Method and device for predicting service life of chemical filter

Abstract

The present invention relates to the technical field of life prediction, and discloses a method and device for predicting the service life of a chemical filter, wherein the method comprises: collecting historical multi-dimensional usage status data of the chemical filter and performing standardized preprocessing to obtain a time series feature data set; performing load feature clustering analysis to obtain a filter load mode feature vector; constructing a double-layer performance attenuation prediction model, and outputting a target performance attenuation prediction sequence through physical attenuation coefficient calculation and deep learning network training; performing filter state probability transfer calculation to obtain a state transfer prediction sequence; performing multi-objective joint optimization on the state transfer prediction sequence, and calculating the optimal operation control parameters; collecting real-time multi-dimensional usage status data based on the optimal operation control parameters, and performing parameter tuning on the double-layer performance attenuation prediction model to obtain a service life prediction model, wherein the method improves the accuracy and reliability of chemical filter life prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of life prediction, and in particular to a method and device for predicting the service life of a chemical filter. Background Art

[0002] Traditional chemical filter life prediction methods mainly rely on expert experience and simple mathematical and statistical models. It is difficult to accurately capture the performance degradation law of filters under complex working conditions, resulting in insufficient prediction accuracy and inability to provide reliable support for equipment maintenance decisions.

[0003] In recent years, with the development of industrial sensing technology and data acquisition systems, the operating status data of chemical filters has become multidimensional and large-scale. However, these data often suffer from complex time series correlations, variable load patterns, and severe noise interference. Traditional single modeling methods have difficulty effectively utilizing multi-source heterogeneous data for accurate predictions. In particular, the generalization ability of the prediction model decreases significantly when operating conditions change drastically. In addition, the performance degradation process of chemical filters is affected by both physical mechanisms and environmental factors. Existing prediction methods either over-rely on mechanistic models and ignore actual operating data, or adopt a completely data-driven approach without physical constraints, resulting in a large deviation between the prediction results and actual engineering requirements. Summary of the Invention

[0004] The present invention provides a method and device for predicting the service life of a chemical filter, which are used to improve the accuracy and reliability of the prediction of the service life of the chemical filter.

[0005] In a first aspect, the present invention provides a method for predicting the service life of a chemical filter, the method comprising:

[0006] Collect historical multi-dimensional usage status data of chemical filters and perform standardized preprocessing to obtain a time series feature dataset;

[0007] Performing load feature cluster analysis on the time series feature data set to obtain a filter load pattern feature vector;

[0008] A two-layer performance degradation prediction model is constructed based on the filter load pattern feature vector, and a target performance degradation prediction sequence is output through physical degradation coefficient calculation and deep learning network training;

[0009] Performing filter state probability transfer calculation on the target performance attenuation prediction sequence to obtain a state transfer prediction sequence;

[0010] Performing multi-objective joint optimization on the state transition prediction sequence and calculating optimal operation control parameters;

[0011] Based on the optimal operation control parameters, real-time multi-dimensional usage status data is collected, and parameters of the double-layer performance degradation prediction model are tuned to obtain a service life prediction model.

[0012] In a second aspect, the present invention provides a device for predicting the service life of a chemical filter, the device comprising:

[0013] The acquisition module is used to collect the historical multi-dimensional usage status data of the chemical filter and perform standardized preprocessing to obtain a time series feature data set;

[0014] A clustering module, configured to perform load feature clustering analysis on the time series feature data set to obtain a filter load pattern feature vector;

[0015] A construction module is used to construct a two-layer performance degradation prediction model based on the filter load pattern feature vector, and output a target performance degradation prediction sequence through physical degradation coefficient calculation and deep learning network training;

[0016] a calculation module, configured to perform filter state probability transfer calculation on the target performance attenuation prediction sequence to obtain a state transfer prediction sequence;

[0017] An optimization module, configured to perform multi-objective joint optimization on the state transition prediction sequence and calculate optimal operation control parameters;

[0018] A parameter tuning module is used to collect real-time multi-dimensional usage status data based on the optimal operation control parameters, and to perform parameter tuning on the double-layer performance degradation prediction model to obtain a service life prediction model.

[0019] The third aspect of the present invention provides a device for predicting the service life of a chemical filter, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the device for predicting the service life of the chemical filter executes the above-mentioned method for predicting the service life of the chemical filter.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium having instructions stored therein, which, when executed on a computer, enables the computer to execute the above-mentioned method for predicting the service life of a chemical filter.

[0021] In the technical solution provided by the present invention, a double-layer performance degradation prediction model is constructed, which combines physical mechanism constraints with deep learning, thereby improving the model accuracy while ensuring the interpretability of the prediction; a dual-time window load feature clustering analysis method is adopted to realize the accurate identification of load patterns of filters at different time scales, and enhance the model's adaptability to changes in working conditions; a state probability transfer calculation method based on a dual hidden Markov prediction model accurately captures the dynamic characteristics of filter performance degradation; a multi-objective joint optimization strategy including life maximization, cost minimization and efficiency maximization is designed to realize intelligent regulation of filter operating parameters; through a dynamic parameter tuning mechanism, the prediction model has the ability to adaptively update, which ensures the stability of the model's prediction performance during long-term operation; the present invention improves the accuracy and reliability of filter life prediction through multi-level feature analysis and prediction strategies.

[0022] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or understood by practicing the present invention. The purposes and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0023] 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

[0024] Figure 1 A schematic diagram of an embodiment of a method for predicting the service life of a chemical filter according to an embodiment of the present invention;

[0025] Figure 2 A schematic diagram of an embodiment of a device for predicting the service life of a chemical filter according to an embodiment of the present invention;

[0026] Figure 3 Schematic diagram of an embodiment of a device for predicting the service life of a chemical filter in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0028] The terms "including," "having," and any variations thereof, as used in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device comprising a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to the process, method, product, or device.

[0029] To facilitate understanding of this embodiment, a method for predicting the service life of a chemical filter disclosed in an embodiment of the present invention is first introduced in detail. Figure 1 As shown, this method includes the following steps:

[0030] 101. Collect historical multi-dimensional usage status data of chemical filters and perform standardized preprocessing to obtain a time series feature dataset;

[0031] It is understandable that the execution subject of the present invention may be a device for predicting the service life of a chemical filter, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0032] Specifically, historical multi-dimensional usage data of chemical filters is collected, including monitoring of multiple key parameters such as inlet pressure, outlet pressure, flow rate, filter media concentration, filtration temperature, filtration efficiency, and filter saturation. This data collectively constitutes the comprehensive usage status of the filter. Data cleaning is performed on this historical multi-dimensional usage data to remove outliers and noise, fill in missing values, and ensure data integrity and accuracy, generating a preprocessed dataset. This preprocessed dataset is then normalized. Different status parameters have different dimensions and numerical ranges. These parameters are converted to a standard range to eliminate the impact of dimensional differences and ensure comparability in subsequent model construction. Normalization methods, including Z-score normalization and Min-Max normalization, effectively transform the data into a uniform distribution, making subsequent analysis and modeling more stable and efficient. The normalized dataset is then time-windowed, dividing the data into multiple time periods to capture dynamic characteristics over time. When dividing the time windows, overlapping time windows are set, meaning that each time period has a certain overlap, to ensure that data continuity and changing trends are captured. In this way, multiple time series data segments are obtained, representing the filter's usage status during different time periods. These time series data segments are then labeled to convert them into time series feature datasets. This process relies on expert knowledge or prior experience to identify the characteristic status of each time period, such as normal filter operation, performance degradation, or near-saturation.

[0033] 102. Perform load feature cluster analysis on the time series feature data set to obtain a filter load pattern feature vector;

[0034] Specifically, the cumulative filtering volume for the first and second time windows of the time series feature dataset is calculated. These time windows are set to 24 hours and 48 hours, respectively. Based on these time windows, the first and second load time series indicators are obtained, reflecting the load status of the filter in different time periods. Segmented thresholds are calculated for the first and second load time series indicators. The first load time series indicator is divided into a first low-load segment, a first medium-load segment, and a first high-load segment, while the second load time series indicator is divided into a second low-load segment, a second medium-load segment, and a second high-load segment. This segmentation process more clearly identifies the filter's performance under different load conditions, resulting in first and second load segmentation results. Based on the segmentation results, feature value extraction is performed on the first and second load segmentation results, including calculating the first pressure change rate, the first temperature distribution feature, the first efficiency decay rate, as well as the second pressure change rate, the second temperature distribution feature, and the second efficiency decay rate. This calculation results in the first and second feature parameter sets. A two-layer feature correlation matrix is ​​constructed based on the first and second feature parameter sets to capture the correlations between different features and obtain multi-scale feature correlation results. Principal component analysis is performed on the multi-scale feature association results to extract the feature combination of the first principal component and the second principal component to form a multi-dimensional dimensionality reduction feature set. Double clustering calculation is performed on the multi-dimensional dimensionality reduction feature set to obtain the first cluster center and the second cluster center respectively. Through these two cluster centers, the first distance matrix and the second distance matrix are established to perform pattern discrimination. The pattern discrimination process effectively classifies the load pattern and obtains multi-level load pattern labels that reflect the load characteristics of the filter under different working conditions. The multi-level load pattern labels are combined with the multi-dimensional dimensionality reduction feature set, and the filter load pattern feature vector is finally obtained through the cascade transformation of the first feature map and the second feature map.

[0035] 103. A two-layer performance degradation prediction model is constructed based on the filter load pattern feature vector. Through physical degradation coefficient calculation and deep learning network training, the target performance degradation prediction sequence is output;

[0036] Specifically, the filter load pattern characteristic vector is divided into a first characteristic subvector and a second characteristic subvector. The first characteristic subvector includes the pressure change rate and efficiency decay rate, which directly reflect the filter's load during operation. The second characteristic subvector contains temperature distribution characteristics and the saturation change rate, which affect filter performance. Based on the first characteristic subvector, a physical attenuation layer is constructed. This layer consists of an attenuation parameter calculation unit and a physical constraint unit. The attenuation parameter calculation unit uses a polynomial fitting structure and a third-order polynomial to perform parameter fitting. This method effectively describes the filter's attenuation behavior under different load conditions. Furthermore, an adaptive weight module is implemented to dynamically adjust the attenuation parameter weights based on feedback from different operating conditions, ensuring flexible model adaptability. The physical constraint unit utilizes a boundary constraint structure to ensure that the prediction results are consistent with physical reality. Maximum pressure drop and minimum efficiency constraints are introduced in the physical constraint unit to ensure that the predicted performance decay does not exceed the actual operating range of the filter. After the physical attenuation layer is constructed, parameter training is performed on the attenuation parameter calculation unit. During training, the first feature subvector is input into the physical attenuation layer. By calculating the attenuation coefficient, a first performance decay prediction sequence is generated, which contains the performance decay trend of the filter over a period of time. A deep learning layer is constructed based on the second feature subvector and the first performance decay prediction sequence. This layer primarily consists of two modules: a time series feature extractor and a performance predictor. The time series feature extractor uses a bidirectional long short-term memory (Bi-LSTM) architecture to capture temporal dependencies in time series data. The time series feature extractor comprises three layers of bidirectional LSTM units, each with 128 hidden nodes, with dropout layers between layers to reduce the risk of overfitting. The performance predictor uses a residual attention network architecture, which effectively extracts deep features. This module consists of five residual blocks, each containing two fully connected layers and a channel attention module. This mechanism uses the attention mechanism to enhance the weights of key features and improve the model's predictive ability. The first performance decay prediction sequence and the second feature subvector are input into the deep learning layer for training. The model learns the complex relationship between the two to generate a second performance decay prediction sequence. This sequence provides information on the expected performance changes of the filter under different operating conditions. To improve the accuracy and reliability of the predictions, a fusion optimization layer is constructed. This layer includes an adaptive weight module and a timing calibration module. It dynamically weights and fuses the first and second performance decay prediction sequences to form the final performance prediction results. The adaptive weight module performs online updates based on the historical performance decay data of the chemical filter, ensuring that the model remains adaptable to new data. The timing calibration module is used to align the fusion results in time, ensuring that predictions at different time points can be properly connected.Finally, the output results of the fusion optimization layer are converted into a target performance attenuation prediction sequence through time series reconstruction, reflecting the future performance trend of the filter.

[0037] 104. Perform filter state probability transfer calculation on the target performance attenuation prediction sequence to obtain a state transfer prediction sequence;

[0038] Specifically, the target performance degradation prediction sequence is segmented into first and second time windows according to time windows. First and second segmented sequence samples are generated, respectively reflecting the characteristics of filter performance variations at different time scales. Dual feature extraction is performed on the first and second segmented sequence samples to analyze state characteristics within different time periods, thereby better understanding the filter's operating state and potential performance degradation mechanisms. Feature extraction yields first and second state feature sequences, respectively. Based on these two state feature sequences, a two-layer state discrimination model is constructed. The first layer performs local feature analysis to capture state changes within a specific time window, while the second layer performs global feature analysis to identify long-term trends and patterns throughout the entire usage cycle. This two-layer analysis calculates the first and second state probability vectors for the chemical filter, reflecting the probability distribution of the filter under different states. These state probability vectors are input into a dual hidden Markov prediction model. This model consists of two components: a first prediction sub-model and a second prediction sub-model. The first prediction sub-model processes the first state probability vector to calculate the corresponding transition probabilities, while the second prediction sub-model processes the second state probability vector to calculate its transition probabilities. Obtain state transition probabilities at two time scales and analyze the dynamic relationships between different states. Perform weighted fusion and multi-scale rolling prediction on the first and second transition probability matrices. Combining information from different time scales enables more accurate state transition predictions. Weighted fusion assigns different weights to each time scale based on its reliability and effectiveness, ensuring the accuracy and rationality of the final results. Multi-scale rolling predictions take into account the continuity and dynamic changes of the time series during prediction, enabling the model to continuously adapt to new data. Double smoothing is performed on the generated multi-scale state prediction sequence to reduce noise interference on the prediction results, making the final state transition prediction sequence smoother and more stable. Appropriate smoothing algorithms, such as sliding average or exponential smoothing, can effectively improve the accuracy and reliability of state transition predictions.

[0039] 105. Perform multi-objective joint optimization on the state transition prediction sequence and calculate the optimal operation control parameters;

[0040] Specifically, a first objective function and a second objective function are constructed for the state transition prediction sequence. The first objective function includes lifespan maximization and efficiency maximization terms, reflecting a direct focus on filter performance; the second objective function includes operating cost minimization and replacement cost minimization terms, reflecting a pursuit of economic benefits. This ensures that both performance and cost are considered during the optimization process, forming a dual-objective optimization function. Based on the dual-objective optimization function, a first constraint set and a second constraint set are set. The first constraint set primarily defines the filter's pressure and efficiency variation ranges to ensure that the optimization process complies with the physical limitations and safety standards of the equipment. The second constraint set defines the filter's lifespan and cost ranges to ensure that the optimization results are economically feasible. Based on the dual-objective optimization function and its constraints, first and second weight coefficients are assigned. The analytic hierarchy process is used to calculate the weights of the objective function and constraints, forming a multi-level weight matrix. Based on this multi-level weight matrix, a first and second Pareto-optimal solution set are constructed, corresponding to the short-term and long-term optimization objectives, respectively. This approach simultaneously considers optimization requirements across different timescales, providing more options for the final decision. The double-layer Pareto solution set is screened for non-inferior solutions. The solution set is sorted using the first and second objective criteria to select a sequence of candidate solutions. This screening process selects the most promising solutions, forming a more refined set of solutions. Based on the candidate solution sequence, an operating parameter mapping relationship is constructed, and the optimal solution is mapped to the dynamic adjustment parameters of the filter's pressure, filtration temperature, and filtration flow rate, resulting in a specific set of parameter adjustment sequences. The parameter adjustment sequence is then optimized for time-series coordination. Parameter smoothing is performed based on the dual constraints of the first and second time scales to ensure that, in actual applications, the adjustment process does not experience drastic fluctuations, maintaining stable equipment operation and obtaining a coordinated control sequence. The coordinated control sequence is converted into a combined instruction set of the filter's pressure control instructions, temperature control instructions, and flow control instructions to obtain the optimal operating control parameters.

[0041] 106. Based on the optimal operation control parameters, real-time multi-dimensional usage status data is collected, and the parameters of the double-layer performance degradation prediction model are tuned to obtain a service life prediction model.

[0042] Specifically, based on optimal operating control parameters, chemical filters are monitored in real time to collect real-time, multi-dimensional usage data, including inlet pressure, outlet pressure, flow rate, filter media concentration, filtration temperature, filtration efficiency, and filter saturation. By continuously monitoring key parameters, the filter's operating status is understood. This real-time, multi-dimensional usage data is input into a two-layer performance degradation prediction model. The physical degradation layer performs inference operations, combining real-time data to evaluate the filter's physical performance. Simultaneously, the deep learning layer performs online inference, analyzing the real-time input data to generate a real-time performance degradation prediction sequence reflecting the filter's performance trends under current operating conditions. The real-time performance degradation prediction sequence is compared with actual performance data to calculate prediction errors. These errors are categorized and statistically analyzed according to pressure error, efficiency error, and lifespan error to construct an error characteristic matrix. This matrix reflects the accuracy of the model's predictions and reveals any deviations or deficiencies in the prediction process. The cause of the error characteristic matrix is ​​analyzed. Key influencing parameters in both the physical degradation layer and the deep learning layer are identified. By identifying these influencing parameters, clear parameter optimization directions are identified, ensuring that subsequent model updates can effectively improve prediction accuracy. Based on the determined parameter optimization direction, the physical and network parameters in the two-layer performance degradation prediction model are updated. Adjustments to the physical parameters typically involve factors such as the attenuation coefficient, efficiency, and pressure change, while network parameter updates involve adjustments to the weights, learning rate, and network structure within the deep learning model. This update process optimizes the model's performance, enabling it to better adapt to real-time data and improving overall predictive capabilities. Through these steps, a service life prediction model is obtained. This model can reflect the filter's performance degradation in real time and dynamically adjust based on the latest operating data, ensuring consistently high prediction accuracy in practical applications.

[0043] In an embodiment of the present invention, a double-layer performance degradation prediction model is constructed, which combines physical mechanism constraints with deep learning, thereby improving the model accuracy while ensuring the interpretability of the prediction; a dual-time window load feature clustering analysis method is adopted to achieve accurate identification of load patterns of filters at different time scales, thereby enhancing the model's adaptability to changes in working conditions; a state probability transfer calculation method based on a dual hidden Markov prediction model accurately captures the dynamic characteristics of filter performance degradation; a multi-objective joint optimization strategy including life maximization, cost minimization and efficiency maximization is designed to achieve intelligent regulation of filter operating parameters; through a dynamic parameter tuning mechanism, the prediction model has adaptive update capabilities, thereby ensuring the stability of the model's prediction performance during long-term operation; the present invention improves the accuracy and reliability of filter life prediction through multi-level feature analysis and prediction strategies.

[0044] In a specific embodiment, the process of executing step 101 may specifically include the following steps:

[0045] Collect data on the chemical filter's inlet pressure, outlet pressure, flow rate, filter medium concentration, filter temperature, filter efficiency, and filter saturation to obtain historical multi-dimensional usage status data;

[0046] Perform data cleaning on historical multi-dimensional usage status data to obtain a preprocessed data set;

[0047] Perform standardized mapping on the preprocessed data set, convert the state parameters of different dimensions into a standard range, and obtain a standardized data set;

[0048] Divide the standardized data set into time windows, set overlapping time windows to segment the standardized data set, and obtain time series data segments;

[0049] The time series data segments are labeled to obtain the time series feature dataset.

[0050] Specifically, data collection was performed on the chemical filter, focusing on several key parameters, including inlet pressure ( )、Outlet pressure( ),flow( ), filter medium concentration ( ), filtration temperature ( ), filtration efficiency ( ) and filter saturation ( Through effective data collection, historical multi-dimensional data on these parameters under different usage conditions is obtained. The data collection process is achieved by installing sensors and a data acquisition system. Sensors monitor the filter's inlet and outlet pressures in real time, flow meters measure flow rates, chemical analyzers detect filter media concentration, and temperature sensors monitor filtration temperature. These sensors send the collected data to the data processing system, forming a multi-dimensional raw data set containing information about the filter's status under various operating conditions. This historical multi-dimensional status data is cleaned to remove noise and erroneous data. For example, outliers caused by sensor failure or environmental factors are eliminated. This is achieved by setting reasonable thresholds or utilizing statistical methods (such as standard deviation and mean). For example, for the filter's outlet pressure data, its mean and standard deviation are calculated. Data is valid within a range of ±3 standard deviations from the mean. Data outside this range is considered an outlier and is eliminated. After data cleaning, a preprocessed data set is obtained. This preprocessed data set is normalized and mapped. State parameters of different dimensions are uniformly converted to a standard range for effective comparison and analysis. Normalization uses the Z-score normalization formula:

[0051] ;

[0052] in, is the normalized value, is the original data value, is the mean of the data, is the standard deviation. By standardization, the values ​​of various parameters are adjusted to the same scale so that they are within the same standard range, eliminating the impact of dimension on analysis. After standardization, the data set is divided into time windows. The standardized data is segmented according to the time series. Overlapping time windows are set to capture the time correlation of the data. Set a window size of The time window is expressed by the following formula:

[0053] ;

[0054] in, Indicates the time range of the segment. is the starting point of the time series, The window size is 1. Data labeling is performed on the segmented time series data to extract the time series feature dataset. Data labeling is based on actual application requirements. For example, filter performance indicators such as filtration efficiency and saturation can be set as labels. The following formula is used to calculate filtration efficiency:

[0055] ;

[0056] in, is the filtration efficiency, is the inlet concentration, is the outlet concentration. Through this step, the characteristic data in each time period are obtained.

[0057] In a specific embodiment, the process of executing step 102 may specifically include the following steps:

[0058] Calculate the cumulative filtering amount of the first time window and the second time window for the time series feature data set respectively to obtain the first load time series index and the second load time series index. The length of the first time window is 24 hours, and the length of the second time window is 48 hours.

[0059] Performing segmentation threshold calculation on the first load timing indicator and the second load timing indicator, dividing the first load timing indicator into a first low load segment, a first medium load segment, and a first high load segment, and dividing the second load timing indicator into a second low load segment, a second medium load segment, and a second high load segment, to obtain a first load segmentation result and a second load segmentation result;

[0060] Performing feature value extraction on the first load segmentation result and the second load segmentation result respectively, calculating the first pressure change rate, the first temperature distribution characteristic, the first efficiency decay rate and the second pressure change rate, the second temperature distribution characteristic, and the second efficiency decay rate, to obtain a first feature parameter set and a second feature parameter set;

[0061] A two-layer feature correlation matrix is ​​constructed based on the first feature parameter set and the second feature parameter set. A multi-scale feature correlation result is obtained by first-layer local correlation calculation and second-layer global correlation calculation. The multi-scale feature correlation result is subjected to principal component analysis. The feature combination of the first principal component and the second principal component is extracted to obtain a multi-dimensional reduced dimension feature set.

[0062] Perform double clustering calculation on the multi-dimensional dimensionality reduction feature set to obtain the first cluster center and the second cluster center respectively, and establish the first distance matrix and the second distance matrix through the first cluster center and the second cluster center for pattern discrimination to obtain multi-level load pattern labels;

[0063] The multi-level load pattern labels are combined with the multi-dimensional dimension reduction feature set, and the filter load pattern feature vector is obtained through the cascade transformation of the first feature map and the second feature map.

[0064] Specifically, the cumulative filtering amount of the first time window and the second time window is calculated to obtain the first load time series index and the second load time series index. The length of the first time window is 24 hours, and the length of the second time window is 48 hours. The filtering amount in each window is extracted from the time series feature data set and calculated using a cumulative method. The cumulative filtering amount is expressed as:

[0065] ;

[0066] in, is the cumulative filtration volume, For in time The flow rate value at The end time of the time window. The first load timing indicator and the second load timing indicator are segmented and calculated. The first load timing indicator is divided into a first low load segment, a first medium load segment and a first high load segment, and the second load timing indicator is divided into a second low load segment, a second medium load segment and a second high load segment. The segmentation process is implemented according to the set threshold. For example, the threshold of the load indicator is defined as and ,in is the upper limit of low load, is the lower limit of high load. The specific segmented formula is expressed as:

[0067] ;

[0068] in, Represents the value of the load timing indicator, is the minimum load value. Different load segments are identified to obtain the first load segmentation results and the second load segmentation results. Feature value extraction is performed on the load segmentation results. The first pressure change rate, the first temperature distribution characteristics, the first efficiency decay rate, and the second pressure change rate, the second temperature distribution characteristics, and the second efficiency decay rate are calculated. The pressure change rate is expressed as:

[0069] ;

[0070] in, is the pressure change rate, and are the pressure values ​​at the end and beginning of the time period, respectively, and and is the corresponding time point. By calculating these eigenvalues, the first feature parameter set and the second feature parameter set are obtained. A two-layer feature correlation matrix is ​​constructed based on the first feature parameter set and the second feature parameter set. The multi-scale feature correlation result is obtained by calculating the first layer local correlation and the second layer global correlation. Assuming that there is a eigenvector and , their correlation is expressed as:

[0071] ;

[0072] in, is the eigenvector and The correlation between and are the norms of the vectors, respectively. Principal component analysis is used to reduce the dimensionality of the multi-scale feature association results, extracting the feature combination of the first and second principal components to obtain a multi-dimensional reduced dimensionality feature set. Double clustering is performed on the multi-dimensional reduced dimensionality feature set. By analyzing the distribution of feature data, the first and second cluster centers are obtained, respectively. These two cluster centers are used to establish the first and second distance matrices for pattern discrimination. The distance matrix is ​​expressed by the following formula:

[0073] ;

[0074] in, For samples With sample The distance between and The samples are and samples In the feature dimension In this way, a multi-level load pattern label is obtained. The multi-level load pattern label is combined with the multi-dimensional dimensionality reduction feature set, and the filter load pattern feature vector is obtained through the cascade transformation of the first feature map and the second feature map.

[0075] In a specific embodiment, the process of executing step 103 may specifically include the following steps:

[0076] Dividing the filter load pattern characteristic vector into a first characteristic sub-vector and a second characteristic sub-vector, wherein the first characteristic sub-vector includes a pressure change rate and an efficiency decay rate, and the second characteristic sub-vector includes a temperature distribution feature and a saturation change rate;

[0077] Constructing a physical attenuation layer according to the first characteristic subvector, the physical attenuation layer includes an attenuation parameter calculation unit and a physical constraint unit, the attenuation parameter calculation unit adopts a polynomial fitting structure, and the physical constraint unit adopts a boundary constraint structure;

[0078] Parameter training is performed on the attenuation parameter calculation unit, which includes a third-order polynomial fitting module and an adaptive weight module, and the physical constraint unit includes a maximum pressure drop constraint and a minimum efficiency constraint;

[0079] Input the first characteristic sub-vector into the physical attenuation layer, and obtain the first performance attenuation prediction sequence by calculating the attenuation coefficient;

[0080] A deep learning layer is constructed based on the second feature sub-vector and the first performance decay prediction sequence. The deep learning layer includes a time series feature extractor and a performance predictor. The time series feature extractor adopts a bidirectional long short-term memory network structure, and the performance predictor adopts a residual attention network structure.

[0081] The temporal feature extractor consists of three layers of bidirectional LSTM units, each containing 128 hidden nodes, and a Dropout layer is set between the LSTM units; the performance predictor consists of five residual blocks, each containing two fully connected layers and a channel attention module;

[0082] Inputting the first performance decay prediction sequence and the second feature sub-vector into the deep learning layer, and obtaining the second performance decay prediction sequence through network training;

[0083] Constructing a fusion optimization layer, which includes an adaptive weight module and a timing calibration module, to dynamically weight the first performance decay prediction sequence and the second performance decay prediction sequence;

[0084] The weight parameters of the adaptive weight module are updated online according to the historical performance attenuation data of the chemical filter, and the fusion results are time-aligned through the timing calibration module to obtain the output results of the fusion optimization layer. The output results of the fusion optimization layer are time-reconstructed to generate the target performance attenuation prediction sequence.

[0085] Specifically, the filter load pattern characteristic vector is divided into a first characteristic sub-vector and a second characteristic sub-vector, where the first characteristic sub-vector includes the pressure change rate and the efficiency decay rate, while the second characteristic sub-vector contains the temperature distribution characteristics and the saturation change rate. By analyzing and selecting the dimensionality of the characteristic vector, it is ensured that each sub-vector represents a specific physical phenomenon. When constructing the physical attenuation layer, the attenuation parameter calculation unit and the physical constraint unit are designed based on the first characteristic sub-vector. The attenuation parameter calculation unit adopts a polynomial fitting structure and uses a third-order polynomial to represent the change in filter performance. The formula is:

[0086] ;

[0087] in, For time Performance indicators when are the coefficients of the polynomials, which are obtained by fitting historical data. At the same time, the physical constraint unit adopts a boundary constraint structure to ensure the rationality of the model under actual physical conditions, for example, setting the maximum pressure drop constraint and the minimum efficiency constraint. When training the parameters of the attenuation parameter calculation unit, a third-order polynomial fitting module and an adaptive weight module are constructed. Through the adaptive weight module, the weights of each parameter are dynamically adjusted according to the characteristics of the input data, making the model more flexible and effective. The goal of attenuation parameter training is to minimize the error between the predicted value and the true value, which is expressed as:

[0088] ;

[0089] in, is the predicted performance value, is the actual performance value, and is the number of data points. This process is iterated through an optimization algorithm (such as gradient descent) until convergence. After the first feature sub-vector is input into the physical attenuation layer, the first performance attenuation prediction sequence is obtained by calculating the attenuation coefficient, which reflects the performance change of the filter over a period of time. A deep learning layer is constructed based on the second feature sub-vector and the first performance attenuation prediction sequence. The deep learning layer consists of a time series feature extractor and a performance predictor, where the time series feature extractor adopts a bidirectional long short-term memory network (Bi-LSTM) structure to effectively capture the front-end and back-end dependencies in time series data. The time series feature extractor contains three layers of bidirectional LSTM units, each layer of LSTM units contains 128 hidden nodes, and a Dropout layer is set between the LSTM units to prevent overfitting. The output of LSTM is expressed as:

[0090] ;

[0091] in, is the hidden state at the current moment, is the current input, is the hidden state of the previous moment. Through this structure, the network can effectively extract time series features. The performance predictor adopts a residual attention network structure, which contains five residual blocks, each of which consists of two fully connected layers and a channel attention module. This design helps to capture the deep correlation between features and enhance the model's learning ability for complex data. By inputting the first performance decay prediction sequence and the second feature sub-vector into the deep learning layer, the second performance decay prediction sequence is obtained after network training. A fusion optimization layer is constructed. In this layer, the fusion optimization layer contains an adaptive weight module and a timing calibration module, which aims to dynamically weighted fuse the first performance decay prediction sequence and the second performance decay prediction sequence. The adaptive weight module is updated online through the historical performance decay data of the chemical filter to ensure the timeliness of the model parameters. Dynamic weighted fusion is expressed by the following formula:

[0092] ;

[0093] in, For the final fusion result, and are the first and second performance decay prediction sequences, respectively, and and The dynamically adjusted weight coefficients are used. The timing calibration module aligns the fusion results to ensure that the predictions are consistent with the actual application scenario, ultimately yielding the output of the fusion optimization layer. This result, after time reconstruction, generates the target performance degradation prediction sequence.

[0094] In a specific embodiment, the process of executing step 104 may specifically include the following steps:

[0095] Segment the target performance attenuation prediction sequence into a first time window and a second time window to obtain a first segmented sequence sample and a second segmented sequence sample;

[0096] Performing dual feature extraction on the first segment sequence sample and the second segment sequence sample respectively to obtain a first state feature sequence and a second state feature sequence;

[0097] A two-layer state discrimination model is constructed based on the first state feature sequence and the second state feature sequence, and the first state probability vector and the second state probability vector of the chemical filter are calculated through the first layer local feature analysis and the second layer global feature analysis;

[0098] Inputting the first state probability vector and the second state probability vector into the dual hidden Markov prediction model, calculating the state transition probabilities of the two time scales respectively through the first prediction sub-model and the second prediction sub-model, and obtaining the first transition probability matrix and the second transition probability matrix;

[0099] The first transfer probability matrix and the second transfer probability matrix are weightedly fused and multi-scale rolling predicted to obtain a multi-scale state prediction sequence, and the multi-scale state prediction sequence is double-smoothed to obtain a state transfer prediction sequence.

[0100] Specifically, the target performance attenuation prediction sequence is segmented into the first time window and the second time window. Assume that the length of the first time window is , the length of the second time window is , use the time index method to generate the first segment sequence sample and the second segment sequence sample, for example by selecting the time range to and to Sample data is used to obtain two segmented sequences. Dual feature extraction is performed on the first and second segmented sequence samples. Key state features are extracted from each time window. Feature extraction uses various methods, such as statistical analysis and frequency domain analysis, to obtain characteristic values ​​related to the filter state. The results of feature extraction form the first state feature sequence and the second state feature sequence, respectively. The following formula describes the state feature extraction:

[0101] ;

[0102] in, is the extracted eigenvalue, For the state samples, and is the feature extraction function. After this step, the required state feature sequence is obtained. Based on the first state feature sequence and the second state feature sequence, a two-layer state discrimination model is constructed. In this model, the first layer focuses on analyzing local features, while the second layer focuses on analyzing global features. Through local feature analysis, the first state probability vector of the chemical filter is calculated, which is expressed as:

[0103] ;

[0104] in, is the probability vector of the first state, is the number of first state samples, and is the total number of samples. Accordingly, the second state probability vector The calculation is performed in a similar way. The two state probability vectors are input into the dual hidden Markov prediction model, and the state transition probabilities of the two time scales are calculated respectively by the first prediction sub-model and the second prediction sub-model. The state transition probability is expressed as:

[0105] ;

[0106] in, From the state Transfer to state The probability of Indicates time This model uses historical state data to learn the transition patterns between states, generating a first transition probability matrix and a second transition probability matrix. Weighted fusion and multi-scale rolling prediction are performed on the first and second transition probability matrices. This integration of information from two time scales improves prediction accuracy. Weighted fusion is expressed as follows:

[0107] ;

[0108] in, is the final fused transition probability matrix, and are the first and second transition probability matrices, respectively, and and is a dynamically adjusted weight coefficient. Through the weighted fusion method, the transition rules at different time scales are comprehensively considered to obtain a more comprehensive state prediction. Double smoothing is performed on the multi-scale state prediction sequence to remove noise, making the prediction results more stable and reliable. Double smoothing is achieved through the moving average method or exponential smoothing method, and the state transition prediction sequence is obtained, which is expressed as:

[0109] ;

[0110] in, is the smoothed state value, is the original state value, is the size of the smoothing window. This step ensures the reliability and accuracy of the predicted sequence.

[0111] In a specific embodiment, the process of executing step 105 may specifically include the following steps:

[0112] A first objective function and a second objective function are constructed for the state transition prediction sequence, wherein the first objective function includes a life maximization term and an efficiency maximization term, and the second objective function includes an operating cost minimization term and a replacement cost minimization term, thereby obtaining a dual-objective optimization function;

[0113] Setting a first set of constraints and a second set of constraints according to a dual-objective optimization function, wherein the first set of constraints defines a filter pressure variation range and an efficiency variation range, and the second set of constraints defines a filter life range and a cost range, to obtain optimization constraints;

[0114] Perform first and second weight coefficient assignments on the dual-objective optimization function and the optimization constraints, calculate the objective function weights and the constraint weights based on the hierarchical analysis method, and obtain a multi-level weight matrix;

[0115] Based on the multi-level weight matrix, the first Pareto optimal solution set and the second Pareto optimal solution set are constructed, corresponding to the short-term optimization goal and the long-term optimization goal respectively, and a double-level Pareto solution set is obtained;

[0116] The double-layer Pareto solution set is screened for non-inferior solutions, and the solution set is sorted by the first and second objective criteria to obtain a candidate solution sequence. An operating parameter mapping relationship is constructed based on the candidate solution sequence, and the optimal solution is mapped to dynamic adjustment parameters of filtration pressure, filtration temperature, and filtration flow to obtain a parameter adjustment sequence.

[0117] The parameter adjustment sequence is optimized in a time-series coordination manner, and parameter smoothing is performed based on the dual constraints of the first time scale and the second time scale to obtain a coordinated control sequence. The coordinated control sequence is then converted into a combined instruction set of pressure control instructions, temperature control instructions, and flow control instructions of the filter to obtain the optimal operating control parameters.

[0118] Specifically, the first objective function and the second objective function are constructed for the state transition prediction sequence. The first objective function mainly includes the life maximization term and the efficiency maximization term, which are specifically expressed as:

[0119] ;

[0120] in, is the first objective function, For the expected life of the filter, is the efficiency of the filter, and are the corresponding weight coefficients, which are used to adjust the importance of the two objectives. The second objective function focuses on minimizing the operating cost and the replacement cost, and its expression is:

[0121] ;

[0122] in, is the second objective function, For operating costs, For replacement cost, and is also the weight coefficient. The combination of the two objective functions forms a dual-objective optimization function, which is expressed as:

[0123] ;

[0124] Trade-offs are made between different objectives to achieve comprehensive optimization. Set corresponding constraint sets. The first constraint set defines the pressure variation range and efficiency variation range of the filter. The formula can be expressed as:

[0125] ;

[0126] in, Represents the actual pressure of the filter, and are the minimum and maximum values ​​allowed, respectively, and and The second set of constraints defines the life and cost range of the filter, which is expressed as:

[0127] ;

[0128] in, and are the minimum and maximum limits on filter life, and is the cost range. The setting of constraints ensures the feasibility of the optimization process in actual operation. The first weight coefficient and the second weight coefficient are assigned to the dual-objective optimization function and the optimization constraints. The objective function weight and the constraint weight are calculated based on the hierarchical analysis method. By scoring the relative importance of the objectives and constraints, a multi-level weight matrix is ​​formed, in which the objective weight and constraint weight are expressed as follows:

[0129] ;

[0130] in, Representative The first goal and The relative importance weights between the constraints. Based on the multi-level weight matrix, the first Pareto optimal solution set and the second Pareto optimal solution set are constructed, corresponding to the short-term optimization goal and the long-term optimization goal respectively. Different objective functions are combined with constraints to form a set of non-dominated solutions. For example, through an optimization algorithm (such as a genetic algorithm or a particle swarm algorithm), multiple potential optimal solutions are obtained, and these solutions together form a double-layer Pareto solution set. The double-layer Pareto solution set is screened for non-inferior solutions. The solution set is sorted by the first objective criterion (such as minimizing a certain objective function) and the second objective criterion (such as maximizing another objective function) to identify a sequence of candidate solutions. These candidate solutions will be used to construct a mapping relationship between the operating parameters, and the optimal solution will be mapped to the dynamic adjustment parameters of the filter's pressure, temperature and flow. The parameter adjustment sequence is established using the following formula:

[0131] ;

[0132] in, 、 and represent the optimized pressure, temperature and flow respectively, and Dynamic adjustment parameters are constructed based on the optimization results. Time-coordinated optimization is performed on the parameter adjustment sequence. Based on the dual constraints of the primary and secondary time scales, the parameters are smoothed to ensure consistency and stability in actual operation. Smoothing uses a sliding average or other smoothing techniques to ultimately generate a coordinated control sequence. This sequence is converted into a combined set of pressure, temperature, and flow control commands for the filter, resulting in the optimal operating control parameters.

[0133] In a specific embodiment, the process of executing step 106 may specifically include the following steps:

[0134] Real-time monitoring of chemical filters based on optimal operating control parameters, collecting real-time multi-dimensional usage status data including inlet pressure, outlet pressure, flow rate, filter medium concentration, filter temperature, filter efficiency, and filter saturation;

[0135] Input real-time multi-dimensional usage status data into the two-layer performance degradation prediction model, perform online inference operations on the physical degradation layer and deep learning layer, and obtain a real-time performance degradation prediction sequence;

[0136] The prediction error is calculated based on the real-time performance degradation prediction sequence and actual performance data, and classified and counted according to pressure error, efficiency error and life error to obtain the error characteristic matrix;

[0137] The cause of the error characteristic matrix is ​​analyzed to determine the key influencing parameters in the physical attenuation layer and the deep learning layer, and the parameter optimization direction is obtained. The physical parameters and network parameters in the double-layer performance attenuation prediction model are updated according to the parameter optimization direction to obtain the service life prediction model.

[0138] Specifically, based on the optimal operating control parameters, real-time, multi-dimensional usage status data, including inlet and outlet pressures, flow, filter medium concentration, filtration temperature, filtration efficiency, and filter saturation, is collected. This data collection is accomplished through sensors installed at key nodes of the filter. The sensors record these parameters in real time and send the information to the central control system via a data transmission system. The collected data is structured as a multi-dimensional array, where each dimension corresponds to a monitoring parameter, for example:

[0139] ;

[0140] in, For in time The collected multi-dimensional data vector, 、 、 、 、 、 and Representing inlet pressure, outlet pressure, flow rate, filter medium concentration, filter temperature, filter efficiency, and filter saturation, respectively. Real-time, multi-dimensional usage status data is input into a two-layer performance degradation prediction model. Through online inference operations in the physical degradation layer and deep learning layer, a real-time performance degradation prediction sequence is generated. The physical degradation layer processes degradation models based on physical principles, such as predicting performance degradation by calculating changes in filter pressure and flow rate. The deep learning layer processes nonlinear relationships and complex features through neural networks. For example, the prediction model is described by the following formula:

[0141] ;

[0142] in, Indicates time The performance degradation prediction value of is the performance degradation prediction model function. This process ensures that the system obtains real-time feedback on filter performance from real-time monitoring data. The prediction error is calculated based on the real-time performance degradation prediction sequence and actual performance data. The error calculation is divided into multiple aspects, including pressure error, efficiency error, and life error. The calculation formula is as follows:

[0143] ;

[0144] ;

[0145] ;

[0146] in, 、 and represent pressure error, efficiency error and life error respectively, and 、 and is the actual measured parameter. Through the statistics of these errors, an error characteristic matrix is ​​constructed for subsequent cause analysis. The cause analysis is performed on the error characteristic matrix. By analyzing the source of each error, the key influencing parameters in the physical attenuation layer and the deep learning layer are determined. For example, if the pressure error is significant and the efficiency error is small, it means that the flow rate of the filter is improperly set. Use statistical analysis methods (such as correlation analysis) to identify which parameters have the greatest impact on the error and obtain the direction of parameter optimization. For example, if it is found that the fluctuation of the filtration temperature is related to the efficiency error, the temperature control strategy should be optimized. According to the parameter optimization direction, the physical parameters and network parameters in the double-layer performance attenuation prediction model are updated to improve the accuracy and reliability of the prediction. Adjust the identified key parameters and retrain the deep learning model. For example, the model parameters are updated using the following formula:

[0147] ;

[0148] in, and Represent the model parameters before and after the update, is the learning rate, is the gradient of the loss function. Updates help make the model better adapt to the current operating conditions.

[0149] Among them, the cause analysis of the error characteristic matrix is ​​performed, the key influencing parameters in the physical attenuation layer and the deep learning layer are determined, the parameter optimization direction is obtained, and the physical parameters and network parameters in the double-layer performance attenuation prediction model are updated according to the parameter optimization direction to obtain the service life prediction model, including the following steps: performing correlation decomposition on the error characteristic matrix, extracting the pressure error vector, efficiency error vector and life error vector, and obtaining a classification error set; constructing an error propagation map based on the classification error set, calculating the error contribution of the physical attenuation layer and the deep learning layer, and obtaining the hierarchical error distribution; performing sensitivity analysis on the hierarchical error distribution, calculating the influence of each parameter on the prediction result, and obtaining the parameter sensitivity matrix; screening the key influencing parameters according to the parameter sensitivity matrix, setting the parameter importance threshold for descending sorting, and obtaining the parameter optimization sequence; the key parameters in the physical attenuation layer are sorted. Perform variation analysis, update the polynomial fitting coefficients and boundary constraint parameters based on the physical constraints, and obtain the optimized physical parameter set; perform gradient calculation on the key parameters in the deep learning layer, and perform local fine-tuning on the network weights and biases to obtain the optimized network parameter set; input the optimized physical parameter set into the physical attenuation layer, recalculate the attenuation coefficient, and obtain the updated physical prediction result; input the optimized network parameter set into the deep learning layer, update the parameters of the feature extractor and predictor, and obtain the updated deep learning prediction result; perform fusion optimization on the updated physical prediction result and the updated deep learning prediction result, update the fusion weight coefficient, and obtain the corrected prediction result; perform an overall performance evaluation on the double-layer performance attenuation prediction model based on the corrected prediction result, and when the prediction error meets the convergence condition, obtain the service life prediction model.

[0150] The above describes the method for predicting the service life of the chemical filter in the embodiment of the present invention. The following describes the device for predicting the service life of the chemical filter in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a device for predicting the service life of a chemical filter includes:

[0151] The acquisition module 201 is used to collect historical multi-dimensional usage status data of the chemical filter and perform standardized preprocessing to obtain a time series feature data set;

[0152] Clustering module 202, configured to perform load feature clustering analysis on the time series feature data set to obtain a filter load pattern feature vector;

[0153] A construction module 203 is used to construct a two-layer performance degradation prediction model based on the filter load pattern feature vector, and output a target performance degradation prediction sequence through physical degradation coefficient calculation and deep learning network training;

[0154] A calculation module 204 is configured to perform filter state probability transfer calculation on the target performance attenuation prediction sequence to obtain a state transfer prediction sequence;

[0155] Optimization module 205, used to perform multi-objective joint optimization on the state transition prediction sequence and calculate the optimal operation control parameters;

[0156] The parameter tuning module 206 is used to collect real-time multi-dimensional usage status data based on the optimal operation control parameters, and perform parameter tuning on the double-layer performance degradation prediction model to obtain a service life prediction model.

[0157] Through the collaborative cooperation of the above-mentioned components, a double-layer performance degradation prediction model is constructed, which combines physical mechanism constraints with deep learning, and improves the model accuracy while ensuring the interpretability of the prediction; the load feature clustering analysis method of the dual time window is adopted to realize the accurate identification of the load patterns of the filter at different time scales, and enhance the adaptability of the model to the changes in working conditions; the state probability transfer calculation method based on the dual hidden Markov prediction model accurately captures the dynamic characteristics of the filter performance degradation; a multi-objective joint optimization strategy including life maximization, cost minimization and efficiency maximization is designed to realize the intelligent regulation of the filter operating parameters; through the dynamic parameter tuning mechanism, the prediction model has the ability of adaptive updating, which ensures the stability of the model's prediction performance during long-term operation; the present invention improves the accuracy and reliability of filter life prediction through multi-level feature analysis and prediction strategies.

[0158] above Figure 2 The device for predicting the service life of the chemical filter in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The device for predicting the service life of the chemical filter in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0159] Figure 3This is a schematic diagram of the structure of a device for predicting the service life of a chemical filter, provided in an embodiment of the present invention. The device 300 may vary significantly depending on its configuration or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors), memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing applications 333 or data 332. The memory 320 and storage media 330 may be either transient or persistent storage. The program stored in the storage medium 330 may include one or more modules (not shown), each of which may include a series of instructions for operating on the device 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, allowing the device 300 to execute the series of instructions stored in the storage medium 330 to implement the steps of the method for predicting the service life of a chemical filter.

[0160] The chemical filter service life prediction device 300 may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The structure of the device for predicting the service life of a chemical filter shown does not constitute a limitation on the device for predicting the service life of a chemical filter provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0161] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the method for predicting the service life of a chemical filter.

[0162] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0163] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0164] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the service life of a chemical filter, characterized in that: The method comprises: Collect historical multi-dimensional usage status data of chemical filters and perform standardized preprocessing to obtain a time series feature data set; Perform load feature clustering analysis on the time series feature data set to obtain a filter load mode feature vector; specifically comprising: calculating the cumulative filtering amount of the first time window and the second time window for the time series feature data set respectively to obtain a first load time series index and a second load time series index, the length of the first time window is 24 hours, and the length of the second time window is 48 hours; performing segmentation threshold calculation on the first load time series index and the second load time series index, dividing the first load time series index into a first low load segment, a first medium load segment and a first high load segment, and dividing the second load time series index into a second low load segment, a second medium load segment and a second high load segment, to obtain a first load segmentation result and a second load segmentation result; performing feature value extraction on the first load segmentation result and the second load segmentation result respectively to calculate a first pressure change rate, a first temperature distribution feature, a first efficiency decay rate and a second pressure drop. force change rate, second temperature distribution characteristics, second efficiency attenuation rate, to obtain a first characteristic parameter set and a second characteristic parameter set; construct a double-layer characteristic association matrix based on the first characteristic parameter set and the second characteristic parameter set, obtain a multi-scale characteristic association result through a first-layer local association calculation and a second-layer global association calculation, and perform principal component analysis on the multi-scale characteristic association result, extract the characteristic combination of the first principal component and the second principal component, and obtain a multi-dimensional dimension reduction feature set; perform double clustering calculation on the multi-dimensional dimension reduction feature set to obtain a first cluster center and a second cluster center respectively, and establish a first distance matrix and a second distance matrix through the first cluster center and the second cluster center for pattern discrimination to obtain a multi-level load pattern label; perform feature combination on the multi-level load pattern label and the multi-dimensional dimension reduction feature set, and obtain a filter load pattern feature vector through a cascade transformation of the first feature map and the second feature map; Based on the filter load mode feature vector, a double-layer performance attenuation prediction model is constructed, and a target performance attenuation prediction sequence is output through physical attenuation coefficient calculation and deep learning network training; specifically, the model comprises: dividing the filter load mode feature vector into a first feature sub-vector and a second feature sub-vector, wherein the first feature sub-vector comprises a pressure change rate and an efficiency attenuation rate, and the second feature sub-vector comprises a temperature distribution feature and a saturation change rate; constructing a physical attenuation layer according to the first feature sub-vector, the physical attenuation layer comprises an attenuation parameter calculation unit and a physical constraint unit, the attenuation parameter calculation unit adopts a polynomial fitting structure, and the physical constraint unit adopts a boundary constraint structure; performing parameter training on the attenuation parameter calculation unit, the attenuation parameter calculation unit comprises a third-order polynomial fitting module and an adaptive weight module, and the physical constraint unit comprises a maximum pressure drop constraint and a minimum efficiency constraint; inputting the first feature sub-vector into the physical attenuation layer, and obtaining a first performance attenuation prediction sequence through attenuation coefficient calculation; constructing a deep learning layer according to the second feature sub-vector and the first performance attenuation prediction sequence, the deep learning layer comprises A time series feature extractor and a performance predictor, wherein the time series feature extractor adopts a bidirectional long short-term memory network structure, and the performance predictor adopts a residual attention network structure; the time series feature extractor comprises three layers of bidirectional LSTM units, each layer of LSTM units comprises 128 hidden nodes, and a Dropout layer is set between the LSTM units; the performance predictor comprises five residual blocks, each residual block comprises two fully connected layers and a channel attention module; the first performance decay prediction sequence and the second feature sub-vector are input into the deep learning layer, and the second performance decay prediction sequence is obtained through network training; a fusion optimization layer is constructed, the fusion optimization layer comprises an adaptive weight module and a timing calibration module, and the first performance decay prediction sequence and the second performance decay prediction sequence are dynamically weighted fused; the weight parameters of the adaptive weight module are updated online according to the historical performance decay data of the chemical filter, and the fusion results are time-aligned through the timing calibration module to obtain the output result of the fusion optimization layer, and the output result of the fusion optimization layer is time-reconstructed to generate a target performance decay prediction sequence; Performing filter state probability transfer calculation on the target performance attenuation prediction sequence to obtain a state transfer prediction sequence; Performing multi-objective joint optimization on the state transition prediction sequence and calculating the optimal operation control parameters; Based on the optimal operation control parameters, real-time multi-dimensional usage status data is collected, and the parameters of the double-layer performance degradation prediction model are tuned to obtain a service life prediction model.

2. The method for predicting the service life of a chemical filter according to claim 1, characterized in that: The historical multi-dimensional usage status data of the chemical filter is collected and standardized preprocessed to obtain a time series feature data set, including: Collect data on the inlet pressure, outlet pressure, flow rate, filter medium concentration, filter temperature, filter efficiency and filter saturation of the chemical filter to obtain historical multi-dimensional usage status data; Performing data cleaning on the historical multi-dimensional usage status data to obtain a preprocessed data set; Performing standardized mapping on the preprocessed data set, uniformly converting state parameters of different dimensions into a standard range, and obtaining a standardized data set; Dividing the standardized data set into time windows, setting overlapping time windows to segment the standardized data set, and obtaining time series data segments; The time series data segments are data labeled to obtain a time series feature data set.

3. The method for predicting the service life of a chemical filter according to claim 1, characterized in that: The performing filter state probability transfer calculation on the target performance attenuation prediction sequence to obtain a state transfer prediction sequence includes: Performing a first time window segmentation and a second time window segmentation on the target performance attenuation prediction sequence to obtain a first segmented sequence sample and a second segmented sequence sample; Performing dual feature extraction on the first segmented sequence samples and the second segmented sequence samples respectively to obtain a first state feature sequence and a second state feature sequence; Building a double-layer state discrimination model based on the first state feature sequence and the second state feature sequence, and calculating a first state probability vector and a second state probability vector of the chemical filter through a first-layer local feature analysis and a second-layer global feature analysis; Inputting the first state probability vector and the second state probability vector into a dual hidden Markov prediction model, respectively calculating the state transition probabilities of two time scales through the first prediction sub-model and the second prediction sub-model, to obtain a first transition probability matrix and a second transition probability matrix; The first transfer probability matrix and the second transfer probability matrix are weightedly fused and multi-scale rolling predicted to obtain a multi-scale state prediction sequence, and the multi-scale state prediction sequence is double-smoothed to obtain a state transfer prediction sequence.

4. The method for predicting the service life of a chemical filter according to claim 3, characterized in that: The multi-objective joint optimization of the state transition prediction sequence and calculating the optimal operation control parameters include: Constructing a first objective function and a second objective function for the state transition prediction sequence, wherein the first objective function includes a life maximization term and an efficiency maximization term, and the second objective function includes an operation cost minimization term and a replacement cost minimization term, to obtain a dual-objective optimization function; According to the dual-objective optimization function, a first set of constraints and a second set of constraints are set, wherein the first set of constraints defines a filter pressure variation range and an efficiency variation range, and the second set of constraints defines a filter life range and a cost range, to obtain optimization constraints; Performing a first weight coefficient allocation and a second weight coefficient allocation on the dual-objective optimization function and the optimization constraint condition, calculating the objective function weight and the constraint condition weight based on the hierarchical analysis method, and obtaining a multi-level weight matrix; Based on the multi-level weight matrix, a first Pareto optimal solution set and a second Pareto optimal solution set are constructed, which correspond to the short-term optimization target and the long-term optimization target respectively, to obtain a double-layer Pareto solution set; The double-layer Pareto solution set is screened for non-inferior solutions, the solution set is sorted by the first objective criterion and the second objective criterion to obtain a candidate solution sequence, and an operation parameter mapping relationship is constructed based on the candidate solution sequence, and the optimal solution is mapped into dynamic adjustment parameters of filtration pressure, filtration temperature and filtration flow rate to obtain a parameter adjustment sequence; The parameter adjustment sequence is optimized in time coordination, and parameter smoothing is performed based on the dual constraints of the first time scale and the second time scale to obtain a coordinated control sequence, and the coordinated control sequence is converted into a combined instruction set of pressure control instructions, temperature control instructions and flow control instructions of the filter to obtain optimal operation control parameters.

5. The method for predicting the service life of a chemical filter according to claim 4, characterized in that: The real-time multi-dimensional usage status data is collected based on the optimal operation control parameters, and the parameters of the double-layer performance degradation prediction model are tuned to obtain a service life prediction model, including: Based on the optimal operation control parameters, the chemical filter is monitored in real time to collect real-time multi-dimensional usage status data including inlet pressure, outlet pressure, flow rate, filter medium concentration, filter temperature, filter efficiency and filter saturation; Input the real-time multi-dimensional usage status data into the two-layer performance degradation prediction model, perform online reasoning operations on the physical degradation layer and the deep learning layer, and obtain a real-time performance degradation prediction sequence; Calculate the prediction error according to the real-time performance decay prediction sequence and the actual performance data, and classify and count the pressure error, efficiency error and life error to obtain an error feature matrix; The cause analysis of the error characteristic matrix is ​​performed to determine the key influencing parameters in the physical attenuation layer and the deep learning layer, and the parameter optimization direction is obtained. The physical parameters and network parameters in the double-layer performance attenuation prediction model are updated according to the parameter optimization direction to obtain a service life prediction model.

6. A device for predicting the service life of a chemical filter, characterized in that: The device is used to perform the method for predicting the service life of a chemical filter according to any one of claims 1 to 5, the device comprising: The acquisition module is used to collect the historical multi-dimensional usage status data of the chemical filter and perform standardized preprocessing to obtain a time series feature data set; A clustering module, used for performing load feature clustering analysis on the time series feature data set to obtain a filter load pattern feature vector; A construction module is used to construct a double-layer performance attenuation prediction model based on the filter load pattern feature vector, and output a target performance attenuation prediction sequence through physical attenuation coefficient calculation and deep learning network training; A calculation module, used for performing filter state probability transfer calculation on the target performance attenuation prediction sequence to obtain a state transfer prediction sequence; An optimization module, used for performing multi-objective joint optimization on the state transition prediction sequence and calculating the optimal operation control parameters; The parameter tuning module is used to collect real-time multi-dimensional usage status data based on the optimal operation control parameters, and to perform parameter tuning on the double-layer performance attenuation prediction model to obtain a service life prediction model.

7. A device for predicting the service life of a chemical filter, characterized in that: The device for predicting the service life of a chemical filter comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory to enable the device for predicting the service life of the chemical filter to perform the method for predicting the service life of the chemical filter as described in any one of claims 1-5.

8. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by a processor, a method for predicting the service life of a chemical filter as described in any one of claims 1 to 5 is implemented.

Images