Digital metrology method based on internet of things and digital twinning

Abstract

This invention discloses a digital metrology method based on the Internet of Things (IoT) and digital twins, belonging to the field of digital metrology technology. This method collects multi-dimensional parameters of measuring instruments in real time through the IoT sensing layer, and stores the data in a health monitoring database after data preprocessing to ensure data quality. A digital twin model is constructed by combining the actual parameters of the instrument, realizing real-time linkage between the physical instrument and the virtual model. An artificial intelligence model is used to complete instrument anomaly detection, health assessment, and prediction of remaining service life, and the metrology results and equipment status are presented intuitively through a visualization platform. This invention realizes the transformation of metrology from instrument management to data management, improves metrological accuracy and efficiency, reduces operation and maintenance costs, and is suitable for high-precision metrology scenarios such as pharmaceutical company laboratories.

Description

Technical Field

[0001] This invention relates to the field of digital metrology technology, specifically to a digital metrology method based on the Internet of Things and digital twins. Background Technology

[0002] In today's pharmaceutical industry, research and application of digital metrology technology in pharmaceutical laboratory settings is of great significance. With the booming development of the global pharmaceutical industry, digital transformation has become an inevitable trend. Under these industry conditions, the requirements for the accuracy of experimental data and the efficiency of experimental procedures are becoming increasingly stringent. However, the application level of digital metrology among pharmaceutical companies is currently uneven, and most companies have not yet built a complete digital metrology system.

[0003] In practical applications, pharmaceutical companies have an extremely urgent need for accurate and reliable experimental data. Drug research and development and production are complex processes, and even minor deviations in experimental data can affect drug quality, thereby jeopardizing patient treatment outcomes. For example, in experiments determining drug component content, inaccurate measurement may result in the drug's active ingredient levels falling short of standards. Simultaneously, traditional manual recording and operation methods are inefficient and prone to data errors and loss. Pharmaceutical companies face problems such as data inconsistency and difficulty in data retrieval in data management, severely hindering research and development and production progress.

[0004] Existing technologies have many shortcomings. Traditional metrological methods are prone to significant human error, making it difficult to control errors in real time during high-precision experiments. The measurement process also lacks transparency, making traceability and monitoring difficult. Furthermore, existing technologies are deficient in multi-parameter integrated detection, with poor data compatibility between different brands and models of measuring instruments, resulting in data silos and failing to meet the pharmaceutical process's needs for simultaneous monitoring of multiple physical quantities and integrated data analysis.

[0005] Currently, metrological instruments play a crucial role in scientific research, technological development, product quality control, and health and safety. However, with increasing instrument usage time, issues such as component wear and improper operation can lead to a gradual increase in the deviation between measurement results and expected values, seriously affecting metrological accuracy. Traditional metrological instrument health monitoring mainly relies on manual analysis and periodic testing, which has the following problems:

[0006] The detection lag is prominent: the traditional mode of relying on manual analysis and periodic testing cannot achieve real-time monitoring, resulting in the accumulation of instrument errors and difficulty in timely detection of sudden failures, which seriously affects the continuity and reliability of metrology tasks.

[0007] High maintenance costs: Manual testing requires frequent and complex comparative experiments, which consumes a lot of professional human resources, and the operation process is cumbersome and inefficient.

[0008] Technical limitations: Existing methods (such as grey theory modeling) rely on linear assumptions, making it difficult to capture the nonlinear dynamic characteristics of instruments and resulting in insufficient prediction accuracy; at the same time, they lack the ability to fuse and analyze multi-source heterogeneous data (such as sensor data and environmental parameters).

[0009] Data value not being realized: massive amounts of historical measurement data have not been effectively mined, and there is a lack of intelligent analysis tools, making it impossible to provide data-driven decision support for instrument health assessment. Summary of the Invention

[0010] The purpose of this invention is to provide a digital measurement method based on the Internet of Things and digital twins, which can effectively solve the problems in the background art mentioned above.

[0011] To achieve the above objectives, the technical solution adopted by this invention is: a digital measurement method based on the Internet of Things and digital twins, comprising the following steps:

[0012] S1: Construct an IoT sensing layer, deploy sensor nodes and gateway devices adapted to biological and chemical measuring instruments, establish communication connections between instruments and data transmission networks, and collect instrument operating parameters and detection environment parameters in real time.

[0013] S2: Preprocess the collected parameter data, including cleaning up pseudo-anomaly data, data amplification and standardization, and store the preprocessed data in the constructed health monitoring database;

[0014] S3: Based on the actual geometric dimensions and performance parameters of the measuring instrument, a digital twin model of the instrument is constructed using a 3D modeling tool to realize real-time data interaction between the digital twin model and the physical instrument;

[0015] S4: Based on data from the health monitoring database, construct an artificial intelligence metrological analysis model that includes a pressure anomaly detection model, a whole machine health calculation model, and a health age and remaining service life prediction module to complete the instrument operation status analysis;

[0016] S5: Map the output of the AI ​​metrology analysis model to the digital twin model, and realize the twin visualization of instrument operation status, metrology results and abnormal warning information through the visualization platform.

[0017] Preferably, in step S1, the sensor nodes include a pressure sensor, a flow sensor, a temperature sensor, a humidity sensor, and a pressure sensor, which respectively collect the pump pressure, infusion flow rate, ambient temperature, ambient humidity, and chamber pressure of the measuring instrument; the gateway device supports RS-232, USB, and Ethernet communication protocols, and can be adapted to different brands and models of biological and chemical measuring instruments, ensuring that the instrument networking rate is not less than 60%.

[0018] Preferably, the pseudo-anomaly data cleaning in step S2 employs a combined detection method:

[0019] First, determine whether the fluctuation range of data at adjacent time points exceeds an adaptive threshold, and then mark the data with excessive fluctuation.

[0020] Then, by calculating the degree of variation of the data within the sliding window, data with excessive variation are marked; the two types of marked data are identified as pseudo-anomalies and removed.

[0021] Preferably, the pressure anomaly detection model is a RePre-LSTM model, which consists of two sub-networks: an LSTMA reconstruction module and an LSTMS prediction module.

[0022] The LSTM reconstruction module learns the time series through an encoder-decoder structure; the encoder part extracts temporal features by two stacked LSTM layers and outputs the hidden state; the hidden state output by the upper layer is mapped to a low-dimensional space through a fully connected layer; the decoder part restores the original sequence from the hidden vector in the upper low-dimensional space through decoding by a fully connected layer and two LSTM layers; finally, the decoder output is converted into a reconstructed value with the same dimension as the input sequence through a fully connected layer.

[0023] The LSTM prediction module predicts the data for the next n time points of the input sequence. The module consists of two parts: an encoder and a predictor. The encoder part uses two layers of LSTM to extract temporal features, and the predictor part connects the hidden output state of the encoder to a fully connected layer to predict the data for the next n time points.

[0024] Preferably, the overall health calculation model in step S4 is based on the principle of relative combined uncertainty, with relative combined uncertainty, pressure anomaly rate, and qualitative / quantitative repeatability as core indicators, and uses a weighted scoring method to calculate the health score;

[0025] The health score ranges from 0 to 100. The higher the score, the more stable the instrument's operation. The weight of each core indicator is adaptively adjusted according to the characteristics of different types of measuring instruments.

[0026] Preferably, the specific process of constructing the digital twin model in step S3 is as follows:

[0027] Using 3ds Max and Maya 3D modeling tools, a 3D model is constructed based on the actual geometric data of the measuring instrument, and an FBX format file is generated.

[0028] Import the FBX file into the Unity3D platform to develop interactive functions for the instrument's movable parts, including switching, rotation, and dragging / scaling.

[0029] Design the measurement information data structure and data interface to achieve the connection with sensor data of the Internet of Things sensing layer and complete the synchronization of the operating status of physical instruments and digital twin models.

[0030] Preferably, in step S3, the real-time data interaction frequency between the digital twin model and the physical instrument is not less than 1 time / second, ensuring that the mapping delay of the digital twin model to the operating status of the physical instrument is less than 100ms.

[0031] In step S5, the visualization platform displays the following content: instrument 3D model, real-time operating parameters, health score change curve, health age change trend, and abnormal warning list. The abnormal warning list indicates the abnormality type, the time of occurrence of the abnormality, and preliminary handling suggestions.

[0032] Preferably, the dynamic weights are calculated using the Sigmoid function for adjustment. The adjustment process comprehensively considers the data fluctuation amplitude, local high-frequency energy, the slow trend of prediction deviation, and the slope of the data within the sliding window, so that the dynamic weights can be adaptively adjusted according to the characteristics of the instrument's operating data, thereby improving the accuracy of anomaly detection.

[0033] Preferably, it also includes data security protection steps:

[0034] The data in the health monitoring database is encrypted using the AES-256 encryption algorithm, and the data transmitted between the IoT sensing layer and the data transmission network is encrypted using the TLS protocol.

[0035] Set up tiered user access permissions to differentiate the operation permissions of administrators, inspectors, and observers, and record the data access and operation logs of all users to ensure the integrity and traceability of measurement data.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] 1. A New Practice in the Transformation of Metrology from Instrument Management to Data Management: This invention achieves a transformation in metrology management from traditional instrument management to data management through the deep integration of IoT and digital twin technologies. By collecting experimental data in real time using intelligent sensors and combining this with digital twin technology for simulation and analysis, a mathematical model reflecting the performance of the measurement system throughout its entire lifecycle is constructed, enabling real-time monitoring, evaluation, and early warning of the system's effectiveness. This transformation not only improves the accuracy and reliability of data but also optimizes data acquisition and experimental efficiency, providing a new practical path for intelligent metrology management. Furthermore, the project has established a data sharing platform, established data protocols, and adopted edge computing technology for data storage and analysis, further promoting the digitalization and intelligentization of metrology management.

[0038] 2. Application of Small-Sample Fine-Tuning Technology in Instrument Health Analysis Model: In constructing the instrument health analysis model, the project innovatively adopted small-sample fine-tuning technology. Initial data was accumulated through a cold-start machine learning model, and the model's accuracy was gradually improved by combining it with a deep neural network, forming a progressive optimization path from a small amount of data to a large amount of data. This technology not only solved the problem of insufficient initial data but also created a positive feedback loop by feeding back the model's predictions with subsequent data, significantly improving the model's predictive ability and adaptability. Furthermore, the project optimized the model's operating efficiency through model compression technology, ensuring that prediction speed was improved while maintaining accuracy, providing efficient technical support for instrument health analysis.

[0039] 3. Instrument Health Twin Visualization: The project utilizes digital twin technology to visualize the health status of instruments. Professional 3D software is used to precisely construct 3D models of instruments, equipment, and the laboratory environment, and platforms such as Unity3D enable natural interaction between users and the virtual model. Through sensor technology and industrial IoT communication protocols, deep integration and real-time synchronization of the physical and information models are achieved, ensuring that the virtual model accurately reflects changes in the physical equipment's status. This visualization technology not only provides users with intuitive monitoring and decision-making support but also achieves dynamic consistency between the virtual model and the physical equipment through a two-way real-time interactive mechanism, offering a novel visualization solution for instrument health management. Attached Figure Description

[0040] Figure 1 This is a diagram of the algorithm framework of this invention;

[0041] Figure 2 This is a diagram showing the effect of training data preprocessing—pseudo-anomaly data filtering in this invention;

[0042] Figure 3 This is a structural diagram of the anomaly detection model of the present invention;

[0043] Figure 4 This is a diagram showing the effect of the detection model of this invention;

[0044] Figure 5 This is a structural diagram of the health age prediction module of the device of the present invention;

[0045] Figure 6 This is a health and lifespan curve of the present invention. Detailed Implementation

[0046] Please refer to Figure 1-6 This embodiment is mainly divided into the following five parts:

[0047] 1. Software Platform Development and Construction: One of the core aspects of this invention is the development and construction of a fully functional software platform. This platform integrates IoT, edge computing, and digital twin technologies to achieve automated acquisition and processing of key input and output quantities of measuring instruments. The platform adopts a modular design to ensure loose coupling between functional modules, facilitating subsequent expansion and maintenance. By establishing data protocols, the platform enables real-time data storage and analysis, providing a fundamental guarantee for intelligent measurement system monitoring. Furthermore, the platform supports multi-terminal synchronization, allowing users and supervisors to view instrument status and data analysis results in real time, improving management efficiency.

[0048] 2. Selection and Integration of Key Parameter Sensors: In selecting and integrating key parameter sensors, the project prioritized sensor accuracy, stability, and compatibility. By using high-precision sensors and combining them with edge computing technology, real-time monitoring of multiple parameters such as pressure, temperature, flow rate, and current was achieved. Sensor data is automatically collected through an IoT platform and compared with standard models to ensure the accuracy and reliability of measurement results. Furthermore, the project developed standardized traceability technology, further improving the quality of sensor data through real-time calibration and error correction.

[0049] 3. Instrument and Equipment Data Exchange Protocol Standard: To achieve data interoperability between different instruments and equipment, the project established a unified data exchange protocol standard. This standard defines the data transmission format, interface specifications, and communication protocols, ensuring seamless access to the IoT platform for various instruments and equipment. By establishing a unified data protocol, the platform enables the automated acquisition and import of key instrument parameters, providing a foundation for subsequent data analysis and model building. Furthermore, the standardized protocol supports compatibility with devices from multiple vendors, facilitating system expansion and application promotion.

[0050] 4. Establishing an Instrument Health Data Model: This invention constructs a mathematical model reflecting the instrument's performance throughout its entire lifecycle, based on an instrument health data model. Initial data is accumulated through cold-start data amplification, and the model's accuracy is progressively optimized using deep neural networks, achieving precise prediction of the instrument's overall error. Furthermore, the project establishes a correlation model between key parameters and errors, providing a scientific basis for real-time monitoring and early warning of the instrument's health status.

[0051] 5. Instrument Health Information Twin Visualization Application: In the area of ​​instrument health information twin visualization, the project utilized professional 3D software to construct high-precision 3D models of instruments, equipment, and the laboratory environment. Through platforms such as Unity3D, it achieved natural interaction between users and the virtual model. By employing sensor technology and industrial IoT communication protocols, it achieved deep integration and real-time synchronization between the physical and information models, ensuring that the virtual model accurately reflects the status changes of the physical equipment. This visualization technology not only provides users with intuitive monitoring and decision-making support but also achieves dynamic consistency between the virtual model and the physical equipment through a two-way real-time interaction mechanism, providing a novel visualization solution for instrument health management.

[0052] As a preferred embodiment, please refer to Figure 1 It demonstrates the framework structure for pressure anomaly detection, instrument health assessment, and lifespan prediction.

[0053] The data layer primarily contains sensor data (pressure, flow rate, etc.) and chromatographic analysis data (such as peak area, retention time, etc.). This data provides multi-source heterogeneous time-series input for subsequent model analysis.

[0054] The model layer is the core processing part of the system, divided into two stages: model training and model testing. Before model training, the training data (historical data) should be preprocessed, including anomaly cleaning, data augmentation, and standardization. Data is crucial for improving model training effectiveness and generalization ability. An anomaly detection model is trained to identify abnormal points during instrument operation. A multi-feature fusion health assessment model is designed by combining instrument uncertainty and anomaly detection results. Based on the health estimation results, uncertainty, and other information, a health age prediction module is constructed.

[0055] The model is deployed on a server and integrated into an online detection platform. Real-time test data is standardized and then input into the model for anomaly detection, health assessment, and lifespan prediction.

[0056] The visualization layer refers to the intuitive display of the model's output results in the data twin platform, helping users to achieve anomaly monitoring and model result reliability assessment. The main content displayed includes: anomaly detection results, health assessment scores, and performance degradation prediction curves.

[0057] The process of constructing a mathematical model reflecting the performance throughout the instrument's entire lifecycle is as follows:

[0058] 1. Algorithm Selection: To accurately identify outliers in time-series data, this paper comprehensively considers the model's generalization ability, its ability to model temporal dependencies, and its performance in anomaly detection. The following commonly used time-series anomaly detection methods are selected for comparative analysis:

[0059] Methods based on moving mean and standard deviation:

[0060] This method dynamically detects outliers by calculating local mean and standard deviation within a time window. Its principle is simple, implementation is convenient, and it is suitable for time-series data scenarios with low noise and stable changes. However, this method is essentially a local statistical method and cannot capture long-term dependent patterns. It has weak detection capabilities for trending or non-stationary time-series signals, and there is a risk of false positives and false negatives.

[0061] Isolation forest-based methods:

[0062] Isolation forest is a typical tree-based unsupervised anomaly detection algorithm that constructs multiple random binary trees to identify easily isolated points as anomalies. This method has strong nonlinear modeling capabilities and can adapt to data distributions of a certain complexity. However, it ignores the sequentiality and dependencies of time sequences, cannot utilize temporal structure characteristics, and has limited performance in data with obvious temporal dependencies (such as equipment operating status and sensor monitoring data).

[0063] Prediction module based on Seq2SeqLSTM:

[0064] The Seq2Seq model uses historical time-series data as input to an encoder, employs an LSTM encoder to capture long-term dependent features, and a decoder to predict data for a future period. If the error between the predicted and observed values ​​is significant, anomalies are considered to exist. This method can model non-linear time patterns, but it relies solely on prediction errors to identify anomalies, and may be insensitive in scenarios where outliers do not significantly affect predictions (such as slow drift).

[0065] To address the shortcomings of the above methods, this paper proposes a joint detection method called RePre-LSTM, which combines LSTM autoencoder reconstruction error and LSTM prediction error.

[0066] LSTM effectively captures long-term time dependencies, making it particularly suitable for time-series data. This study selects an LSTM-based approach due to its strong ability to model time-series data. Combining reconstruction and prediction methods with LSTM effectively reduces the possibility of false alarms and missed detections. Since devices may exhibit potential anomalies during degradation, relying solely on the prediction module cannot detect them, while the reconstruction module can capture internal pattern changes. Therefore, this study chooses RePre-LSTM, which combines an LSTM reconstruction subnetwork and an LSTM prediction subnetwork, as the stress anomaly detection model.

[0067] 2. Pump Pressure Anomaly Detection Module: The following mainly introduces the data preprocessing method, model training method, and model testing method in the pump pressure anomaly detection module.

[0068] Pseudo-abnormal data processing:

[0069] The raw data often contains some pseudo-anomalies caused by human manipulation. Pseudo-anomalies include, for example,... Figure 2 As shown in (a), the pseudo-anomaly data after filtering is as follows: Figure 2 As shown in (b), Pressure and Flow vs Time represents the step change in pressure and flow rate values. Pressure (kPa) is the pressure (kilopascals), Flow (mL / min) is the flow rate (milliliters per minute), and Time is the time.

[0070] Current data cleaning methods for step data in training data combine temporal continuity assessment and anomaly filtering. If time points are discontinuous, the sequence is segmented. Anomaly filtering methods in training data primarily combine differential mutation detection and sliding window standard deviation volatility detection.

[0071] In the outlier removal method based on differential mutation detection, the adaptive threshold calculation method is as follows:

[0072] ;

[0073] in, This indicates the fluctuation percentage, with a default value of 0.015. It uses the top 10% of the largest values ​​in this time series as the adaptive threshold. It calculates the first-order difference sequence; if the change in value at the current time point compared to the value at the previous time point exceeds the threshold... If so, the data at that point in time is marked as abnormal.

[0074] The first-order difference sequence of the original sequence can be represented as:

[0075] ;

[0076] All Sort in ascending order to obtain the sorted sequence. Take the top 10% of the maximum values ​​in the sorted sequence and multiply them by the fluctuation ratio. As an adaptive threshold .

[0077] In the sliding window-based standard deviation volatility anomaly detection method, the ratio of the standard deviation to the mean of the pressure values ​​within each window is calculated to determine whether a set threshold is exceeded, thereby identifying abnormal time points. The calculation method for anomaly detection within the sliding window is as follows:

[0078] ;

[0079] ;

[0080] ;

[0081] in, This represents the pressure value at time point t. , Indicates the size of the sliding window. This indicates the total length of the time series. This represents the standard deviation within the t-th window. This represents the mean within the t-th window. This represents the coefficient of variation within the window. This is considered an anomaly; default setting. It is 0.015.

[0082] Data augmentation:

[0083] In the data preprocessing method, linear interpolation is used for data augmentation. When the length of the input sequence is less than 100, interpolation is performed to supplement the time points. The linear interpolation calculation method is as follows:

[0084] ;

[0085] Known The corresponding value for the time is , The corresponding value for the time is Then at time t (t is in) and The corresponding values ​​between (between) The calculation formula is as above.

[0086] Data standardization processing:

[0087] MinMaxScaler is used as a method for standardizing training and testing data, scaling each time series data to a specified interval [0,1] for the original features. The standardized calculation method is as follows:

[0088] ;

[0089] This represents the minimum value in the current input sequence. Corresponding to the maximum value, It corresponds The standardized result, after normalization, has a value between 0 and 1.

[0090] Anomaly detection model:

[0091] Please refer to Figure 3The anomaly detection model uses the improved algorithm RePre-LSTM, which mainly consists of two sub-networks: the LSTMAutoencoder reconstruction module and the LSTMSequencePredictor prediction module.

[0092] The LSTM reconstruction module learns time series data using an encoder-decoder structure. The encoder consists of two stacked LSTM layers extracting temporal features and outputting hidden states. Fully connected layers map the hidden states from the upper layers to a low-dimensional space. The decoder uses fully connected layers and two LSTM layers to decode the hidden vectors from the lower-dimensional space and reconstruct the original sequence. Finally, fully connected layers convert the decoder output into reconstructed values ​​of the same dimension as the input sequence.

[0093] The LSTM prediction module predicts the data for the next n time points from the input sequence. The model structure consists of two parts: an encoder and a predictor. The encoder uses a two-layer LSTM to extract temporal features, while the predictor connects the hidden output state of the encoder to a fully connected layer to predict the data for the next n time points.

[0094] Dual-module fusion determination mechanism:

[0095] The reconstruction module is more sensitive to intermittent fluctuations in the input sequence. Using the reconstruction method alone may significantly increase the reconstruction error when the instrument performance degrades. The prediction error is more sensitive to trend drift and slow degradation. Combining the two can serve as a complementary judgment, improve detection stability, and avoid missed or false detections caused by using the reconstruction module or prediction module alone.

[0096] The fusion method of the two modules is as follows: First, anomaly detection analysis is performed on the input temporal features to determine whether they are in the stage of sudden anomaly (reconstruction-oriented) or gradual degradation (prediction-oriented). Then, the weight parameters are dynamically adjusted based on the anomaly detection results. If the abnormality is a sudden anomaly, it will manifest as: As the error increases, the proportion of reconstruction error also increases. If the anomalous feature is a gradual degradation, it will manifest as... Decreasing the value increases the prediction error. Specifically:

[0097] I. Extracting 4 types of core quantization features

[0098] Based on input time-series data (such as pressure, flow rate), quickly calculate key features:

[0099] Instantaneous fluctuation amplitude This reflects sudden and abrupt changes;

[0100] Local high frequency energy Variance of data within a sliding window (default 30 steps), reflecting the intensity of local fluctuations;

[0101] Prediction Deviation Trend Exponentially weighted average of prediction errors ( ), capturing long-term degradation;

[0102] Window slope The absolute value of the slope of the linear fit of the sliding window data reflects the overall direction of change.

[0103] Rules for determining Type II and Type III anomalies

[0104] The thresholds are all derived from historical normal data statistics (such as the 95th percentile) and are dynamically updated.

[0105] Sudden abnormality: Exceeding the threshold (historical normal data) (95th percentile), and high, Low (no long-term trend);

[0106] Gradual degradation: Exceeding the threshold and satisfying the condition for 3 consecutive windows, Stable (0.02-0.2) Low (no jump);

[0107] Intermittent fluctuations: High but Low, Low (no degradation), with high-frequency fluctuations accounting for 30%-70% within the window.

[0108] In extreme cases, the optimal type is selected based on "feature matching degree" (core condition weight 0.6, auxiliary 0.4).

[0109] III. Confidence Validation and Weight Fitting

[0110] Confidence level: Total weight of satisfying conditions / Total weight of rules; ≥0.8 confirms the type, otherwise a second verification is required.

[0111] Weight Adjustments: Sudden anomalies (0.7-0.9, reconstruction), gradual degradation (0.1-0.3, reprediction), intermittent fluctuations (0.4-0.6, equilibrium).

[0112] Then obtain the fused error result. .

[0113] ;

[0114] ;

[0115] ;

[0116] and These represent the reconstruction error and prediction error at time t, respectively. and This represents the standardized reconstruction error and prediction error. , , , These represent the mean and standard deviation of the error of the reconstruction module on normal data, and the mean and standard deviation of the error of the prediction module on normal data, respectively. This represents the fusion error result at time point t.

[0117] Dynamic parameters The calculation method is as follows:

[0118] ;

[0119] in, Represents the data value at time t Compared to time t-1 Fluctuation range This is an adaptive threshold in differential mutation detection. This indicates a sudden change switch.

[0120] It represents local high-frequency energy, calculated using local variance.

[0121] It represents the slow-changing trend of prediction bias and is obtained by calculating the exponentially weighted average of the prediction error series. EMA is a method for analyzing trend changes.

[0122] This represents the absolute value of the slope within the window.

[0123] Representing the Sigmoid function:

[0124] ;

[0125] in, It is a fixed parameter used to adjust the degree of change.

[0126] To distinguish between different anomalous features in the input time series characteristics, such as sudden anomalies, gradual degradation, and intermittent fluctuations, a dynamic parameter fusion mechanism based on anomalous features is designed. This mechanism employs a complementary fusion mechanism for the output errors of the reconstruction and prediction modules, enabling the system to detect various anomalous changes and thus improving detection coverage and reliability.

[0127] Figure 4The image shows the performance of the pressure detection model. The left axis represents the normalized pressure value, and the right axis represents the prediction error. The horizontal axis indicates the number of time-series samples. The red curve (PressureTrueValue) represents the normalized actual pressure value time-series data. The blue dashed line (PressurePredict) represents the prediction result of the LSTM prediction module at the corresponding time point, and the orange dashed line (PressureReconstructed) represents the reconstruction result of the LSTM reconstruction module at the corresponding time point. The black curve (ReconstructionError) represents the reconstruction error, and the green curve (PredictionError) represents the prediction error. The black dashed line (ReconstructionThreshold) and the green dashed line (PredictionThreshold) represent the adaptive thresholds. Time points with errors exceeding the dashed lines can be considered outliers.

[0128] Anomaly detection index calculation:

[0129] RePre-LSTM uses a weighted fusion of reconstruction error and prediction error to achieve anomaly detection:

[0130] ;

[0131] in, and The two numbers represent the reconstruction error and prediction error at time t, respectively. The error calculation method uses MSE (mean squared error). This represents the fusion error result at time point t. The weighting coefficients represent the reconstruction and prediction errors.

[0132] The method for identifying outliers is the same as the method for identifying abnormal data in data preprocessing, that is, based on differential mutation detection and sliding window standard deviation volatility detection methods, a combination of two adaptive threshold judgment criteria is used to evaluate the model results. Perform anomaly detection.

[0133] By employing dual modeling of two sub-networks—reconstruction and prediction—and an adaptive thresholding method for outlier detection, the sensitivity to complex anomalies is enhanced.

[0134] Anomaly detection focuses on outlier data points. Therefore, outlier data points are associated with positive examples in the sample, denoted as P, while non-outlier points are associated with negative examples in the sample, denoted as N. Precision and recall are calculated as follows:

[0135] ;

[0136] ;

[0137] Among them, TP represents a true anomaly, that is, an anomaly is detected and is actually an anomaly; FP represents a false positive anomaly, that is, an anomaly is detected but is actually normal; FN represents a false negative, that is, an anomaly is actually detected but was not detected (missed detection).

[0138] Test metric comparison results:

[0139] This study compares commonly used anomaly detection methods based on moving mean and standard deviation in statistical data analysis, the isolated forest anomaly detection method in traditional machine learning, and the Seq2SeqLSTM prediction method commonly used in deep learning. The comparison results are as follows:

[0140] Table 1 Comparison of anomaly detection model results indicators

[0141]

[0142] The RePre-LSTM fusion reconstruction and prediction method described above is a further improvement on Seq2Seq and outperforms traditional machine learning methods such as isolated forest and moving standard deviation statistical methods.

[0143] 3. Overall health assessment: During operation, the key parameters of the chromatograph may be affected by factors such as environmental fluctuations, instrument aging, or component failure, resulting in unstable measured values. The overall health assessment uses statistical methods to analyze the fluctuations of each parameter, thereby obtaining the overall health result.

[0144] To achieve quantitative monitoring and health assessment of the chromatograph's operating status, a system of overall health indicators based on multi-feature fusion was constructed. This system mainly considers factors related to performance indicators and operating status, with key indicators including uncertainty analysis, pump pressure anomaly rate, peak area, and quantitative repeatability of retention time.

[0145] Minimum detection concentration:

[0146] The uncertainty of a liquid chromatograph is calculated based on the uncertainty at the minimum detectable concentration. The method for calculating the minimum detectable concentration is as follows:

[0147] ;

[0148] In the formula: The minimum detectable concentration (g / mL); Baseline noise peak height (AU); The standard solution concentration is expressed in g / mL. The standard solution chromatographic peak height (AU) is used. The injection volume is (μL).

[0149] The synthesis uncertainty of the minimum detection concentration requires calculating the uncertainties of baseline noise peak height, standard solution concentration, injection volume, and standard substance peak height.

[0150] Qualitative / quantitative repeatability calculation:

[0151] Qualitative / quantitative repeatability calculations primarily involve calculating the repeatability of peak area and retention time. This typically includes calculating the relative standard deviation. :

[0152] ;

[0153] In the formula: For qualitative / quantitative measurement repeatability relative standard deviation; For the first The retention time or peak area measured in this study; The arithmetic mean of the six measurements; Measurement serial number; For the number of times measured.

[0154] The uncertainty propagation rate is:

[0155] ;

[0156] Relative combination uncertainty:

[0157] ;

[0158] In the formula, each term represents a source of combined uncertainty in the liquid chromatography detection results: Indicates the uncertainty introduced by the standard reference material; This represents the uncertainty introduced by baseline noise; Indicates the uncertainty introduced by the chromatographic peak height; This indicates the uncertainty introduced by the microsyringe;

[0159] Relative expanded uncertainty:

[0160] ;

[0161] In the formula: Indicates the inclusion factor, taking =2;

[0162] Synthetic health calculation:

[0163] Let the health index be H. The health index is calculated as follows:

[0164] ;

[0165] in, Let F represent the relative combined uncertainty, F be the pressure anomaly rate, and R be the qualitative / quantitative repeatability. The default weight values ​​are... =0.55, =0.3, =0.15, and + + =1.

[0166] A comprehensive health assessment model is constructed by integrating relative combined uncertainty, pump pressure anomaly detection results, and the relative standard deviation of other key factors. Through weighted calculation of multiple indicators, a scoring index reflecting the overall health level of the equipment is output. This index can be used for long-term trend analysis, preventative maintenance, and equipment life prediction.

[0167] Health assessment test results:

[0168] The health score calculated by this model is compared with the experience-weighted scoring method used in a commercial system, and evaluated using 30 days of operational data from multiple devices. The indicators are as follows:

[0169] Table 2. Chromatograph health assessment stability

[0170]

[0171] Note: CV is the coefficient of variation, which is the ratio of the standard deviation to the mean.

[0172] The fluctuation rate of the health score indicates that the health score of this system is more stable, the accuracy of the instrument's overall fault warning is higher, and the false alarm rate is lower.

[0173] 4. Method for predicting the remaining service life of equipment:

[0174] Please refer to Figure 6 ,exist Figure 6 middle:

[0175] Health Index: The health index ranges from 0 to 1. A higher value indicates a more stable operating status of the device, while a lower value indicates a more significant performance degradation.

[0176] Year: A time dimension, covering the entire lifecycle of the device from 2020 to 2050;

[0177] Failure Threshold: A fixed threshold line. When the device health index falls below this line, the device is considered to have reached a failure state and needs maintenance or replacement.

[0178] Device A Curve: This refers to the health evolution trend of device A. It is characterized by a stable health index above 0.8 in the early stage, a gradual decline in the later stage, and approaching the failure threshold around 2038. The overall performance degradation rate is the slowest.

[0179] Device B Curve: Refers to the health evolution trend of device B. It is characterized by a rapid drop in the health index to around 0.6 in the early stage, a slower rate of decline in the middle stage, and the first to approach the failure threshold around 2032, with the fastest overall performance degradation.

[0180] Device C Curve: This curve represents the health evolution trend of device C. It is characterized by a uniform decline in the health index, falling between Device A and Device B. It approaches the failure threshold around 2035, reflecting a moderate rate of performance degradation.

[0181] Data source:

[0182] Based on real-time data from anomaly detection, HI time series health assessment results, and equipment replacement records during daily use, a module for predicting the remaining service life of equipment is established.

[0183] Device health age prediction method:

[0184] Because equipment performance exhibits a trend of relative stability in the early stages, slow decline in the middle stages, and significant performance degradation in the later stages, a combination of statistical and data-driven models is employed. Please refer to [the relevant documentation / reference]. Figure 5 The system performs trend fitting based on time-series health scores and uses deep learning models (such as LSTM) to predict future health evolution. The LSTM model layer learns and predicts short-term performance degradation trends, while the statistical layer uses Wiener to derive remaining lifespan. This approach leverages the multi-feature learning capabilities of deep learning networks while preserving the interpretability of the remaining lifespan from the statistical layer.

[0185] The model layer uses a two-layer LSTM to extract multi-source features from the nearest W steps. Extract features, historical window length W, prediction length H, and output the short-term state drift predicted by the LSTM:

[0186] ;

[0187] ;

[0188] Given the current time t, the above LSTM model obtains the state drift result. , This indicates the average rate of performance degradation within window H.

[0189] The statistical layer uses the Wiener degradation method:

[0190] ;

[0191] in The sampling time interval is fixed.

[0192] The remaining useful life prediction point estimate can be expressed as:

[0193] ;

[0194] in, This represents the predicted remaining useful life estimate, where L represents the failure threshold, calibrated from a threshold in the verification standard. The degradation state is denoted as... .

[0195] The overall remaining useful life estimation process is as follows: Input the latest real-time window LSTM output Aggregate to obtain the average degradation rate at the current time. Read the latest health status. , Obtained by fitting a mapping to the health score. Calculate the current distance to the failure threshold L, and finally calculate the estimated remaining usage time. .

[0196] In this method, remaining useful life prediction utilizes LSTM to learn the short-term average degradation rate, followed by the Wiener method to provide an estimate of remaining useful life. By combining anomaly monitoring and health score time-series data, a remaining useful life prediction module is constructed, compatible with both data-driven and statistically-driven methods, possessing both the predictive power of time-series models and the interpretability of statistical analysis.

[0197] Online updates and self-correction:

[0198] During equipment operation, based on the collected real-time data and verification test results, the health score, anomaly detection results, and RUL remaining service life prediction results are updated in real time.

[0199] The basic principles, main features, and advantages of the present invention have been shown and described. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A digital measurement method based on the Internet of Things and digital twins, characterized in that, Includes the following steps: S1: Construct an IoT sensing layer, deploy sensor nodes and gateway devices adapted to biological and chemical measuring instruments, establish communication connections between instruments and data transmission networks, and collect instrument operating parameters and detection environment parameters in real time. S2: Preprocess the collected parameter data, including cleaning up pseudo-anomaly data, data amplification and standardization, and store the preprocessed data in the constructed health monitoring database; S3: Based on the actual geometric dimensions and performance parameters of the measuring instrument, a digital twin model of the instrument is constructed using a 3D modeling tool to realize real-time data interaction between the digital twin model and the physical instrument; S4: Based on data from the health monitoring database, construct an artificial intelligence metrological analysis model that includes a pressure anomaly detection model, a whole machine health calculation model, and a health age and remaining service life prediction module to complete the instrument operation status analysis; The pressure anomaly detection model is a RePre-LSTM model, which consists of two sub-networks: an LSTMA reconstruction module and an LSTMS prediction module. The LSTM reconstruction module learns the time series through an encoder-decoder structure; the encoder part extracts temporal features by two stacked LSTM layers and outputs the hidden state; the hidden state output by the upper layer is mapped to a low-dimensional space through a fully connected layer; the decoder part restores the original sequence from the hidden vector in the upper low-dimensional space through decoding by a fully connected layer and two LSTM layers; finally, the decoder output is converted into a reconstructed value with the same dimension as the input sequence through a fully connected layer. The LSTM prediction module predicts the data for the next n time points of the input sequence. The module structure consists of two parts: an encoder and a predictor. The encoder part uses two layers of LSTM to extract temporal features, and the predictor part connects the hidden output state of the encoder to a fully connected layer to predict the data for the next n time points. S4 also includes a dual-module fusion determination mechanism: The fusion method of the two modules is as follows: First, perform anomaly detection analysis on the input time-series features to determine whether they are in the form of sudden anomalies or gradual degradation; then, dynamically adjust the weight parameters based on the results of the anomaly detection. If the abnormality is a sudden anomaly, it will manifest as: Increased size leads to a greater proportion of reconstruction error; if the anomalous feature is progressive degradation, it manifests as... Decreasing the value increases the prediction error; obtain the fused error result. ; ； ； ； and These represent the reconstruction error and prediction error at time t, respectively. and This represents the standardized reconstruction error and prediction error; , , , These represent the mean and standard deviation of the error of the reconstruction module on normal data, and the mean and standard deviation of the error of the prediction module on normal data, respectively. This represents the fusion error result at time point t; Dynamic parameters The calculation method is as follows: ； in, Represents the data value at time t Compared to time t-1 Fluctuation range An adaptive threshold for differential mutation detection; Indicates a sudden change switch; Local high-frequency energy is represented by local variance and calculated. It represents the slow-changing trend of prediction bias and is obtained by calculating the exponentially weighted average of the prediction error series. EMA is a method for analyzing trend changes. This represents the absolute value of the slope within the window. Representing the Sigmoid function: ； in, These are fixed parameters used to adjust the degree of change. S5: Map the output of the AI ​​metrology analysis model to the digital twin model, and realize the twin visualization of instrument operation status, metrology results and abnormal warning information through the visualization platform.

2. The digital measurement method based on the Internet of Things and digital twins according to claim 1, characterized in that, In step S1, the sensor nodes include pressure sensors, flow sensors, temperature sensors, humidity sensors, and air pressure sensors, which respectively collect the pump pressure, infusion flow rate, ambient temperature, ambient humidity, and air pressure of the measuring instrument; the gateway device supports three communication protocols: RS-232, USB, and Ethernet, and can be adapted to different brands and models of biological and chemical measuring instruments to ensure that the instrument network connection rate is not less than 60%.

3. The digital measurement method based on the Internet of Things and digital twins according to claim 1, characterized in that, In step S2, the pseudo-anomaly data cleaning employs a combined detection method: First, determine whether the fluctuation range of data at adjacent time points exceeds an adaptive threshold, and then mark the data with excessive fluctuation. Then, by calculating the degree of variation of the data within the sliding window, data with excessive variation are marked; The two types of labeled data were identified as pseudo-anomalies and removed.

4. The digital measurement method based on the Internet of Things and digital twins according to claim 1, characterized in that, In step S4, the overall health calculation model is based on the principle of relative combined uncertainty. It uses relative combined uncertainty, pressure anomaly rate, and qualitative / quantitative repeatability as core indicators and calculates the health score using a weighted scoring method. The uncertainty of a liquid chromatograph is calculated based on the uncertainty of the minimum detectable concentration; the method for calculating the minimum detectable concentration is as follows: ； In the formula: The minimum detectable concentration (g / mL); Baseline noise peak height (AU); The standard solution concentration is expressed in g / mL. The standard solution chromatographic peak height (AU) is used. The injection volume is (μL). The synthesis uncertainty of the minimum detection concentration requires calculating the uncertainties of baseline noise peak height, standard solution concentration, injection volume, and standard substance peak height. Qualitative / quantitative repeatability calculation: Qualitative / quantitative repeatability calculations were performed, including repeatability calculations of peak area and retention time, and calculation of relative standard deviation. : ； In the formula: For qualitative / quantitative measurement repeatability relative standard deviation; For the first The retention time or peak area measured in this study; The arithmetic mean of the six measurements; Measurement serial number; For the number of measurements; The uncertainty propagation rate is: ； Relative combination uncertainty: ； In the formula, each term represents a source of combined uncertainty in the liquid chromatography detection results: Indicates the uncertainty introduced by the standard reference material; This represents the uncertainty introduced by baseline noise; Indicates the uncertainty introduced by the chromatographic peak height; This indicates the uncertainty introduced by the microsyringe; Relative expanded uncertainty: ； In the formula: Indicates the inclusion factor, taking =2; Synthetic health calculation: Let the health index be H. The health index is calculated as follows: ； in, The relative combined uncertainty is given by F, where F represents the pressure anomaly rate and R represents the qualitative / quantitative repeatability; the default weight values ​​are... =0.55, =0.3, =0.15, and + + =1; By integrating relative composite uncertainty, pump pressure anomaly detection results, and the relative standard deviation of other key factors, a whole-machine health assessment model is constructed. Through multi-index weighted calculation, a scoring index reflecting the overall health level of the equipment is output. This index can be used for long-term trend analysis, preventive maintenance, and equipment life prediction.

5. The digital measurement method based on the Internet of Things and digital twins according to claim 4, characterized in that, The health score ranges from 0 to 100. The higher the score, the more stable the instrument's operation. The weight of each core indicator is adaptively adjusted according to the characteristics of different types of measuring instruments.

6. The digital measurement method based on the Internet of Things and digital twins according to claim 1, characterized in that, The specific process of constructing the digital twin model in step S3 is as follows: Using 3ds Max and Maya 3D modeling tools, a 3D model is constructed based on the actual geometric data of the measuring instrument, and an FBX format file is generated. Import the FBX file into the Unity3D platform to develop interactive functions for the instrument's movable parts, including switching, rotation, and dragging / scaling. Design the measurement information data structure and data interface to achieve the connection with sensor data of the Internet of Things sensing layer and complete the synchronization of the operating status of physical instruments and digital twin models.

7. The digital measurement method based on the Internet of Things and digital twins according to claim 1, characterized in that, In step S3, the real-time data interaction frequency between the digital twin model and the physical instrument is no less than 1 time / second, ensuring that the mapping delay of the digital twin model to the operating status of the physical instrument is less than 100ms. In step S5, the visualization platform displays the following content: instrument 3D model, real-time operating parameters, health score change curve, health age change trend, and abnormal warning list. The abnormal warning list indicates the abnormality type, the time of occurrence of the abnormality, and preliminary handling suggestions.

Images