Wetting time missing data filling model and fine characterization method

Abstract

The invention discloses a wetting time missing data filling model and a fine characterization method. Environmental factor data of a test station is collected, and a meteorological station environmental factor data set is collected; multi-source data granularity alignment is completed; performing feature analysis on data of the test station and the weather station based on statistical indexes, and evaluating temperature and humidity differences of the test station and the weather station; constructing an LSTM-based environmental factor mapping model and model parameters with optimal curing performance; on the basis of the optimal mapping model, missing data is filled, a complete data set is formed, and by comparing and analyzing characteristics before and after data filling, it is ensured that the data rule is unified; and the wetting time definition is given for different metals and different environments, the features of different time ranges such as annual wetting time and monthly wetting time are generated based on the complete data set, the features of different mechanism meanings are generated, fine representation of the wetting time is completed, and a basis is provided for subsequent metal environment adaptability evaluation.

Description

Technical Field

[0001] The present invention belongs to the field of natural environment testing technology, and specifically relates to a wetting time missing data filling model and a refined characterization method. Background Art

[0002] As one of the most commonly used material categories in modern industry, the study of metal performance degradation and corrosion resistance is crucial. The essence of metal corrosion is electrochemical corrosion. Wetting time is a direct factor that determines the speed at which the metal surface contacts the medium, affecting the corrosion rate and protective effect. Therefore, conducting metal environmental adaptability assessments requires a detailed characterization of the wetting time. However, existing natural environment tests only collect environmental factors such as temperature and humidity. Some of the collection equipment is relatively simple and there are cases of missing data. It is necessary to first obtain complete data and then characterize the wetting time. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a missing data filling model and a refined characterization method for wetting time, which fills the missing data by using an environmental factor mapping model, and then uses the filled data to perform a refined characterization of the wetting time.

[0004] The purpose of the present invention is achieved through the following technical solutions:

[0005] A method for filling missing data of experimental stations based on an environmental factor mapping model includes:

[0006] S1: Acquire test site environmental factor data and weather station environmental factor data, wherein the test site environmental factor data includes data of various environmental factors under different test modes collected at a preset test site, and the weather station environmental data includes data of various environmental factors at the weather station, and the environmental factor data includes data corresponding to at least two environmental factors: temperature and humidity;

[0007] S2: Perform granular alignment on the test site environmental factor data and the weather station environmental factor data, where granular alignment includes site granularity alignment, time zone granularity alignment, and time granularity alignment. The aligned test site environmental factor data and weather station environmental factor data are then integrated. The integrated test site environmental factor data all have corresponding weather station environmental factor data. However, some weather station environmental factor data after integration may not have corresponding test site environmental factor data, i.e., there may be missing test site environmental factor data.

[0008] S3: Constructing an environmental factor mapping model, wherein the environmental factor mapping model is a dual-input dual-output model based on LSTM, wherein one input of the environmental factor mapping model is a time series consisting of the environmental factor data of the test station, and the other input is a time series consisting of the environmental factor data of the corresponding weather station; one output is the predicted environmental factor data of the test station at other times, and the other output is the predicted environmental factor data of the weather station at other times;

[0009] S4: Use the integrated results in S2 to train the environmental factor mapping model and fill in the missing environmental factor data of the experimental sites.

[0010] Furthermore, the environmental factor mapping model is a forward prediction model or a reverse prediction model, which realizes the prediction of the next moment corresponding to the latest moment in the input time series through the forward prediction model, and realizes the prediction of the previous moment corresponding to the earliest moment in the input time series through the reverse prediction model.

[0011] Furthermore, the method further comprises:

[0012] S5: Perform feature analysis on the environmental factor data of the experimental station and the environmental factor data of the meteorological station after integration in S2 to generate a first statistical feature to evaluate the difference in environmental factors between the experimental station and the meteorological station;

[0013] S6: After S4, feature analysis is performed using the completed test site environmental factor data and the meteorological station environmental factor data to generate a second statistical feature;

[0014] S7: Compare the second statistical feature with the first statistical feature, and determine whether the data before and after the completion are consistent based on the comparison result; if they are consistent, the training of the environmental factor mapping model is completed; if they are inconsistent, return to S4, and re-train the environmental factor mapping model by introducing the randomness of the test site environmental factor data and the meteorological station environmental factor data, and use the re-trained environmental factor mapping model to re-complete the test site environmental factor data with missing data until the data before and after the completion are consistent. The environmental factor mapping model obtained by this training is the final environmental factor mapping model.

[0015] Furthermore, the first statistical feature and the second statistical feature include a combination of one or more of the mean absolute error, mean square error, maximum deviation, minimum deviation, correlation coefficient, etc. between the test station environmental factor data and the meteorological station environmental factor data.

[0016] The environmental factor mapping model is a dual-input dual-output model based on LSTM, wherein the dual-input dual-output model based on LSTM includes two branches.

[0017] The first branch uses the experimental station data and the meteorological station data as input to form a strong coupling relationship and predict the missing experimental station environmental factor data;

[0018] The second branch uses the weather station data as input to predict the weather station data at other times;

[0019] The first branch is the main method for achieving prediction, and the second branch is the self-iteration of weather station data, which is used to correct the weight of the second branch to improve the prediction accuracy of the first branch.

[0020] The present invention also provides a method for fine characterization of wetting time, comprising:

[0021] Obtain the completed test site environmental factor data obtained according to the above data completion method;

[0022] Based on the completed test site environmental factor data, count the amount of first data in which the temperature exceeds a first temperature threshold and the humidity exceeds a first humidity threshold within a time period corresponding to a preset time dimension, and the amount of second data in which the temperature exceeds a second temperature threshold and the humidity exceeds a second humidity threshold within a time period corresponding to a different time dimension, where the first humidity threshold and the second humidity threshold are different;

[0023] determining a first wetting time corresponding to a preset time dimension based on the first data volume and a time interval corresponding to data in the test site environmental factor data;

[0024] Determine a second wetting time corresponding to a preset time dimension based on the second data volume and the time interval corresponding to the data in the test site environmental factor data,

[0025] Through different time dimensions and different wetting times, a refined characterization of wetting time is achieved.

[0026] Furthermore, the preset time dimension includes a year time dimension and a month time dimension.

[0027] Furthermore, the first humidity threshold, the first temperature threshold, the second humidity threshold and / or the second temperature threshold may be adjusted according to different test subjects.

[0028] The beneficial effects of the present invention are:

[0029] The present invention collects environmental factor data of test sites and environmental factor data sets of meteorological stations; unifies the environmental factor data of test sites and meteorological stations to the same caliber and completes the granularity alignment of multi-source data; performs feature analysis on the data of test sites and meteorological stations based on statistical indicators, and evaluates the temperature and humidity differences between the test sites and meteorological stations; constructs an environmental factor mapping model based on LSTM and solidifies the model parameters with the best performance; fills in missing data based on the optimal mapping model to form a complete data set, and ensures the uniformity of data rules by comparing and analyzing the features before and after data filling; defines wetting time for different metals and different environments, generates features of different time ranges such as annual wetting time and monthly wetting time based on the complete data set, and generates features of different mechanism meanings such as first critical wetting time and second critical wetting time, to complete the refined characterization of wetting time.

[0030] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0032] Figure 1 It is a schematic diagram of the environmental factor mapping model;

[0033] Figure 2 It is a schematic flow chart of the missing data filling method and the wetting time refinement characterization method based on the environmental factor mapping model test station;

[0034] Figure 3 It is a line graph of experimental station data and weather station data. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure. The following disclosure provides many different embodiments or examples for implementing different schemes of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described below. Of course, they are merely examples and are not intended to limit the present invention. In addition, the present invention may repeat reference numbers and / or letters in different examples. This repetition is for the purpose of simplicity and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed.

[0036] The following will describe the embodiments of the present disclosure with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be understood that the preferred embodiments are only for the purpose of illustrating the present disclosure and are not intended to limit the scope of protection of the present disclosure. The present invention will be further described in detail below with reference to the accompanying drawings.

[0037] Figure 1 It is a schematic diagram of the environmental factor mapping model, combined with Figure 1 A method for filling missing data of experimental stations based on an environmental factor mapping model (also known as a "wetting time missing data filling model" or a "missing data filling model") may include the following steps:

[0038] S1: Obtain test site environmental factor data ("test site data" for short) and meteorological station environmental factor data ("weather station data" for short), wherein the test site environmental factor data includes data of various environmental factors under different test modes collected at a preset test site; the meteorological station environmental data includes data of various environmental factors collected at the meteorological station site.

[0039] Environmental factor data includes at least data corresponding to temperature and humidity. This is because the subsequent wetting time selection principle is only related to temperature and humidity. Therefore, in some embodiments, only temperature and humidity are selected as data items. In some embodiments, other environmental factors such as solar radiation, wind direction and speed may also be included.

[0040] S2: Perform granular alignment on the test site environmental factor data and the meteorological station environmental factor data, where granular alignment includes site granularity alignment, time zone granularity alignment, and time granularity alignment, and integrate the aligned test site environmental factor data and meteorological station environmental factor data. The integrated test site environmental factor data all have corresponding meteorological station environmental factor data, but some of the integrated meteorological station environmental factor data may not have corresponding test site environmental factor data, that is, there is missing test site environmental factor data.

[0041] Since the experimental stations and meteorological stations are output by different units and are located in different locations, their station codes and names are different. The station granularity is based on the experimental station and meteorological station with the closest longitude and latitude. It can be roughly assumed that the environmental factors of the two are similar. A mapping relationship is established between the two to fill in the missing data of the experimental stations.

[0042] In addition, different data sources may use different time zone bases, that is, the time zone granularity is different. It is necessary to clarify whether the time zone uses the Coordinate Universal Times (UTC) or the GMT+8 time zone, and align the time with the time zone granularity according to the time zone base.

[0043] In addition, the data collected by different sites and different equipment are at the daily, three-hour, hourly, and minute levels, that is, the time granularity is different. The coarsest granularity of all data can be selected as the benchmark, and the timestamps of the test station data and the weather station data can be aligned.

[0044] S3: Construct an environmental factor mapping model, wherein the environmental factor mapping model is a dual-input dual-output model based on LSTM, such as Figure 1 As shown, one input of the environmental factor mapping model is a time series consisting of the environmental factor data of the experimental station, and the other input is a time series consisting of the environmental factor data of the corresponding meteorological station; one output is the predicted environmental factor data of the experimental station at other times, and the other output is the predicted environmental factor data of the meteorological station at other times.

[0045] like Figure 1 As shown, the LSTM-based dual-input dual-output model includes two branches:

[0046] The first branch uses the experimental station data and the meteorological station data as input to form a strong coupling relationship and predict the missing experimental station environmental factor data;

[0047] The second branch uses the weather station data as input to predict the weather station data at other times;

[0048] The first branch is the main method for achieving prediction, and the second branch is the self-iteration of weather station data, which is used to correct the weight of the second branch to improve the prediction accuracy of the first branch.

[0049] Since the prediction using the experimental station data is the focus of the model, and the prediction using the weather station data is only for fine-tuning the weight, the weight W1 of the first branch is set much larger than the weight W2 of the second branch. If W2 is too large, the final prediction result will be tilted towards the weather station data prediction, which will reduce the accuracy of the experimental station data prediction.

[0050] In this invention, through this dual input, the missing data of the test station is filled in using the weather station data trend with the test station data as the benchmark, and the missing data is predicted together. The effect is far better than using only the weather station data to predict the test station data (i.e., a single sequence predicts another sequence), or using only the test station data to predict the missing test station data (i.e., a single sequence predicts the unknown). At the same time, an additional task is added to predict the weather station environmental factor data at other times, that is, to predict the known weather station environmental factor data at other times, thereby correcting the model parameters, promoting mutual learning, and avoiding overfitting caused by the model only selecting the test station data to correct the weight.

[0051] S4: Use the integrated results in S2 to train the environmental factor mapping model and fill in the missing environmental factor data of the experimental sites.

[0052] The environmental factor mapping model is a forward prediction model or a reverse prediction model. The forward prediction model is used to predict the moment after the latest moment in the input time series, and the reverse prediction model is used to predict the moment before the earliest moment in the input time series.

[0053] Specifically, the training of the environmental factor mapping model may include:

[0054] Data cutting: The length of the input and output sequence k should cover 24 hours. If the data interval is δ, then k ≥ 24 / δ. The data cutting principle is to ensure that the model can learn the periodicity of the daily cycle.

[0055] Data normalization and denormalization: The mean and standard deviation of the two sequences are recorded separately for data normalization and denormalization. Normalization ensures model training and rapid convergence, while denormalization restores the predicted data to a normal scale.

[0056] Hyperparameter setting: Reasonably determine the hyperparameters of each layer of the network structure based on the data volume and sequence length.

[0057] Data randomization and cross-validation: Since environmental factor data has annual and daily periodicity, scrambling the data sequence (only the order between sequences is scrambled, not the order within the sequence) to generate random sequences helps the model learn associated features rather than cyclical features. At the same time, cross-validation ensures that the model will not lead to deviations in results due to a random result. By splitting the data set multiple times to average these deviations, more stable results are obtained.

[0058] Model evaluation: Divide the dataset into 7:2:1, with 70% as the training set, 20% as the validation set, and 10% as the test set. Evaluate the model prediction effect on the test set, including at least MAE, MSE, and correlation coefficient R 2 , maximum deviation and other indicators.

[0059] In some embodiments, the method further comprises:

[0060] S5: Perform feature analysis on the environmental factor data of the experimental station and the environmental factor data of the meteorological station after integration in S2 to generate a first statistical feature to evaluate the difference in environmental factors between the experimental station and the meteorological station;

[0061] S6: After S4, feature analysis is performed using the completed test site environmental factor data and the meteorological station environmental factor data to generate a second statistical feature;

[0062] S7: Compare the second statistical feature with the first statistical feature, and determine whether the data before and after the completion are consistent based on the comparison result; if they are consistent, the training of the environmental factor mapping model is completed; if they are inconsistent, return to S4, and re-train the environmental factor mapping model by introducing the randomness of the test site environmental factor data and the meteorological station environmental factor data, and use the re-trained environmental factor mapping model to re-complete the test site environmental factor data with missing data until the data before and after the completion are consistent. The environmental factor mapping model obtained by this training is the final environmental factor mapping model.

[0063] The first statistical feature and the second statistical feature include one or more combinations of the mean absolute error, mean square error, maximum deviation (for example, the maximum deviation of the weather station is higher than that of the test station), minimum deviation (for example, the minimum deviation of the weather station is lower than that of the test station), correlation coefficient, etc. between the test station environmental factor data and the weather station environmental factor data.

[0064] By comparing the features before and after data completion to ensure consistency, we can determine the optimal mapping model for repeated use. Based on the optimal mapping model, we fill in missing data and consolidate all data into a single table, creating a consistent, complete data set with consistent time intervals. We then compare and analyze the features before and after data completion to ensure consistent data patterns.

[0065] In some embodiments of the present invention, line graphs of the experimental station data and the weather station data are drawn as a posteriori test of the aforementioned granularity alignment, and the trend and followability of the data are visually checked at the same time.

[0066] After the complete data is finally generated, the present invention also provides a method for fine-tuning the wetting time, including:

[0067] Obtain the completed test site environmental factor data obtained according to the above data completion method, including temperature and humidity data;

[0068] Based on the completed test site environmental factor data, the amount of first data in which the temperature exceeds the first temperature threshold and the humidity exceeds the first humidity threshold within the time period corresponding to the preset time dimension, and the amount of second data in which the temperature exceeds the second temperature threshold and the humidity exceeds the second humidity threshold within the time period corresponding to a different time dimension are counted, wherein the first humidity threshold and the second humidity threshold are different, for example, the first humidity threshold is set to 50% and the second humidity threshold is set to 80%. The first temperature threshold and the second temperature threshold can be the same (for example, both are 0°C) or different. The preset time dimension can include a year time dimension and a month time dimension;

[0069] Determining a first wetting time corresponding to a preset time dimension based on the first data volume and a time interval corresponding to data in the test site environmental factor data, for example, determining a first year wetting time, a first month wetting time, etc.;

[0070] Determining a second wetting time corresponding to a preset time dimension according to the second data volume and a time interval corresponding to data in the test site environmental factor data, for example, determining a second year wetting time, a second month wetting time, etc.;

[0071] Through different time dimensions and different wetting times, a refined characterization of wetting time can be achieved.

[0072] It should be noted that to ensure a refined representation of wetting time, the highest acquisition frequency data should be retained as much as possible. If hourly data is available, daily data should not be used; if daily data is available, monthly data should not be used. Due to different test methods, even at the same site, different test methods will lead to different results. Therefore, data collection also needs to cover different test methods, such as outdoor exposure, offshore outdoor, near-shore outdoor, and offshore platform exposure.

[0073] In some embodiments, the first humidity threshold, the first temperature threshold, the second humidity threshold and / or the second temperature threshold can be adjusted according to different test subjects.

[0074] The following is explained through specific examples.

[0075] In this example, the weather station data used are the federal climate integrated surface hourly data (i.e., Integrated Surface Hourly Data, ISH data), as shown in Table 1; the temperature and humidity data of the experimental station are shown in Tables 2 and 3.

[0076] Table 1. Raw data of weather stations

[0077]

[0078]

[0079] It should be noted that since the ISH data is international data, the station names of each station in the original ISH are all in English. For convenience, the Chinese names corresponding to the English station names of each station are listed here.

[0080] Table 2 Raw data of test station - temperature

[0081] Place Test method time Temperature (℃) Beijing Outdoor exposure 2019 / 4 / 1713:00 23.9 Mohe Outdoor exposure 2019 / 4 / 1713:00 1.29 Dunhuang Outdoor exposure 2019 / 4 / 1713:00 24.9 Lhasa Outdoor exposure 2019 / 4 / 1713:00 13.7 Xishuangbanna Outdoor exposure 2019 / 4 / 1713:00 33.3 Wanning Coastal Outdoors 2019 / 4 / 1713:00 29.1 Wanning Coastal outdoor 2019 / 4 / 1713:00 29.6 Beijing Outdoor exposure 2019 / 4 / 1712:00 21.8 Mohe Outdoor exposure 2019 / 4 / 1712:00 0.9 Dunhuang Outdoor exposure 2019 / 4 / 1712:00 22.2 Lhasa Outdoor exposure 2019 / 4 / 1712:00 11.9 Xishuangbanna Outdoor exposure 2019 / 4 / 1712:00 30.6 Wanning Coastal Outdoors 2019 / 4 / 1712:00 28.9 Wanning Coastal outdoor 2019 / 4 / 1712:00 29.5

[0082] Table 3 Raw data of test station - humidity

[0083] Place Test method time relative humidity Wanning Coastal outdoor 2019 / 5 / 174:00 87 Wanning Coastal Outdoors 2019 / 5 / 174:00 88 Beijing Outdoor exposure 2019 / 5 / 173:00 57 Mohe Outdoor exposure 2019 / 5 / 173:00 100 Lhasa Outdoor exposure 2019 / 5 / 173:00 75 Jiangjin Outdoor exposure 2019 / 5 / 173:00 98 Dunhuang Outdoor exposure 2019 / 5 / 173:00 25 Xishuangbanna Outdoor exposure 2019 / 5 / 173:00 81 Wanning Coastal outdoor 2019 / 5 / 173:00 87 Wanning Coastal Outdoors 2019 / 5 / 173:00 87 Beijing Outdoor exposure 2019 / 5 / 172:00 49 Mohe Outdoor exposure 2019 / 5 / 172:00 100 Lhasa Outdoor exposure 2019 / 5 / 172:00 79 Jiangjin Outdoor exposure 2019 / 5 / 172:00 97 Dunhuang Outdoor exposure 2019 / 5 / 172:00 27

[0084] The “station” in Table 1 is the name of the meteorological station; the “location” in Tables 2 and 3 is the name of the experimental station; the time in Tables 1-3 is the time of data collection.

[0085] After the data is acquired, the multi-source data of the test site and the weather station environmental factor data are aligned in granularity, where the granularity includes at least dimensions such as site granularity, time zone granularity, and time granularity.

[0086] (1) Site granularity. During the research process, given that the experimental stations and meteorological stations were established by different institutions, the experimental station data were generated by the station network management unit, and the meteorological station data used ISH data. In addition, their specific locations were different, resulting in different coding and naming between the stations. In this context, considering that the environmental factors of experimental stations and meteorological stations with adjacent longitude and latitude are highly consistent, adjacent stations were screened and the closest experimental stations and meteorological stations were selected as research objects. The site mapping is shown in Table 4.

[0087] Table 4 Weather station-experimental station site mapping

[0088]

[0089] In Table 4, “Station Name” refers to the name of the meteorological station; “Longitude / Latitude” refers to the location of the meteorological station; “Station Network” refers to the source of the experimental station data; “Site” refers to the experimental station site.

[0090] (2) Time zone granularity. The weather station data uses Universal Time as the benchmark, while the experimental station data uses Beijing Time, that is, Beijing Time (GMT+8). The time zone is unified to the GMT+8 time zone, and the time of all weather station data is increased by eight hours.

[0091] (3) Time granularity. Given that the data recorded by multiple source sites and equipment have different time resolutions, such as daily, three-hour, hourly, and minute-level data, the lowest time resolution, namely daily data, was selected as the reference benchmark. On this basis, the time series of the data from the experimental station and the meteorological station were synchronized to ensure their temporal consistency. For the temperature and humidity data from the meteorological station, the time was directly aligned and merged together.

[0092] (4) The environmental factor data of the test sites and meteorological stations were integrated into one table, as shown in Table 5. There were cases where data from some test stations were missing.

[0093] Table 5 Weather station-experimental station integrated data set

[0094]

[0095]

[0096] In Table 5, “experimental station” refers to the experimental station; “experimental station humidity” and “weather station humidity” both refer to relative humidity; blank positions in Table 5 indicate missing data.

[0097] Experimental station data[v1,v2,...,v n ]、Weather station data[u1,u2,...,u n ] are regarded as two sequences, and the differences between the data of the experimental station and the meteorological station are calculated based on indicators such as mean absolute error (MAE), mean squared error (MSE), maximum deviation, and correlation coefficient, which serve as the benchmark for subsequent mapping model performance evaluation and data filling feature comparison.

[0098] Table 6 Analysis of weather station-test station data characteristics

[0099]

[0100] In this embodiment, a line graph of experimental station data and weather station data is drawn (taking Lhasa as an example, Figure 3 As shown, Figure 3 The horizontal axis in the figure corresponds to different experimental station data or meteorological station data). As a posteriori test of the aforementioned granularity alignment, the data alignment is visually checked. The temperature and humidity alignment of each station is good, and there is no offset.

[0101] Next, the detailed construction process of the dual-input and dual-output environmental factor mapping model based on LSTM is as follows:

[0102] (1) Data cutting. Experimental station data [v1, v2, ..., v n ]、Weather station data[u1,u2,...,u n ] Two long sequences are cut into multiple sequences of equal length as input. The sequence length k should cover 24 hours. The data interval is three hours, so k is 8. The data cutting principle is to ensure that the model can learn the periodicity with a daily cycle.

[0103] (2) Data normalization and denormalization. The mean and standard deviation of the experimental station data and the weather station data are recorded separately. The data needs to be normalized before model training to ensure rapid convergence of the model training process. During the prediction process, the data also needs to be restored to a normal scale through the mean and standard deviation.

[0104] (3) Determination of model structure. The model structure adopts a dual-input dual-output structure, with two sequences of weather station data and experimental station data as input, jointly predicting the missing data of the experimental station. At the same time, an additional task is added. Since the weather station data is complete, the known data of the predicted weather station is added to correct the weight of the weather station prediction branch to promote mutual learning. If there is only dual input and single output, the weight of the weather station data will be corrected by the experimental station data during the training process, which is unscientific. Therefore, this branch is added to correct this part of the weight. Therefore, this method can avoid the model only using the experimental station data to correct the weight, resulting in overfitting.

[0105] (4) Hyperparameter setting. First, we sort out the amount of data after cutting. Since the data of the experimental stations are incomplete, the amount of data at different stations varies. The specific data amounts are shown in Table 7.

[0106] Table 7 Statistics of data volume after data cutting

[0107] Experimental Station Test method Temperature series data volume Humidity series data volume Wanning Platform exposure site 8371 8118 Wanning Coastal Outdoors 12675 12664 Wanning Coastal outdoor 12822 12828 Beijing Outdoor exposure 12399 13180 Lhasa Outdoor exposure 18318 17484 Dunhuang Outdoor exposure 18349 17817 Jiangjin Outdoor exposure 18137 20970 Mohe Outdoor exposure 17358 17346 Xishuangbanna Outdoor exposure 11424 13519

[0108] In this example, a three-layer LSTM is used, with the number of hidden units in each layer ranging from 4 to 32. Because the model employs a dual-output structure, weights must be set for both. For example, the weight for the experimental station data prediction can be set to 1, while the weight for the weather station data prediction can be set to 0.2 to avoid excessive weight bias toward the weather station data prediction. Training and testing have demonstrated prediction accuracy exceeding 95%.

[0109] (5) Data randomization and cross-validation. The random operation disrupts the data order, which only disrupts the order of the input data, and does not disrupt the order of the data within each input sequence. For example, the existing input data [v1,v2,...,v8], [v2,v3,...,v9], [v3,v4,...,v 10 ]…, by scrambling the order to get [v 50 ,v 51 ,...,v 57 ]、[v 199 ,v 200 ,...,v 206 ], [v1, v2, ..., v8]…, using data randomness to ensure the model learns correlation features rather than just periodic features. At the same time, combined with a cross-validation strategy, we obtain a model with minimal deviation and achieve stable results.

[0110] (6) Model evaluation. The dataset was divided into 7:2:1, 70% was used as the training set to train the model, 20% was used as the validation set to measure the quality of the model during the training process, and the last 10% was used as the test set. MAE, MSE, and correlation coefficient R were used on the test set. 2 , maximum deviation and other indicators to evaluate the prediction effect of the model.

[0111] In the actual model construction process, the missing value may be before the known data or after the known data, so in addition to distinguishing the site, temperature and humidity, the model also needs to distinguish the direction. k ,v k+1 ,...,v k+7 ]、Weather station data[u k ,u k+1 ,...,u k+7 ] is input, predict v k+8 , when making reverse predictions, the experimental station data [v k+7 ,v k+6 ,...,v k ]、Weather station data[u k+7 ,u k+6 ,...,u k ] is input, predict v k-1 .

[0112] In this example, the data volume in Table 7 is selected for prediction. The specific prediction results are shown in Table 8.

[0113] Table 8 Model prediction results

[0114]

[0115] The model obtained through this training fills in the missing data of the experimental stations to form a complete data set.

[0116] When performing model prediction to fill the data set, if there are multiple consecutive missing data, the sliding window strategy is used to fill the missing data one by one. As shown in Table 9, the data of the Wanning platform exposure field on January 1, 2020 is missing. Based on the Wanning-platform-temperature-forward prediction model, the temperature data series of the experimental station [23.8, 23.8, 22.4, 23.7, 24, 23.5, 23.3, 23.2] and the temperature data series of the meteorological station [22.2, 21.7, 21.8, 23.1, 23.8, 22.5, 21.9, 22. The temperature at 2:00 on January 1st is predicted to be 22.76. The temperature data of the experimental station [23.8, 22.4, 23.7, 24, 23.5, 23.3, 23.2, 22.76] and the temperature data sequence of the weather station [21.7, 21.8, 23.1, 23.8, 22.5, 21.9, 22, 21.2] are used to predict the temperature at 5:00 on January 1, 2020. The predicted result is 22.48, and so on.

[0117] Next, the statistical characteristics of the experimental station data and the meteorological station data after data filling were calculated, as shown in Table 10. A joint analysis of the results in Table 10 and the characteristics before data filling (see Table 6) shows that the characteristics of Wanning coastal temperature, Beijing temperature, and Beijing humidity change significantly after data filling. In theory, the data characteristics before and after data filling should remain basically consistent. If the characteristics differ significantly, it is because of long-term data loss. Using a sliding window strategy for prediction filling will result in errors and require model correction.

[0118] Table 10 Feature analysis after data filling

[0119]

[0120] Consider retraining the model, that is, repeating the model training steps to address the above problem by introducing new randomness. After testing, the problem is basically alleviated after retraining, as shown in Table 11.

[0121] Table 11 Customization analysis of model features after retraining

[0122]

[0123] All data filling is completed based on the above model to generate a test station dataset with no missing items.

[0124] Next, several methods for fine-grained characterization of wetting time were proposed. From the time dimension, it can be divided into annual wetting time and monthly wetting time. From the corrosion mechanism dimension, it can be divided into the first critical humidity / temperature and the second critical humidity / temperature.

[0125] The wetting time is defined as the duration of exposure to temperatures exceeding a certain humidity level. The complete dataset obtained through the above steps calculates the wetting time Tw (which can also be considered the number of data points, i.e., the data volume). The dataset interval δ is 3 hours, calculated as T = δ·Tw. Furthermore, the corrosion rate of metals increases significantly with increasing relative humidity. This phenomenon is often closely related to the so-called "critical relative humidity." Specifically, when the relative humidity exceeds a certain threshold, electrochemical reactions on the metal surface accelerate, leading to a sharp increase in the corrosion rate. The first critical relative humidity is approximately 50%. Under this condition, redox reactions on the metal surface are promoted, significantly exacerbating the corrosion rate. The second critical relative humidity corresponds to more stringent environmental conditions, characterized by a specific relative humidity value (e.g., 80%, which may vary depending on the metal type; see Table 12) and a temperature above 0°C. At this stage, the corrosion rate of the metal increases significantly, indicating that the influence of humidity on the corrosion process has reached a critical point.

[0126] Table 12 Second critical humidity of different metals

[0127]

[0128] When refining the wetting time characterization, we first identify the metal type. For example, when studying the wetting time associated with aluminum alloys, we select 80% as the second critical humidity. Combined with the resulting dataset, for example, in February 2018, Lhasa recorded three points with temperatures exceeding 0°C and relative humidity exceeding 50%. Therefore, the monthly wetting time at the first critical humidity is 9 hours. The resulting refined wetting time characterization is shown in Tables 13 and 14.

[0129] Table 13 Characterization of monthly wetting time of aluminum alloy

[0130]

[0131] Table 14 Characterization of annual wetting time of aluminum alloy

[0132]

[0133] The “first critical humidity wetting time” and the “second critical humidity wetting time” in Table 13 and Table 14 are also referred to as the “first wetting time” and the “second wetting time”.

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for filling missing data of experimental stations based on environmental factor mapping model, characterized in that: include: S1: Acquire test site environmental factor data and weather station environmental factor data, wherein the test site environmental factor data includes data of various environmental factors under different test modes collected at a preset test site, and the weather station environmental data includes data of various environmental factors at the weather station, and the environmental factor data includes data corresponding to at least two environmental factors: temperature and humidity; S2: Perform granular alignment on the test site environmental factor data and the weather station environmental factor data, where granular alignment includes site granularity alignment, time zone granularity alignment, and time granularity alignment. The aligned test site environmental factor data and weather station environmental factor data are then integrated. The integrated test site environmental factor data all have corresponding weather station environmental factor data. However, some weather station environmental factor data after integration may not have corresponding test site environmental factor data, i.e., there may be missing test site environmental factor data. S3: Constructing an environmental factor mapping model, wherein the environmental factor mapping model is a dual-input dual-output model based on LSTM, wherein one input of the environmental factor mapping model is a time series consisting of the environmental factor data of the test station, and the other input is a time series consisting of the environmental factor data of the corresponding weather station; one output is the predicted environmental factor data of the test station at other times, and the other output is the predicted environmental factor data of the weather station at other times; S4: Use the integrated results in S2 to train the environmental factor mapping model and fill in the missing environmental factor data of the experimental sites.

2. The data completion method based on the environmental factor mapping model according to claim 1, characterized in that: The environmental factor mapping model is a forward prediction model or a reverse prediction model. The forward prediction model is used to predict the moment after the latest moment in the input time series, and the reverse prediction model is used to predict the moment before the earliest moment in the input time series.

3. The data completion method based on the environmental factor mapping model according to claim 1, characterized in that: The method further comprises: S5: Perform feature analysis on the environmental factor data of the experimental station and the environmental factor data of the meteorological station after integration in S2 to generate a first statistical feature to evaluate the difference in environmental factors between the experimental station and the meteorological station; S6: After S4, feature analysis is performed using the completed test site environmental factor data and the meteorological station environmental factor data to generate a second statistical feature; S7: Compare the second statistical feature with the first statistical feature, and determine whether the data features before and after the completion are consistent based on the comparison result; if they are consistent, the training of the environmental factor mapping model is completed; if they are inconsistent, return to S4, and re-train the environmental factor mapping model by introducing the randomness of the test site environmental factor data and the meteorological station environmental factor data, and use the re-trained environmental factor mapping model to re-complete the test site environmental factor data with missing data until the data before and after completion are consistent. The environmental factor mapping model obtained by this training is the final environmental factor mapping model.

4. The data completion method based on the environmental factor mapping model according to claim 3 is characterized in that: The first statistical feature and the second statistical feature include one or more combinations of mean absolute error, mean square error, maximum deviation, minimum deviation, correlation coefficient, etc. between the test station environmental factor data and the weather station environmental factor data.

5. The data completion method based on the environmental factor mapping model according to claim 1, characterized in that: The environmental factor mapping model is a dual-input dual-output model based on LSTM, wherein the dual-input dual-output model based on LSTM includes two branches. The first branch uses the experimental station data and the meteorological station data as input to form a strong coupling relationship and predict the missing experimental station environmental factor data; The second branch uses the weather station data as input to predict the weather station data at other times; The first branch is the main method for achieving prediction, and the second branch is the self-iteration of weather station data, which is used to correct the weight of the second branch to improve the prediction accuracy of the first branch.

6. A method for fine characterization of wetting time, characterized in that: include: Obtaining the completed test site environmental factor data generated according to any one of claims 1 to 6; Based on the completed test site environmental factor data, count the amount of first data in which the temperature exceeds a first temperature threshold and the humidity exceeds a first humidity threshold within a time period corresponding to a preset time dimension, and the amount of second data in which the temperature exceeds a second temperature threshold and the humidity exceeds a second humidity threshold within a time period corresponding to a different time dimension, where the first humidity threshold and the second humidity threshold are different; determining a first wetting time corresponding to a preset time dimension based on the first data volume and a time interval corresponding to data in the test site environmental factor data; Determine a second wetting time corresponding to a preset time dimension based on the second data volume and the time interval corresponding to the data in the test site environmental factor data, Through different time dimensions and different wetting times, a refined characterization of wetting time is achieved.

7. A method for fine characterization of wetting time according to claim 7, characterized in that: The preset time dimensions include the year time dimension and the month time dimension.

8. A method for fine characterization of wetting time according to claim 7, characterized in that: The first humidity threshold, the first temperature threshold, the second humidity threshold and / or the second temperature threshold may be adjusted according to different test subjects.

Images