in the data acquisition layer or the decision application layer, greatly improving the system's maintainability and scalability.

[0031] Example 2 This embodiment provides a skincare product end-to-end testing device, including a central processing unit and a computer program. When the central processing unit executes the computer program, it implements the production control method of the skincare product end-to-end testing device as described above.

[0032] All content not described in detail in this specification is prior art known to those skilled in the art, and the model parameters of each electrical appliance are not specifically limited; conventional equipment can be used. Electrical control components not mentioned in this technical solution are not shown in the figures because they are prior art, and will not be described further here.

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A production control method for a skincare product end-to-end testing equipment, characterized in that, include: Step 1: Acquire multi-source detection data throughout the entire production process. By deploying a sensor network at the raw material feeding, mixing and reaction, and filling and packaging stages, detection data from each process is collected and multi-source detection data is generated. Step 2: Preprocess and extract features from the multi-source detection data, use data cleaning algorithms to remove outliers, and use feature engineering methods to extract feature vectors that characterize the process state to form a standardized feature dataset; Step 3: Construct a dynamic process status assessment model. Based on the standardized feature dataset, use machine learning algorithms to establish a nonlinear mapping relationship between process parameters and quality indicators, and output the process status score and deviation in real time. Step 4: Perform cross-process collaborative analysis and risk warning, integrate the process status assessment results of each process, identify potential quality risk propagation paths through multi-source information fusion algorithms, and generate graded warning signals; Step 5: Implement adaptive production control and quality traceability. Dynamically adjust relevant process parameters based on the risk warning signals, record the data correlation of the entire process, and construct a traceable quality archive.

2. The production control method for a skincare product end-to-end testing equipment according to claim 1, characterized in that: The sensor network includes: The sensor network includes a high-precision flow sensor, an online spectral analyzer, a machine vision inspection unit, and a temperature and pressure sensor array. The high-precision flow sensor is deployed at the inlet of the raw material feeding pipeline to monitor the instantaneous flow rate and cumulative feeding amount of various liquid or paste raw materials in real time. The online spectrometer is installed in the observation window or bypass flow cell of the mixing reactor and uses near-infrared or Raman spectroscopy to perform non-contact online analysis of the component concentration, moisture content and active ingredient ratio of the materials in the reactor. The machine vision inspection unit includes a high-resolution industrial camera, a ring LED light source, and an image processing industrial control computer, which are respectively deployed at the raw material barrel label recognition station, the mixing kettle liquid level observation point, below the filling head, and the finished product sealing inspection area. It is used to capture static or dynamic high-definition images and to identify label information, liquid level height, filling volume, and sealing integrity. The temperature and pressure sensor array is distributed and embedded in the wall of the mixing reactor, pipe connections and filling chamber, for real-time monitoring of temperature and pressure changes; The data collected by the sensor network is uploaded to the central processing unit in real time via the industrial Ethernet protocol and constructed as multi-source detection data.

3. The production control method for a skincare product end-to-end testing equipment according to claim 1, characterized in that: The data cleaning algorithm includes: For text data, the data cleaning algorithm adopts a statistical process control method based on sliding windows; an independent sliding window is established for each sensor channel, and the window size is not a fixed value, but is adaptively adjusted according to the current production line's process cycle time; For image data, a method combining morphological filtering and background subtraction is used to remove noise and irrelevant background interference.

4. The production control method for a skincare product end-to-end testing equipment according to claim 1, characterized in that: The feature engineering method includes: Feature engineering methods include principal component analysis and time-domain and frequency-domain feature extraction; wherein, the principal component analysis is used for dimensionality reduction, and the time-domain and frequency-domain feature extraction is used to capture dynamic characteristics; For spectral data or vectors composed of physical parameters, principal component analysis is applied to retain principal components with a cumulative contribution rate of not less than 95%, and the original data is projected into a low-dimensional feature space to eliminate redundant information and reduce computational complexity. For time-series data, time-domain feature extraction and frequency-domain feature extraction are performed. By performing a fast Fourier transform on the data, the energy proportion and dominant frequency in a specific frequency band are obtained.

5. The production control method for a skincare product end-to-end testing equipment according to claim 2, characterized in that: The dynamic process status assessment model includes: The dynamic process status evaluation model adopts the gradient boosting decision tree algorithm, takes the standardized feature dataset generated in step 2 as input, and outputs a process stability index and a component uniformity score. The process stability index is used to reflect the smoothness of the current process operation, and the component uniformity score is used to quantify the uniformity level of the mixture. The dynamic process status assessment model establishes a mapping between features and quality indicators through supervised learning. In addition, the dynamic process status assessment model introduces an online learning mechanism. When new production batch data is generated, the new production batch data is added to the training set, and the incremental update of the dynamic process status assessment model is triggered. The update process adopts the mini-batch gradient descent method, and only performs a limited number of iterations on the new data to avoid overfitting of the dynamic process status assessment model.

6. The production control method for a skincare product end-to-end testing equipment according to claim 5, characterized in that: Step 4 includes: The multi-source information fusion algorithm adopts a Bayesian network structure; the network nodes of the Bayesian network correspond to the state variables of the raw material feeding, mixing reaction, filling and packaging processes, and the state of the node corresponds to the process stability index and component uniformity score; the edges of the Bayesian network correspond to the influence relationship between processes. During real-time operation, the latest status assessment results of each process are input into the Bayesian network, and the posterior probability of each node is calculated using Bayes' theorem. When the posterior probability of a node is lower than the preset safety threshold, a quality risk is determined, and a graded warning signal is generated based on the degree to which the posterior probability deviates from the threshold. The warning levels are usually divided into three levels: Level 1 warning indicates that there is a potential risk and requires operator attention; Level 2 warning indicates that the risk is relatively high and process fine-tuning is recommended; Level 3 warning indicates that the risk is extremely high and the machine must be stopped immediately for investigation.

7. The production control method for a skincare product end-to-end testing equipment according to claim 6, characterized in that: Step 5 includes: The adaptive production control adopts a fuzzy logic controller; the input of the fuzzy logic controller is the process status score and the early warning level, and the output is the adjustment amount sent to the actuator. The fuzzy logic controller includes fuzzy rules to convert numerical inputs into linguistic variables; after fuzzy inference, it generates control outputs through defuzzification to ensure that process parameters are maintained within the optimal operating range. In addition, a globally unique identifier is assigned to each production batch; the identifier is associated with all data of the corresponding batch and stored in the distributed database of the central processing unit; A data association map is constructed based on the identification code. The data association map supports forward tracing and backward tracing, providing a data foundation for quality accident analysis and continuous process improvement.

8. The production control method for a skincare product end-to-end testing equipment according to claim 7, characterized in that: The process knowledge base includes: The process knowledge base includes historical production data, empirical rules, and process optimization cases. During the production process, the current production situation is matched with the process optimization cases in the process knowledge base using case reasoning technology. The best case is retrieved, and its corresponding optimization strategy or risk avoidance measures are used as auxiliary decision-making suggestions for the fuzzy logic controller to refer to.

9. The production control method for a skincare product end-to-end testing equipment according to claim 8, characterized in that: The central processing unit includes: The central processing unit adopts a distributed computing architecture, including a data acquisition layer, a business logic layer, and a decision application layer. The data acquisition layer is responsible for interfacing with the underlying sensor network to receive, cache, and perform preliminary verification of data. The business logic layer carries the core data processing, model reasoning, and risk analysis tasks. The decision application layer is responsible for knowledge base management, human-computer interaction, issuing control commands, and constructing quality archives.

10. A skincare product end-to-end testing device, characterized in that: It includes a central processing unit and a computer program, wherein when the central processing unit executes the computer program, it implements the production control method of a skin care product full-process testing equipment as described in any one of claims 1-9.

Images