A ship coating dosage data asset library construction method, electronic equipment, storage medium and product

Abstract

This invention discloses a method, electronic device, storage medium, and product for constructing a marine coating usage data asset library, including: acquiring historical static parameter data, historical dynamic parameter data, and corresponding actual coating usage; performing SSA feature extraction on the historical dynamic parameter data according to the dynamic parameter type; extracting the maximum singular value after SSA decomposition of each dynamic parameter data and calculating the basic weight of each feature; calculating the basic weight of each static parameter using the coefficient of variation method; calculating the mutual information between all features extracted from each dynamic parameter data and the actual coating usage, and taking the maximum mutual information value as the mutual information entropy of that feature; adjusting the basic weights based on the mutual information entropy to obtain the influence coefficient; designating features with influence coefficients greater than a set threshold as key features; optimizing the key features through variance inflation factor analysis; and saving the optimized key features and corresponding parameter data to construct a marine coating usage data asset library.

Description

Technical Field

[0001] This invention relates to the field of ship coating technology, and in particular to a method for constructing a ship coating usage data asset database, electronic equipment, storage medium, and products. Background Technology

[0002] Ship painting is a crucial aspect of shipbuilding and maintenance. It refers to the amount of paint consumed to complete a specific painting task and serves as an important basis for enterprises to formulate production plans, evaluate performance, and calculate costs, thus contributing to scientific management. Its core function lies not only in extending the service life of ship structures such as the hull, outfitting components, and compartments through anti-corrosion and anti-rust coatings, but also in meeting environmental and safety standards through exterior coatings. However, currently, with the upgrading of painting equipment and the updating of painting processes, the lack of accurate paint consumption databases means that the original methods for determining paint consumption for ship painting tasks are no longer sufficient to meet current requirements for rational paint usage.

[0003] By using the actual paint usage data automatically collected from coating equipment to intelligently predict and iteratively optimize paint usage for coating tasks, companies can continuously optimize their paint usage calculation system for ship coating tasks, ultimately achieving the goal of continuously reducing paint consumption. However, this intelligent paint usage prediction process requires a complete and accurate historical paint usage database. Current technologies typically only record partial process nodes and results, lacking effective technical means to form a complete paint usage data asset library. This fails to provide companies with effective digital assets, nor can it offer effective guidance for subsequent operations, or help companies reduce costs and increase efficiency in their coating operations. Summary of the Invention

[0004] Purpose of the invention: To address the problem that existing coating usage data asset databases only record partial process nodes and result information, this invention proposes a method for constructing a marine coating usage data asset database, electronic equipment, storage media, and products.

[0005] Technical Solution: In the first aspect, this invention proposes a method for constructing a database of ship coating usage data, comprising the following steps:

[0006] Step 1: Obtain historical static parameter data, historical dynamic parameter data, and corresponding actual paint usage;

[0007] Step 2: Based on the dynamic parameter type, perform SSA feature extraction on the historical dynamic parameter data to obtain trend features, periodic features, and statistical features;

[0008] Step 3: Extract the maximum singular value after SSA decomposition of each dynamic parameter data, and calculate the basic weight of each feature based on the maximum singular value: calculate the basic weight of each static parameter using the coefficient of variation method;

[0009] Step 4: Calculate the mutual information between all features extracted from each dynamic parameter data and the actual amount of paint used, and take the maximum mutual information value as the mutual information entropy of that feature; calculate the mutual information entropy between each static parameter data and the actual amount of paint used; based on the mutual information entropy, adjust the basic weights to obtain the influence coefficient;

[0010] Step 5: Select features with an influence coefficient greater than a set threshold as key features;

[0011] Step 6: Optimize the key features through variance inflation factor analysis to obtain the optimized key features;

[0012] Step 7: Save the optimized key features and corresponding parameter data to build a marine coating usage data asset library.

[0013] Furthermore, the static parameter data is obtained based on design parameter data, material parameter data, task priority, and planned operation period; the dynamic parameter data is obtained based on external environment data, real-time paint viscosity, spray gun moving speed, and paint flow rate.

[0014] Furthermore, both the historical static parameter data and the historical dynamic parameter data are obtained through the following preprocessing steps:

[0015] Based on the parameter type and its criticality to the calculation and prediction of coating usage, differentiated missing value processing is implemented;

[0016] By integrating equipment status and process rules, outlier identification and correction can be performed.

[0017] Z-Score standardization is used to standardize the dynamic parameter data;

[0018] Min-Max normalization is used to normalize the static parameter data.

[0019] Furthermore, the fusion of equipment status and process rules involves outlier identification and correction, including:

[0020] For all numerical time series parameters in the most recent Calculate the mean of the sequence over a period of time. with standard deviation If the value of the current numerical time series parameter satisfies If so, it is marked as an anomaly;

[0021] If the equipment status signal is off, faulty, or in standby mode at the time corresponding to the anomaly point, it is determined to be invalid noise and replaced by the mean smoothing method of the adjacent time period.

[0022] If an outlier matches a specific process instruction time period in the task dispatching system, it will be retained, and a special process business tag will be added to this data segment. This business tag will be treated as an independent feature.

[0023] Anomalies that do not fall into either of the above two categories are processed as noise smoothing and marked as pending verification. Once a certain number are accumulated, they are reviewed by process experts.

[0024] Furthermore, the differential missing value processing based on the parameter type and its criticality to the calculation and prediction of coating usage includes:

[0025] The forward imputation method is used to fill in missing values ​​for continuous parameters;

[0026] Missing values ​​of static parameters were filled using the average values ​​of outfitting components of the same type.

[0027] Secondly, the present invention provides an electronic device, the electronic device comprising:

[0028] At least one processor;

[0029] and a memory communicatively connected to the at least one processor;

[0030] The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to execute a method for constructing a marine coating usage data asset database.

[0031] Thirdly, the present invention proposes a computer-readable storage medium storing computer instructions for causing a processor to execute a method for constructing a database of marine coating usage data.

[0032] Fourthly, the present invention proposes a computer program product, which includes a computer program that, when executed by a processor, implements a method for constructing a database of ship coating usage data.

[0033] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0034] (1) This invention collects the amount of paint used for ship outfitting parts, establishes the correlation between paint usage and the coating object and external environment, and combines historical paint usage data with SSA singular spectrum analysis to not only achieve data denoising and eliminate the fluctuation of usage data caused by sensor noise, but also quantifies the parameter influence coefficient, clarifies the weight of key parameters, and avoids blind optimization.

[0035] (2) This invention performs intelligent analysis on the amount of paint used in historical coating processes and related influencing parameters, obtains the influence coefficients of related parameters, generates a data asset library of paint usage information, and performs preprocessing, feature extraction and optimization to form an optimized feature database, providing enterprises with intangible digital assets, and also providing support for the subsequent accurate calculation of coating usage. Furthermore, through integration with related spraying equipment, it can further realize real-time analysis and operation guidance for workers in the coating process. Attached Figure Description

[0036] Figure 1 This is a flowchart of a method for constructing a marine coating usage data asset database as proposed in Example 2;

[0037] Figure 2 This is a flowchart of an intelligent prediction method for ship coating consumption based on data acquisition from coating equipment, as proposed in Example 3.

[0038] Figure 3 The flowchart shows the overall structure of the PLATM model designed for Example 3. Detailed Implementation

[0039] The technical solution of this embodiment will now be further described in conjunction with the accompanying drawings and examples.

[0040] Example 1:

[0041] This invention proposes a method for constructing a database of marine coating usage data, comprising the following steps:

[0042] Step 1: Obtain historical static parameter data, historical dynamic parameter data, and corresponding actual paint usage. Specifically, static parameter data is obtained based on design parameter data, material parameter data, task priority, and planned operation period; dynamic parameter data is obtained based on external environment data, real-time paint viscosity, spray gun movement speed, and paint flow rate.

[0043] Step 2: Based on the dynamic parameter type, perform SSA feature extraction on the historical dynamic parameter data to obtain trend features, periodic features, and statistical features;

[0044] Step 3: Extract the maximum singular value after SSA decomposition of each dynamic parameter data, and calculate the basic weight of each feature based on the maximum singular value: calculate the basic weight of each static parameter using the coefficient of variation method;

[0045] Step 4: Calculate the mutual information between all features extracted from each dynamic parameter data and the actual amount of paint used, and take the maximum mutual information value as the mutual information entropy of that feature; calculate the mutual information entropy between each static parameter data and the actual amount of paint used; based on the mutual information entropy, adjust the basic weights to obtain the influence coefficient;

[0046] Step 5: Select features with an influence coefficient greater than a set threshold as key features;

[0047] Step 6: Optimize the key features through variance inflation factor analysis to obtain the optimized key features;

[0048] Step 7: Save the optimized key features and corresponding parameter data to build a marine coating usage data asset library.

[0049] In this embodiment of the invention, both historical static parameter data and historical dynamic parameter data are obtained through the following preprocessing steps:

[0050] Based on the parameter type and its criticality to the calculation and prediction of coating usage, differentiated missing value processing is implemented, including: using the forward filling method to fill missing values ​​of continuous parameters; and using the average value of outfitting parts of the same type to fill missing values ​​of static parameters.

[0051] By integrating equipment status and process rules, outlier identification and correction can be performed.

[0052] Z-Score standardization is used to standardize the dynamic parameter data;

[0053] Min-Max normalization is used to normalize the static parameter data.

[0054] This includes integrating equipment status and process rules to identify and correct outliers, including:

[0055] For all numerical time series parameters in the most recent Calculate the mean of the sequence over a period of time. with standard deviation If the value of the current numerical time series parameter satisfies If so, it is marked as an anomaly;

[0056] If the equipment status signal is off, faulty, or in standby mode at the time corresponding to the anomaly point, it is determined to be invalid noise and replaced by the mean smoothing method of the adjacent time period.

[0057] If an outlier matches a specific process instruction time period in the task dispatching system, it will be retained, and a special process business tag will be added to this data segment. This business tag will be treated as an independent feature.

[0058] Anomalies that do not fall into either of the above two categories are processed as noise smoothing and marked as pending verification. Once a certain number are accumulated, they are reviewed by process experts.

[0059] Example 2:

[0060] like Figure 1 As shown in the figure, this invention proposes a method for constructing a marine coatings usage data asset database, including the following steps:

[0061] S1. Taking each outfitting part to be painted as an object, the painting task is assigned through the task assignment system, and the equipment data and external environment data of the painting equipment operation process are collected and stored in the server.

[0062] S2. Clean the collected painting process data to obtain the actual paint consumption data of the outfitting parts painting process;

[0063] S3. Establish a correlation between the actual paint usage data and the outfitting component quantity data of the task dispatch system to form a data set of each outfitting component's quantity, process, paint, external environmental parameters, and paint usage, and establish a paint usage information data asset library.

[0064] S4. Preprocess and extract features from the data in the paint usage information database to form a feature database. Optimize the feature database by filtering the data based on the paint usage impact coefficient. Specific operations include:

[0065] Historical coating data is preprocessed to remove outliers and missing values, and Z-score normalization is performed. This includes:

[0066] Missing value handling: For missing values ​​of continuous parameters, the forward imputation method is used. t represents the missing time. The value is the valid value from the previous moment;

[0067] For missing values ​​of static parameters, fill them with the average value of outfitting components of the same type:

[0068]

[0069] k represents the quantity of outfitting components of the same type. The surface area of ​​similar outfitting components;

[0070] Use the 3σ criterion to identify outliers: , The mean of the parameters, If the standard deviation is zero, then replace the outlier with the mean of the next 5 minutes:

[0071] ;

[0072] The parameters collected from the coating equipment and external environment during the operation process are processed using the Z-Score standardization method to eliminate dimensions while preserving data distribution characteristics. The calculation formula is as follows:

[0073] ;

[0074] in The mean of the parameters, Let be the standard deviation, and the standardized data follow . distributed.

[0075] Singular spectral analysis (SSA) is used to decompose and denoise the preprocessed data. Based on the contribution rate of singular values, data with singular values ​​less than a set value are considered noise and removed.

[0076] The influence coefficients of each data parameter were calculated, and parameters with influence coefficients higher than the set values ​​were retained as the final optimized feature data, including: surface area A, number of coating layers N, dry film thickness T, and coating viscosity. Diluent content Solid content specific gravity V, surface roughness F, air temperature C, atmospheric humidity H, and wind force level W.

[0077] In this embodiment of the invention, the SSA singular spectrum analysis method includes: trajectory matrix construction, singular value decomposition, denoised sequence grouping and diagonal averaging. Through SSA singular spectrum analysis, trend components, periodic components and noise components that characterize the core law of paint consumption are separated from complex multi-source time series data, forming enhanced and denoised feature data.

[0078] The methods for constructing the trajectory matrix include:

[0079] One-dimensional raw data According to the length of the sliding window The values ​​are sliced ​​to divide the original time series data into several vectors. trajectory matrix The mathematical expression is as follows:

[0080]

[0081] Among them, parameters and Constraints must be met: ,and Each column of matrix X Both represent a length of The time segment is N, where N is the length of the original time series data. This matrix simultaneously captures both the short-term dynamics and long-term evolution patterns of the flow sequence.

[0082] Singular value decomposition methods include: using singular values... The original signal data sequence is denoised to process the short-term, nonlinear, and irregular coating data.

[0083] The denoising sequence grouping method includes: using singular values ​​as the grouping basis, intelligently grouping the decomposed components according to the contribution rate of singular values ​​to achieve signal enhancement and noise filtering, and treating data with singular values ​​less than a set value as noise to be removed, thereby extracting effective feature components.

[0084] The diagonal averaging method includes embedding the original time series data into a trajectory matrix. In this process, several sub-matrices are obtained through SVD singular value decomposition and denoising sequence grouping. Each sub-matrix corresponds to a group, and each group matrix is ​​transformed back into a one-dimensional time series component with the same length as the original sequence. By averaging the anti-diagonal elements, the information from the two-dimensional matrix is ​​fused to obtain a one-dimensional time series component.

[0085] This invention collects data on the amount of paint used for ship outfitting components using IoT technology, integrates a task assignment system, establishes a correlation between paint usage and the object being painted and the external environment, and combines historical paint usage data with SSA singular spectrum analysis to not only denoise the data and eliminate fluctuations in usage data caused by sensor noise, but also quantify the parameter influence coefficients, clarify the weights of key parameters, and avoid blind optimization.

[0086] Example 3:

[0087] Accurate calculation of paint usage during the painting process in shipbuilding enterprises is crucial for controlling painting costs. Paint usage is influenced by factors such as the surface roughness of the object to be painted, the technical parameters of the paint itself, and the external environment during the painting process. Furthermore, the influence level of each factor is difficult to quantify accurately manually, making precise cost control in shipbuilding enterprises challenging. This invention intelligently analyzes historical paint usage and related influencing parameters to obtain their influence coefficients, enabling accurate calculation of paint usage. Through integration with relevant spraying equipment, it provides real-time analysis and operational guidance for workers during the painting process.

[0088] Based on Example 2, this invention proposes an intelligent prediction method for ship coating usage based on data acquisition from coating equipment, such as... Figure 2 As shown, it includes the following steps:

[0089] Step 1: For each area to be coated, a coating task is assigned through a task dispatching system. Coating operation data from the coating equipment and external environmental data are collected and stored on the server. For example, real-time data collection of the coating process, including surface area, spray gun pressure, coating speed, temperature, humidity, viscosity, and solids content, is conducted using IoT devices to create a database of tens of millions of samples, which are then stored on the server.

[0090] Step 2: Based on the historical coating consumption database, coating operation data, and external environmental data, collect coating consumption data in real time;

[0091] Step 3: Establish a correlation between the real-time collected coating usage data and the coating object quantity data from the task dispatch system to form a data set of each coating object's quantity, process, coating, external environmental parameters, and coating usage.

[0092] Step 4: Preprocess the dataset obtained in Step 3, including but not limited to handling outliers and missing values, performing Z-score standardization, and establishing a corporate coating usage data asset library. Specific operations include:

[0093] To ensure the high quality, consistency, and comparability of paint consumption and related influencing parameters used for intelligent prediction, invalid data was removed, abnormal interference was smoothed, and efficient features were constructed from the multi-source heterogeneous data collected during painting operations. This provides high-quality input for the subsequent SSA-PLSTM prediction model. This includes:

[0094] S400: Handling missing values ​​based on parameter importance and business logic:

[0095] Based on the parameter type and its criticality to the calculation and prediction of coating usage, a differentiated filling strategy is implemented to ensure data continuity and prediction reliability.

[0096] Key process timing parameters are missing (such as instantaneous flow rate of coating). Painting speed Spray gun switch status These parameters are the direct basis for calculating real-time paint consumption rate and cumulative usage; their absence will seriously affect the accuracy of cumulative usage and model input. The following hierarchical processing strategy is adopted:

[0097] Short-term gaps (≤30s): Linear interpolation is used to fill in these gaps to more accurately restore job continuity. The formula is:

[0098]

[0099] Long-term missing data (>30s): Considered an invalid job segment. The system automatically marks this time period as "invalid data," and the corresponding time series data will not be included in the cumulative usage calculation and will be excluded during model training to avoid introducing unexplained errors.

[0100] Missing auxiliary environmental time-series parameters (e.g., air temperature C, atmospheric humidity H): These parameters indirectly affect coating volatilization and film formation quality, and their continuity is more important to the prediction model than absolute accuracy. Therefore, the forward imputation method is used. ( For missing moments, (The value is the valid value from the previous time step), ensuring the continuity of the sequence in the simplest way, which makes it easier for the model to capture environmental trends.

[0101] Missing static / material parameters (e.g., surface area A of the object to be coated, designed dry film thickness T): These parameters are fundamental for calculating theoretical usage and differentiating work tasks. Fill in the gaps using the average value of similar objects to be coated.

[0102]

[0103] In the formula, k is the number of objects of the same type that are to be painted. The surface area of ​​objects with the same type of coating.

[0104] S410: Outlier identification and correction integrating equipment status and process rules:

[0105] To effectively distinguish between sensor noise, equipment malfunctions, and actual special operating conditions, a two-stage processing flow of "statistical detection + business verification" is adopted to avoid the accidental deletion of valuable operating condition samples.

[0106] Phase 1: Statistical Initial Screening: For all numerical time series parameters, an improved method is used. The criteria (based on a moving window) are used for initial screening. For parameters that are most recently... a point in time (e.g.) , representing a sequence within a 5-minute window, calculate its mean. with standard deviation If the current value satisfy If it is, then it is marked as an "outlier".

[0107] Phase Two: Business Logic Validation and Classification Processing

[0108] Equipment noise or invalid operation: If the equipment status signal is "off," "fault," or "standby" at the time corresponding to the "anomaly," it is determined to be invalid noise. Replace it using the mean smoothing method of adjacent time periods: Replace the anomaly value with the mean of valid data within the preceding and following 2 minutes.

[0109]

[0110] Compliant special process operations: If outliers match the time period of special process instructions (such as "local touch-up coating" or "thickening spraying") in the task dispatch system, they will be retained. A "special process" business tag will be added to this data segment. This tag will be used as an independent feature input to the prediction model, enabling the model to learn and adapt to the usage patterns under such specific working conditions.

[0111] Pending anomalies: Anomalies that do not belong to the above two categories are temporarily treated as noise smoothing and marked as "to be checked" in the dataset. After accumulating a certain number, they are reviewed by process experts to optimize the anomaly identification rules.

[0112] S430: To eliminate the dimensional differences between different parameters and to construct features that are more direct and effective for usage prediction, Z-Score normalization and feature engineering based on parameter characteristics are employed.

[0113] Data standardization includes:

[0114] Dynamic timing parameter standardization: for speed ,temperature ,humidity Parameters that change continuously are standardized using Z-Score. The Z-Score is calculated using global statistics from the historical training dataset, and the formula is as follows:

[0115]

[0116] in, This is the mean of the parameter in the historical training set. Let $\frac{ ... distributed.

[0117] Static parameter normalization: For static parameters such as surface area S and design film thickness T, the following normalization method is adopted: Normalization, scaling it to the [0,1] interval, is calculated using the following formula:

[0118]

[0119] Predictive-oriented feature engineering: Based on the cleaned and standardized basic data, derived features that directly reflect paint consumption patterns and operational efficiency are constructed as key inputs to the predictive model.

[0120] Efficiency characteristics include:

[0121] Theoretical spraying efficiency (m² / min) = Spray gun moving speed (coating speed) ) × Spraying width (Constant or refer to the table according to the nozzle model).

[0122] Overall equipment efficiency (%) = (Theoretical coating usage) / Actual cumulative usage ) ×100%. Of which, theoretical usage = Surface area A × Dry film thickness T × Number of spray coats / (Specific gravity of paint volume solids S×10).

[0123] Stability characteristics include:

[0124] Flow volatility = Instantaneous flow rate within a sliding window (e.g., 5 minutes) Standard deviation / Instantaneous flow rate within the window The mean.

[0125] Operation start-stop ratio =Total spray gun operating time Total task duration .

[0126] Cumulative trend characteristics include:

[0127] Acceleration consumption =Cumulative usage at current moment The rate of change of the first derivative with respect to time (instantaneous consumption rate).

[0128] Through the data preprocessing process closely integrated with the coating business described above, the system can automatically generate a clean, well-organized, and predictive data asset library of coating usage, laying a solid and reliable data foundation for subsequent intelligent prediction of coating usage based on SSA-PLSTM.

[0129] Step 5: Use SSA singular spectral analysis to decompose and denoise the preprocessed data, based on the singular value contribution rate. Smaller outliers are treated as noise and removed. Specific operations include:

[0130] After preprocessing the time-series data, singular spectral analysis (SSA) is performed on the time-series data obtained from different enterprise information systems during the coating operation. SSA can separate the trend component, periodic component, and noise component that characterize the core laws of paint consumption from complex multi-source time-series data, providing enhanced and denoised high-quality input features for the subsequent PLSTM prediction model, thereby significantly improving prediction accuracy and model robustness.

[0131] First, identify the sources of the core time-series data to be analyzed. This data is automatically acquired through enterprise system interfaces: Time-series static parameters obtained from the Management Execution System (MES), including task status (e.g., "Preparing," "Painting," "Completed"), can be converted into time series; process planning times can serve as a reference baseline. Dynamic process parameters obtained from the coating equipment data acquisition system (SCADA / IoT platform): core parameters include instantaneous paint flow rate. (L / min), spray gun on / off status Spray gun movement speed Compressed air pressure AP (MPa), with auxiliary parameters including paint pipeline pressure D and pump frequency P. External environmental parameters obtained from an environmental monitoring system (or sensor network), such as work area temperature C, atmospheric humidity H, and paint viscosity B. Related parameters obtained from the enterprise task dispatch system: specific gravity of fixed matter S, spray width W (which can be associated with the spray gun model). For any of the above key parameters (based on instantaneous paint flow rate)... Taking this as an example, singular spectral analysis is performed, and the main steps include: trajectory matrix construction, singular value decomposition, sequence grouping, and diagonal averaging. The specific calculations are as follows:

[0132] ① Trajectory Matrix Construction

[0133] Given the inherent continuity of time series, the trajectory matrix It is usually fixed. Its core lies in the length of the sliding window. The settings for window length The selection should be related to the physical cycle of the painting operation, rather than a simple data segmentation. Recommended approach:

[0134] Based on process knowledge: the number of sampling points is set to correspond to the average time of a typical spray gun's single reciprocating motion. For example, if the average reciprocating cycle is 30 seconds and the data sampling frequency is 1Hz, then... =30;

[0135] Based on data analysis: historical instantaneous flow rate of coatings Autocorrelation analysis was performed on the sequence, and the lag time corresponding to the first significant peak was taken as... Reference values.

[0136] make sure It can usually be initially set to , Conduct the experiment.

[0137] One-dimensional raw data ,according to The values ​​are sliced ​​to divide the original time series data into several vectors. The trajectory matrix The mathematical expression of the (Hankell matrix) is shown in formula (4) (from the original sequence) (Obtained through sliding window slicing). Parameters and Constraints must be met: ,and Each column of the trajectory matrix X Both represent a length of The time segment is N, where N is the length of the original time series data. This matrix simultaneously captures the short-term dynamics (row direction) and long-term evolution patterns (column direction) of the flow sequence.

[0138]

[0139] ② Singular Value Decomposition

[0140] Singular value decomposition (SVD) is the most crucial step in Singular Spectral Analysis (SSA), capable of handling short-term, nonlinear, and irregular data. The core objective of this step is to decompose the singular values... The original signal data sequence is denoised, where The principal component term contains the left eigenvector. Right eigenvector Left singular matrix Right singular matrix and diagonal matrix .

[0141] Column vectors of the left singular matrix U It is a symmetric matrix The There are orthogonal eigenvectors that satisfy... ; Column vectors of the right singular matrix V It is a symmetric matrix The eigenvectors, satisfying The eigenvectors are obtained through eigenvalue decomposition. (Diagonal matrix) It is a symmetric matrix and symmetric matrix non-zero eigenvalues = composition.

[0142] Left eigenvector It is a symmetric matrix The Each eigenvector represents a different temporal evolution pattern, such as... This typically corresponds to a trend pattern (such as the average flow rate change trend from the start to the end of a spraying operation). The eigenvector is obtained through eigenvalue decomposition, and the calculation steps are as follows:

[0143] S500: Calculate the row covariance matrix

[0144] First, based on the trajectory matrix Calculate its product with its own transpose. The formula is:

[0145]

[0146] in The first of the trajectory matrix Columns, dimensions .

[0147] S510: Solve eigenvalues ​​and eigenvectors

[0148] For symmetric matrices Eigenvalue decomposition is performed, satisfying:

[0149]

[0150] in for The 1 eigenvalue (arranged in descending order) ); correspond Feature vectors (dimensions) And satisfy orthogonality

[0151] Right eigenvector It is a symmetric matrix The Each feature vector represents a sequence of weights (principal components) of these patterns over time. For example... and By combining these factors, the main trend components of traffic can be reconstructed. (Subsequently...) This typically corresponds to periodic or quasi-periodic fluctuations (such as the periodic fluctuations in flow rate caused by the reciprocating motion of the spray gun) and higher-frequency random noise (such as sensor noise and minute disturbances). The eigenvector is obtained through eigenvalue decomposition, and the calculation steps are as follows:

[0152] S510-1: Calculate the column covariance matrix

[0153] First, based on the trajectory matrix Calculate the product of its transpose and itself. The formula is:

[0154]

[0155] in The first of the trajectory matrix Rows, dimensions are .

[0156] S510-2: Solving eigenvalues ​​and eigenvectors

[0157] For symmetric matrices Eigenvalue decomposition is performed, satisfying:

[0158]

[0159] in for The eigenvalues ​​(and) (The non-zero eigenvalues ​​are exactly the same). , These are SVD singular values; correspond Feature vectors (dimensions) And satisfy orthogonality

[0160] when For a column full-rank matrix ( ≤ and When >0), it can be obtained through the left eigenvector. With singular values Directly derive the right eigenvector To avoid double counting The eigenvalues ​​are given by the formula:

[0161]

[0162] SVD decomposition Both sides ride together ,get X= Then ride together on both sides ( diagonal elements are The diagonal elements of the inverse matrix are )get After transposition, Therefore, the i-th column ;

[0163] By using symmetric matrices Diagonalization yields a matrix. Singularity Through the The square root operation is used to obtain ( , for The (i.e., eigenvalues), these singular values ​​form a diagonal matrix. ,in satisfy and Approaching 0. Diagonal matrix. for:

[0164]

[0165] Fundamental matrix It is a trajectory matrix The first one obtained after SVD decomposition The physical meaning of the nth component matrix is ​​the nth component matrix in the original time series. A matrix representation of a feature mode (such as trend, periodicity, noise), the basic matrix As shown in equation (11), the set of basic matrices is ,in .

[0166]

[0167] ③ Grouping of denoised sequences

[0168] After completing the singular value decomposition (SVD), a series of basic matrices are obtained. Then, the matrix is ​​grouped. By separating components with different physical meanings (trend, period, noise), the main signal components are retained, random noise is eliminated, and meaningful signal patterns are extracted, laying the foundation for subsequent analysis.

[0169] Based on the singular value contribution rate, the decomposed components are intelligently grouped to achieve signal enhancement and noise filtering; smaller singular values ​​are considered noise. Therefore, the preceding... Large singular values Represents the main information, the rest are singular values. This represents the noise component. To achieve better feature extraction and training results, the remaining components are discarded. Values ​​are determined by the contribution rate of singular values. Decision. Set the cumulative contribution rate threshold to 90%, select the components corresponding to the first m singular values ​​as "effective feature components", and discard the rest as "noise components".

[0170]

[0171] in This represents the total number of valid singular values.

[0172]

[0173] in These are the eigenvalues.

[0174] Business-oriented group validation:

[0175] Trend component: Check whether the sequence reconstructed from the first 1-2 components smoothly reflects the overall intensity change of the operation and is consistent with the logic of the spray gun switch state sequence. The trend component verification reflects the overall intensity change trend of the spraying operation.

[0176] Periodic component: Check whether the periodicity of subsequent components matches the known reciprocating frequency of the spray gun or periodic fluctuations in the environment (such as workshop ventilation cycle). Periodic component verification reflects periodic factors such as the reciprocating motion of the spray gun.

[0177] Grouping matrix calculation:

[0178] fundamental matrix Divide into m mutually exclusive subsets For each molecule Its corresponding grouping matrix We obtain the following by summing all the basic matrices within the group:

[0179]

[0180] in Let k be the grouping matrix with dimension . It has the same dimension as the original trajectory matrix X.

[0181] Reconstruction matrix calculation:

[0182] Summing all the grouped matrices yields the final reconstructed matrix. :

[0183]

[0184] If the reconstructed matrix contains all grouped matrices (m equals the total number of basic matrices d), then the reconstructed matrix... =Original trajectory matrix X; if it only contains partial groups (discarding noise groups), then This is the reconstructed matrix after denoising.

[0185] ④ Diagonal average method

[0186] Embedding raw time series data into a trajectory matrix In the (Hankell matrix), several sub-matrices are obtained through SVD singular value decomposition and denoising sequence grouping (each sub-matrix corresponds to a group, such as trend, period, etc.). However, these sub-matrices (except for the original trajectory matrix itself) are not necessarily Hankel matrices. Each grouping matrix (sub-matrix) needs to be transformed back into a one-dimensional time series component with the same length as the original sequence to facilitate subsequent intelligent prediction of paint usage. That is, each grouping matrix... Convert to length of The time series. Let the k-th grouping matrix be... Its elements are ,in , The grouping matrix has OK Columns. Grouping matrix By averaging the anti-diagonal elements, the information from the two-dimensional matrix is ​​fused to obtain the one-dimensional time series components. ,in The formula for calculating the t-th element in the sequence is:

[0187]

[0188] in, The k-th grouping matrix (obtained by adding the elementary matrices belonging to the same group after SVD decomposition); Grouping matrix The element in the i-th row and j-th column; It is the length of the sliding window (number of rows in the trajectory matrix); It is the number of columns in the trajectory matrix. N is the length of the original time series; It is the value of the k-th reconstructed sequence at time point t; Let N represent the k-th reconstructed sequence with length N; t is the time index, with a value range of 1≤t≤N; n is the summation variable, which is the row index in the traversal matrix that satisfies the anti-diagonal condition.

[0189] The original time series is After SSA singular spectral decomposition, it is decomposed into Sub-component sequence The original time series representation:

[0190]

[0191] in This represents the value of the original time series at time point t. It is the value of the k-th reconstructed component sequence at time t, where n is the number of groups (i.e., the number of subsequences decomposed); N is the sequence length.

[0192] Advanced feature extraction (for denoising flow sequences) (For example)

[0193] Trend characteristics: calculation The slope of the linear fit is used as the "intensity of the overall trend of flow change";

[0194] Periodic characteristics: Perform a Fourier transform to extract the dominant frequency and corresponding amplitude, which are used as the "periodic fluctuation intensity".

[0195] Statistical characteristics: calculation The mean, standard deviation, peak factor, etc., are used as "flow stability indicators".

[0196] Multi-parameter SSA and feature fusion: The above SSA process is repeated for other key parameters such as spray gun movement speed v and ambient humidity H, extracting their respective trends, periods, and statistical features. Finally, these high-level features of all parameters are combined with static parameters (such as surface area A and design film thickness T) obtained from the task assignment system to form the feature input matrix of the PLSTM model.

[0197] In the SSA decomposition process, the first step is data integration: multi-source time-series and static data required for preprocessing are automatically retrieved from systems such as task dispatching, SCADA, and environmental monitoring via API or data middleware. The second step is parameterized configuration: SSA analysis parameters (such as window length) are preset in the system or automatically determined by the analysis module. 1. Contribution rate threshold (cumulative contribution rate, set to 90%); For each painting task, the system automatically performs SSA decomposition, denoising and feature extraction processes on preset key parameters; Finally, the extracted high-level features (trend strength, period amplitude, stability index, etc.) are associated with the task ID and stored in the feature database for model training and real-time prediction.

[0198] Step 6: Calculate the influence coefficients of each parameter and retain the influence coefficients. Parameters ≥0.05 are used as input features for the PLSTM model; specific operations include:

[0199] Based on the completion of SSA decomposition and feature extraction, the impact of each parameter on coating usage is quantitatively assessed, and key predictive features are selected. A scientific feature selection mechanism is constructed using a comprehensive evaluation method combining "data significance + business relevance" to provide the optimal combination of input features for the PLSTM prediction model.

[0200] ① Obtain complete datasets from various enterprise information systems and perform preprocessing:

[0201] Static parameters obtained from the task dispatch system:

[0202] Design parameters: Surface area A, number of coating layers Dry film thickness T;

[0203] Material parameters: Coating viscosity Diluent content , solid content specific gravity S, surface roughness F.

[0204] Static parameters obtained by the task dispatching system:

[0205] Task priority and planned work periods;

[0206] Dynamic time-series parameters obtained from the environmental monitoring system:

[0207] Air temperature C, atmospheric humidity H, wind force level Bf.

[0208] Dynamic timing parameters obtained from the device's IoT platform:

[0209] Real-time paint viscosity B, spray gun movement speed v, paint flow rate .

[0210] Target variable:

[0211] Actual paint usage Actual consumption data confirmed from the task dispatch system.

[0212] ② Parameter data preparation and SSA feature extraction:

[0213] The above parameters are categorized and processed as follows:

[0214] Dynamic time-series parameters: For time-series parameters such as air temperature C, atmospheric humidity H, wind force Bf, and real-time coating viscosity B, the SSA process described above is applied to extract:

[0215] Trend characteristics: linear trend slope, quadratic trend coefficient;

[0216] Periodic characteristics: dominant frequency amplitude, period length;

[0217] Statistical characteristics: mean, standard deviation, extreme values.

[0218] Static parameters: For static parameters such as surface area A and dry film thickness T, the values ​​are recorded directly, and some can be converted into features through business rules.

[0219] For example, surface area A can be classified as "small (<1㎡)", "medium (10-50㎡)" and "large (>50㎡)";

[0220] Surface roughness F can be mapped to roughness level (1-5).

[0221] ③ Basic weight calculation

[0222] Extracting the maximum singular values ​​of each dynamic parameter after SSA decomposition This represents the significance of the time-series characteristics of the parameter, and the basic weights are calculated based on the singular value contribution.

[0223]

[0224] in ,and (such as surface area A) maximum, (Potentially reaching 0.3), where K is the total number of dynamic parameters involved in the assessment. For example, atmospheric humidity H typically has a high [value / value]. The value reflects significant changes during the operation.

[0225] For static parameters, singular values ​​cannot be obtained through SSA; therefore, the coefficient of variation method is used to calculate the basic weights. The coefficient of variation (standard deviation / mean) of each static parameter in the historical dataset is calculated, normalized, and used as the basis weights. For example, the surface area A varies greatly across different tasks, and its coefficient of variation and basic weights are usually high.

[0226] ④ Nonlinear correlation correction (quantifying the business correlation between parameters and paint usage)

[0227] Using mutual information entropy The nonlinear correlation between quantification parameters and paint usage is used to adjust the base weights. Mutual information entropy is used to measure the strength of the association between two random variables; the calculation formula is:

[0228]

[0229] in Denoising sequences for parameters Compared with actual usage The joint probability distribution, , These are the marginal probability distributions of the two, respectively;

[0230] Dynamic parameters: Calculate all SSA extracted features (trend, period, statistics) for each dynamic parameter and... The mutual information is used, and the maximum value is taken as the parameter. .

[0231] Static parameters: Parameter values ​​are directly calculated and... Mutual information.

[0232] The purpose of adjusting the influence coefficient is to normalize the mutual information value and ensure the rationality of the weights.

[0233]

[0234] in The maximum value of the mutual information entropy of all parameters, ultimately ∈[0,1], ;

[0235] ⑤ Key parameter screening and business verification

[0236] Threshold filtering: retain influence coefficient Parameters ≥0.05 are used as input features for the PLSTM model (such as surface area A, humidity H, paint viscosity B, and diluent content, which are typically selected). As a key parameter, The values ​​are 0.32, 0.25, 0.18, and 0.12 respectively.

[0237] Business expert verification: The screening results are submitted to coating process experts for review. Based on their experience, the experts can propose corrections for obviously unreasonable weights (such as the diluent content under specific processes). (This should be more important). The system records expert correction factors for subsequent model iterations.

[0238] Feature combination optimization: For the parameters that pass the screening, further analyze the multicollinearity between features (e.g., temperature C and humidity H may be correlated). Through variance inflation factor (VIF) analysis, if VIF>10, consider removing or merging correlated features.

[0239] ⑥ Output and Model Integration

[0240] Feature list output: Generates a list of key features, including:

[0241] Feature names (e.g., "humidity_trend slope", "surface area_value");

[0242] Influence coefficient ;

[0243] Data source system (such as "environmental monitoring system" or "PDM system").

[0244] Model configuration file: Contains the filtered feature list and preprocessing parameters (such as SSA window length). The standardized parameters are saved as a configuration file for use during PLSTM model training and prediction.

[0245] Regular update mechanism: Establish a mechanism for regularly re-evaluating the importance of features (e.g., quarterly or after accumulating 100 new tasks):

[0246] The influence coefficients of each parameter were recalculated using the new data.

[0247] If the ranking of feature importance changes significantly (e.g., the top 5 features change by more than 2 places), the model retraining process is triggered.

[0248] In the parameter SSA decomposition and influence coefficient calculation steps, the required parameter data is first automatically extracted from systems such as task dispatching, PDM, and environmental monitoring via the Enterprise Service Bus (ESB). Then, the system automatically performs SSA decomposition, feature extraction, mutual information calculation, and weight calculation according to a preset cycle (e.g., daily). After calculation, the feature list that meets the set requirements is automatically updated to the prediction model configuration, ensuring that the model uses the latest and most relevant feature combinations. Finally, the system generates a feature importance analysis report, showing the changing trends of the influence of each parameter, providing data support for process optimization. This systematic and business-integrated influence coefficient calculation method not only achieves data-driven feature selection but also ensures that the selection results conform to the physical laws of actual coating processes, laying a solid foundation for building a high-precision, interpretable paint usage prediction model.

[0249] (4) Coating usage prediction model based on PLSTM

[0250] During the execution of the predictive model, data needs to be acquired in real time from various enterprise information systems and preprocessed, feature extracted, and influence coefficients calculated according to the methods described above, ultimately forming the model input. The specific data flow is as follows:

[0251] ① Obtain design parameters from task assignment and PDM system:

[0252] Static parameters: Surface area A, dry film thickness T, number of coating layers Paint grade (used to correlate solids content, specific gravity S, and diluent content) The default value), surface roughness F (which can be obtained from the process card).

[0253] Acquisition method: Obtained through the API of the task dispatch system and PDM system, based on the product number and process number in the task dispatch order.

[0254] ② Obtain task and process parameters from the task dispatching system:

[0255] Static parameters: diluent content τ (may be adjusted on-site and may differ from the design value), work team, and planned work time (used to estimate environmental conditions).

[0256] Dynamic parameter: Actual operation time (used to correlate with environmental data).

[0257] Acquisition method: Obtained in real time based on task number.

[0258] ③ Obtain environmental data from the environmental monitoring system:

[0259] Dynamic parameters: air temperature (C), atmospheric humidity (H), wind speed (Bf).

[0260] Acquisition method: Time-series data is acquired through the data interface of the environmental monitoring system, according to the task operation time period (or in real time).

[0261] ④ Obtain device data from the device IoT platform:

[0262] Dynamic parameters: paint viscosity B (online sensor), spray gun movement speed v, paint flow rate wait.

[0263] Acquisition method: Query time-series data within the task time period through a real-time database.

[0264] ⑤ Obtain actual usage from historical databases:

[0265] Target variable: Actual paint usage .

[0266] Acquisition method: Obtained from historical task records and used for model training and validation.

[0267] For each painting task (whether a historical task or a task to be predicted), the following inputs need to be constructed:

[0268] ① Construction of dynamic parameter time series

[0269] For each key dynamic parameter (for example, after screening by influence coefficients, atmospheric humidity H, coating viscosity B, and air temperature C are identified as key dynamic parameters), a time series of length T needs to be constructed. Here, the time step T refers to the length of time the model reviews in the past, for example, T=30 (meaning 30 time points, each time point represents 1 minute, i.e., reviewing the past 30 minutes).

[0270] For historical tasks, a time series can be formed by taking a point every minute from the start to the end of the task. Then, a sliding window (window size T) is used to generate multiple training samples.

[0271] For the task to be predicted, it is necessary to obtain dynamic parameter data from the past T minutes in real time to form a sequence of length T.

[0272] For each dynamic parameter at each time point t (t=1,2,...,T), the input features are the parameter's value at time t and the features extracted through SSA (such as trend components, periodic components, etc.). Therefore, for a dynamic parameter, the input at each time point may be a multi-dimensional vector. Assuming the use of three features—the original value, the trend component, and the periodic component—the input dimension of each dynamic parameter at each time point is 3.

[0273] If There are 1 key dynamic parameters, so the total dimension of the dynamic parameter input at each time point is 1. * 3. The dimension of the entire dynamic parameter timing input matrix is: T × ( * 3).

[0274] ② Static parameters (such as surface area A, dry film thickness T, etc.) remain unchanged during the task, but in order to align with the dynamic parameter sequence in the time dimension, the static parameters need to be repeated T times to form a T× The matrix, where It is the feature dimension of static parameters (note that static parameters may also undergo feature engineering, such as one-hot encoding of class parameters).

[0275] ③ Input merging

[0276] The dynamic parameter time series matrix and the static parameter time series matrix are concatenated along the feature dimension to obtain the input matrix of the entire model, with dimensions T × ( * 3 + Then, according to the requirements of the PLSTM model, the dynamic parameters are input into the dynamic LSTM subnetwork, and the static parameters are input into the static LSTM subnetwork.

[0277] In recurrent neural networks (RNNs), excessively long distances between nodes can lead to severe gradient explosion and vanishing problems. Long Short-Term Memory (LSTM) networks were developed to address these shortcomings of RNNs. The core improvement of LSTM lies in optimizing the repetitive modules in RNNs from the original single-layer neural network structure into four interactive modules. Therefore, LSTMs are particularly suitable for handling transactions with long-term dependencies and excel at predicting highly volatile and abruptly changing time-series data. PLSTMs (Parallel LSTM Subnetworks) are designed to handle different time-series characteristics of dynamic parameters (such as humidity H, temperature C) and static parameters (such as surface area A, dry film thickness T). The structure includes an input layer, a parallel LSTM subnetwork, a fusion layer, and an output layer. Figure 3 As shown.

[0278] Step 1: Preparation and preprocessing of input layer coating object data

[0279] Input static and dynamic parameters. Static parameters include surface area A and dry film thickness T; dynamic parameters include atmospheric humidity H and coating viscosity. The data is preprocessed to expand the static parameters into a time series static parameter matrix. Obtain the dynamic parameter sequence of the past T time steps. Dynamic parameters are processed. The input layer receives denoised time-series data containing key parameters, such as the input dimension: ( (where K is the number of key parameters, e.g., K=4; T is the time step, e.g., T=24 minutes).

[0280] Step 2: Parallel LSTM Subnetwork Computation

[0281] The parallel LSTM subnetwork comprises subnetwork 1 and subnetwork 2. Subnetwork 1 captures the real-time changes in dynamic parameters such as humidity and temperature, while subnetwork 2 captures the fixed feature dependencies of static parameters such as surface area and film thickness. The Long Short-Term Memory (LSTM) network structure includes three gates: an input gate, an output gate, and a forget gate. Each subnetwork is configured with 32 hidden layer nodes, using the tanh activation function, and the input, output, and forget gates are activated using the sigmoid function.

[0282] Subnetwork 1 (Dynamic Parameter LSTM) calculation process:

[0283] 1) Forget Gate Calculation:

[0284]

[0285] The forget gate determines the proportion of historical information of dynamic parameters (such as previous humidity change patterns) that is retained.

[0286] 2) Input gate calculation:

[0287]

[0288] The input gate controls the proportion of current dynamic parameters (current humidity, viscosity) entering the memory system.

[0289] 3) Calculation of temporary cell states:

[0290]

[0291] Temporary cell states generate new dynamic feature representations based on current input and historical information.

[0292] 4) Cell state update:

[0293]

[0294] Cellular state updates combine forgetting and input gates to update dynamic features in long-term memory.

[0295] 5) Output gate calculation:

[0296]

[0297] Hidden layer state output:

[0298]

[0299] The hidden layer state output extracts dynamic features that have a significant impact on the amount of paint used at the current moment.

[0300] Final output: (The hidden state at the last time step, with a dimension of 32×1)

[0301] in The activation function is Sigmoid, with a range of [0,1]. , , , These are the dynamic subnetwork weight matrices for the forget gate, input gate, candidate state, and output gate, respectively, with a dimension of 32 (32+2). It is the hidden state of the dynamic subnetwork at time t-1, with a dimension of 32×1; The dynamic parameter input vector (such as humidity, viscosity) represents time step t, with a dimension of 2×1; , , , It is the dynamic subnetwork bias vector of the forget gate, input gate, candidate state, and output gate, with a dimension of 32×1. This is element-wise multiplication.

[0302] The calculation process for subnetwork 2 is the same as that for subnetwork 1, and the final output is... The dimension is 32×1.

[0303] Step 3: Calculation of the fusion layer

[0304] The fusion layer integrates the outputs of two sub-networks, learning the interactions between parameters, such as the combined effect of humidity and viscosity on paint usage. The fully connected layer has an input dimension of 64 and an output dimension of 32, using the ReLU activation function.

[0305] 1) Feature splicing

[0306]

[0307] Feature concatenation combines the time-dependent features of dynamic parameters with the fixed features of static parameters.

[0308] 2) Fully connected transformation

[0309]

[0310]

[0311] Fully connected transformations learn the interaction between dynamic and static parameters (such as the combined effect of "high humidity + large surface area" on dosage). It is the weight matrix of the fusion layer, with a dimension of 32×64; It is a concatenated vector of the outputs of the two sub-networks, with a dimension of 64×1; It is the bias vector of the fusion layer, with a dimension of 32×1; This indicates that the fusion layer outputs a feature vector with a dimension of 32×1.

[0312] Step 4: Output layer prediction

[0313]

[0314] in It is the output layer weight matrix, with a dimension of 1×32; It is the output layer bias scalar, with a dimension of 1×1; It predicts the amount of paint used (unit: L), with a dimension of 1×1.

[0315] Step 5: Summary of Complete Prediction Formulas

[0316] Dynamic subnetworks:

[0317] Static subnetwork:

[0318] Fusion layer:

[0319] Output layer:

[0320] (5) Model training and hyperparameter optimization

[0321] The dataset is split into training and testing sets in an 8:2 ratio. The input is the sequence of key parameters for denoising, and the labels are the actual usage. Input the key parameter sequence for denoising, labeled with the actual usage. .

[0322] Training parameter configuration: Adam optimizer is used, initial learning rate... =0.001 (Dynamically adjusted: if the validation set loss does not decrease for 3 consecutive rounds, the learning rate is halved).

[0323] The training rounds are set to 50, and the batch size is 32.

[0324] Regularization: Regularization techniques are used to prevent model overfitting. The Dropout ratio is set to 0.2 (when the Dropout ratio is set to 0.2, it means that during the model training process, for the neurons in the current layer, 20% of the neurons will be randomly "dropped (temporarily disabled). That is, these 20% of the neurons will not participate in the forward and backward propagation calculations in the current iteration, nor will their weights be updated, while the remaining 80% of the neurons will work normally).

[0325] The L2 regularization coefficient is 1. The L2 regularization coefficient (also known as the weight decay coefficient) is a hyperparameter that controls the strength of L2 regularization and is used to prevent the model from overfitting. L2 regularization limits the magnitude of the model parameter weights by adding a "sum of squared weights" term to the model's loss function. Assume the original loss function is... ( (where all weight parameters of the model are given), then the loss function after adding L2 regularization is:

[0326]

[0327] in That is Regularization coefficient (1 here) The added regularization strength is relatively weak—it can suppress excessively large weights without overly restricting the model from learning effective features, making it suitable for scenarios that require a balance between model complexity and fitting ability. The larger the value, the stronger the "penalty" on the weights, forcing the weight parameters closer to 0 (suppressing excessively large weights), thereby preventing the model from overfitting details (such as noise) in the training data. It is the sum of squares of all weight parameters.

[0328] The loss function is calculated using the mean squared error (MSE), and the formula is as follows:

[0329]

[0330] Hyperparameter optimization: Key hyperparameters are adjusted using Bayesian optimization, with the goal of minimizing the MSE on the test set. The hyperparameter search range is as follows:

[0331] Time step T: 12, 24, 36 minutes;

[0332] LSTM hidden layer node count: 16, 32, 64;

[0333] Learning rate α: 1 3 5

[0334] (6) Predicting actual paint usage

[0335] For a new batch of objects to be coated, input their key parameters (such as...). The PLSTM model, once trained, outputs the predicted usage. (like =28.5L)

[0336] 5. Automatically collect actual paint consumption data during the coating process and transmit it to the server database to update the enterprise's coating data asset library. Based on the actual paint consumption data of the object being coated, correct the paint consumption predicted by the SSA-PLSTM algorithm, iteratively optimize the model, and ultimately achieve the goal of continuously reducing paint consumption.

[0337] (1) Calculation of prediction deviation

[0338] Once the new batch of objects to be painted is completed, collect the actual amount used. Calculate two types of deviation indices:

[0339] Absolute error:

[0340] Relative error: (Reflects the relative severity of the deviation)

[0341] (2) Model Iterative Optimization Strategy

[0342] The model is adjusted according to the magnitude of the deviation for different scenarios to ensure continuous improvement in prediction accuracy.

[0343]

[0344] (3) Optimization effect verification

[0345] After each iteration, the model performance is verified using a test set, and the evaluation metrics are as follows (the target value is the minimum requirement after iterative optimization):

[0346] Root Mean Square Error (RMSE):

[0347] Mean Absolute Percentage Error (MAPE):

[0348] Coefficient of determination ( ): ( (Average actual usage)

[0349] VI. The spraying process can provide prompts or warnings based on actual consumption, guiding construction personnel to control the pace appropriately.

[0350] (1) Real-time prompts and early warnings of actual consumption

[0351] ① Real-time consumption monitoring

[0352] During the painting process, the cumulative consumption is collected in real time through the equipment's flow sensor. (Updated every minute), calculate the "progress consumption ratio":

[0353]

[0354] Predicting remaining usage by displaying the "Current Progress" (r%) in real time: This helps operators control the amount and timing of paint application.

[0355] ② Design of a multi-level early warning mechanism

[0356] Based on the consumption progress ratio r, three threshold levels are set to trigger corresponding prompts / warning actions, as follows:

[0357]

[0358] ③ Closed-loop processing after early warning

[0359] When a warning / emergency alert is triggered, the system automatically connects to the task dispatch system and records the parameter data (such as pressure and humidity) at the time of the alert.

[0360] After management intervenes, they enter the "treatment measures (such as adjusting the pressure to P=0.4MPa)" and "changes in usage after treatment" into the system as supplementary data for subsequent model iterations.

[0361] Monthly statistics on the number of warning triggers and their effectiveness are compiled to optimize warning thresholds (e.g., if humidity significantly affects a particular workshop, the warning threshold can be lowered to a lower level). ).

[0362] This invention introduces machine learning algorithms (such as random forests and LSTM neural networks) to replace traditional linear formulas, capturing the nonlinear relationships between multiple variables. After each batch of coating is completed, the system automatically compares the estimated usage with the actual usage. When the deviation exceeds a set value, it triggers model parameter updates (such as adjusting the weight of the influence of environmental humidity on the usage). At the same time, the "parameter-usage" data of the new batch is incorporated into the enterprise's data asset library, realizing the autonomous iteration of the estimation system.

[0363] The symbols used in this invention are explained in the following table:

[0364]

[0365]

[0366]

[0367] .

Claims

1. A method of building a ship coating usage data asset library, characterized by: Includes the following steps: Step 1: Obtain historical static parameter data, historical dynamic parameter data, and corresponding actual paint usage; Step 2: Based on the dynamic parameter type, perform SSA feature extraction on the historical dynamic parameter data to obtain trend features, periodic features, and statistical features; Step 3: Extract the maximum singular value after SSA decomposition of each dynamic parameter data, and calculate the basic weight of each feature based on the maximum singular value: calculate the basic weight of each static parameter using the coefficient of variation method; Step 4: Calculate the mutual information between all features extracted from each dynamic parameter data and the actual amount of paint used, and take the maximum mutual information value as the mutual information entropy of that feature; calculate the mutual information entropy between each static parameter data and the actual amount of paint used; based on the mutual information entropy, adjust the basic weights to obtain the influence coefficient; Step 5: Select features with an influence coefficient greater than a set threshold as key features; Step 6: Optimize the key features through variance inflation factor analysis to obtain the optimized key features; Step 7: Save the optimized key features and corresponding parameter data to build a marine coating usage data asset library.

2. A method of building a data asset library of marine coating usage data according to claim 1, characterised in that: The static parameter data is obtained based on design parameter data, material parameter data, task priority, and planned operation period; the dynamic parameter data is obtained based on external environment data, real-time paint viscosity, spray gun moving speed, and paint flow rate.

3. The method for constructing a marine coating usage data asset database according to claim 2, characterized in that: Both the historical static parameter data and the historical dynamic parameter data are obtained through the following preprocessing steps: Based on the parameter type and its criticality to the calculation and prediction of coating usage, differentiated missing value processing is implemented; By integrating equipment status and process rules, outlier identification and correction can be performed. Z-Score standardization is used to standardize the dynamic parameter data; Min-Max normalization is used to normalize the static parameter data.

4. A method of building a data asset library of paint usage data for a marine vessel according to claim 3, characterised in that: The fusion of equipment status and process rules includes outlier identification and correction, including: For all numerical time series parameters in the most recent Calculate the mean of the sequence over a given time point. with standard deviation If the value of the current numerical time series parameter satisfies If so, it is marked as an anomaly; If the equipment status signal is off, faulty, or in standby mode at the time corresponding to the anomaly point, it is determined to be invalid noise and replaced by the mean smoothing method of the adjacent time period. If an outlier matches a specific process instruction time period in the task dispatching system, it will be retained, and a special process business tag will be added to this data segment. This business tag will be treated as an independent feature. Anomalies that do not fall into either of the above two categories are processed as noise smoothing and marked as pending verification. Once a certain number are accumulated, they are reviewed by process experts.

5. A method of building a data asset library for marine coating usage data according to claim 3, characterized in that: The process of handling differential missing values ​​based on parameter type and its criticality to coating usage calculation and prediction includes: The forward imputation method is used to fill in missing values ​​for continuous parameters; Missing values ​​of static parameters were filled using the average values ​​of outfitting components of the same type.

6. An electronic device, comprising: The electronic device includes: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, which is then executed by the at least one processor to enable the at least one processor to perform the method for constructing a marine coating usage data asset database as described in any one of claims 1-5.

7. A computer readable storage medium characterized by The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the method for constructing a marine coating usage data asset database as described in any one of claims 1-5.

8. A computer program product, characterised in that, The computer program product includes a computer program that, when executed by a processor, implements the method for constructing a marine coating usage data asset database as described in any one of claims 1-5.

Images