High-pile wharf durability monitoring data prediction analysis method based on GABP neural network

Abstract

The invention relates to the technical field of structural health monitoring, and discloses a GABP neural network-based high-pile wharf durability monitoring data prediction analysis method, which comprises the following steps of: acquiring multi-parameter monitoring data acquired by a high-pile wharf anode ladder sensor system; correlation analysis is carried out, and input features used for predicting predetermined target parameters are selected; preprocessing the selected input feature data and the measured data of the corresponding predetermined target parameters; constructing and training a GABP neural network prediction model, and optimizing an initial weight and a threshold value of a BP neural network through a genetic algorithm; and using the trained GABP neural network prediction model to input new monitoring data to predict predetermined target parameters. According to the method, the BP network is optimized through the GA, the accuracy of durability monitoring data prediction and the generalization ability of the model are remarkably improved, and a reliable technical means is provided for accurate evaluation and early warning of the durability state of the long-piled wharf.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of structural health monitoring, in particular to a high-pile wharf durability monitoring data prediction and analysis method based on a GABP neural network. BACKGROUND

[0002] As important port infrastructure, high-pile wharfs serve in harsh marine environments for a long time, and the concrete structures thereof are inevitably subjected to the continuous effects of multiple factors such as chloride ion erosion, carbonation, freeze-thaw cycles and marine bio-attachment. These factors collectively accelerate the corrosion process of the internal steel bars of the concrete, thereby weakening the bearing capacity of the structure, shortening the service life thereof, and in severe cases, even causing structural failure, causing huge economic losses and safety hazards. Therefore, long-term and effective durability monitoring of high-pile wharf structures, accurate assessment of the health state thereof and scientific prediction of the future performance degradation trend have important practical significance for guaranteeing the safe operation of the wharf, guiding maintenance and reinforcement decisions and prolonging the service life of the structure.

[0003] At present, various technical means have been developed for the durability monitoring of concrete structures such as high-pile wharfs. Among them, the electrochemical monitoring technology based on embedded anodic ladder sensors has been widely applied because it can reflect the corrosion activity state of the environment in which the internal steel bars of the concrete are located in real time and in situ. By collecting parameters such as the potential, current and environmental temperature of the anode rod, key durability indicators such as the chloride ion penetration depth and the steel bar corrosion rate can be indirectly inferred. However, these monitoring systems often generate massive, multi-dimensional time series data. How to effectively extract key information representing the durability state of the structure from these complex data and establish an accurate prediction model is a major challenge currently faced by the field.

[0004] Traditional durability data analysis methods, such as inference based on empirical formulas or simple statistical regression analysis, often fail to fully capture the complex nonlinear degradation process under the coupling action of multiple factors, resulting in low prediction accuracy and difficulty in meeting the needs of refined operation and maintenance management. In recent years, machine learning methods, especially neural network technology, have shown great potential in the field of structural health monitoring data analysis due to their strong nonlinear mapping ability and self-learning ability. For example, BP (back propagation) neural networks have been tried for structural response prediction.

[0005] Although BP neural network has advantages in dealing with nonlinear problems, it also has some inherent defects. BP algorithm is essentially a gradient descent method, and its training results are highly dependent on the selection of network initial weights and thresholds. If the initial parameters are not properly selected, the network may easily fall into local optimal solution in the training process, resulting in slow learning speed, and the generalization ability and prediction accuracy of the final model may not be ideal. How to effectively optimize the initial parameters of BP neural network, improve its global search ability, and obtain a more stable and accurate prediction model is the key problem to be solved in current research, and is also the technical bottleneck to further improve the prediction and analysis level of high-pile wharf durability monitoring data. SUMMARY

[0006] The technical problem to be solved by the present application is that in the existing analysis method of high-pile wharf durability monitoring data, directly interpreting the monitoring data affected by various environmental factors is easy to produce misjudgment, and the traditional prediction model is easy to fall into local optimum when dealing with long-term monitoring data with complex nonlinear relationship, resulting in low prediction accuracy and limited generalization ability, which is difficult to fully meet the needs of accurate evaluation and early warning of high-pile wharf structure health state.

[0007] To solve the above technical problems, the present application provides a high-pile wharf durability monitoring data prediction and analysis method based on GABP neural network, which aims to optimize the construction and training process of neural network model, and significantly improve the accuracy and reliability of the prediction of key parameters of high-pile wharf durability.

[0008] The technical scheme provided by the present application specifically includes the following steps:

[0009] First, obtain the high-pile wharf durability monitoring data. An anode ladder sensor system is pre-installed at the key structural parts of the high-pile wharf, which can monitor and record various parameters related to structural durability. The collected data at least covers the voltage values between each anode rod (such as anode rods A1 to A6) and the cathode, reference electrode, and internal steel bars, the current values between each anode rod and the cathode and internal steel bars, and the environmental temperature values. These data are the basis for subsequent analysis and prediction.

[0010] The second step is to conduct correlation analysis of monitoring data and select input features. In order to improve the efficiency and accuracy of the prediction model, it is necessary to screen out the factors with the strongest correlation with the target parameter to be predicted (usually the voltage value or current value of a specific anode rod that characterizes the durability status of the dock) from the many monitoring parameters obtained. This step uses the Pearson correlation coefficient method to calculate the correlation coefficient between each monitoring parameter and between each monitoring parameter and the predetermined target parameter. By analyzing the size of the correlation coefficient, the main factors affecting the change of the target parameter are identified. For example, when predicting the voltage value of a specific anode rod (such as A1), other parameters that show a strong correlation with the voltage value can be selected, such as its corresponding current value, the voltage value between the anode rod and the reference electrode, the voltage value between the anode rod and the internal steel bar, the current value between the anode rod and the internal steel bar, and the current ambient temperature value, which together constitute the input feature set of the neural network model.

[0011] The third step is to preprocess the monitoring data. In order to eliminate the adverse effects of different physical dimensions on model training and accelerate model convergence, it is necessary to standardize the screened input feature data and the corresponding target parameter measured data. First, the data set is divided into a training set and a test set. For example, the first 80% of the data can be selected as the training set according to the time series, and the last 20% can be selected as the test set. Then, the maximum and minimum normalization method is used for the data in the training set and the test set, and it is linearly mapped to a preset numerical interval (such as [0,1]). The normalization formula can be:

[0012]

[0013] Among them, X i is the original data value, X min and X max are the minimum and maximum values ​​in the original data set, Y min and Y max are the lower and upper bounds of the target normalization interval (e.g., 0 and 1).

[0014] The fourth step is to build and train the GABP neural network prediction model. This step is the core innovation of the present invention.

[0015] First, design a basic BP (back propagation) neural network structure, including an input layer, at least one hidden layer, and an output layer. The number of neurons in the input layer is consistent with the number of input features selected in the second step, and the number of neurons in the output layer is usually 1 (corresponding to a predetermined target parameter). The number of layers and neurons in the hidden layer can be calculated according to empirical formulas (such as Where n is the number of input units, m is the number of output units, and a is an empirical constant) is initially determined and then fine-tuned through experimentation. The activation function for the hidden layer neurons can be a Sigmoid function or other suitable nonlinear function.

[0016] Secondly, and also the key, the genetic algorithm (GA) is introduced to globally optimize the initial connection weights and neuron thresholds of the BP neural network. The traditional BP neural network uses gradient descent method to update the weights and thresholds, which is easy to fall into local minimum. Genetic algorithm has strong global search ability by simulating the selection, crossover and mutation operations in the biological evolution process. In this method, the initial weights and thresholds of the BP neural network are encoded as individuals (chromosomes) of the genetic algorithm, and the initial population size, the maximum evolution (iteration) number, the crossover probability and the mutation probability, etc. are set as the genetic algorithm running parameters. The minimum prediction error (such as mean square error) of the BP neural network on the training set is taken as the target, or a fitness function reflecting the prediction accuracy is designed, and through the iterative optimization of the genetic algorithm, a set of weights and thresholds combination that can make the initial performance of the BP neural network best is found.

[0017] Finally, the best initial weights and thresholds optimized by the genetic algorithm are assigned to the BP neural network, and then the preprocessed training set data is used to further train and fine-tune the BP neural network through the standard back propagation algorithm until the error between the predicted value and the true value of the network output reaches the preset accuracy requirement or the preset training period is completed, thereby obtaining the final GABP neural network prediction model.

[0018] Fifthly, the trained GABP model is applied for durability monitoring data prediction. The preprocessed test set data or new real-time monitoring data (after the same input feature selection and preprocessing as the training data) is input into the trained GABP neural network model, and the model can output the prediction result of the predetermined target parameter (such as the future voltage or current value of a specific anode rod). If the prediction result is a normalized value, it needs to be denormalized to restore it to the original data scale.

[0019] Through the above steps, the global optimization ability of the genetic algorithm is used to improve the initial parameter setting of the BP neural network, effectively avoiding the defect that the traditional BP network is easy to fall into local optimum, thereby improving the prediction accuracy of the high-pile wharf durability monitoring data and the generalization ability of the model. The prediction result can provide more reliable data support for evaluating the current corrosion state of the wharf structure, predicting the future durability trend and formulating a scientific maintenance strategy.

[0020] The present application provides a GABP neural network-based high-pile wharf durability monitoring data prediction and analysis method, which has the following beneficial effects:

[0021] 1. The application adopts genetic algorithm (GA) to globally optimize the initial connection weight and neuron threshold of BP neural network when constructing the neural network model, effectively avoids the problem that the traditional BP neural network is easy to fall into local optimal solution in the training process, thereby significantly improves the precision of the prediction result of the key durability parameters such as the anode rod voltage and current of the high-pile wharf, and makes the prediction value closer to the real monitoring value.

[0022] 2. The GABP neural network model optimized by the genetic algorithm has more reasonable initial parameter setting, and when facing new monitoring data that has not participated in the training, can still maintain a high prediction accuracy. This shows that the model constructed by the method has better stability and adaptability to different working condition data.

[0023] 3. The application adopts Pearson correlation coefficient for input feature selection, ensures that the model input is highly correlated with the target prediction parameter, and in combination with the high-precision prediction of the GABP model, can more accurately reflect the current corrosion state and future development trend of the structure. This provides a scientific basis for timely discovering potential durability risks.

[0024] 4. The introduction of the genetic algorithm not only optimizes the initial parameters of the BP neural network, but also indirectly improves the convergence speed and final performance in the subsequent BP algorithm fine-tuning stage. By finding a better starting point, the possibility of the BP algorithm falling into a poor local optimum or taking too long to train is reduced, making the training process of the entire model more efficient. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is a method flowchart of the application;

[0026] Figure 2 is a schematic diagram of the anode ladder sensor installation position of the application;

[0027] Figure 3 is a schematic diagram of the anode ladder sensor of the application;

[0028] Figure 4 is a schematic diagram of the BP neural network structure of the application;

[0029] Figure 5 is a schematic diagram of the GABP neural network structure of the application;

[0030] Figure 6 is a comparison chart of the anode rod A1 voltage and current prediction values predicted by the BP, PSO-BP and GABP three models of the application; wherein (a) is the anode rod A1 voltage prediction value, and (b) is the anode rod A1 current prediction value;

[0031] Figure 7Figure 2 is a comparative diagram of errors in predicting the voltage and current of the anode rod A1 using the BP, PSO-BP and GABP three models of the application; wherein (a) is the error value of the anode rod A1 voltage prediction, and (b) is the error value of the anode rod A1 current prediction. DETAILED DESCRIPTION

[0032] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described clearly and completely below in combination with specific embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0033] Please refer to the accompanying Figure 1 - the accompanying Figure 7 The present application provides a kind of based on GABP neural network's high-pile wharf durability monitoring data prediction analysis method.

[0034] As Figure 1 Indicated, in one preferred embodiment, in order to realize the accurate prediction and analysis of high-pile wharf durability monitoring data, the method can include the following steps:

[0035] Step S1: systematic collection and standardization preparation of high-pile wharf durability monitoring data;

[0036] This step is mainly related to deploying anode ladder sensor system on target high-pile wharf, and continuously collecting multi-dimensional original monitoring data including anode rod voltage, current and environmental temperature, then preliminarily arranging, checking and structuring storing these data, laying data foundation for subsequent analysis.

[0037] Step S2: statistical monitoring data correlation depth analysis and core prediction feature screening;

[0038] This step aims to quantify the linear correlation strength between each monitoring parameter and between each monitoring parameter and the predetermined prediction target (such as the voltage or current of a specific anode rod) by statistical analysis methods such as Pearson correlation coefficient, so as to identify and screen out key input features that have significant influence on the prediction target, and optimize the effectiveness of model input.

[0039] Step S3: neural network model-oriented training and test data set construction and standardized preprocessing;

[0040] This step includes dividing the historical monitoring data into training and testing sets according to a certain proportion (such as 80% training and 20% testing), and performing normalization operations on the selected input feature data and corresponding target parameter data, such as linear mapping to the [0, 1] interval, to eliminate dimensional differences and improve model training efficiency and stability.

[0041] Step S4: Construction and iterative training of the genetic algorithm optimized BP neural network (GABP) prediction model

[0042] This is the core step. First, the topology of the basic BP neural network is designed (including the number of neurons in the input layer, hidden layer, and output layer, and the activation function, etc.), and then the key operating parameters of the genetic algorithm (GA) are set (such as population size, evolution generation, crossover rate, mutation rate, etc.). The global search capability of GA is used to optimize the initial connection weights and neuron thresholds of the BP neural network. Finally, based on the optimized initial parameters and training set data, the network is supervised trained through the back propagation algorithm until the model converges or the preset training target is reached, forming the GABP prediction model.

[0043] Step S5: Comprehensive evaluation of the performance of the GABP prediction model and verification of the generalization ability in actual application

[0044] This step involves using the reserved test set data to test the prediction effect of the trained GABP model. By calculating and analyzing multiple evaluation indicators such as root mean square error (RMSE), mean absolute error (MAE), and determination coefficient (R2), the prediction accuracy and goodness of fit of the model are comprehensively evaluated, and the generalization performance and actual application value of the model on different data sets can be verified by applying the model to data prediction of other similar monitoring points.

[0045] In the following, the specific implementation details of each of the above steps S1 to S5, the technical means used and their principles will be described in more detail.

[0046] Step S1: Systematic collection and standardization of high-pile wharf durability monitoring data

[0047] The core of this step is to establish a stable and reliable data source to provide high-quality raw data for subsequent prediction analysis.

[0048] S1.1 Planning and deployment of sensor system

[0049] In order to effectively monitor the durability of the concrete structure of the high-pile wharf, especially the depth of chloride ion erosion and the risk of steel corrosion, the anode ladder sensor system is selected for monitoring in this embodiment.

[0050] Sensor composition and layout principle:

[0051] The anode ladder sensor system is mainly composed of a sensor body embedded in the concrete structure and an external data acquisition and test interface. The sensor body at least includes:

[0052] Multiple anode rods: In this embodiment, each anode ladder contains 6 anode rods made of carbon steel, numbered A1, A2, A3, A4, A5, and A6. These anode rods are embedded in different preset depths of the concrete protective layer by adjusting the inclination angle of the specially designed support, so as to monitor the electrochemical state of concrete at different depths.

[0053] Cathode: As a corresponding electrode in the macro-battery corrosion monitoring system.

[0054] Reference electrode: used to provide a stable potential reference for measuring the potential of the anode rod.

[0055] Counter reinforcement (CR): can be used as another reference object to monitor the electrochemical parameters between the anode rod and the actual structure reinforcement.

[0056] Temperature probe: In this embodiment, PT1000 type temperature probe is used to monitor the environmental temperature at the location of the sensor in real time, because temperature is an important factor affecting the electrochemical reaction rate and monitoring data.

[0057] The basic principle of its monitoring is that as the external corrosive medium (such as chloride ions) gradually penetrates into the concrete, the anode rods at different depths will reach the critical condition of rusting in turn. By measuring the potential, current and other parameters of these anode rods over time, the depth of chloride ion reaching can be determined, and the time of steel bar starting to rust can be evaluated.

[0058] Sensor specific layout position:

[0059] In this example, a high-pile wharf in Tianjin Port is located in the Nangang Port Area of Tianjin Port and mainly undertakes bulk cargo handling operations. The wharf structure is designed for 300,000 DWT bulk carriers. The total length of the wharf is 390 meters and the width is 75 meters. The project consists of a main wharf platform and two lateral approach bridges, with the main platform being 390 meters long and each approach bridge being 73.3 meters long. The main wharf platform is divided into 9 structural sections, each 65 meters long, and further subdivided into front and rear pile caps. The front pile cap is 36.5 meters wide, while the rear pile cap is slightly wider at 38.5 meters. In the front pile cap section, each structural section consists of 17 row frames, each containing 9 steel pipe piles, including 4 forked piles and 5 straight piles. In the rear pile cap section, each row frame consists of 9 prestressed reinforced concrete square piles with dimensions of 650 mm x 650 mm. The design load distribution of the wharf is segmented, with a uniform load of 30 kiloPascals (kPa) for the first 18.5 meters from the wharf front, a load of 50 kPa for the next 18 meters (from 18.5 meters to 36.5 meters), and a load of 80 kPa from 36.5 meters to the end of the wharf (75 meters). In addition, the wharf is equipped with a structural health monitoring system, which is installed on the second structural section. The specific arrangement of the durability sensors on the wharf is indicated by arrows in Figure 2 .

[0060] The wharf initially used an anode ladder sensor made in Germany for structural durability monitoring. This system consists of several key components: an anode ladder embedded in the concrete, an external test interface, and specialized data acquisition equipment. The anode ladder sensor itself consists of several components, including an anode rod, a cathode, a reference electrode, a anode steel rod (CR), and a temperature probe (PT1000). The anode ladder is composed of 6 anodes made of carbon steel, labeled A1-A6. These anode rods are embedded at different depths of the concrete cover, achieved by adjusting the inclination angle of the bracket. The specific installation method and side view are shown in Figure 3 .

[0061] S1.2 Continuous acquisition of original monitoring data

[0062] After the main structure of the wharf is completed and the monitoring conditions are met, the data acquisition of the anode ladder sensor is started.

[0063] Data acquisition start and cycle:

[0064] Taking the Tianjin Port Terminal as an example, after the main structure construction was completed in 2017, the data collection of the anode ladder sensor was officially started in January 2018, and long-term and continuous monitoring was carried out. In the research background of the present application, continuous monitoring data for 4 years has been accumulated. The frequency of data collection can be once a day, several times or other periods set according to actual needs, to ensure the continuity and effectiveness of the data.

[0065] Key parameters collected:

[0066] In each collection period, the system automatically or through manual inspection collects and records up to 40 different parameters. For the subsequent prediction analysis of the present application, the following core electrochemical parameters and environmental parameters related to each anode rod (A1-A6) are focused on:

[0067] Voltage value between anode rod and cathode (U): for example, U1 is the voltage between anode rod A1 and cathode, and so on to U6.

[0068] Voltage value between anode rod and reference electrode (U'): for example, U1' is the voltage between anode rod A1 and reference electrode, and so on to U6'.

[0069] Voltage value between anode rod and internal steel bar (U"): for example, U1" is the voltage between anode rod A1 and internal steel bar, and so on to U6".

[0070] Current value between anode rod and cathode (I): for example, I1 is the current between anode rod A1 and cathode, and so on to I6.

[0071] Current value between anode rod and internal steel bar (I'): for example, I1' is the current between anode rod A1 and internal steel bar, and so on to I6'.

[0072] Environmental temperature value (T): directly measured by PT1000 temperature probe.

[0073] The typical units of these parameters are: the units of voltage (U, U', U") are usually millivolts (mV), the units of current (I, I') are usually microamperes (pA), and the unit of temperature (T) is Celsius (°C).

[0074] Through the systematic implementation of step S1, a set of time series long, parameter dimension rich, and reliable quality high-pile wharf durability original monitoring data set can be obtained. This is indispensable for subsequent correlation analysis, feature selection, model training and prediction verification.

[0075] Step S2: Statistical monitoring data correlation depth analysis and core prediction feature selection

[0076] After obtaining a large amount of original monitoring data, not all parameters have equal importance on the specific durability indicators of interest (such as the voltage or current of a certain anode rod). In order to build a more efficient and accurate prediction model, it is necessary to perform correlation analysis on the data and select the input features most relevant to the prediction target. The core of this step is to quantify the relationship between different monitoring parameters and the strength of their association with the prediction target.

[0077] S2.1 Determination of target prediction parameters

[0078] First of all, it is necessary to clarify the core target of this prediction analysis. In the durability assessment of high-pile wharfs, the electrochemical parameters of each anode rod in the anode ladder sensor are the key to determining whether the steel reinforcement has started to corrode or is in an activated state. Therefore, in this embodiment, the following two types of parameters of a specific anode rod in the anode ladder are selected as the main predetermined target parameters:

[0079] Voltage between anode rod and cathode (U series): For example, the voltage U1 between anode rod A1 and cathode, the voltage U2 between anode rod A2 and cathode, etc. The change in voltage value can reflect the electrochemical state of the concrete environment in which the anode rod is located.

[0080] Current between anode rod and cathode (I series): For example, the current I1 between anode rod A1 and cathode, the current I2 between anode rod A2 and cathode, etc. The size of the macrocell corrosion current is directly related to the corrosion rate of the steel reinforcement.

[0081] In some specific analysis scenarios, the voltage between anode rod and reference electrode (U' series) or the voltage / current between anode rod and internal steel reinforcement (U'' / I' series) can also be selected as the prediction target, depending on the focus of the study. This embodiment mainly describes the U series and I series parameters as prediction targets.

[0082] S2.2 Data correlation calculation and analysis

[0083] In order to quantitatively evaluate the linear correlation degree between each monitoring parameter, this embodiment uses the Pearson Correlation Coefficient (Pearson Correlation Coefficient) for calculation. The Pearson Correlation Coefficient r is a statistical measure of the strength and direction of linear correlation between two continuous variables, and its calculation formula is as follows:

[0084]

[0085] Where:

[0086] n is the number of sample points (i.e. the number of monitoring data).

[0087] x i and yi respectively are the observation values of two different monitoring parameters at the i-th time point.

[0088] and respectively are the average values of the two monitoring parameters at all observation time points.

[0089] The value range of the correlation coefficient r is between -1 and +1:

[0090] r>0: indicates that the two variables are positively correlated.

[0091] r<0: indicates that the two variables are negatively correlated.

[0092] r=0: indicates that there is no linear correlation between the two variables.

[0093] The closer |r| is to 1, the stronger the linear correlation; the closer it is to 0, the weaker the linear correlation.

[0094] The general correlation strength classification standard can be referred to (the specific threshold may be slightly different depending on the field and research purpose):

[0095] 0.0≤|r|<0.2: very weak correlation or no correlation.

[0096] 0.2≤|r|<0.4: weak correlation.

[0097] 0.4≤|r|<0.6: moderate correlation.

[0098] 0.6≤|r|<0.8: strong correlation.

[0099] 0.8≤|r|≤1.0: very strong correlation.

[0100] In specific implementation, the Pearson correlation coefficient between each pair of monitoring data (U1-U6, I1-I6, U1'-U6', U1''-U6'', I1'-I6', T) obtained in step S1 is calculated by substituting the above formula. For example, the correlation coefficient of U1 and I1, the correlation coefficient of U1 and U1', the correlation coefficient of U1 and T, etc.

[0101] S2.3 Determination of neural network model input features

[0102] Based on the correlation coefficients calculated in S2.2, analysis is performed to screen out monitoring parameters that have a significant impact on the predetermined target parameter as input features of the subsequent GABP neural network model. The principle of selecting input features is:

[0103] High correlation: Prefer parameters that have strong or very strong correlation with the predetermined target parameter.

[0104] Clear physical meaning: Choose parameters with clear physical meaning in combination with the electrochemical corrosion mechanism and sensor working principle.

[0105] Information complementarity and avoiding high collinearity: Try to choose parameters that can provide different aspects of information, and pay attention to avoid choosing multiple parameters with too high correlation (e.g. |r|>0.9 or higher) as inputs to reduce the redundancy and multicollinearity problem of the model, although neural networks are less sensitive to this than traditional linear models, but good feature selection is still beneficial.

[0106] Include key influencing factors: For example, temperature is usually an important environmental factor that affects the rate of electrochemical reactions, even if its direct linear correlation with the target parameter is not the highest, it should be considered for inclusion.

[0107] Based on practical experience, some typical feature selection strategies can be drawn:

[0108] There is a very strong positive correlation between anode voltage U1-U6 and corresponding current I1-I6 (correlation coefficients are all above 0.85). This means that when predicting voltage, the corresponding current is a very important input feature, and vice versa.

[0109] There is a strong correlation between anode voltage U1-U6 and the voltage value of the corresponding reference electrode U1'-U6' (correlation coefficients are all above 0.7). This shows that the voltage value of the reference electrode can provide valuable information for predicting the anode voltage.

[0110] The voltage / current value of the anode rod relative to the cathode also shows a certain positive correlation with the voltage / current value relative to the steel bar, which is consistent with the electrochemical principle and indirectly verifies the reliability of the monitoring data. These parameters can also be used as candidate input features.

[0111] The environmental temperature (T) is an important environmental parameter that will affect the concrete resistivity, ion diffusion rate and electrochemical reaction kinetics, so it should also be selected as one of the input features.

[0112] Specific example (taking the prediction of anode rod A1 voltage value U1 as an example):

[0113] According to the above principles and analysis, when the predetermined target parameter is the voltage value U1 of the anode rod A1, a reasonable input feature set can include:

[0114] I1: The current value between anode rod A1 and cathode (strongly correlated with U1).

[0115] U1': Voltage value between anode rod A1 and reference electrode (strongly related to U1).

[0116] U1": Voltage value between anode rod A1 and internal reinforcement (provides supplementary information).

[0117] I1': Current value between anode rod A1 and internal reinforcement (provides supplementary information).

[0118] T: Ambient temperature value (key environmental influencing factor).

[0119] Therefore, the input layer of the neural network model in this scenario will have 5 neurons, corresponding to the 5 selected input features. For predicting the voltage or current values of other anode rods, a similar analysis method can be used to determine their corresponding input feature sets.

[0120] Through the implementation of step S2, the original multi-dimensional monitoring data is transformed into a set of input features that are highly relevant to the prediction target and rich in information. This not only helps to improve the prediction accuracy of the subsequent GABP neural network model, but also reduces the complexity of the model and improves the training efficiency.

[0121] Step S3: Construction and normalization preprocessing of training and testing data sets for neural network model

[0122] After determining the prediction target and corresponding input features, the next step is to prepare the data set suitable for training and evaluating the neural network model. The original monitoring data cannot be directly used for model training and needs to go through a series of preprocessing steps to improve the training efficiency, convergence speed and final prediction performance of the model. This step mainly includes reasonable division of the data set and normalization of the data.

[0123] S3.1 Data set division

[0124] In order to objectively evaluate the performance of the trained neural network model, especially its generalization ability, the input feature data selected in step S2 and its corresponding target parameter measured data (hereinafter referred to as "complete data set") need to be divided into two independent subsets: training set and test set.

[0125] Training set (Training Set):

[0126] Used to train the neural network model. The model learns the data patterns in the training set, adjusts its internal connection weights and thresholds to minimize the difference between the predicted output and the true target value.

[0127] Test set (Test Set):

[0128] It is used to evaluate the performance of the model after the training is completed. The data of the test set does not participate in the training process of the model, so it can be used to test the adaptability and prediction accuracy of the model to new data.

[0129] Division method and proportion:

[0130] For time series data, such as durability monitoring data in the present application, the method of division in chronological order is usually adopted to avoid future data "leaking" into historical training data, thereby ensuring the effectiveness of the evaluation.

[0131] In an embodiment of the present application, based on the acquired continuous monitoring data for 4 years, the division can be carried out according to the following proportions:

[0132] Training set: 80% of the total amount of data is selected as the training set. For example, if there are N data records in total, the first 0.8xN records (i.e. data in the earlier time period) are used for training.

[0133] Test set: 20% of the total amount of data is selected as the test set. That is, the last 0.2xN records (i.e. data in the later time period) are used for testing.

[0134] This division proportion (such as 80 / 20, or 70 / 30, 75 / 25) is commonly used in machine learning practice. The specific division point should be selected to ensure that the training set contains enough diverse data patterns for the model to learn, and the test set should also be representative, reflecting the data characteristics that the model may encounter in future actual applications.

[0135] Suppose we take the prediction of the voltage value U1 of the anode rod A1 as an example, and the input features are I1, U1', U1'', I1', T. Then the training set will include the observed values of these 5 input features in the selected time period (e.g. the first 80% of the time), as well as the true observed values of U1. The test set is the same.

[0136] S3.2 Data Normalization

[0137] Since different monitoring parameters (such as voltage, current, temperature) collected from the anode ladder sensor system have different physical dimensions and value ranges, it is necessary to normalize (Normalization) the input feature data and target output data, and map them to a specific, smaller value interval, usually [0, 1] or [-1, 1].

[0138] This embodiment adopts the Min-Max Normalization method to linearly scale the data to the [0, 1] interval. The calculation formula is as follows:

[0139]

[0140] If mapping to a custom interval [Y_min, Y_max] is desired, it can be represented as:

[0141]

[0142] where:

[0143] X i (or X) is the original data value.

[0144] X min is the minimum value of the feature in the training set.

[0145] X max is the maximum value of the feature in the training set.

[0146] X norm (or X') is the normalized data value.

[0147] Y min and Y max are the lower and upper limits of the target normalization interval (in this embodiment, Y min = 0, Y max = 1, in which case the formula simplifies to the former form).

[0148] Through the implementation of step S3, the original, scale-inconsistent monitoring data is converted into a standardized data set suitable for training and testing of the GABP neural network model. This lays a solid foundation for stable training of subsequent models, rapid convergence, and obtaining high-precision prediction results.

[0149] Step S4: Construction and iterative training of the genetic algorithm-optimized BP neural network (GABP) prediction model

[0150] After completing the data preparation, this step will focus on how to design a basic BP neural network and introduce genetic algorithm (GA) to optimize its initial parameters, and finally obtain the GABP prediction model through supervised learning training.

[0151] S4.1 BP neural network basic structure design

[0152] The BP (Back Propagation) neural network is a classic multi-layer feedforward neural network that learns through the error backpropagation algorithm. The BP neural network structure of this embodiment is shown in Figure 4 , and its basic structure includes an input layer, at least one hidden layer (also known as a hidden layer), and an output layer.

[0153] 1. Input layer (Input Layer):

[0154] The number of input layer neurons is determined by the number of input features determined in step S2.

[0155] For example, when predicting the voltage value U1 of the anode rod A1 and the selected input features are I1, U1', U1", I1', and T, the input layer will include 5 neurons, each corresponding to an input feature.

[0156] The input layer neurons usually do not perform calculations and are only responsible for receiving external input data (feature values after normalization in step S3) and passing them to the hidden layer.

[0157] 2. Hidden Layer(s):

[0158] The hidden layer is the core of the BP neural network and is responsible for non-linear transformation and feature extraction of input information.

[0159] Number of layers: For most problems, a BP network with a single hidden layer can already approximate any continuous function with arbitrary precision (universal approximation theorem). In this embodiment, a single hidden layer is first considered. If the problem is particularly complex, the number of hidden layers can also be increased, but too many hidden layers can lead to training difficulties and overfitting.

[0160] Number of neurons: The number of hidden layer neurons is a key hyperparameter, and there is no fixed theoretical formula to directly determine the optimal value. It is usually set according to experience, experiment, or some heuristic rules. A commonly used empirical formula is: where N h is the number of hidden layer neurons, N i is the number of input layer neurons, N o is the number of output layer neurons, and a is an empirical constant, usually an integer between 1 and 10. For example, if the input layer has 5 neurons and the output layer has 1 neuron, and a = 5 is taken, then or 8 neurons. The final number of hidden layer neurons needs to be determined by repeated trials and performance evaluation to determine the optimal value.

[0161] Activation Function: Hidden layer neurons usually use a non-linear activation function to give the network the ability to learn complex non-linear relationships. In this embodiment, the Sigmoid function is chosen as the activation function for the hidden layer neurons. Its common form (Logistic function) is: The output range of the Sigmoid function is between (0, 1), which matches the data normalization to the [0, 1] interval. Other optional activation functions include the Tanh function, the ReLU function and its variants, etc.

[0162] 3. Output Layer:

[0163] The number of output layer neurons is determined by the number of prediction targets.

[0164] In this application, usually a single durability parameter (e.g. voltage or current of a certain anode rod) is predicted, so the output layer usually contains 1 neuron.

[0165] Activation function: For regression problems (e.g. predicting continuous voltage or current values), the output layer neuron can use the same Sigmoid function as the hidden layer, or use a linear activation function (i.e. no activation function, directly output the weighted sum) if the output data has been normalized to the [0, 1] interval. If the target output range is different, the corresponding activation function is selected.

[0166] Brief description of the working principle of BP neural network:

[0167] Data is input from the input layer, processed by the weighted sum and activation function of the hidden layer neurons, and then transmitted to the output layer. The output layer neurons also perform weighted sum and activation function processing to obtain the network's predicted output. The training process compares the error (e.g. mean square error) between the predicted output and the true target value, uses the gradient descent method and error backpropagation mechanism to adjust the weights and biases of each connection in the network from the output layer to the front layer, to minimize the error.

[0168] S4.2 Genetic Algorithm (GA) parameter configuration and initialization

[0169] In traditional BP neural networks, the random setting of initial weights and thresholds has a great impact on their final performance and convergence speed. Poor initial values can easily make the network fall into a local optimal solution. Genetic Algorithm (Genetic Algorithm, GA) is a global optimization search algorithm that simulates natural selection and genetic mechanisms, which can effectively solve this problem.

[0170] Basic idea of GA optimizing BP network parameters:

[0171] A set of initial weights and thresholds of BP neural network is encoded into an individual (chromosome), and the population composed of these individuals is iteratively evolved through genetic operations (selection, crossover, mutation) to find the initial weight and threshold combination that can make the BP network perform optimally on a specific data set.

[0172] GA parameter configuration:

[0173] Before applying GA to optimize BP network, the following key GA running parameters need to be set:

[0174] 1. Population Size (pop_size): refers to the number of individuals (chromosomes) contained in each generation. The value range is generally wide, such as 20-200. Larger population size increases the diversity of search space, but the computational cost is also higher.

[0175] 2. Maximum Generations (max_gen): refers to the total number of iterations of genetic algorithm. The algorithm stops after reaching the maximum number of generations.

[0176] 3. Crossover Rate (pc): refers to the proportion of individuals that undergo crossover operation in each generation. Usually take a higher value, such as 0.6-0.9.

[0177] 4. Mutation Rate (pm): refers to the probability of performing mutation operation on individual genes in each generation. Usually take a lower value, such as 0.01-0.1, to avoid destroying the existing good gene structure while ensuring population diversity.

[0178] 5. Encoding method: it is necessary to connect all the weights and thresholds of BP neural network to be optimized to form a one-dimensional vector, which is used as chromosome coding in genetic algorithm. Usually use real number coding.

[0179] 6. Fitness Function: this is the core of GA for individual evaluation and selection. Fitness function is used to measure the pros and cons of each individual (i.e. each set of initial weights and thresholds of BP network). In this invention, the goal of fitness function is to minimize the prediction error of BP neural network on the training set. Specifically, it can be defined as:

[0180] After taking the weights and thresholds coded by the current individual as BP network for one (or several iterations) forward propagation, the reciprocal of the mean squared error (MSE) calculated on the training set is obtained (because GA usually maximizes fitness).

[0181] Or directly use MSE as evaluation index, the goal of genetic algorithm is to find the individual that minimizes MSE. Where N is the number of training samples, y i is the true value, is the predicted value obtained by the network using the current weights and thresholds.

[0182] Initialize the population:

[0183] After the parameters are set, an initial population is randomly generated. Each individual (chromosome) is composed of a set of randomly generated real-valued numbers, which correspond to the initial weights and thresholds of the BP neural network. The range of random numbers also needs to be pre-set.

[0184] S4.3 Optimization of initial weights and thresholds of BP neural network based on GA

[0185] This step is the core optimization process of the GA BP model:

[0186] 1. Fitness evaluation: For each individual in the current population, decode it into the initial weights and thresholds of the BP network. Then, initialize the BP network with this set of parameters and perform one (or a small number of iterations) of training on the training set or directly forward propagation, calculate the error (such as MSE) between its predicted output and the true target. According to the pre-defined fitness function, calculate the fitness value of each individual.

[0187] 2. Selection: According to the size of the individual fitness value, select good individuals into the next generation. Common selection strategies include roulette wheel selection, tournament selection, elite preservation strategy (copy the top individuals in the current generation directly to the next generation to ensure that the optimal solution is not lost) and so on.

[0188] 3. Crossover: With a crossover rate pc, randomly select pairs from the selected parent individuals, exchange part of their gene segments, and generate new offspring individuals. Common crossover methods include single-point crossover, multi-point crossover, arithmetic crossover, etc.

[0189] 4. Mutation: With a mutation rate pm, randomly change some gene sites of the offspring individuals by a small amount, introduce new genes, increase population diversity, and help jump out of local optimum. For real number coding, it can be uniform mutation, Gaussian mutation, etc.

[0190] 5. Iterative evolution: Repeat steps 1-4 until the maximum evolution generation max_gen is reached, or the average fitness of the population reaches a certain threshold, or the fitness of the optimal individual does not significantly improve for several generations in a row.

[0191] 6. Get the optimal solution: When the GA iteration stops, the set of weights and thresholds coded by the individual with the highest fitness in the population is the best initial weights and thresholds of the BP neural network optimized by the genetic algorithm.

[0192] S4.4 Training of GA BP neural network and model generation

[0193] 1. Load optimized initial parameters: Decode the best individual found by the genetic algorithm in S4.3 to obtain a set of weights and thresholds as the initial parameters for the BP neural network.

[0194] 2. Fine-tune the BP algorithm:

[0195] Initialize the BP neural network with these optimized initial parameters.

[0196] Then, use the training set data prepared in step S3 to further train the BP network through the standard backpropagation algorithm for more precise training.

[0197] During training, the network adjusts weights and thresholds based on input and desired output from training samples to further reduce prediction error.

[0198] Set training parameters such as learning rate, number of training iterations (epochs), target error for the BP network.

[0199] 3. Model convergence and saving:

[0200] The training process continues until the network's prediction error on the training set meets the preset accuracy requirement (e.g., MSE is less than a certain threshold), or the maximum number of training iterations is completed.

[0201] At this point, the trained neural network model (i.e., GABP model, which contains the weights and thresholds optimized by GA and fine-tuned by BP) is saved for subsequent prediction applications.

[0202] Through step S4, a GABP neural network is constructed. The structure of the constructed GABP neural network is shown in Figure 5 The core advantage is that the global search capability of the genetic algorithm provides a better training starting point for the BP network, avoiding the defect of the BP algorithm itself easily falling into local minimum, so that the final trained GABP model has higher prediction accuracy and better generalization performance.

[0203] Step S5: Comprehensive evaluation of GABP prediction model performance and verification of generalization ability in actual application

[0204] After successfully constructing and training the GABP neural network model in step S4, the core task of this step is to comprehensively and objectively evaluate the prediction performance of the model and verify its generalization ability and reliability in actual application.

[0205] S5.1 Model prediction based on test set

[0206] 1. Load the trained GABP model:

[0207] First, load the GABP neural network model that has been trained in S4.4.

[0208] 2. Prepare test data:

[0209] Use the test set divided in step S3.1. The test set contains the same structure of input features as the training set, and the corresponding true target parameter values.

[0210] 3. Perform prediction:

[0211] Input the normalized test set feature data piece by piece or in batches into the loaded GABP model. The model will perform forward propagation calculation according to the learned patterns, and output the predicted value of the target parameter.

[0212] 4. Reverse normalization processing:

[0213] In order to restore the prediction results of the model to the original data scale, facilitate comparison with the true monitoring value and practical application, it is necessary to perform reverse normalization operation on the normalized prediction value output by the model.

[0214] After the above steps, the original scale prediction value of the target parameter of all samples in the test set by the GABP model can be obtained.

[0215] S5.2 Performance evaluation of prediction results

[0216] Comparison with benchmark model:

[0217] When predicting the voltage value U1 of the anode rod A1, use the current value I1, the voltage value U1', the voltage value U1", the current value I1' and the environmental temperature T as the original input, among the 56 groups of original data, select the first 80% of the data as the training set, and select the last 20% as the test set. When predicting the current value I1 of the anode rod A1, the same five related factors are selected as the original input for prediction. In order to better judge the effectiveness and optimization effect of the model, BP neural network, particle swarm optimization BP neural network and genetic algorithm optimized BP neural network model are used for prediction, and the prediction results are compared and analyzed, among which the voltage value and current value prediction results of the anode rod A1 are shown in Tables 1 and Figure 6 .

[0218] Table 1 Model prediction value summary table

[0219]

[0220]

[0221] Based on the above prediction results, the root mean square error RMSE, the mean absolute error MAE, the determination coefficient R2 and the calculation results of mean square relative error MSRE are used for error analysis of the model.

[0222] The prediction results of the three prediction methods are shown in Table 2, and error values are compared as shown in Table 2. Figure 7 It can be seen from Table 2 that the determination coefficient R 2 of the voltage value predicted by the GABP model is increased by 14% compared with the BP neural network model and by 5% compared with the PSO-BP neural network model, and the determination coefficient R 2 of the current value is increased by 21% compared with the BP neural network model and by 9% compared with the PSO-BP neural network model. Therefore, the prediction effect of the GABP method is significantly better than that of other methods, and the feasibility and prediction effect of the prediction method are verified.

[0223] Table 2 Error summary table of model prediction values

[0224]

[0225] S5.3 Model generalization performance verification

[0226] In order to verify the generalization performance of the GABP model, the prediction method of the anode rod A1 monitoring data is used to predict and verify the voltage and current measured values of the anode rods A2-A6, and then the RMSE, MAE, R 2 , MSRE are used for error analysis, and the specific results are shown in Table 3.

[0227] Table 3 Error calculation summary table of A2-A6

[0228]

[0229]

[0230] From the error analysis of the voltage and current prediction results of the anode rods A2-A6 in Table 3, the determination coefficient R 2 is all above 0.80, which indicates that the GABP model has high prediction accuracy and also reflects the practicability and generalization performance of the model.

[0231] Through the comprehensive evaluation and verification of step S5, it can be fully proved that the GABP neural network-based durability monitoring data prediction analysis method for high-pile wharfs proposed in the application is effective, accurate and has superior generalization ability, which indicates that it can provide reliable technical support for health state evaluation and maintenance decision of high-pile wharfs.

[0232] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A high-pile wharf durability monitoring data prediction and analysis method based on GABP neural network is characterized by: The following steps are involved: Obtain multi-parameter monitoring data collected by the anode ladder sensor system of the high-pile dock; performing a correlation analysis based on the multi-parameter monitoring data to select input features for predicting a predetermined target parameter, wherein the predetermined target parameter is a monitoring parameter characterizing the durability of the wharf; Preprocessing the selected input feature data and the corresponding measured data of the predetermined target parameters; Constructing and training a GABP neural network prediction model, wherein the training of the GABP neural network prediction model utilizes the preprocessed input feature data and measured data of predetermined target parameters, and optimizes the initial weights and thresholds of the BP neural network through a genetic algorithm; The trained GABP neural network prediction model is used to input new monitoring data to predict the predetermined target parameters.

2. The high-pile wharf durability monitoring data prediction and analysis method based on GABP neural network according to claim 1 is characterized in that: The multi-parameter monitoring data obtained includes at least: One or more combinations of the voltage value of the anode rod relative to the cathode, the voltage value of the anode rod relative to the reference electrode, the voltage value of the anode rod relative to the internal steel bar, the current value of the anode rod relative to the cathode, the current value of the anode rod relative to the internal steel bar, and the ambient temperature value.

3. The high-pile wharf durability monitoring data prediction and analysis method based on GABP neural network according to claim 1 is characterized in that: The correlation analysis uses the Pearson correlation coefficient method to determine the degree of correlation between each monitoring parameter and the predetermined target parameter.

4. The high-pile wharf durability monitoring data prediction and analysis method based on GABP neural network according to claim 1 is characterized in that: The pretreatment includes: The input feature data and measured data of predetermined target parameters are divided into a training set and a test set, and the data are normalized.

5. The high-pile wharf durability monitoring data prediction and analysis method based on GABP neural network according to claim 4 is characterized in that: The normalization process maps the data to a preset numerical range using the following formula: Among them, X i is the original data, X min and X max are the minimum and maximum values ​​in the original data set, respectively, min and Y max are the lower limit and upper limit of the preset value range.

6. The high-pile wharf durability monitoring data prediction and analysis method based on GABP neural network according to claim 1 is characterized in that: The structure of the BP neural network includes an input layer, at least one hidden layer and an output layer, wherein the activation function of the neurons in the hidden layer is a Sigmoid function.

7. The method for predicting and analyzing high-pile wharf durability monitoring data based on GABP neural network according to claim 1 is characterized in that: When the genetic algorithm optimizes the initial weights and thresholds of the BP neural network, the initial population size, maximum evolutionary generations, crossover rate and mutation rate parameters are preset.

8. The method for predicting and analyzing high-pile wharf durability monitoring data based on GABP neural network according to claim 1 is characterized in that: The predetermined target parameter is a voltage value or a current value of a specific anode rod in the anode ladder sensor system.

9. The method for predicting and analyzing high-pile wharf durability monitoring data based on GABP neural network according to claim 1 or 8, characterized in that: When the predetermined target parameter is the voltage value of a specific anode rod, the selected input features include at least: The current value corresponding to the specific anode rod, the voltage value between the specific anode rod and the reference electrode, the voltage value between the specific anode rod and the internal steel bar, the current value between the specific anode rod and the internal steel bar, and the ambient temperature value.

10. The method for predicting and analyzing high-pile wharf durability monitoring data based on GABP neural network according to claim 1 is characterized in that: The genetic algorithm optimizes the initial weights and thresholds of the BP neural network with the goal of minimizing the prediction error of the BP neural network on the training set or optimizing a preset fitness function.

Images