Liquid chromatograph state monitoring and predictive maintenance method and system

Abstract

The invention provides a liquid chromatograph state monitoring and predictive maintenance method and a liquid chromatograph state monitoring and predictive maintenance system, and relates to the technical field of data processing. Liquid chromatograph sensor data (continuous parameters such as flow, temperature, pressure and the like) and log data (discrete information such as alarm state, equipment event and the like) are subjected to joint modeling; and data dimension reduction is carried out based on the geometric similarity of the data points, so that the local structure and global information of the data of the liquid chromatograph are kept, and meanwhile, information loss in the dimension reduction process is reduced. The state monitoring model is further utilized for data processing, the context sensing ability of the model is improved, when abnormity or faults occur, multiple factors can be comprehensively considered for accurate diagnosis, and higher accuracy and reliability are shown when complex state changes of the liquid chromatograph are monitored.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and system for monitoring and predicting the state of a liquid chromatograph. Background Art

[0002] As a precision analytical equipment, liquid chromatograph is widely used in chemical analysis, biopharmaceuticals, food safety testing and other fields. With the continuous development of technology, the application scope and frequency of liquid chromatograph are increasing. However, equipment failure and downtime are still the main factors affecting its operating efficiency and reliability. During the operation of liquid chromatograph, real-time monitoring of multiple parameters is involved, such as flow, pressure, temperature, solvent concentration, etc. Slight fluctuations in these parameters may affect the performance of the equipment and even cause failure.

[0003] With the advancement of technology, modern liquid chromatographs are able to monitor multiple parameters (such as flow, pressure, temperature, solvent concentration, etc.) in real time, and these data constitute a multi-dimensional time series. However, most existing methods rely on manual experience and judgment, lacking automated data analysis and early warning mechanisms, resulting in untimely problem discovery and missed optimal processing opportunities. In addition, the amount of data generated by each run of the liquid chromatograph is huge, especially in the case of high-frequency sampling, and traditional storage and processing methods cannot efficiently manage this data. Moreover, the interaction between multiple parameters makes data analysis more complicated. For example, changes in flow may affect pressure and temperature, and changes in these variables in turn affect the separation effect. In summary, traditional data processing methods are challenging when processing high-dimensional data. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a liquid chromatograph state monitoring and predictive maintenance method and system, which can realize the intelligent operation and maintenance management of the liquid chromatograph, so as to increase the service life of the equipment, optimize the allocation of maintenance resources, reduce the operating costs, and ensure the stability and reliability of the analysis work.

[0005] In a first aspect, an embodiment of the present invention provides a method for condition monitoring and predictive maintenance of a liquid chromatograph, the method comprising: obtaining operation monitoring data of a liquid chromatograph to be monitored; performing distance measurement on data points of the operation monitoring data, and determining the geometric similarity of the data points corresponding to the operation monitoring data; using a pre-trained autoencoder to perform dimensionality reduction processing on the operation monitoring data based on the geometric similarity of the data points, and determining the reduced dimensionality data representation corresponding to the operation monitoring data; loading the reduced dimensionality data representation into a pre-trained state monitoring model, and outputting a state analysis result corresponding to the reduced dimensionality data representation; wherein the state monitoring model updates model parameters based on a fractional-order corrected cross entropy loss function; and based on the state analysis result, evaluating the working state of the liquid chromatograph.

[0006] In combination with the first aspect, an embodiment of the present invention also provides a first implementation method of the first aspect, wherein the state monitoring model is constructed based on a high-order neural network; the step of loading the reduced dimensionality data representation into a pre-trained state monitoring model and outputting the state analysis result corresponding to the reduced dimensionality data representation includes: forward propagating the reduced dimensionality data representation through multiple hidden layers of the high-order neural network; and using the deep structure of the high-order neural network to fuse the reduced dimensionality data representation layer by layer to determine the state analysis result corresponding to the reduced dimensionality data representation.

[0007] In combination with the first aspect, an embodiment of the present invention also provides a second implementation of the first aspect, wherein the method also includes: training a preset initial high-order neural network using a preset training sample set, and determining a cross-entropy loss function corresponding to the initial high-order neural network; performing a fractional-order correction on the cross-entropy loss function using a preset fractional-order calculus operator, and determining a fractional-order loss gradient; back-propagating and updating the model parameters of the initial high-order neural network based on the fractional-order loss gradient and the cross-entropy loss function; until the initial high-order neural network meets the preset iteration conditions, a state monitoring model is constructed based on the initial high-order neural network.

[0008] In combination with the first aspect, an embodiment of the present invention also provides a third implementation of the first aspect, wherein the steps of performing distance measurement on data points of the operation monitoring data and determining the geometric similarity of the data points corresponding to the operation monitoring data include: measuring the distance between the data points of the operation monitoring data based on the dimension-by-dimensional difference of the data points of the operation monitoring data; calculating the geometric similarity matrix corresponding to the operation monitoring data based on the data point distance and a preset similarity weight coefficient, and determining the geometric similarity of the data points corresponding to the operation monitoring data.

[0009] In combination with the first aspect, an embodiment of the present invention also provides a fourth implementation of the first aspect, wherein a step of using a pre-trained autoencoder to perform dimensionality reduction processing on the operation monitoring data based on the geometric similarity of data points to determine the reduced dimensionality data representation corresponding to the operation monitoring data includes: based on the geometric similarity of the data points, using a preset autoencoder to perform weighted mapping on the operation monitoring data to determine the initial reduced dimensionality features; performing interaction measurement on the data points of the initial reduced dimensionality features to determine a feature mapping from the initial reduced dimensionality features; and performing adaptive smoothing on the feature mapping to obtain the reduced dimensionality data representation corresponding to the operation monitoring data.

[0010] In combination with the first aspect, an embodiment of the present invention also provides a fifth implementation of the first aspect, wherein the step of performing interaction measurement on the data points of the initial dimensionality reduction features and determining the feature mapping from the initial dimensionality reduction features includes: determining the gradient change corresponding to the data points of the initial dimensionality reduction features, and the geometric distance between the data points; determining the interaction energy of the data points of the initial dimensionality reduction features based on the gradient change and the geometric distance; and minimizing the interaction energy to determine the feature mapping corresponding to the initial dimensionality reduction features.

[0011] In combination with the first aspect, an embodiment of the present invention also provides a sixth implementation of the first aspect, wherein the training method of the autoencoder includes: training a preset autoencoder through a preset training sample set, and calculating the loss function of the autoencoder; introducing a constraint condition of minimizing the path difference into the loss function, and calculating the gradient of the loss function with respect to the weight of the autoencoder; and updating the parameters of the autoencoder based on the gradient and the pre-calculated adaptive weight adjustment coefficient.

[0012] In combination with the first aspect, an embodiment of the present invention also provides a seventh implementation method of the first aspect, wherein a method for calculating an adaptive weight adjustment coefficient comprises: measuring the local density of each data point of the operation monitoring data in its neighborhood, and dynamically calculating the adaptive weight adjustment coefficient.

[0013] In the second aspect, an embodiment of the present invention also provides a liquid chromatograph state monitoring and predictive maintenance device, which includes: a data acquisition module, used to obtain operation monitoring data of the liquid chromatograph to be monitored; a calculation module, used to perform distance measurement on data points of the operation monitoring data, and determine the geometric similarity of data points corresponding to the operation monitoring data; a data processing module, used to use a pre-trained autoencoder to perform dimensionality reduction processing on the operation monitoring data based on the geometric similarity of the data points, and determine the reduced dimensionality data representation corresponding to the operation monitoring data; an execution module, used to load the reduced dimensionality data representation into a pre-trained state monitoring model, and output a state analysis result corresponding to the reduced dimensionality data representation; wherein the state monitoring model updates model parameters based on a fractional-order corrected cross entropy loss function; a predictive maintenance module, used to evaluate the working state of the liquid chromatograph based on the state analysis result.

[0014] In a third aspect, an embodiment of the present invention further provides a liquid chromatograph condition monitoring and predictive maintenance system, which is configured with the above-mentioned device and is used to execute the method of any of the above-mentioned embodiments.

[0015] The embodiments of the present invention bring the following beneficial effects: The present invention provides a method and system for monitoring and predictive maintenance of a liquid chromatograph state, which relates to the field of data processing technology. By jointly modeling the liquid chromatograph sensor data (such as continuous parameters such as flow, temperature, and pressure) and log data (such as discrete information such as alarm status and equipment events), and performing data dimension reduction based on the geometric similarity of data points, the local structure and global information of the liquid chromatograph data are maintained, while reducing the information loss in the dimension reduction process. Further using the state monitoring model for data processing improves the context perception ability of the model. When an abnormality or failure occurs, it can comprehensively consider multiple factors for accurate diagnosis, and show higher accuracy and reliability when monitoring the complex state changes of the liquid chromatograph.

[0016] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 A flow chart of a method for condition monitoring and predictive maintenance of a liquid chromatograph provided in an embodiment of the present invention; Figure 2 A flow chart of another method for monitoring and predictive maintenance of a liquid chromatograph provided by an embodiment of the present invention; Figure 3 A schematic diagram showing the performance of the dimensionality reduction algorithm provided by an embodiment of the present invention under different hyperparameter settings; Figure 4 A schematic diagram showing a comparison of classification accuracy of the dimensionality reduction algorithms provided in an embodiment of the present invention; Figure 5 A schematic diagram of the structure of a liquid chromatograph condition monitoring and predictive maintenance device provided by an embodiment of the present invention; Figure 6 A logical schematic diagram of a liquid chromatograph condition monitoring and predictive maintenance system provided by an embodiment of the present invention; Figure 7 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described in combination with the embodiments below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Traditional maintenance methods for liquid chromatographs usually rely on regular inspections and manual intervention, which cannot detect potential equipment failures and performance degradation in a timely and accurate manner, resulting in the risk of over-maintenance or under-maintenance. This not only increases maintenance costs, but may also lead to unplanned downtime of equipment, affecting overall production efficiency and the accuracy of analytical results. Therefore, how to achieve real-time status monitoring and predictive maintenance of liquid chromatographs has become a key issue in improving equipment operating efficiency, extending service life, and reducing maintenance costs.

[0022] With the rapid development of big data and artificial intelligence technologies, especially the application of machine learning and deep learning algorithms, new ideas and solutions have been provided for the condition monitoring and fault prediction of liquid chromatographs. By collecting the operating data of the equipment and using machine learning methods to establish the equipment status model, it is not only possible to monitor the working status of the liquid chromatograph in real time, but also to warn of potential faults in advance through the analysis of historical data, thereby achieving more intelligent and precise maintenance management. However, there are many problems with existing monitoring and maintenance methods, such as computational redundancy caused by high data dimensions, the inability of traditional dimensionality reduction methods to retain local and global information in the data, and overfitting during model training. These problems limit the effectiveness of existing technologies in practical applications.

[0023] The present invention provides a method and system for monitoring the state of a liquid chromatograph and predictive maintenance, which can realize intelligent operation and maintenance management of the liquid chromatograph to increase the service life of the equipment, optimize the allocation of maintenance resources, reduce operating costs, and ensure the stability and reliability of analytical work.

[0024] To facilitate understanding of this embodiment, a liquid chromatograph condition monitoring and predictive maintenance method disclosed in an embodiment of the present invention is first introduced in detail, see Figure 1 The schematic diagram of a process of a liquid chromatograph condition monitoring and predictive maintenance method shown in FIG. 1 includes the following steps: Step S102, obtaining operation monitoring data of the liquid chromatograph to be monitored.

[0025] Most of the existing technologies are for single data source processing, which cannot effectively reflect the complex mode of the liquid chromatograph operation process. The operation monitoring data of the embodiment of the present invention includes the equipment operation parameters of the liquid chromatograph and the state operation information in the log of the liquid chromatograph, and uses the real-time physical parameters provided by the sensor and the event information in the log record to obtain a more comprehensive and richer data set. By jointly modeling the two types of data, abnormal situations can be identified more accurately. For example, a slight change in a parameter may not attract attention alone, but if it is accompanied by an abnormal record in the log (such as an alarm or a fault), potential problems can be discovered earlier. Among them, data acquisition and storage can be performed by a data acquisition unit, including sensor data acquisition and log data analysis. In addition, the data can be pre-processed as operation monitoring data. Among them, sensor data acquisition is used to collect equipment operation parameters such as flow, pressure, temperature, and solvent composition; log data analysis is used to extract the operation status information in the instrument's own log, including start time, stop time, fault record, alarm information, etc.; data preprocessing is used for denoising, normalization and outlier processing to improve data quality; storage function ensures that data is stored in a database or data lake in an efficient and secure manner for subsequent analysis.

[0026] Step S104: performing distance measurement on the data points of the operation monitoring data to determine the geometric similarity of the data points corresponding to the operation monitoring data.

[0027] Step S106, using a pre-trained autoencoder to perform dimensionality reduction processing on the operation monitoring data based on the geometric similarity of the data points, and determining the dimensionality reduction data representation corresponding to the operation monitoring data.

[0028] The embodiment of the present invention measures the distance of each data point, and forms a similarity matrix with the distance calculation results of each data point and all other data points, which reflects the geometric similarity between data points and quantifies the spatial relationship between different data points. Based on this, data dimensionality reduction is performed to ensure that the data points after dimensionality reduction still maintain their original geometric structure in the low-dimensional space. These low-dimensional data not only retain the key features of the original data, but also reduce redundant information and significantly reduce storage requirements. It can also be more easily visualized to help researchers better understand the distribution and characteristics of the data.

[0029] Step S108, loading the dimension-reduced data representation into a pre-trained state monitoring model, and outputting a state analysis result corresponding to the dimension-reduced data representation.

[0030] Step S110: Based on the status analysis result, the working status of the liquid chromatograph is evaluated.

[0031] With the rapid development of big data and artificial intelligence technology, the embodiment of the present invention uses a state monitoring model established by machine learning and deep learning algorithms to realize state monitoring and fault prediction of liquid chromatographs. Not only can the working state of the liquid chromatograph be monitored in real time, but also potential fault problems can be warned in advance through the analysis of historical data, thereby realizing more intelligent and precise maintenance management.

[0032] Among them, the results of the model output include evaluations of different states such as normal operation, potential problem warning, and fault detection. According to the state analysis results, the current working state of the liquid chromatograph (such as normal, warning, fault) can be judged, and corresponding maintenance suggestions or warning information can be proposed. In one embodiment, the liquid chromatograph state monitoring and predictive maintenance method proposed in the embodiment of the present invention relies on the liquid chromatograph state monitoring and predictive maintenance system to realize real-time monitoring, abnormality detection and fault prediction of the equipment operation state. The system works together through multiple units to ensure the stable operation of the liquid chromatograph, improve maintenance efficiency, and reduce the risk of unplanned downtime. For example, the state monitoring unit can perform state analysis based on real-time collected data to determine whether the equipment is in normal working state; the state monitoring unit performs data comparison, abnormality detection and alarm processing. Data comparison is used to monitor the deviation between the current operating parameters and the historical normal state in real time; abnormality detection determines whether there is an operation abnormality based on a machine learning model or a rule threshold; alarm processing is used to trigger an alarm mechanism, and reminds the operation and maintenance personnel to take corresponding measures through a visual interface or a notification system.

[0033] Furthermore, the predictive maintenance unit can also be used to predict faults, estimate remaining life, and recommend maintenance plans. The health of the equipment can be evaluated based on historical operating data and prediction models, and maintenance recommendations can be provided. Fault prediction uses time series analysis and machine learning methods to infer the type and time of possible future faults; remaining life estimation is based on equipment aging models and historical fault data to calculate the expected service life of key components; maintenance plan recommendations combine equipment usage, inventory status, and maintenance resources to generate optimized maintenance plans to avoid over-maintenance or under-maintenance.

[0034] Furthermore, the status display, alarm management and interactive control can be performed through visualization and user interaction units. Intuitive monitoring results and interactive functions are provided to improve user experience and decision-making efficiency. Among them, the status display is used to present the equipment operation status in the form of charts, dashboards or heat maps; alarm management provides historical alarm record query, alarm confirmation and processing tracking functions; interactive control allows users to adjust monitoring thresholds, select different prediction models or export analysis reports to support personalized needs.

[0035] In summary, the embodiments of the present invention realize intelligent operation and maintenance management of the liquid chromatograph through the coordinated operation of various functional units, thereby increasing the service life of the equipment, optimizing the allocation of maintenance resources, reducing operating costs, and ensuring the stability and reliability of the analysis work.

[0036] Furthermore, the state monitoring detection model of the embodiment of the present invention updates the model parameters based on the fractional-order modified cross entropy loss function, which can enhance convergence stability and suppress overfitting. When updating the model parameters, not only the instantaneous gradient information is considered, but also the memory characteristics of the historical gradient are used to improve the fitting stability.

[0037] In one embodiment, the state monitoring model is constructed based on a high-order neural network. The reduced-dimensional data representation is forward propagated through multiple hidden layers of the high-order neural network; and the reduced-dimensional data representation is layer-by-layer fused using the deep structure of the high-order neural network to determine the state analysis result corresponding to the reduced-dimensional data representation.

[0038] In the specific implementation, the liquid chromatograph data is loaded into the high-order neural network model, and the liquid chromatograph data is forward propagated through multiple hidden layers to calculate the output of the network based on the input data. The output of the network is finally obtained by passing the input data to each layer of the neural network layer by layer and applying the activation function. This output can be a classification label, regression value, or other form of predicted value.

[0039] Specifically, each hidden layer uses the ReLU activation function to increase the nonlinear processing capability, so as to better characterize the complex mapping relationship. For example, the flow and pressure parameters of the liquid chromatograph often show piecewise nonlinear changes during operation (such as sudden changes in flow during the gradient elution stage), and ReLU can effectively capture such local features through its sparse activation characteristics. At the same time, the deep structure of the high-order neural network can fuse sensor data (continuous numerical type) and log parsing data (discrete state labels) layer by layer. For example, the continuous fluctuation of the temperature sensor and the "column temperature limit alarm" event recorded in the log are jointly modeled to improve the context perception ability of state classification. It is expressed as:

[0040] In the formula, is the linear transformation output of the qth layer of the high-order neural network; is the output of the qth layer of the high-order neural network after being processed by the ReLU activation function, It is the output of the q-1th layer of the high-order neural network after being processed by the ReLU activation function; is the weight matrix of the qth layer of the high-order neural network; Represents the maximum value function.

[0041] Traditional neural network training methods may suffer from slow convergence or overfitting problems in long-term trend prediction of equipment degradation or fault detection, especially when dealing with complex and nonlinear data. The embodiment of the present invention updates model parameters based on a fractional-order corrected cross-entropy loss function to construct a state monitoring model to enhance convergence stability and suppress overfitting. The network is able to remember historical gradient information, and performs well when features are offset due to equipment aging. It avoids the instability caused by traditional gradient descent methods that rely only on instantaneous gradients, and exhibits higher accuracy and reliability when monitoring complex state changes of liquid chromatographs. The specific steps are as follows: 1) Use the preset training sample set to train the preset initial high-order neural network, and determine the cross entropy loss function corresponding to the initial high-order neural network.

[0042] Initialize the parameters of the high-order neural network model, including the weights and biases of the high-order neural network. Use the Glorot method to initialize the weights and biases of the high-order neural network to ensure the rationality of the initial distribution, thereby improving the convergence speed and the stability of the fitting performance. For example, the solvent composition parameters usually have a nonlinear coupling relationship. A reasonable initial weight distribution can accelerate the model's ability to extract features of complex solvent concentration changes in the early stages of training, and avoid gradient disappearance or explosion due to excessive or too small initial parameter values. The initialization method is expressed as:

[0043] In the formula, is the element in the i-th row and j-th column of the q-th layer weight matrix of the high-order neural network; is the bias vector of the qth layer of the high-order neural network; Indicates parameter update operation; and is the number of neurons in the q-1th layer and the qth layer of the high-order neural network; r is a random number drawn from the uniform distribution U(-1,1); i is a positive integer, j is a positive integer, and q is a positive integer.

[0044] Furthermore, the training data is forward propagated, and the above steps can be referred to. Furthermore, the cross entropy loss function is used to evaluate the difference between the model output and the actual label, and the model parameters are back-propagated and updated in combination with the fractional calculus improvement to achieve a more stable fitting process.

[0045] Specifically, first define the conventional cross entropy loss function and the calculation formula of its integer gradient. In the embodiment of the present invention, the fractional calculus operator is used to perform fractional order correction on the gradient during the back propagation process to enhance the convergence stability of the model and suppress the overfitting trend. The calculation method of the conventional cross entropy loss function is expressed as:

[0046] In the formula, p c is the probability distribution predicted by the model; y c is the true label distribution; C is the total number of categories, representing the number of categories in the classification problem; L r is the cross entropy loss function, which is used to evaluate the difference between the predicted probability and the actual probability; is the symbol of partial derivative.

[0047] Among them, the integer-order gradient of the loss function relative to the network output is expressed as:

[0048] In the formula, p c is the probability distribution predicted by the model, y c is the true label distribution.

[0049] And, the integer derivative of the ReLU activation function is expressed as:

[0050] In the formula, It is the linear transformation output of the qth layer of the high-order neural network.

[0051] Moreover, the integer-order gradient of the linear transformation with respect to the weight is approximately equal to the output of the previous layer, expressed as:

[0052] In the formula, h (q-1) It is the output of the q-1th layer of the high-order neural network after being processed by the ReLU activation function.

[0053] 2) Use the preset fractional-order calculus operator to perform fractional-order correction on the cross entropy loss function and determine the fractional-order loss gradient.

[0054] In order to enhance the fitting stability of high-order neural networks, the present invention further adopts fractional-order calculus operators based on the original gradient calculation. The convergence stability of the model is improved by utilizing its memory characteristics of historical gradient information in parameter updating. Since the operating parameters of liquid chromatographs (such as flow rate and temperature) have time continuity and state inertia, the traditional integer-order gradient only depends on the current instantaneous gradient, while the fractional-order calculus enhances the model's fitting stability for the gradual change of the equipment state by integrating historical gradient information (such as the gradient change trend of the pressure parameter in the past 100 iterations). For example, when detecting slow cumulative faults such as "chromatographic column aging", fractional-order correction can suppress misclassification caused by fluctuations in single sampling data and smooth the impact of instantaneous noise through the historical gradient memory characteristics.

[0055] Preferably, a Caputo-type fractional derivative is used (or other equivalent fractional-order operators), perform fractional-order gradient correction on the loss function, and define the fractional-order loss gradient as:

[0056] in, is the order of the fractional derivative, 0<α≤1, preferably, α=0.9; Characterizes the process of transforming integer-order gradients into fractional-order ones.

[0057] To further reflect the combination of the operator and the network structure, Preferably, the following formula is used to perform fractional order correction on the weight gradient of the i-th row and j-th column, which can retain historical information and smooth the fluctuation of instantaneous gradient during back propagation, thereby enhancing the stability of the network fitting process, which is expressed as:

[0058] In the formula, is the Gamma function; is the upper limit of the number of discrete training steps; Representation is in the training step number The integer gradient at time instant; is a positive integer; Characterize the conventional integer order gradient The corresponding gradient coefficient (which includes the forward propagation derivative information and the output of the previous layer); and They are The predicted probability and true probability corresponding to each training step.

[0059] 3) Based on the fractional-order loss gradient and cross-entropy loss functions, the model parameters of the initial high-order neural network are back-propagated and updated.

[0060] The embodiment of the present invention combines the parameter update of adaptive sub-gradient and fractional order correction to dynamically adjust the learning rate of each parameter, suppress overfitting and balance the convergence speed. Define a conventional adaptive sub-gradient update method, and then further add a fractional order correction term to achieve high-order neural network classifier model training with enhanced fitting stability. First, the weight update method of the conventional adaptive sub-gradient is given, which is expressed as:

[0061] In the formula, It is the parameter update operation; is the global learning rate; is the square of the accumulated gradient before the update; To avoid division by zero, a small constant is used. Preferably, Set to 0.001.

[0062] Furthermore, the cumulative squared gradient update is expressed as:

[0063] In the formula, Update operation for parameters.

[0064] However, in order to consider the gradient correction effect of fractional calculus based on the adaptive sub-gradient, we further By integrating it into the update formula and performing adaptive sub-gradient updates with fractional-order correction, the high-order neural network not only considers the instantaneous integer-order gradient information during the update process, but also considers the memory and smoothing effect of the fractional order on the historical gradient, thereby improving the overall fitting stability and generalization performance. For example, there is a dimensional difference between the numerical range of solvent composition parameters (such as methanol proportion) and the binary state mark in the log (such as "pump start / stop"), and the traditional unified learning rate may cause imbalance in parameter updates. By adopting the adaptive mechanism of fractional-order correction terms, the model can assign different gradient weights to the high-frequency fluctuation data of the flow sensor (requires fast response) and the low-frequency events of the log state (requires long-term memory). In the scenario of sudden changes in solvent composition, fractional-order gradient correction can slow down the update speed of related parameters to avoid overfitting caused by instantaneous concentration jumps, while the state mark of log parsing maintains stable feature associations through adaptive learning rates.

[0065] In summary, the update combined with fractional gradient correction is expressed as:

[0066] In the formula, is the updated weight after combining the fractional gradient correction; is the balance coefficient between the fractional order gradient and the integer order gradient, preferably, =0.1.

[0067] 4) Until the initial high-order neural network meets the preset iteration conditions, a state monitoring model is constructed based on the initial high-order neural network.

[0068] Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0069] In summary, the embodiment of the present invention adopts a high-order neural network model classification algorithm based on a fractional-order update strategy, and combines fractional-order calculus to perform gradient correction during the back-propagation process, which can enhance convergence stability and suppress overfitting. At the same time, combined with the adaptive sub-gradient update method, the fractional-order gradient correction method is adopted, so that when updating the model parameters, not only the instantaneous gradient information is considered, but also the memory characteristics of the historical gradient can be used to improve the fitting stability.

[0070] Furthermore, on the basis of the above-mentioned embodiments, the embodiments of the present invention also provide another method for condition monitoring and predictive maintenance of liquid chromatographs, which ensures the stable operation of liquid chromatographs, improves maintenance efficiency, and reduces the risk of unplanned downtime through the coordinated work of multiple functional units such as data collection, feature extraction, model training, predictive analysis, and maintenance suggestions. Among them, feature engineering, model training, model evaluation, and model updating can be achieved through machine learning modeling units. A condition monitoring and prediction model is established based on historical data to identify equipment operation trends and provide early warning of potential faults. Among them, feature engineering is used to extract representative feature variables, including time series features, frequency domain features, and statistical features; model training uses supervised learning or unsupervised learning methods to train classification, regression, or clustering models to achieve anomaly detection, life prediction, and health assessment; model evaluation evaluates the accuracy and generalization ability of the model through indicators such as cross-validation, confusion matrix, and regression error; model update continuously optimizes the training model based on the latest operating data to adapt to the dynamic changes in equipment status. Figure 2 A flowchart showing the method is shown in FIG. Figure 2 , the method comprises the following steps: Step S202, obtaining operation monitoring data of the liquid chromatograph to be monitored.

[0071] Among them, the key parameters of the liquid chromatograph during operation can be collected in real time by the data acquisition unit of the system, that is, the equipment operating parameters of the liquid chromatograph and the status operating information in the log of the liquid chromatograph, to ensure the integrity and real-time nature of the data. The operation monitoring data can refer to the above embodiment and will not be repeated here.

[0072] Furthermore, the data is reduced in dimensionality. Among them, liquid chromatographs are usually equipped with multiple sensors to monitor various operating parameters in real time. These sensors generate data every second or every few seconds, resulting in a rapid increase in data dimensions. In order to capture instantaneous changes and small fluctuations, liquid chromatographs usually adopt high-frequency data acquisition methods, which generate a large amount of time series data, increasing the dimension and sample size of the data. In addition, the long-term operation of the equipment will generate a large amount of historical data, which not only contains monitoring information from different time periods, but may also involve samples from different batches, further increasing the complexity of the data. Moreover, each parameter does not exist in isolation, and there are complex interactions between them. Due to the different requirements and processing methods for data in different fields, the data is highly complex.

[0073] However, high-dimensional data will significantly increase the consumption of computing resources, including memory and processing time, when processed and analyzed. By reducing the dimension, the computational complexity can be significantly reduced, making data analysis more efficient. Existing dimensionality reduction methods, such as PCA, mainly compress data through linear relationships, which tends to ignore the nonlinear characteristics of the data, especially when facing high-dimensional liquid chromatography data, which is easy to lose local structure and global information.

[0074] The embodiment of the present invention adopts an autoencoder algorithm based on geometric similarity, obtains the relative relationship between input data points through geometric similarity calculation, and performs feature dimensionality reduction in combination with the strategy of path reconstruction. When reducing the dimensionality of high-dimensional data, the local structure and global information of the liquid chromatograph data are maintained, while reducing the information loss in the dimensionality reduction process. Moreover, in the process of feature dimensionality reduction, the adaptive weight adjustment coefficient of the feature is adopted, and different update rates are assigned to features of different importance by dynamically calculating the local density. For high-density areas (such as the stable operation stage), the features are accelerated to learn, while for the features of low-density areas (such as the startup transient stage), a more conservative update strategy is adopted, which enhances the robustness and flexibility of the model when dealing with the nonlinear and time-continuous characteristics of the liquid chromatograph. Specifically, refer to the following steps S204-S212.

[0075] Step S204: measuring the distance between the data points of the operation monitoring data based on the dimension-by-dimensional difference between the data points of the operation monitoring data.

[0076] Step S206, based on the data point distance and the preset similarity weight coefficient, a geometric similarity matrix corresponding to the operation monitoring data is calculated to determine the geometric similarity of the data points corresponding to the operation monitoring data.

[0077] If the feature dimension of the collected liquid chromatograph is too high, it will easily lead to computational redundancy. In order to solve the problem that it is difficult to retain both local and global information during the dimensionality reduction process of high-dimensional liquid chromatograph data, an autoencoder algorithm based on geometric similarity measurement is adopted. The geometric similarity measurement is combined with the path reconstruction strategy to perform feature dimensionality reduction. While compressing the dimension, the key information of the liquid chromatograph data is preserved as much as possible, thereby improving the representation ability and accuracy after dimensionality reduction.

[0078] The present invention obtains the relative relationship between input data points through geometric similarity calculation, and then uses the pre-trained autoencoder to reduce the dimension of the data based on the relative relationship. The calculation method of the geometric similarity matrix is ​​expressed as:

[0079] In the formula, is the geometric similarity matrix, representing the input liquid chromatograph data The similarity between the data points in ; is the input liquid chromatography data of the autoencoder; is the weight coefficient of similarity, representing the data point and the relative importance of Input the i-th data point of the liquid chromatograph data to the autoencoder; Input the jth data point of the liquid chromatography data to the autoencoder; For data points and The distance between The number of data points in the liquid chromatography data input to the autoencoder; i is a positive integer; j is a positive integer.

[0080] Furthermore, in order to better characterize the relative importance between data points, the weight coefficient of similarity is calculated based on the square of the distance between data points, expressed as:

[0081] In the formula, is a hyperparameter to control similarity decay; For data points and The square of the distance; For data points and The square of the distance; Input the HPLC data to the autoencoder data points; is a positive integer. Preferably, Set to 0.95.

[0082] Furthermore, in the process of geometric similarity calculation, for example, by measuring the local structural relationship between multidimensional parameters such as flow, pressure, and temperature collected by sensors, in the scenario of dynamic changes in solvent composition, temperature and solvent concentration may show nonlinear correlation. The model can strengthen the similarity of adjacent data points and weaken the influence of outliers, thereby retaining the local characteristics of solvent gradient changes during the operation of the chromatograph and solving the problem of similarity misjudgment caused by high-dimensional noise in sensor data. The calculation method for measuring the distance between data points is expressed as:

[0083] In the formula, is the value of the i-th data point in the k-th feature dimension; m r is the number of feature dimensions of the liquid chromatography data input to the autoencoder; Input the value of the j-th data point of the liquid chromatograph data to the autoencoder on the k-th feature dimension.

[0084] Step S208, based on the geometric similarity of the data points, a preset autoencoder is used to perform weighted mapping on the operation monitoring data to determine the initial dimensionality reduction features.

[0085] Based on geometric similarity, the autoencoder is used to perform weighted mapping on the data to reduce the dimensionality of the data. The geometric relationship between data points can be retained when passing between layers of the autoencoder, so that the dimensionality reduction process can take into account both the local structure and high-dimensional characteristics of the liquid chromatograph data. The calculation method is expressed as:

[0086] In the formula, is the liquid chromatograph data representation after weighted mapping, i.e., the initial dimension reduction feature; is the weight matrix of the autoencoder; is the bias term of the autoencoder.

[0087] Furthermore, in order to dynamically obtain the weight matrix and reduce the similarity deviation, the weight matrix of the autoencoder is calculated by a diagonal matrix, which is expressed as:

[0088] In the formula, An adjustment factor to control the weight ratio; is a diagonal matrix with diagonal elements , is the element in the i-th row and i-th column of the diagonal matrix; is the element in the i-th row and j-th column of the geometric similarity matrix; i is a positive integer; j is a positive integer; The number of data points in the liquid chromatograph data is input to the autoencoder. Preferably, Set to 0.1.

[0089] Furthermore, traditional machine learning algorithms usually adopt a fixed learning rate and a uniform update rate, and the importance of different liquid chromatograph data features is relatively simple and lacks flexibility. The embodiment of the present invention not only performs data dimensionality reduction based on geometric similarity, but also optimizes the path difference between data points before and after dimensionality reduction, and also trains the autoencoder through the following steps, adopts the adaptive weight adjustment coefficient of the feature, and assigns different update rates to features of different importance by dynamically calculating the local density. For high-density areas (such as the stable operation stage), the features are accelerated to learn, while for the features of low-density areas (such as the startup transient stage), a more conservative update strategy is adopted, which enhances the robustness and flexibility of the model when dealing with the nonlinear and time-continuous characteristics of the liquid chromatograph.

[0090] The specific steps are as follows: 1) Train the preset autoencoder using the preset training sample set and calculate the loss function of the autoencoder.

[0091] 2) Introduce the constraint of minimizing the path difference into the loss function and calculate the gradient of the loss function with respect to the weight of the autoencoder.

[0092] Among them, the above-mentioned steps of calculating the initial dimensionality reduction features can be used to process the data as a training sample set for training the autoencoder. In order to achieve path reconstruction and dimensionality reduction optimization from high-dimensional mapping to low-dimensional mapping, the global structure and topological relationship of the liquid chromatograph data are retained after dimensionality reduction to reduce the global information loss that occurs when high-dimensional mapping is reduced to low-dimensional mapping. The embodiment of the present invention improves the overall consistency of dimensionality reduction by optimizing the path differences between data points before and after dimensionality reduction. For example, the instrument operating status (such as the pump pressure fluctuation period) obtained by parsing log data usually has a global topological relationship in time series. By minimizing the path differences of data points before and after dimensionality reduction, the model can maintain the overall shape of the pressure-flow curve and avoid the break of the time series pattern caused by dimensionality reduction. During the iterative training process of the autoencoder, the constraints are:

[0093] In the formula, Represents the weight matrix and bias term of the autoencoder when the condition is minimized; is the i-th liquid chromatograph data representation after weighted mapping; is the j-th liquid chromatograph data representation after weighted mapping; is the L2 norm.

[0094] Furthermore, the gradient of the loss function is calculated to update the weights of the autoencoder. The calculation method of the gradient of the loss function to the weight is expressed as:

[0095] In the formula, is the symbol of partial derivative; is the reconstruction loss function of the autoencoder.

[0096] 3) Update the autoencoder parameters based on the gradient and pre-computed adaptive weight adjustment coefficients.

[0097] Traditional machine learning algorithms usually use a fixed learning rate and a uniform update rate. The importance of different liquid chromatograph data features is relatively simple to learn and lacks flexibility. In combination with the above steps, the embodiment of the present invention adjusts the feature weights through dynamic learning and updates the parameters in combination with the adaptive weight adjustment coefficient. Specifically, through the adaptive weight adjustment coefficient of the feature, different update rates are adaptively assigned to features of different importance during the training process, so that the model can accelerate learning for key features, while using more stable updates for secondary features, thereby improving the flexibility and accuracy of dimensionality reduction. For example, in solvent component analysis, the absorbance characteristics at certain specific wavelengths may be crucial for the detection of target substances, while the baseline drift state recorded in the log may correspond to secondary features. By accelerating the learning of features in high-density areas (such as the stable operation stage) and conservatively updating features in low-density areas (such as the transient state of instrument startup), the model can more accurately capture the changes in the main components of the solvent composition while suppressing temporary equipment noise recorded in the log. The calculation method is expressed as:

[0098] In the formula, is the weight matrix of the updated autoencoder; is the learning rate of the autoencoder; is the adaptive weight adjustment coefficient of the feature.

[0099] Furthermore, the adaptive weight adjustment coefficient of the feature is dynamically calculated by local density, and the calculation method is expressed as:

[0100] In the formula, is the adjustment factor of the autoencoder; is the local density of the input liquid chromatography data of the autoencoder; is an exponential function. Preferably, Set to 0.2.

[0101] Furthermore, the local density of the input HPLC data of the autoencoder measures the importance of the feature in the training HPLC data, and the calculation method is expressed as:

[0102] In the formula, is the density adjustment factor parameter; Characterizes the L2 norm between the i-th data point and the target data point. Preferably, Set to 0.5.

[0103] Step S210 , performing interaction measurement on the data points of the initial dimensionality reduction features, and determining a feature mapping from the initial dimensionality reduction features.

[0104] In combination with the above embodiments, the embodiments of the present invention also define an energy function as a measure of the interaction energy between a data point and other data points during the dimension reduction process in the feature dimension reduction process. The energy function takes into account the influence of distance, similarity, and local density to maintain the local structure of the liquid chromatograph data and ensure the smoothness of the dimension reduction process to avoid excessive stretching or compression. For example, if there are short-term abnormal jumps in the temperature data collected by the sensor (such as transient sensor failure), the gradient smoothing strategy in the energy function can suppress the stretching effect of such abnormal points on the overall dimension reduction result.

[0105] In the specific implementation, the gradient change corresponding to the data points of the initial dimensionality reduction feature and the geometric distance between the data points are determined; based on the gradient change and the geometric distance, the interaction energy of the data points of the initial dimensionality reduction feature is determined; the interaction energy is minimized to determine the feature mapping corresponding to the initial dimensionality reduction feature. Specifically, the calculation method is expressed as:

[0106] In the formula, is the energy function; is the weight coefficient of similarity, representing the data point and the relative importance of represents the Euclidean distance between two points, is the density adjustment factor parameter; It is a data point The gradient of , characterizes the local deformation of the liquid chromatography data.

[0107] Furthermore, by minimizing the energy function as the goal, a feature mapping that minimizes the energy is sought to optimize the feature dimension reduction effect. In one embodiment, the energy function is optimized using a gradient descent method, and the positions of the data points are gradually adjusted by updating the gradient of the energy function to achieve overall energy minimization. In the gradient descent method, the gradient calculation method of the energy function with respect to the data point is expressed as:

[0108] In the formula, Characterize the energy function for The gradient of the data point; Indicates The second derivative of each data point characterizes the acceleration of local changes in liquid chromatography data.

[0109] Furthermore, the embodiment of the present invention can also measure the interaction of data points based on the energy function to train the autoencoder. Among them, the embodiment of the present invention also adopts a dynamic adaptive adjustment mechanism to adjust the learning rate, automatically adjust the learning rate of the autoencoder according to the current energy value and gradient changes, and avoid the convergence problem caused by the fixed learning rate, which is expressed as:

[0110] In the formula, is the initial learning rate of the autoencoder; is a hyperparameter that controls the learning rate decay. Preferably, Set to 0.95, Set to 0.01.

[0111] Furthermore, the convergence conditions of the autoencoder used in the embodiment of the present invention are as follows: Dynamically determine whether convergence is satisfied based on loss changes and weight stability to prevent premature stopping or overfitting. At the same time, in order to examine the stability of weights at the same time, add a measure of weight differences when calculating loss changes. For example, when storing months of historical chromatograph data, equipment performance degradation may cause slow drift of feature distributions. Traditional fixed threshold convergence is prone to early stopping. After using the weight difference term, the model can continue to track small changes in the weight matrix. Based on this, when the loss function changes very little, it is determined that the model training has reached a relatively stable process, and the model performance is basically stable, and training can be stopped.

[0112] In the scenario where data distribution changes gradually, extend the training rounds to ensure that the dimensionality reduction model fully adapts to the feature shift caused by device aging. The judgment conditions are:

[0113] In the formula, is the weight matrix of the autoencoder of the previous iteration; is the convergence adjustment factor; is the loss function value of the i-th iteration of the autoencoder; is the loss function value of the i-1th iteration of the autoencoder. Preferably, Set to 0.1. In summary, the above steps are repeated and iterated until the convergence conditions are met, which means that the autoencoder model training is completed.

[0114] Step S212, performing adaptive smoothing processing on the feature map to obtain a dimension-reduced data representation corresponding to the operation monitoring data.

[0115] The embodiment of the present invention uses the trained autoencoder parameters to apply to the liquid chromatograph data to obtain a more compact and effective feature representation. The output feature representation after dimensionality reduction is as follows:

[0116] In the formula, It is the low-dimensional feature representation after dimensionality reduction; is the weight matrix of the updated autoencoder; is the updated bias term of the autoencoder, and the update method is the same as the weight matrix of the autoencoder; is the activation function of the autoencoder.

[0117] In order to enhance the expression ability of nonlinear liquid chromatograph data, the embodiment of the present invention dynamically calculates the adaptive weight adjustment coefficient by measuring the local density of each data point of the operation monitoring data in its neighborhood. Specifically, the smoothness is controlled by the adaptive parameter in the activation function, and the calculation method is expressed as:

[0118] In the formula, is an adaptive parameter for controlling the smoothness of the activation function. Preferably, Set to 0.1.

[0119] Step S214, loading the dimension-reduced data representation into a pre-trained state monitoring model, and outputting a state analysis result corresponding to the dimension-reduced data representation.

[0120] Step S216: Based on the status analysis result, the working status of the liquid chromatograph is evaluated.

[0121] In summary, another method for liquid chromatograph status monitoring and predictive maintenance provided by an embodiment of the present invention jointly models liquid chromatograph sensor data (such as continuous parameters such as flow, temperature, and pressure) and log data (such as discrete information such as alarm status and equipment events), and uses a deep high-order neural network for data fusion and processing, thereby improving the context perception ability of the model. When an abnormality or failure occurs, it can comprehensively consider multiple factors for accurate diagnosis.

[0122] In addition, an autoencoder algorithm based on geometric similarity is used to obtain the relative relationship between input data points through geometric similarity calculation, and the path reconstruction strategy is combined to perform feature dimensionality reduction. When reducing the dimensionality of high-dimensional data, the local structure and global information of the liquid chromatograph data are maintained, while reducing the information loss in the dimensionality reduction process. In the process of feature dimensionality reduction, the adaptive weight adjustment coefficient of the feature is used, and different update rates are assigned to features of different importance by dynamically calculating the local density. For high-density areas (such as the stable operation stage), the features are accelerated to learn, while for low-density areas (such as the startup transient stage), a more conservative update strategy is adopted. When dealing with the nonlinear and time-continuous characteristics of the liquid chromatograph, the robustness and flexibility of the model are enhanced.

[0123] Furthermore, a high-order neural network classification algorithm combined with fractional calculus for gradient correction is adopted. The historical gradient is corrected by fractional order during model training to enhance convergence stability and suppress overfitting, so that the network can remember the historical gradient information. It performs well when the feature shift is caused by equipment aging, avoids the instability brought by the traditional gradient descent method that only relies on the instantaneous gradient, and shows higher accuracy and reliability when monitoring the complex state changes of the liquid chromatograph.

[0124] Further, refer to Figure 3 , showing the performance of the geometric similarity measurement-based autoencoder (GS-AE) and the traditional principal component analysis (PCA) algorithm under different hyperparameter settings of the embodiment of the present invention. Hyperparameters such as geometric similarity weight coefficient, feature learning rate, etc. may have a significant impact on the dimensionality reduction results. In order to evaluate the performance stability and superiority of GS-AE under different hyperparameter settings, GS-AE can better preserve the global structure and local features of liquid chromatograph data under different hyperparameter conditions. PCA, on the other hand, shows a greater sensitivity to hyperparameter changes, and the reconstruction error after dimensionality reduction is larger. The experimental results show that GS-AE can obtain a lower reconstruction error during the hyperparameter optimization process, showing its advantage in maintaining key information of the data.

[0125] Further, Figure 4The accuracy of various common dimensionality reduction algorithms such as GS-AE, PCA, t-SNE and local linear embedding (LLE) in classification tasks after dimensionality reduction of liquid chromatography data is shown. By verifying the effects of different dimensionality reduction algorithms in data classification tasks after dimensionality reduction, it is shown that GS-AE can retain more useful information and improve classification accuracy in the complex dimensionality reduction process of liquid chromatography data. GS-AE showed the highest classification accuracy among all methods, especially when faced with complex high-dimensional data, it can better maintain the classification structure of the data. Traditional PCA and t-SNE methods have low classification accuracy after dimensionality reduction due to over-simplification of data or failure to fully capture global information. The advantage of GS-AE is that it can better maintain the detailed features in liquid chromatography data, thereby significantly improving the classification effect.

[0126] Furthermore, based on the above embodiment, the embodiment of the present invention also provides a liquid chromatograph state monitoring and predictive maintenance device, Figure 5 FIG. 4 shows a schematic diagram of a liquid chromatograph state monitoring and predictive maintenance device provided by an embodiment of the present invention, such as Figure 5 As shown, the device includes: a data acquisition module 100, used to obtain the operation monitoring data of the liquid chromatograph to be monitored; a calculation module 200, used to perform distance measurement on the data points of the operation monitoring data, and determine the geometric similarity of the data points corresponding to the operation monitoring data; a data processing module 300, used to use a pre-trained autoencoder to perform dimensionality reduction processing on the operation monitoring data based on the geometric similarity of the data points, and determine the reduced-dimensional data representation corresponding to the operation monitoring data; an execution module 400, used to load the reduced-dimensional data representation into a pre-trained state monitoring model, and output a state analysis result corresponding to the reduced-dimensional data representation; wherein the state monitoring model updates the model parameters based on the fractional-order corrected cross entropy loss function; a predictive maintenance module 500, used to evaluate the working state of the liquid chromatograph based on the state analysis result.

[0127] An embodiment of the present invention provides a liquid chromatograph condition monitoring and predictive maintenance device, whose implementation principle and technical effects are the same as those of the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the device embodiment, reference may be made to the corresponding contents in the aforementioned method embodiment.

[0128] Among them, the state monitoring model is constructed based on a high-order neural network; the execution module 400 is also used to forward propagate the dimension reduction data representation through multiple hidden layers of the high-order neural network; and, using the deep structure of the high-order neural network to fuse the dimension reduction data representation layer by layer, and determine the state analysis result corresponding to the dimension reduction data representation. The execution module 400 is also used to train the preset initial high-order neural network using a preset training sample set, and determine the cross entropy loss function corresponding to the initial high-order neural network; use a preset fractional-order calculus operator to perform fractional-order correction on the cross entropy loss function and determine the fractional-order loss gradient; based on the fractional-order loss gradient and the cross entropy loss function, back-propagate and update the model parameters of the initial high-order neural network; until the initial high-order neural network meets the preset iteration conditions, the state monitoring model is constructed based on the initial high-order neural network.

[0129] The calculation module 200 is also used to measure the distance between data points of the operation monitoring data based on the dimension-by-dimensional difference between the data points of the operation monitoring data; based on the data point distance and the preset similarity weight coefficient, calculate the geometric similarity matrix corresponding to the operation monitoring data, and determine the geometric similarity of the data points corresponding to the operation monitoring data.

[0130] The data processing module 300 is also used to perform weighted mapping of the operation monitoring data using a preset autoencoder based on the geometric similarity of the data points to determine the initial dimension reduction features; to measure the interaction of the data points of the initial dimension reduction features to determine the feature mapping from the initial dimension reduction features; to perform adaptive smoothing on the feature mapping to obtain the dimension reduction data representation corresponding to the operation monitoring data. The data processing module 300 is also used to determine the gradient change corresponding to the data points of the initial dimension reduction features and the geometric distance between the data points; to determine the interaction energy of the data points of the initial dimension reduction features based on the gradient change and the geometric distance; to minimize the interaction energy and determine the feature mapping corresponding to the initial dimension reduction features. The data processing module 300 is also used to train the preset autoencoder through a preset training sample set and calculate the loss function of the autoencoder; to introduce the constraint condition of minimizing the path difference into the loss function, and calculate the gradient of the loss function to the weight of the autoencoder; to update the parameters of the autoencoder based on the gradient and the pre-calculated adaptive weight adjustment coefficient. The data processing module 300 is also used to measure the local density of each data point of the operation monitoring data in its neighborhood and dynamically calculate the adaptive weight adjustment coefficient.

[0131] Furthermore, an embodiment of the present invention also provides a liquid chromatograph condition monitoring and predictive maintenance system, which is configured with the device of the above embodiment, for executing Figure 1 to Figure 2 The method of any embodiment. Figure 6A logical schematic diagram corresponding to an embodiment of the present invention is shown, wherein the liquid chromatograph state monitoring and predictive maintenance system ensures the stable operation of the liquid chromatograph, improves maintenance efficiency, and reduces the risk of unplanned downtime through the collaborative work of multiple functional units such as data collection, feature extraction, model training, predictive analysis, and maintenance suggestions.

[0132] An embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above Figure 1 to Figure 2 The embodiment of the present invention also provides a computer-readable storage medium on which a computer program is stored. The computer program is executed by a processor to perform the above Figure 1 to Figure 2 The embodiment of the present invention also provides a structural diagram of an electronic device, such as Figure 7 FIG. 1 is a schematic diagram of the structure of the electronic device, wherein the electronic device includes a processor 71 and a memory 70, the memory 70 stores computer executable instructions that can be executed by the processor 71, and the processor 71 executes the computer executable instructions to implement the above Figures 1 to 3 The method shown. Figure 7In the illustrated embodiment, the electronic device further includes a bus 72 and a communication interface 73, wherein the processor 71, the communication interface 73 and the memory 70 are connected via the bus 72. The memory 70 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 73 (which may be wired or wireless), and the Internet, a wide area network, a local area network, a metropolitan area network, etc. may be used. The bus 72 may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. It may also be an AMBA (Advanced Microcontroller Bus Architecture) bus, where AMBA defines three types of buses, including an APB (Advanced Peripheral Bus) bus, an AHB (Advanced High-performance Bus) bus, and an AXI (Advanced Xtensible Interface) bus. The bus 72 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 7 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0133] The processor 71 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit or software instructions in the processor 71. The above processor 71 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiment of the present application can be directly embodied as a hardware decoding processor to execute, or the hardware and software modules in the decoding processor can be executed. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor 71 reads the information in the memory and combines its hardware to complete the above Figure 1 to Figure 2 Any of the methods shown.

[0134] A computer program product of a method and system for condition monitoring and predictive maintenance of a liquid chromatograph provided in an embodiment of the present invention includes a computer-readable storage medium storing a program code, and the instructions included in the program code can be used to execute the method described in the previous method embodiment. The specific implementation can refer to the method embodiment, which will not be repeated here. A person skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the system described above can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here. In addition, in the description of the embodiment of the present invention, unless otherwise clearly specified and limited, the terms "installation", "connection" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in the present invention can be understood in specific circumstances. If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a disk or an optical disk, and other media that can store program codes. In the description of the present invention, it should be noted that the orientation or position relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inside", "outside", etc. is based on the orientation or position relationship shown in the accompanying drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second", and "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.Finally, it should be noted that the above embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above embodiments, those skilled in the art should understand that any person skilled in the art can still modify the technical solutions recorded in the above embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for condition monitoring and predictive maintenance of a liquid chromatograph, characterized in that: The method comprises: Acquiring operation monitoring data of the liquid chromatograph to be monitored; Performing distance measurement on the data points of the operation monitoring data to determine the geometric similarity of the data points corresponding to the operation monitoring data; Using a pre-trained autoencoder to perform dimensionality reduction processing on the operation monitoring data based on the geometric similarity of the data points, and determining a dimensionality reduction data representation corresponding to the operation monitoring data; Loading the reduced-dimensional data representation into a pre-trained state monitoring model, and outputting a state analysis result corresponding to the reduced-dimensional data representation; wherein the state monitoring model updates model parameters based on a fractional-order corrected cross entropy loss function; Based on the status analysis result, the working status of the liquid chromatograph is evaluated.

2. The method according to claim 1, characterized in that The condition monitoring model is constructed based on a high-order neural network; the steps of loading the reduced-dimensional data representation into a pre-trained condition monitoring model and outputting a condition analysis result corresponding to the reduced-dimensional data representation include: The reduced-dimensional data representation is forward-propagated through multiple hidden layers of the high-order neural network; and the reduced-dimensional data representation is layer-by-layer fused using the deep structure of the high-order neural network to determine a state analysis result corresponding to the reduced-dimensional data representation.

3. The method according to claim 2, characterized in that The method further comprises: Training a preset initial high-order neural network using a preset training sample set, and determining a cross entropy loss function corresponding to the initial high-order neural network; Performing fractional-order correction on the cross entropy loss function using a preset fractional-order calculus operator to determine the fractional-order loss gradient; Based on the fractional-order loss gradient and the cross-entropy loss function, back-propagation update is performed on the model parameters of the initial high-order neural network; Until the initial high-order neural network meets the preset iteration condition, a state monitoring model is constructed based on the initial high-order neural network.

4. The method according to claim 1, characterized in that: The step of performing distance measurement on the data points of the operation monitoring data to determine the geometric similarity of the data points corresponding to the operation monitoring data comprises: Measuring the distance between data points of the operation monitoring data based on the dimension-by-dimensional difference between the data points of the operation monitoring data; Based on the data point distance and a preset similarity weight coefficient, a geometric similarity matrix corresponding to the operation monitoring data is calculated to determine the geometric similarity of the data points corresponding to the operation monitoring data.

5. The method according to claim 1, characterized in that: The step of performing dimensionality reduction processing on the operation monitoring data based on the geometric similarity of the data points using a pre-trained autoencoder to determine the dimensionality reduction data representation corresponding to the operation monitoring data comprises: Based on the geometric similarity of the data points, a preset autoencoder is used to perform weighted mapping on the operation monitoring data to determine an initial dimensionality reduction feature; performing interaction measurement on data points of the initial reduced-dimensionality features and determining a feature map from the initial reduced-dimensionality features; The feature map is adaptively smoothed to obtain a dimension-reduced data representation corresponding to the operation monitoring data.

6. The method according to claim 5, characterized in that The step of performing interaction measurement on the data points of the initial dimensionality reduction features and determining a feature map from the initial dimensionality reduction features comprises: Determine the gradient change corresponding to the data points of the initial dimension reduction feature, and the geometric distance between the data points; Determining the interaction energy of the data points of the initial dimensionality reduction feature based on the gradient change and the geometric distance; The interaction energy is minimized to determine a feature map corresponding to the initial dimensionality reduction feature.

7. The method according to claim 5, characterized in that The training method of the autoencoder comprises: Training a preset autoencoder using a preset training sample set, and calculating a loss function of the autoencoder; Introducing a constraint condition of minimizing the path difference into the loss function, and calculating the gradient of the loss function with respect to the weight of the autoencoder; Based on the gradient and the pre-calculated adaptive weight adjustment coefficient, the parameters of the autoencoder are updated.

8. The method according to claim 7, characterized in that The method for calculating the adaptive weight adjustment coefficient includes: The local density of each data point of the operation monitoring data in its neighborhood is measured, and an adaptive weight adjustment coefficient is dynamically calculated.

9. A liquid chromatograph condition monitoring and predictive maintenance device, characterized in that: The device comprises: A data acquisition module, used for acquiring operation monitoring data of the liquid chromatograph to be monitored; A calculation module, used to perform distance measurement on the data points of the operation monitoring data to determine the geometric similarity of the data points corresponding to the operation monitoring data; A data processing module, configured to perform dimensionality reduction processing on the operation monitoring data based on the geometric similarity of the data points using a pre-trained autoencoder, and determine a dimensionality reduction data representation corresponding to the operation monitoring data; An execution module, used for loading the reduced-dimensionality data representation into a pre-trained state monitoring model, and outputting a state analysis result corresponding to the reduced-dimensionality data representation; wherein the state monitoring model updates model parameters based on a fractional-order corrected cross entropy loss function; A predictive maintenance module is used to evaluate the working status of the liquid chromatograph based on the status analysis result.

10. A liquid chromatograph condition monitoring and predictive maintenance system, characterized in that: The system is configured with the apparatus according to claim 9, and is used to execute the method according to any one of claims 1 to 8.

Images