A pipeline inner wall corrosion detection method and system suitable for water conservancy projects

Abstract

The present invention relates to the technical field of pipeline maintenance, and discloses a pipeline inner wall corrosion detection method and system suitable for water conservancy projects, including: obtaining a number of sensor data from a sensor component arranged in the pipeline to be tested; performing feature extraction to obtain detection data features; performing time-space synchronous fusion to obtain fusion features; obtaining historical corrosion data of the inner wall of the pipeline and the fusion features and inputting them into a long short-term memory network model to obtain a first corrosion prediction result; generating a second corrosion prediction result by a random forest algorithm; generating a third corrosion prediction result by a gradient boosting tree algorithm; and inputting the three prediction results into a preset detection model to obtain a detection report. The present invention obtains fusion features by fusing multiple sensor data, comprehensively refers to multiple factors affecting corrosion, performs prediction analysis by multiple algorithms based on the fusion features to obtain three prediction results, generates a detection report, and effectively improves the accuracy of the detection results.

Description

Technical Field

[0001] The present invention relates to the technical field of pipeline maintenance, and in particular to a pipeline inner wall corrosion detection method and system suitable for water conservancy projects. Background Art

[0002] In the field of water conservancy project pipeline maintenance, pipelines are prone to corrosion due to factors such as microorganisms, water flow and oxygen during use, leading to safety hazards in pipelines. The current detection method mainly relies on manual visual inspection on the outside of the pipeline and ultrasonic testing of selected pipelines.

[0003] Traditional pipeline corrosion detection methods only involve physical methods, which require periodic detection. During the detection process, the pipeline is shut down, and corrosion cannot be detected in time. In addition, the data source of traditional detection methods is single, and the parameters required for each detection are basically the same. It cannot comprehensively refer to the various factors that affect corrosion, resulting in inaccurate detection results. The detection results usually need to be manually reviewed, the review cycle is long, and risks cannot be reported in time.

[0004] In summary, traditional pipeline corrosion detection methods have at least one technical problem of low efficiency, insufficient accuracy, high labor costs, and high safety risks. Summary of the invention

[0005] In view of this, an embodiment of the present invention provides a pipeline inner wall corrosion detection method and system applicable to water conservancy projects, so as to solve the technical problems of low efficiency and inaccurate detection results of existing pipeline corrosion detection methods.

[0006] To achieve the above objectives, in a first aspect, the invention provides a method for detecting inner wall corrosion of a pipeline applicable to a water conservancy project, comprising the following steps:

[0007] Acquire a number of sensor data from a sensor assembly disposed in the pipeline to be tested;

[0008] Extracting features from the plurality of sensor data to obtain a plurality of detection data features;

[0009] Performing spatiotemporal synchronous fusion on the plurality of detection data features to obtain fusion features;

[0010] Obtaining a time series of historical corrosion data of the inner wall of the pipeline, inputting the time series of historical corrosion data of the inner wall of the pipeline and the fusion feature into a long short-term memory network model, and obtaining a first corrosion prediction result;

[0011] A second corrosion prediction result is obtained by using a random forest algorithm according to the fusion features;

[0012] A third corrosion prediction result is obtained by using a gradient boosting tree algorithm according to the fusion feature;

[0013] A pipeline inner wall corrosion detection report of the water conservancy project is outputted according to the first corrosion prediction result, the second corrosion prediction result and the third corrosion prediction result.

[0014] In a second aspect, the present invention provides a pipeline inner wall corrosion detection system applicable to water conservancy projects, comprising:

[0015] A sensor component, wherein the sensor component is disposed in the pipeline to be tested and is used to obtain a number of sensor data;

[0016] A feature extraction module, used to extract features from the plurality of sensor data to obtain a plurality of detection data features;

[0017] A feature fusion module, used for performing spatiotemporal synchronous fusion of the plurality of detection data features to obtain fused features;

[0018] A first prediction module is used to obtain a time series of historical corrosion data of the inner wall of the pipeline, input the time series of historical corrosion data of the inner wall of the pipeline and the fusion feature into a long short-term memory network model, and obtain a first corrosion prediction result;

[0019] A second prediction module, used for obtaining a second corrosion prediction result by using a random forest algorithm according to the fusion features;

[0020] A third prediction module, used for obtaining a third corrosion prediction result by using a gradient boosting tree algorithm according to the fusion feature;

[0021] An output module is used to output a pipeline inner wall corrosion detection report of a water conservancy project according to the first corrosion prediction result, the second corrosion prediction result and the third corrosion prediction result.

[0022] In a third aspect, the present invention provides an electronic device, comprising:

[0023] one or more processors;

[0024] a storage device for storing one or more programs,

[0025] When the one or more programs are executed by the one or more processors, the one or more processors implement a pipeline inner wall corrosion detection method applicable to water conservancy projects as described in the first aspect.

[0026] In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, a method for detecting inner wall corrosion of a pipeline applicable to a water conservancy project as described in the first aspect is implemented.

[0027] The above technical solution has the following beneficial technical effects:

[0028] The present invention obtains fusion features by acquiring data from multiple sensors and fusing them, thereby diversifying the data sources. It can comprehensively refer to multiple factors that affect corrosion, and adopts multiple algorithms to perform predictive analysis based on the fusion features to obtain three prediction results. A detection report is generated based on the three prediction results. The use of different types of algorithms to comprehensively generate reports effectively improves the accuracy of the detection results. The present invention generates three results through a long short-term memory network model, a random forest algorithm, and a gradient boosting tree algorithm, and generates a detection report based on the three results, which effectively reduces the risk of overfitting and can also capture the complex nonlinear relationship of corrosion development. Compared with traditional corrosion detection methods, the efficiency and accuracy are greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings are used to better understand the present invention and do not constitute an improper limitation of the present invention.

[0030] Figure 1 It is a flow chart of a pipeline inner wall corrosion detection method applicable to water conservancy projects in an embodiment of the present invention;

[0031] Figure 2 is a cross-sectional schematic diagram of a unit in an ultrasonic sensor according to an embodiment of the present invention;

[0032] Figure 3 is a flow chart of a method for setting the coil spacing and coil density of an eddy current probe in an embodiment of the present invention;

[0033] Figure 4 is a specific method flow chart of step S20 in an embodiment of the present invention;

[0034] Figure 5 is a specific method flow chart of step S21 in an embodiment of the present invention;

[0035] Figure 6 is a specific method flow chart of step S30 in an embodiment of the present invention;

[0036] Figure 7 is a specific method flow chart of step S40 in an embodiment of the present invention;

[0037] Figure 8 is a specific method flow chart of step S50 in an embodiment of the present invention;

[0038] Fig. 9 is a specific method flow chart of step S60 in an embodiment of the present invention;

[0039] Fig.10 is a specific method flow chart of step S70 in an embodiment of the present invention;

[0040] Fig.11It is a structural block diagram of a pipeline inner wall corrosion detection system applicable to water conservancy projects in an embodiment of the present invention;

[0041] Fig.12 It is a schematic diagram of the structure of a computer system in an embodiment of the present invention.

[0042] In the figure: 1. transducer; 2. damping layer; 3. acoustic matching layer; 4. protective layer. DETAILED DESCRIPTION

[0043] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0044] Embodiment 1

[0045] like Figure 1 As shown, an embodiment of the present invention provides a pipeline inner wall corrosion detection method applicable to water conservancy projects, which includes the following steps:

[0046] S10: Acquire some sensor data from the sensor assembly disposed in the pipeline to be tested.

[0047] Specifically, the sensor assembly includes several types of sensors, including but not limited to ultrasonic sensors, electromagnetic induction sensors and / or optical scanning sensors. During operation, the types and quantities of sensors can be increased or decreased according to pipeline conditions.

[0048] Specifically, each sensor in the sensor assembly adopts a high-precision sensor with high sensitivity and wide-band response characteristics. It can accurately capture subtle corrosion changes below 0.1mm, thereby comprehensively evaluating the corrosion condition of the inner wall of the pipeline.

[0049] Specifically, the ultrasonic sensor includes 5 to 16 units, each unit has a diameter of 2mm to 5mm, and each unit has independent focusing and transmitting and receiving functions. The multi-unit design enables the ultrasonic sensor to dynamically adjust the detection parameters according to different pipeline materials and corrosion characteristics, thereby improving the adaptability and accuracy of the detection. By limiting the number of ultrasonic sensor units, it is possible to avoid abnormal data acquisition due to too few units, resulting in inaccurate detection results, and it can also effectively prevent excessive units from being too expensive and wasteful.

[0050] Specifically, Figure 2As shown, each unit is provided with a transducer 1, and the transducer 1 is preferably an electric ceramic transducer. A damping layer 2 is provided outside the transducer 1, and an acoustic matching layer 3 is provided outside the damping layer 2, and a protective layer 4 is provided outside the acoustic matching layer 3. The damping layer 2 is directly attached to the transducer 1, and its main function is to suppress the mechanical oscillation of the transducer 1, quickly attenuate excess mechanical energy, and improve the pulse response characteristics of the ultrasonic sensor; the acoustic matching layer 3 is located on the damping layer 2, and its function is to optimize the propagation of sound waves between media with different acoustic impedances, reduce the reflection of sound waves at the interface, and improve the coupling efficiency and penetration depth of sound waves. By setting the acoustic matching layer 3 and the damping layer 2, the signal penetration ability and signal-to-noise ratio are improved. The acoustic matching layer 3 can optimize the propagation characteristics of sound waves between different media, reduce the reflection and attenuation of sound waves, and enable the ultrasonic signal to penetrate the metal material more deeply and clearly. The damping layer 2 effectively suppresses the sensor oscillation, improves the transient response characteristics of the signal, and enables the tiny corrosion features to be captured more accurately. This multi-layer composite design not only improves signal quality, but also enhances the performance stability of the sensor in complex detection environments.

[0051] Specifically, each transducer 1 is provided with a temperature compensation circuit, which is connected in parallel and arranged on the back or side of the transducer 1. The temperature compensation circuit may include a temperature sensor (precision thermistor or thermocouple), a programmable automation front end (PAFE) and a processing unit. The temperature sensor is used to monitor the temperature change around the transducer 1 in real time, and convert the analog signal of the temperature change into a digital signal through the PAFE. The digital signal is input into the processing unit, and the processing unit dynamically adjusts the bias voltage, gain and excitation signal parameters of the electro-ceramic transducer according to the preset temperature-performance mapping model. By arranging a temperature compensation circuit on each electro-ceramic transducer, each transducer 1 can independently and in real time compensate for the changes in the dielectric constant, electromechanical coupling coefficient and acoustic wave propagation velocity of the piezoelectric material caused by temperature, effectively suppressing the negative impact of temperature fluctuations on the sensitivity and measurement accuracy of the sensor, and ensuring the stability and reliability of the detection system in extreme temperature environments.

[0052] Specifically, all sensors in the sensor assembly are made of piezoelectric materials, which mainly include three types of materials: the first type is piezoelectric composite materials, such as piezoelectric polymer (Polyvinylidene Fluoride, PVDF) and ceramic particles. The precise distribution of nanoscale ceramic particles improves the electrical and mechanical coupling coefficient, which can increase the coupling coefficient from 50% to 60% of traditional piezoelectric ceramics to 70% to 85%; the second type is high-performance piezoelectric crystal materials, such as lithium tantalate crystals (LiTaO3) and lithium germanate crystals (LiGaO3). These materials have lower dielectric loss and a wider frequency response range, and can maintain stable piezoelectric performance in the frequency band of 100kHz to 10MHz; the third type is piezoelectric film materials, such as aluminum nitride (AlN) and lead zirconate titanate (PZT) nanofilms, which achieve higher piezoelectric coefficients and lower mechanical losses through precise film preparation processes, enabling the sensor to detect tiny corrosion changes below 0.1mm. The comprehensive use of these piezoelectric materials improves the detection performance and sensitivity of ultrasonic sensor arrays.

[0053] Specifically, the electromagnetic induction sensor includes a Hall sensor and an eddy current probe. The Hall sensor uses magnetic sensitive materials and has a magnetic field resolution of up to 10 -6 Tesla can accurately detect tiny magnetic field changes on metal surfaces; the eddy current probe is composed of multiple turns of precision coils, and the coil spacing and density can be dynamically adjusted according to the detection object. The electromagnetic induction sensor can quickly identify cracks, corrosion spots and changes in material conductivity on the metal surface, with a detection accuracy of up to 0.05mm. The electromagnetic induction sensor is equipped with a temperature compensation circuit and a signal amplification circuit to improve anti-interference capabilities.

[0054] Specifically, Figure 3 As shown, the setting of the coil spacing and coil density of the eddy current probe includes the following steps:

[0055] S11: Determine material parameters and wavelength of detection signal according to the material of the pipeline to be tested;

[0056] S12: Substituting the material parameters and the wavelength of the detection signal into a preset coil parameter model to obtain the coil spacing and coil density; the preset coil parameter model includes a coil spacing model and a coil density model;

[0057] The coil spacing model is as follows:

[0058] d=k λ;

[0059] Wherein, d is the coil spacing; k is the preset coefficient, k∈(0.3, 0.7); λ is the wavelength of the detection signal.

[0060] The coil density model is as follows:

[0061] ;

[0062] Where ρ is the coil density, σ is the material conductivity, and μ is the magnetic permeability.

[0063] S13: Determine whether the coil spacing and coil density meet the preset constraints; if the coil spacing and coil density do not meet the constraints, adjust the coil parameter model parameters and repeat steps S12 to S13; if the first coil spacing and the first coil density meet the constraints, output the coil spacing and coil density.

[0064] Specifically, the preset constraints are as follows: d∈[0.1mm, 0.6mm]; ρ∈[50 turns / square centimeter, 500 turns / square centimeter].

[0065] Specifically, the detection frequency of the eddy current probe set by the above steps is 100 Hz to 10 kHz, which is higher than the detection frequency of the existing eddy current probe and the data collection is more comprehensive.

[0066] As a preferred implementation of this embodiment, when the material of the pipeline to be tested is a copper alloy with high conductivity, a coil array with a smaller spacing (0.1mm to 0.3mm) and a higher density can be used; while for stainless steel or aluminum alloy with poor conductivity, a coil configuration with a slightly larger spacing (0.3mm to 0.6mm) and a relatively lower density can be used. This change is mainly based on the electromagnetic properties of the material and the expected defect detection accuracy.

[0067] Specifically, the optical scanning sensor includes an optical sensor, an optical lens, and an image processing unit. The pixel density of the optical sensor can reach 20 million pixels / cm², and the spectral response range covers the visible light and near-infrared spectrum. The precision optical lens uses a multi-layer anti-reflective coating with ultra-low distortion characteristics, which can capture surface details with an accuracy of 0.01mm. The image processing unit can identify the corrosion morphology, area, and depth in real time, and automatically generate a high-precision corrosion distribution map.

[0068] Specifically, the sensor assembly synchronously collects the raw signals from each sensor at a sampling rate of 100MHz through a data acquisition card. The data acquisition card has a high resolution of 16 to 24 bits and can accurately capture the tiny signal changes of each sensor. Each sensor converts the analog signal into a digital signal through a dedicated interface and performs preliminary data formatting to prepare for subsequent processing.

[0069] S20: Extracting features from the plurality of sensor data to obtain a plurality of detection data features.

[0070] Specifically, Figure 4 As shown, step S20 includes:

[0071] S21: performing data cleaning on a plurality of sensor data to obtain a plurality of cleaned data;

[0072] S22: Classifying the plurality of cleaning data according to sensor types to obtain a plurality of types of cleaning data;

[0073] S23: According to the several types of cleaned data, feature extraction is performed on each type of cleaned data to obtain several detection data features.

[0074] Specifically, Figure 5 As shown, step S21 includes:

[0075] S211: performing frequency domain analysis on a number of sensor data by using wavelet transform and Fourier transform to remove random noise and power frequency interference, and obtaining sensor data processed by frequency domain analysis;

[0076] S212: Setting an adaptive filter to remove system errors and environmental noise in the output data of step S211 through the adaptive filter, thereby obtaining sensor data processed by the adaptive filter;

[0077] S213: using an outlier detection algorithm to remove abnormal values ​​in the sensor data processed by the adaptive filter to obtain cleaned sensor data.

[0078] Specifically, by first performing wavelet transform and Fourier transform on a number of sensor data for frequency domain analysis, random noise and power frequency interference in the sensor data can be effectively removed, thereby performing large-scale data cleaning. Then, by setting an adaptive filter, targeted processing is performed on the remaining data after large-scale data cleaning to further remove invalid data. Finally, the outlier detection algorithm is used to remove abnormal values ​​in the sensor data to complete data cleaning. Through steps S211 to S213, data cleaning can be effectively performed, the reliability of sensor data is improved, and subsequent detection is facilitated.

[0079] Specifically, in this embodiment, there are three types of sensors, namely ultrasonic sensors, electromagnetic induction sensors and optical scanning sensors.

[0080] Specifically, step S23 specifically includes:

[0081] For ultrasonic sensor data, short-time Fourier transform and wavelet packet transform are used to extract spectrum features;

[0082] For electromagnetic induction sensor data, Hilbert-Huang transform is used to extract instantaneous frequency and energy features;

[0083] For optical scanning sensor data, a deep convolutional neural network is used to extract corrosion texture features and morphological features.

[0084] By adopting different methods to extract features for different types of sensors, the accuracy of the extracted features is ensured.

[0085] As a preferred implementation of this embodiment, in step S213, a standard score (Z-Score) algorithm can be used to remove abnormal values. First, a number of sensor data are classified, and then the mean of each type of sensor data is calculated. The difference between each sensor data and the mean of the sensor data of this type is calculated and recorded as a deviation value. It is determined whether the deviation value is within a preset interval. If the deviation value belongs to the preset interval, the sensor data is not an abnormal value. If the deviation value does not belong to the preset interval, the sensor data is an abnormal value, and the sensor data whose deviation value does not belong to the preset interval is deleted. Specifically, the preset interval is set comprehensively according to factors such as the type of pipeline to be tested, the type of sensor in the sensor assembly, and the environment in which the pipeline to be tested is located.

[0086] S30: Performing spatiotemporal synchronous fusion on the plurality of detection data features to obtain fusion features.

[0087] Specifically, Figure 6 As shown, step S30 includes:

[0088] S31: dynamically adjusting the time of each detection data feature to align the time axis of each detection data feature to obtain a plurality of detection data features with aligned time axes;

[0089] S32: performing coordinate transformation and geometric calibration on each detection data feature to obtain a number of detection data features under the same reference system;

[0090] Specifically, first, collect information such as the installation position, angle, and measurement direction of each sensor to determine the initial coordinate system of each sensor. According to the installation information of the sensor relative to the pipeline, establish the global reference coordinate system of the pipeline, for example, with the center of the pipeline as the origin and the axis of the pipeline as the Z axis. Subsequently, convert each detection data feature from the original coordinate system of the sensor to the global reference coordinate system of the pipeline, and use the coordinate transformation formula to complete the coordinate transformation. The specific formula is X′=R·X+T, where X is the original coordinate vector, R is the rotation matrix, T is the translation matrix, and X′ is the converted coordinate vector. Through this process, the mapping of detection data from the local coordinate system of the equipment to the unified reference coordinate system of the pipeline is realized.

[0091] Then, the transformed data is calibrated according to the geometric parameters of the pipeline (e.g. radius, position of the center axis). The calibration algorithm is used to adjust the offset data for the feature point position offset caused by sensor error in the detection results. For example, these errors can be corrected by fitting the geometric shape of the inner wall of the pipeline (e.g. cylindrical or elliptical). After calibration, ensure that all detection data features accurately reflect the actual geometric shape of the inner wall of the pipeline, and recalculate the spatial coordinates of all feature points based on the global reference coordinate system of the pipeline.

[0092] After calibration, the detection data features in a unified reference coordinate system are output. The output results are checked for integrity to ensure that the data collected by all sensors are calibrated and meet unified geometric standards.

[0093] S33: according to the plurality of detection data features aligned with the time axis and the plurality of detection data features under the same reference system, a dynamic Bayesian probability fusion algorithm is used to obtain a time series confidence and a confidence weight of each detection data feature;

[0094] Specifically, in step S33, the following steps are included:

[0095] First, define the time series of the detection data features. Assume that there are N detection data features, each of which is represented by X i (t), where i = 1, 2, ..., N, and t represents the time series. This step establishes the data foundation for the subsequent probability analysis, ensuring that the detection data features of each sensor can be accurately described and tracked. Secondly, an initial confidence level is set for each detection data feature as the prior probability P(X i ). The initial value can be based on historical data statistics or set to equal weights, such as P(X i ) = 1 / N. This step provides a starting point for Bayesian probability inference, reflecting the initial cognition and trust level of each feature before obtaining new data. Then, according to the detection data feature X aligned with the time axis i (t) and the detection data feature X after geometric calibration i '(t), calculate its conditional probability P(X i (t)|X i '(t)). The calculation of conditional probability can be quantified by measuring the distance or similarity between feature values ​​using Gaussian distribution. This step evaluates the consistency of different sensor data in time and space. Next, the posterior probability (i.e., time-series confidence) of each detection data feature at the current time t is calculated using the Bayesian formula: s i (t)=P(X i(t)|data). This step dynamically updates the degree of trust in each feature by combining prior probability and conditional probability. The posterior probability reflects the updated cognition of feature reliability after observing new data. Based on the calculated time series confidence s i (t), the temporal confidence of all features is normalized to obtain the confidence weight w i (t). Normalization ensures that the sum of weights is 1, providing a standardized weight distribution scheme for subsequent feature fusion. This step converts the probability into a weight coefficient that can be directly used for weighted calculation. Then, for the entire time series [t0, t n ] Repeat the above steps to dynamically update the temporal confidence s of each feature i (t) and confidence weight w i (t). This dynamic update mechanism enables the algorithm to continuously adapt to and reflect the changing characteristics of sensor data, improving the real-time and accuracy of corrosion detection. Finally, the output of each detection data feature X i The time series confidence s over the entire time series i (t) and confidence weight w i (t). These results will be directly used in subsequent weighted fusion calculations to provide a probabilistic basis for the comprehensive assessment of pipeline inner wall corrosion.

[0096] S34: Perform weighted calculation based on each detection data feature and its corresponding time series confidence and confidence weight to obtain a fusion feature. The specific formula is as follows:

[0097] ;

[0098] In the formula, F (t) represents the fusion feature, F i (t) represents the characteristic value of sensor i at time t, which is extracted from the original sensor data by the feature extraction step, that is, the aforementioned detection data feature; w i (t) represents the confidence weight of sensor i at time t, s i (t) represents the temporal confidence of sensor i at time t.

[0099] S40: Obtain a time series of historical corrosion data of the inner wall of the pipeline, input the time series of historical corrosion data of the inner wall of the pipeline and the fusion feature into a long short-term memory network model, and obtain a first corrosion prediction result.

[0100] Specifically, Figure 7 As shown, in step S40, it specifically includes:

[0101] S41: stacking several LSTM network units to form a LSTM network model;

[0102] S42: Obtaining the time series of historical corrosion data of the inner wall of the pipeline;

[0103] S43: Inputting the time series of the historical corrosion data of the inner wall of the pipeline and the fusion features into the long short-term memory network model, and using the gated recurrent unit algorithm to obtain a first prediction result.

[0104] Specifically, each LSTM network unit finely regulates the flow of information through the forget gate, input gate, and output gate. The LSTM network model introduces a gated unit cycle algorithm, which includes three gated units, namely the forget gate, input gate, and output gate. The forget gate is used to filter and discard irrelevant historical information, thereby optimizing the effectiveness of memory; the input gate determines the extent to which new information enters the memory unit to ensure that important information is retained; the output gate regulates the impact of the current memory state on the output, achieving precise control of the model output. This gating mechanism enables the LSTM network model to effectively capture contextual information when processing long sequence data.

[0105] This gating mechanism enables the LSTM network model to effectively solve the long-term dependency problem in traditional neural networks and accurately capture the complex nonlinear relationships in time series. In the pipeline corrosion prediction scenario, the LSTM network model can accurately learn the dynamic evolution of corrosion progress, identify corrosion acceleration, potential turning points and development trends, and provide key information in the deep time dimension for subsequent predictions.

[0106] In this way, the LSTM network model can not only memorize important long-term information, but also dynamically adjust the weights of information in different time periods, thereby realizing intelligent modeling and accurate prediction of the pipeline corrosion development process. Specifically, the LSTM network model captures the key characteristics of corrosion at different time scales, such as corrosion acceleration, inflection points, and potential development trends. Through long-term memory units, the LSTM network model can identify complex dependencies between time steps that are far apart, and distinguish short-term fluctuations from long-term development trends. The training process of the LSTM network model adopts a time series cross-validation method and uses sliding window technology to ensure that the model can accurately capture the laws of the dynamic evolution of corrosion.

[0107] S50: Obtain a second corrosion prediction result using a random forest algorithm according to the fusion features.

[0108] Specifically, in step S50, if Figure 8 As shown, the following steps are included:

[0109] S51: establishing a number of random forest training sets according to the fusion features;

[0110] Specifically, when constructing the random forest training set, the fused features obtained in the previous step are first comprehensively analyzed and processed. Specifically, the fused features are randomly divided into training and test sets according to a predefined ratio (for example, 7:3). The training set data is standardized and preprocessed to eliminate the influence of different feature dimensions. The stratified sampling technique is used to ensure that the training set contains samples representing different corrosion degrees and time periods to avoid uneven data distribution. In addition, the feature selection algorithm is used to screen out the most significant feature dimensions for corrosion prediction to improve the information quality of the training set.

[0111] S52: establishing a number of random forest decision trees using a random forest algorithm according to the number of random forest training sets, wherein the number of the random forest decision trees is the same as the number of the random forest training sets;

[0112] Specifically, in the random forest decision tree construction stage, a random sampling method is first used to perform replacement sampling on each training set to generate multiple sub-training sets. For each sub-training set, an independent decision tree is constructed based on a decision tree generation algorithm (such as a classification and regression tree algorithm). When each node of the tree splits, a partial feature subset is randomly selected to select the best split feature and split point. During the decision tree generation process, the maximum depth of the tree and the minimum number of leaf node samples are controlled to prevent overfitting. A decision tree is generated for each training set, and the number of trees is the same as the number of training sets to ensure the diversity and randomness of the algorithm.

[0113] S53: generating a number of decision results according to the number of random forest decision trees;

[0114] Specifically, in step S53, the input is the fusion feature obtained in the previous step. The specific processing process is to input the fusion feature as the sample to be predicted into each random forest decision tree constructed previously in turn. Each decision tree predicts the input fusion feature according to its internally learned classification rules, and gives a classification result or numerical prediction of the degree of corrosion. For classification problems, the category prediction of each decision tree is recorded; for regression problems, the specific numerical prediction of each decision tree is recorded. By comparing the prediction results of different decision trees, the consistency of the decision trees can be preliminarily evaluated, preparing for the subsequent voting algorithm and high-quality target decision tree screening.

[0115] S54: Voting on a number of decision results using a voting algorithm to obtain a second corrosion prediction result.

[0116] As a preferred implementation of this embodiment, in step S54, the following steps are specifically included:

[0117] S541: Repeat steps S51 to S53 for several rounds, and use a voting algorithm to obtain the voting result of each round according to several decision results of each round;

[0118] S542: Calculate the accuracy of each random forest decision tree according to the voting results of each round;

[0119] S543: According to the accuracy of each random forest decision tree and a preset accuracy threshold, delete the random forest decision trees whose accuracy is lower than the accuracy threshold, and obtain a number of high-quality target random forest decision trees; the high-quality target random forest decision tree refers to a random forest decision tree whose accuracy is higher than the preset accuracy threshold;

[0120] S544: Generate a number of high-quality target decision results based on a number of high-quality target random forest decision trees;

[0121] S545: Using a voting algorithm to vote on a number of high-quality target decision results to obtain a second corrosion prediction result.

[0122] Specifically, the preset correct threshold should not be set too high, so as to avoid the model losing its representativeness due to too few random forest decision trees.

[0123] As a preferred implementation of this embodiment, the accuracy threshold ∈[0.6,0.8], and the accuracy threshold is preferably 0.7.

[0124] Specifically, when using the multi-round voting method, more initial random forest decision trees should be reserved to ensure that there are still enough random forest decision trees for integration even after deleting some low-quality random forest decision trees. Through the multi-round voting method, the random forest algorithm can be dynamically adjusted to improve the prediction performance, which is particularly suitable for complex prediction tasks such as corrosion detection on the inner wall of water conservancy project pipelines.

[0125] S60: Obtain a third corrosion prediction result using a gradient boosting tree algorithm according to the fusion features.

[0126] Specifically, in step S60, if Fig. 9 As shown, the specific steps include:

[0127] S61: preprocessing the fused features, and establishing a gradient boosting tree training set according to the preprocessed fused features;

[0128] Specifically, the preprocessing includes standardization and normalization, which eliminate the dimensional differences between different features and ensure that subsequent algorithms can process input data fairly and accurately.

[0129] S62: Building a gradient boosting decision tree;

[0130] S63: Calculate a numerical residual according to the gradient boosting tree training set and the predicted value;

[0131] Specifically, in steps S62 and S63, the construction of the gradient boosting decision tree is an iterative learning process. First, the preprocessed gradient boosting tree training set is used as input, and the model is constructed using the forward step-by-step algorithm. Specifically, the initial model is a constant prediction value, which is set to the average of the target value of the training set. In each round of iteration, the algorithm calculates the negative gradient information of the current model, which is the first-order derivative of the loss function to the model prediction value. Based on this negative gradient information, a new decision tree is constructed, and its goal is to fit these negative gradient values, that is, to find a decision tree structure that can best approximate the negative gradient. The generation process of the decision tree follows a greedy strategy, selecting features and split points that can minimize the loss function at each node. The generation of the tree limits the maximum depth and the minimum number of leaf node samples to prevent overfitting. The newly generated decision tree is scaled by the specified learning rate (shrinkage coefficient) and combined with the previous cumulative model by additive model. This iterative method allows each decision tree to focus on correcting the prediction error of the previous round of models, thereby gradually optimizing the overall prediction performance. For the complex nonlinear problem of pipeline corrosion prediction, the gradient boosting decision tree can effectively capture the complex interactions between features and continuously improve the prediction accuracy through multiple rounds of iterations.

[0132] S64: Calculate a numerical residual according to the gradient boosting tree training set and the predicted value;

[0133] Specifically, the numerical residual is the deviation between the input value of the gradient boosting tree training set and the predicted value, and the predicted value is generated according to the input value of the gradient boosting tree training set.

[0134] S65: Preset a corrosion risk weight coefficient, and determine a loss function according to the corrosion risk weight coefficient and the numerical residual;

[0135] Specifically, the formula of the loss function is as follows:

[0136] ;

[0137] In the formula, represents the loss function, represents the i input values ​​of the gradient boosting tree training set, represents the i-th predicted value, α i represents the corrosion risk weight coefficient;

[0138] Specifically, the corrosion risk weight coefficient α i The calculation formula is as follows:

[0139] ;

[0140] In the formula, β is the weight amplification parameter, risk_score(x i ) represents the risk score range; x i Indicates The feature vector of the sample, x i It is a multidimensional feature vector that contains features related to corrosion prediction, such as pipeline material, environmental factors, and historical corrosion data.

[0141] Specifically, risk_score(x i ) is set according to the material of the pipeline to be tested, environmental factors and historical corrosion data. As a preferred implementation of this embodiment, risk_score(x i )∈[0,1].

[0142] Specifically, β is used as a weight amplification parameter to control the degree of weight amplification. The more attention is paid to the high corrosion risk area, the larger the value of β is. As a preferred implementation of this embodiment, β∈[1,3], so that the prediction error of the high corrosion risk area is more penalized.

[0143] Specifically, the traditional gradient boosting tree algorithm uses the standard square error as the loss function, which only involves the actual value and the predicted value. However, in the field of corrosion prediction, the degree of corrosion has a greater impact. Therefore, in this embodiment, the corrosion risk weight coefficient is introduced into the loss function to make the loss function pay more attention to the high corrosion risk area, thereby improving the accuracy and practicality of the prediction.

[0144] S66: Obtain environmental data in the pipeline to be tested, and calculate physical residuals and structural residuals according to the environmental data, the gradient boosting tree training set and the predicted value;

[0145] Specifically, in step S66, it specifically includes:

[0146] S661, obtaining environmental data in the pipeline to be tested, and calculating a physical residual according to the environmental data and the Arrhenius formula;

[0147] Specifically, the calculation formula of the physical residual, that is, the Arrhenius formula, is as follows:

[0148] ;

[0149] In the formula, A is the frequency factor, Ea is the activation energy, R is the gas constant, and T is the absolute temperature. The frequency factor, activation energy, gas constant and absolute temperature are all environmental data.

[0150] S662: Calculate feature interaction entropy and conditional mutual information according to the gradient boosting tree training set and the predicted value;

[0151] Specifically, in step S662, feature interaction entropy is an information-theoretic measure of the strength of nonlinear interactions between multiple features. The specific calculation process first requires the definition of joint probability distribution and conditional probability distribution. For the feature set X = {x1, x2, ..., x n}, the feature interaction entropy H(X) can be expressed as:

[0152] ;

[0153] In the formula, P(x1,...,x n ) is the joint probability distribution of the features. Conditional Mutual Information further expresses the information transfer and dependency between features under the given prediction value condition. Its calculation formula is:

[0154] ;

[0155] Among them, X and Y are characteristic variables, and Z is a conditional variable, which is a predicted value in this embodiment. Lowercase x and y represent the specific values ​​of characteristic variables X and Y, respectively, and lowercase z represents the specific value of conditional variable Z. The calculation of these two indicators involves probability estimation and information theory calculation, and advanced statistical techniques such as kernel density estimation and information entropy estimation can be used. For pipeline corrosion prediction, the purpose of this step is to reveal the potential nonlinear correlation between different environmental factors, material properties, historical corrosion data and other features, and provide deep feature interaction information for subsequent structural residual calculation and model optimization. By quantifying the interaction intensity and information transfer between features, the complex mechanism of corrosion development can be understood more accurately, and the interpretability and accuracy of the prediction model can be improved.

[0156] S663: Calculate a structural residual according to the feature interaction entropy and the conditional mutual information;

[0157] Specifically, structural residual is a key indicator that reflects the deviation of the intrinsic structural features of the model, and is used to capture the nonlinear relationship between features and the complexity of the model structure. In the pipeline corrosion prediction scenario, structural residual reveals the model's ability and potential risks in learning complex corrosion mechanisms by quantifying feature interaction entropy and conditional mutual information.

[0158] The calculation formula of the structural residual Rstruct can be expressed as: Rstruct = λ1 H(X) + λ2 I(X;Y|Z), where H(X) is the feature interaction entropy, I(X;Y|Z) is the conditional mutual information, λ1 and λ2 are adjustment weight coefficients, and λ1 + λ2 = 1. The calculation process includes normalizing the feature interaction entropy and conditional mutual information, weighting according to predefined weight coefficients, and introducing penalty terms to control the complexity of feature interaction.

[0159] By calculating the structural residual, the key interactive features that affect pipeline corrosion can be accurately identified, the learning effect of the model on complex corrosion mechanisms can be evaluated, and an important basis for subsequent model structure optimization and feature selection can be provided. The structural residual not only reflects the interdependence between features, but also quantifies the ability of the model to capture nonlinear relationships, which helps reduce the risk of overfitting.

[0160] S67: Obtain a fusion residual by weighted calculation according to the numerical residual, the physical residual and the structural residual;

[0161] Specifically, the calculation formula of the fusion residual is as follows:

[0162] R = ω1 Rnum + ω2 Rstruct + ω3 Rphys;

[0163] Where R is the fusion residual, Rnum is the numerical residual, Rstruct is the structural residual, Rphys is the physical residual, ω1 is the first weight, ω2 is the second weight, ω3 is the third weight, and ω1 + ω2 + ω3 = 1.

[0164] Specifically, this embodiment breaks through the limitations of the traditional single numerical residual by adopting a multi-dimensional residual learning strategy. In addition to calculating numerical residuals, structural residuals and physical residuals are also introduced. Numerical residuals reflect prediction errors, structural residuals capture the nonlinear relationship between features, and physical residuals are constrained by the corrosion physics model. By combining these three residuals, the complex mechanism of corrosion development can be more comprehensively understood and learned, and the accuracy of prediction can be improved.

[0165] S68: Update the gradient boosting decision tree according to the fusion residual and the predicted value, repeat steps S63 to S67, use the corrosion risk weight coefficient and the numerical residual to calculate the loss function in each iteration until the loss function meets the preset cutoff condition, and output the predicted value at this time as the third prediction result.

[0166] Specifically, in step S68, the gradient boosting decision tree is first updated according to the fused residual and the current prediction value calculated in the previous iteration. The fused residual is used as the training target of the new round of decision trees, and the new decision tree reduces the model error by fitting these residuals. For example, assuming that the current prediction value is y_pred and the fused residual is r, the gradient boosting decision tree constructs a new subtree by fitting r, and updates the prediction value through the addition model: y_pred_new = y_pred +η Tree(r). Where η is the learning rate, which is used to control the contribution of each subtree to the model and is set to a small value (e.g., 0.1) to prevent overfitting. After completing a new round of subtree training, the system will repeat the iterative process of steps S63 to S68. In each round of iteration, the gradient boosting decision tree is dynamically adjusted to complete the following steps in sequence: calculate the numerical residual, physical residual, and structural residual, obtain the fused residual by weighted calculation as the next round of training target, fit the fused residual with the new subtree and update the predicted value, and finally evaluate the model performance according to the loss function calculated during this round of iteration. When the loss function changes less than a certain threshold or reaches the maximum number of iterations in multiple consecutive rounds of iterations, the iteration is terminated and the predicted value of the gradient boosting decision tree model at the time of the termination of the iteration is output as the third corrosion prediction result. Through this iterative optimization method, the model can gradually reduce the error, accurately capture the complex characteristics of the inner wall corrosion of the pipeline, and provide reliable decision support for corrosion detection of water conservancy projects.

[0167] The integration of the random forest algorithm and the forward step-by-step algorithm in the gradient boosting tree can not only reduce the risk of overfitting, but also capture the complex nonlinear relationship of corrosion development. The prediction targets include indicators such as corrosion development rate, future corrosion depth, and possible failure time.

[0168] S70: Outputting a pipeline inner wall corrosion detection report of the water conservancy project according to the first corrosion prediction result, the second corrosion prediction result and the third corrosion prediction result.

[0169] Specifically, in step S70, if Fig.10 As shown, the following steps are included:

[0170] S71: Obtain the material type, environmental factors and historical corrosion data of the pipeline to be tested; specifically, the material type refers to the basic constituent material of the pipeline, such as carbon steel, stainless steel, cast iron, copper alloy, etc. Environmental factors refer to the external conditions that affect pipeline corrosion, mainly including temperature, humidity, pH value, salt concentration, redox potential, etc. Historical corrosion data refers to the time series data of corrosion traces, corrosion rate and corrosion degree recorded during the past operation of the pipeline.

[0171] S72: Calculate the root mean square error based on the first prediction result, the second prediction result and the third prediction result; specifically, in step S72, the root mean square error is calculated by comparing the deviation between the prediction results of pipeline inner wall corrosion by three different machine learning models (long short-term memory network LSTM, random forest RF and gradient boosting tree GBT) and the actual observed values. The specific calculation process is as follows: First, obtain the prediction results and actual corrosion depth observations of the three models, then calculate the square error between the predicted value and the actual value of each model, calculate the average of these square errors, and finally take the square root. In this way, the accuracy of the predictions of the three models can be comprehensively evaluated. The smaller the value, the closer the prediction result is to the actual situation, which is helpful to select the best prediction model or perform model integration.

[0172] S73: Calculating the quantitative corrosion risk of the pipeline to be tested according to the material type of the pipeline to be tested, the environmental factors and the historical corrosion data;

[0173] Specifically, in step S73, the formula for quantifying corrosion risk is as follows:

[0174] S = f(M, E, H);

[0175] In the formula, S represents the quantitative corrosion risk, M represents the material type of the pipeline to be tested, E represents the environmental factor, H represents the historical corrosion data, and f represents a functional relationship, which can be a linear weighted model, a nonlinear model, or a model based on machine learning.

[0176] S74: Calculate a theoretical corrosion value according to the historical corrosion data, and calculate a degree of agreement according to the first prediction result, the second prediction result, the third prediction result, and the theoretical corrosion value;

[0177] Specifically, in step S74, the calculation of the theoretical corrosion value is based on the historical corrosion monitoring data of the inner wall of the pipeline, and is performed by linear interpolation and trend extrapolation methods. The specific calculation process is: first collect the corrosion depth data of the inner wall of the pipeline at different time points, draw the corrosion depth-time curve, and use the least squares method to fit the linear corrosion rate model. The theoretical corrosion value calculation formula is: V = V0 + k × t, where V is the predicted corrosion depth, that is, the theoretical corrosion value, V0 is the initial corrosion depth, k is the corrosion rate coefficient, and t is the predicted time. The corrosion rate coefficient k is obtained by linear regression of historical data, which represents the average corrosion depth increment per unit time.

[0178] Specifically, in step S74, the formula for the degree of fit is as follows:

[0179] ;

[0180] Where G represents the degree of fit, P represents the corrosion prediction result, and T represents the theoretical corrosion value.

[0181] Specifically, the corrosion prediction result includes an average value or a weighted average value determined by the first corrosion prediction result, the second corrosion prediction result and the third corrosion prediction result.

[0182] S75: Obtaining a comprehensive corrosion risk index by weighted calculation according to the root mean square error, the quantified corrosion risk and the degree of agreement;

[0183] Specifically, in step S75, the formula of the comprehensive corrosion risk index is as follows:

[0184] CI = root mean square error τ1+ quantifies corrosion risk τ2+ fit τ3;

[0185] Where CI represents the comprehensive corrosion risk index, τ1 represents the root mean square error weight, τ2 represents the quantitative corrosion risk weight, and τ3 represents the fit weight.

[0186] Specifically, the sum of the root mean square error weight, the quantitative corrosion risk weight, and the goodness of fit weight is 1. As a preferred implementation of this embodiment, the root mean square error weight is 0.4, the quantitative corrosion risk weight is 0.3, and the goodness of fit weight is 0.3. The root mean square error weight, the quantitative corrosion risk weight, and the goodness of fit weight are set through cross-validation and historical data. The comprehensive corrosion risk index not only provides numerical predictions, but also integrates risk assessment and model credibility information.

[0187] S76: Outputting a pipeline inner wall corrosion detection report of the water conservancy project according to the comprehensive corrosion risk index.

[0188] Specifically, in step S76, the following steps are included:

[0189] S761: Obtain the corrosion failure probability distribution, and construct a log-normal distribution model using a probability prediction method based on the corrosion failure probability distribution;

[0190] Specifically, in step S761, the lognormal distribution model is as follows:

[0191] ;

[0192] Where μ is the logarithmic space mean, σ is the logarithmic space standard deviation, erf(·) is the error function, X represents the corrosion failure time of a pipeline (random variable), x represents the specific corrosion failure time, and x is a specific value of the random variable X, which is the input value of the lognormal distribution model.

[0193] S762: Establish a risk stratification model based on the lognormal distribution model;

[0194] Specifically, in step S762, the risk grading model is as follows:

[0195] L = f(P(X), J, C);

[0196] Where P(X) is the probability of corrosion failure, J is the comprehensive coefficient of corrosion conditions, C is the potential impact of corrosion, and L is the corrosion risk level.

[0197] Specifically, the comprehensive coefficient of corrosion conditions is generated according to the corrosion depth and the corrosion extension rate. The comprehensive coefficient of corrosion conditions is obtained through comprehensive analysis and quantification of the corrosion depth and the corrosion extension rate, and is used to reflect the overall severity of the corrosion process. Specifically, the corrosion depth is an indicator of the maximum depth of the inner wall material of the pipeline removed by corrosion, which directly reflects the severity of the current corrosion; the corrosion extension rate indicates the rate at which corrosion extends over time, reflecting the dynamic development trend of corrosion. By combining these two indicators, a comprehensive evaluation model can be constructed, such as a weighted average or nonlinear combination, to combine the static influence of the corrosion depth with the dynamic change of the extension rate, and generate a unified comprehensive coefficient of corrosion conditions J, so as to more comprehensively reflect the complexity and changing trend of the corrosion state in risk assessment.

[0198] The potential impact of corrosion, C, is determined by comprehensively evaluating the possible consequences of corrosion. Specifically, it is quantified based on multiple factors, including the use scenario of the pipeline, the sensitivity of the surrounding environment, the importance of the pipeline function, and the safety hazards and environmental impacts caused by corrosion failure. The value of C reflects the severity of the potential consequences of corrosion failure.

[0199] S763: Obtaining the corrosion risk level of the pipeline to be tested according to the risk grading model;

[0200] Specifically, in step S763, the corrosion risk levels include: Level I, Level II, Level III, Level IV and Level V. The risk grading model calculates the corrosion risk level based on the corrosion failure probability P(X), the corrosion situation comprehensive coefficient J and the corrosion potential impact C obtained through actual detection data: L=f(P(X),J,C). The risk levels are divided according to the value of L as follows: Level I (L<1) corresponds to low risk, Level II (1≤L<2) corresponds to lower risk, Level III (2≤L<3) corresponds to medium risk, Level IV (3≤L<4) corresponds to higher risk, and Level V (L≥4) corresponds to high risk. Each risk level is accompanied by a confidence interval to provide more detailed risk assessment information.

[0201] S764: Outputting a pipeline inner wall corrosion detection report of the water conservancy project according to the risk level and the comprehensive corrosion risk index.

[0202] Specifically, this embodiment adopts a probability prediction and risk grading method. The system constructs a corrosion failure probability distribution map and divides the risk into five levels. From negligible risk to high risk level requiring immediate intervention, it provides managers with a clear and intuitive risk assessment. At the same time, it outputs detailed confidence intervals and quantifies the uncertainty of the prediction, helping decision makers better understand and respond to potential corrosion risks.

[0203] Specifically, the analytic hierarchy process can also be used to generate a test report based on the first corrosion prediction result, the second corrosion prediction result, and the third corrosion prediction result. The analytic hierarchy process can systematically decompose the corrosion diagnosis problem into multiple levels and criteria, quantify the qualitative judgment through the pairwise comparison matrix, and finally derive the weight and priority of the diagnostic suggestion. This method can not only improve the interpretability of the diagnostic suggestion, but also continuously optimize the algorithm model and realize the deep integration of artificial intelligence and past experience. In the specific implementation, a pipeline corrosion evaluation system can be constructed, a scientific scoring standard and analytic hierarchy process framework can be designed, the algorithm diagnosis results can be regularly evaluated in multiple dimensions, and the intelligent diagnostic capability of the algorithm engine can be continuously adjusted and improved through the feedback of historical data.

[0204] Specifically, it also includes an adaptive learning mechanism to maintain the real-time and accuracy of the algorithm. Every time new corrosion detection data arrives, the long short-term memory network model, random forest algorithm parameters, gradient boosting tree algorithm parameters and spatiotemporal synchronization fusion parameters will be incrementally learned to adjust the internal parameters.

[0205] The algorithm also introduces an adaptive parameter adjustment mechanism that can dynamically optimize model parameters based on new detection data.

[0206] Specifically, in the parameter adjustment of the LSTM network model, the weight matrices of the forget gate, input gate, and output gate are the key adjustment objects. By analyzing the newly arrived corrosion detection data, the weights of the forget gate, input gate, and output gate are dynamically updated, so that the network can more sensitively capture the subtle changes in the progress of corrosion. By dynamically adjusting the learning rate of the LSTM network, the parameter update step size can be adjusted in real time according to the error of the LSTM network model, ensuring that the network can respond quickly to new features without causing parameter oscillations due to excessive learning rates. The dynamic optimization of the cell state memory vector enables the LSTM network model to track the long-term trend and short-term fluctuations of corrosion development over a long period of time.

[0207] Specifically, the parameters adjusted by the adaptive parameter adjustment mechanism include: the number of random forest decision trees, the maximum tree depth of random forest, the minimum number of split samples, the random selection ratio of features, the learning rate of the gradient boosting tree algorithm, the number of gradient boosting decision trees, the maximum tree depth of gradient boosting and the sub-sample ratio.

[0208] Specifically, in the random forest algorithm, the adaptive parameter adjustment mechanism dynamically adjusts the number of random forest decision trees to balance the bias and variance of the model. The adaptive adjustment of the maximum tree depth of the random forest can control the complexity of a single random forest decision tree and prevent overfitting. The adaptive adjustment of the minimum number of split samples can optimize the splitting strategy of the random forest decision tree to ensure that each node split is statistically significant. The adaptive adjustment of the random selection ratio of features helps to enhance generalization ability and reduce the correlation between features.

[0209] Specifically, in the gradient boosting tree algorithm, by dynamically adjusting the learning rate of the gradient boosting tree algorithm according to the error, rapid learning can be achieved in the early stages of training, and detailed optimization can be performed in the later stages. Adaptive adjustment of the number of gradient boosting decision trees can balance the complexity of the gradient boosting tree algorithm and effectively reduce the computational cost. Adaptive adjustment of the maximum tree depth of gradient boosting can control the complexity of a single gradient boosting decision tree and prevent overfitting. Dynamic adjustment of the subsample ratio can introduce randomness and enhance robustness.

[0210] Specifically, the adaptive learning rate model of the adaptive parameter adjustment mechanism is as follows:

[0211] η(t) = (1 - z t);

[0212] if (e_cur / e_pre) > δ, then η(t) = η(t) k;

[0213] In the formula, is the initial learning rate, z is the error attenuation coefficient; e_cur is the error at the current moment, e_pre is the error at the previous moment, δ is the preset learning threshold, and k is the adaptive factor.

[0214] Specifically, the initial learning rate ∈ [0.01, 0.1], the error attenuation coefficient [0.01, 0.1], the preset learning threshold is used to judge the significance of the error change, and the adaptive factor is used to dynamically adjust the learning rate to prevent the model from falling into the local optimum. The adaptive parameter adjustment mechanism can adjust the learning rate in real time according to the error change trend, thereby improving the model convergence efficiency and prediction accuracy.

[0215] Specifically, the goal of the adaptive parameter adjustment mechanism is to create an intelligent model that can dynamically adapt to complex corrosion environments. By evaluating the model performance in real time, the hyperparameters can be automatically adjusted to keep each algorithm in the best state. The adaptive parameter adjustment mechanism is based on real-time feedback of model errors. The model errors are generated through the detection values ​​of sensor components and the predicted values ​​of each algorithm model. It can quickly respond to new corrosion detection data and continuously optimize the prediction accuracy. At the same time, dynamic parameter adjustment can also effectively balance the bias-variance trade-off of the model, reduce the risk of overfitting, and improve the generalization ability of the model in actual engineering applications. This multi-dimensional, adaptive parameter adjustment strategy can more accurately capture the complex nonlinear characteristics of corrosion development. By intelligently adjusting the key parameters of the random forest algorithm and the gradient boosting tree algorithm, it can automatically adapt to different corrosion environments and detection scenarios, providing more accurate and reliable prediction support for engineering practice.

[0216] The adjustment of spatiotemporal synchronous fusion parameters is a multi-dimensional and multi-scale dynamic optimization process. The weight coefficients of the time dimension and the space dimension will be adjusted in real time according to the latest detection data to balance the importance of features in different dimensions. The dynamic optimization of multi-scale fusion parameters enables the model to capture corrosion features at both micro and macro scales. The real-time adjustment of feature correlation strength can reveal the potential correlation pattern of corrosion development and improve the overall accuracy of the prediction. The dynamic balance of spatiotemporal correlation weights ensures that the model can fully and accurately describe the evolution of corrosion.

[0217] This multi-level, multi-dimensional adaptive parameter adjustment mechanism enables the corrosion prediction model to continuously learn and evolve. By responding to new detection data in real time and dynamically optimizing the internal parameters of the model, the system can always maintain a high degree of prediction accuracy and adaptability, providing intelligent and efficient technical support for corrosion monitoring and prevention. The use of online learning algorithms, such as stochastic gradient descent and adaptive learning rate methods, enables the model to quickly respond to new corrosion development characteristics. At the same time, a model performance tracking system is established to continuously monitor the prediction accuracy and automatically trigger model retraining when performance degradation is detected.

[0218] Embodiment 2

[0219] like Fig.11 As shown, an embodiment of the present invention further provides a pipeline inner wall corrosion detection system suitable for water conservancy projects, comprising:

[0220] A sensor component, wherein the sensor component is disposed in the pipeline to be tested and is used to obtain a number of sensor data;

[0221] A feature extraction module, used to extract features from the plurality of sensor data to obtain a plurality of detection data features;

[0222] A feature fusion module, used for performing spatiotemporal synchronous fusion of the plurality of detection data features to obtain fused features;

[0223] A first prediction module is used to obtain a time series of historical corrosion data of the inner wall of the pipeline, input the time series of historical corrosion data of the inner wall of the pipeline and the fusion feature into a long short-term memory network model, and obtain a first corrosion prediction result;

[0224] A second prediction module, used for obtaining a second corrosion prediction result by using a random forest algorithm according to the fusion features;

[0225] A third prediction module, used for obtaining a third corrosion prediction result by using a gradient boosting tree algorithm according to the fusion feature;

[0226] An output module is used to output a pipeline inner wall corrosion detection report of a water conservancy project according to the first corrosion prediction result, the second corrosion prediction result and the third corrosion prediction result.

[0227] Specifically, the feature extraction module includes: a cleaning unit, which is used to clean a number of sensor data to obtain a number of cleaned data; a classification unit, which is used to classify the number of cleaned data according to sensor type to obtain a number of categories of cleaned data; and an extraction unit, which is used to extract features from each category of cleaned data according to the several categories of cleaned data to obtain a number of detection data features.

[0228] Specifically, the cleaning unit includes:

[0229] The frequency domain analysis subunit is used to perform frequency domain analysis on a number of sensor data by using wavelet transform and Fourier transform, remove random noise and power frequency interference, and obtain sensor data processed by frequency domain analysis;

[0230] The filter subunit is used to set an adaptive filter, which removes the system error and environmental noise in the output data of the frequency domain analysis subunit through the adaptive filter to obtain the sensor data processed by the adaptive filter;

[0231] An outlier elimination subunit is used to remove outliers from the sensor data processed by the adaptive filter by using an outlier detection algorithm to obtain cleaned sensor data;

[0232] The output subunit is used to output the sensor data after cleaning.

[0233] Specifically, the extraction unit includes several extraction subunits, the number of which is the same as the number of sensor types, including an ultrasonic sensor data extraction subunit, an electromagnetic induction sensor extraction subunit and an optical scanning sensor data extraction subunit.

[0234] An ultrasonic sensor data extraction subunit, used for extracting frequency spectrum features in ultrasonic sensor data according to short-time Fourier transform and wavelet packet transform;

[0235] An electromagnetic induction sensor extraction subunit, used for extracting instantaneous frequency and energy features in electromagnetic induction sensor data according to Hilbert-Huang transform;

[0236] The optical scanning sensor data extraction subunit is used to extract corrosion texture features and morphological features in the optical scanning sensor data based on a deep convolutional neural network.

[0237] Specifically, the feature fusion module includes:

[0238] A time synchronization unit, used to dynamically adjust the time of each detection data feature, align the time axis of each detection data feature, and obtain a number of detection data features with aligned time axes;

[0239] A spatial synchronization unit is used to perform coordinate transformation and geometric calibration on each detection data feature to obtain a number of detection data features under the same reference system;

[0240] A Bayesian calculation unit, used to obtain the time series confidence and confidence weight of each detection data feature by using a dynamic Bayesian probability fusion algorithm according to the plurality of detection data features aligned with the time axis and the plurality of detection data features under the same reference system;

[0241] The weighted fusion unit is used to perform weighted calculation according to each detection data feature and its corresponding time series confidence and confidence weight to obtain a fusion feature.

[0242] Specifically, the first prediction module includes: a stacking unit, used to stack a number of long short-term memory network units to form a long short-term memory network model; a data acquisition unit, used to obtain the time series of historical corrosion data of the inner wall of the pipeline; a first prediction unit, used to input the time series of the historical corrosion data of the inner wall of the pipeline and the fusion feature into the long short-term memory network model, and use the gated recurrent unit algorithm to obtain the first prediction result.

[0243] Specifically, the second prediction model includes:

[0244] A random forest training set establishing unit, used for establishing a plurality of random forest training sets according to the fusion features;

[0245] A decision tree building unit, used for building a number of random forest decision trees using a random forest algorithm according to a number of random forest training sets, wherein the number of the random forest decision trees is the same as the number of the random forest training sets;

[0246] A result generating unit, used for generating a number of decision results according to the number of random forest decision trees;

[0247] The second prediction unit is used to vote on a number of decision results according to a voting algorithm to obtain a second corrosion prediction result.

[0248] Specifically, the second prediction unit includes:

[0249] A repeating subunit, used for repeatedly executing the random forest training set establishment unit to the second prediction unit for several rounds, and obtaining the voting result of each round by using a voting algorithm according to several decision results of each round;

[0250] The accuracy calculation subunit is used to calculate the accuracy of each random forest decision tree according to the voting results of each round;

[0251] A screening subunit is used to delete the random forest decision trees whose accuracy is lower than the correct threshold according to the accuracy of each random forest decision tree and a preset correct threshold, so as to obtain several high-quality target random forest decision trees;

[0252] A high-quality target decision subunit, used to generate a number of high-quality target decision results based on a number of high-quality target random forest decision trees;

[0253] The second prediction subunit is used to vote on a number of high-quality target decision results according to a voting algorithm to obtain a second corrosion prediction result.

[0254] Specifically, the third prediction model includes:

[0255] A gradient boosting tree training set establishing unit, used for preprocessing the fusion features and establishing a gradient boosting tree training set according to the preprocessed fusion features;

[0256] Gradient boosting decision tree unit, used to build gradient boosting decision trees;

[0257] A prediction value generating unit, used for inputting the gradient boosting tree training set into a gradient boosting decision tree to generate a prediction value;

[0258] A numerical residual calculation unit, which calculates a numerical residual according to the gradient boosting tree training set and the predicted value;

[0259] Specifically, the numerical residual is the deviation between the input value of the gradient boosting tree training set and the predicted value, and the predicted value is generated according to the input value of the gradient boosting tree training set.

[0260] A loss function calculation unit, used for presetting a corrosion risk weight coefficient and calculating a loss function according to the corrosion risk weight coefficient and the numerical residual;

[0261] The multi-residual calculation unit is used to obtain environmental data in the pipeline to be tested, and calculate physical residuals and structural residuals according to the environmental data, the gradient boosting tree training set and the predicted value.

[0262] A fusion residual calculation unit, used for obtaining a fusion residual by weighted calculation according to the numerical residual, the physical residual and the structural residual;

[0263] The third prediction unit is used to update the gradient boosting decision tree according to the fused residual and the predicted value, repeatedly execute the predicted value generation unit to the fused residual calculation unit, use the corrosion risk weight coefficient and the numerical residual to calculate the loss function in each iteration until the loss function meets the preset cutoff condition, and output the predicted value at this time as the third prediction result.

[0264] Specifically, the multi-residual calculation unit includes:

[0265] A physical residual calculation subunit, used for calculating the physical residual according to the environmental data and the Arrhenius formula;

[0266] The structural residual calculation subunit is used to calculate the feature interaction entropy and the conditional mutual information according to the gradient boosting tree training set and the predicted value, and then calculate the structural residual according to the feature interaction entropy and the conditional mutual information.

[0267] Specifically, the output module includes:

[0268] A data acquisition unit, used to obtain the material type, environmental factors and historical corrosion data of the pipeline to be tested;

[0269] a root mean square error calculation unit, configured to calculate a root mean square error according to the first prediction result, the second prediction result and the third prediction result;

[0270] A quantitative corrosion risk calculation unit, used for calculating the quantitative corrosion risk of the pipeline to be tested according to the material type of the pipeline to be tested, the environmental factors and the historical corrosion data;

[0271] Specifically, the predicted values ​​include a first corrosion prediction result, a second corrosion prediction result, and a third corrosion prediction result.

[0272] A corrosion risk comprehensive index calculation unit, used for obtaining a corrosion risk comprehensive index by weighted calculation according to the root mean square error, the quantified corrosion risk and the degree of agreement;

[0273] The report generating unit is used to output a pipeline inner wall corrosion detection report of the water conservancy project according to the comprehensive corrosion risk index.

[0274] Specifically, the report generation module includes:

[0275] The log-normal distribution model establishment subunit is used to obtain the corrosion failure probability distribution, and the log-normal distribution model is constructed by using the probability prediction method according to the corrosion failure probability distribution;

[0276] A risk grading model building subunit is used to build a risk grading model based on a log-normal distribution model;

[0277] A risk level generating subunit is used to obtain the risk level of the pipeline to be tested according to the corrosion failure probability P(X), the corrosion situation comprehensive coefficient J and the corrosion potential impact C calculated based on the actual detection data, and the risk grading model;

[0278] The report output subunit is used to output a pipeline inner wall corrosion detection report of the water conservancy project according to the risk level and the corrosion risk comprehensive index.

[0279] Specifically, it also includes a parameter adjustment module, which is used to adjust the parameters in the first prediction module, the second prediction module and the third prediction module according to the fusion characteristics, the first corrosion prediction value, the second corrosion prediction value and the third corrosion prediction value.

[0280] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, any one of the above-mentioned methods for detecting inner wall corrosion of a pipeline applicable to a water conservancy project is implemented.

[0281] The present invention also provides an electronic device. The electronic device of the embodiment of the present invention includes: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement a pipeline inner wall corrosion detection method applicable to water conservancy projects provided by the present invention. Fig.12 , which shows a schematic diagram of the structure of a computer system 800 suitable for implementing an electronic device of an embodiment of the present invention. Fig.12 As shown, the computer system 800 includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage part 808 into a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the computer system 800 are also stored. The CPU 801, the ROM 802, and the RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0282] The following components are connected to the I / O interface 805: an input section 806 including a keyboard, a mouse, etc.; an output section 807 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, a modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the I / O interface 805 as needed. A removable medium 811, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 810 as needed so that a computer program read therefrom is installed into the storage section 808 as needed.

[0283] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting inner wall corrosion of a pipeline suitable for a water conservancy project, characterized in that: The following steps are involved: S10: Acquiring a number of sensor data from a sensor assembly disposed in the pipeline to be tested; S20: extracting features from the plurality of sensor data to obtain a plurality of detection data features; S30: Performing spatiotemporal synchronous fusion on the plurality of detection data features to obtain fusion features; S40: Obtain a time series of historical corrosion data of the inner wall of the pipeline, input the time series of historical corrosion data of the inner wall of the pipeline and the fusion feature into a long short-term memory network model, and obtain a first corrosion prediction result; S50: Obtaining a second corrosion prediction result by using a random forest algorithm according to the fusion feature; S60: Obtaining a third corrosion prediction result by using a gradient boosting tree algorithm according to the fusion feature; S70: Outputting a pipeline inner wall corrosion detection report of a water conservancy project according to the first corrosion prediction result, the second corrosion prediction result and the third corrosion prediction result; Wherein, step S60 specifically includes: S61: preprocessing the fused features, and establishing a gradient boosting tree training set according to the preprocessed fused features; S62: Building a gradient boosting decision tree; S63: inputting the gradient boosting tree training set into the gradient boosting decision tree to generate a prediction value; S64: Calculate a numerical residual according to the gradient boosting tree training set and the predicted value; S65: Preset a corrosion risk weight coefficient, and calculate a loss function according to the corrosion risk weight coefficient and the numerical residual; S66: Obtain environmental data in the pipeline to be tested, and calculate physical residuals and structural residuals according to the environmental data, the gradient boosting tree training set and the predicted value; S67: Obtain a fusion residual by weighted calculation according to the numerical residual, the physical residual and the structural residual; S68: updating the gradient boosting decision tree according to the fusion residual and the predicted value, repeatedly executing steps S63 to S67, using the corrosion risk weight coefficient and the numerical residual to calculate the loss function in each iteration, until the loss function meets a preset cutoff condition, and outputting the predicted value at this time as the third corrosion prediction result; Wherein, step S66 specifically includes: S661: Acquire environmental data in the pipeline to be tested, and calculate a physical residual according to the environmental data and the Arrhenius formula; S662: Calculate feature interaction entropy and conditional mutual information according to the gradient boosting tree training set and the predicted value; S663: Calculate a structural residual according to the feature interaction entropy and the conditional mutual information; Wherein, step S70 specifically includes: S71: Obtain material type, environmental factors and historical corrosion data of the pipeline to be tested; S72: Calculating a root mean square error according to the first corrosion prediction result, the second corrosion prediction result, and the third corrosion prediction result; S73: Calculating the quantitative corrosion risk of the pipeline to be tested according to the material type of the pipeline to be tested, the environmental factors and the historical corrosion data; S74: Calculate a theoretical corrosion value according to the historical corrosion data, and calculate a degree of agreement according to the first corrosion prediction result, the second corrosion prediction result, the third corrosion prediction result, and the theoretical corrosion value; The formula for the degree of fit is as follows: ; In the formula, G represents the degree of fit, P represents the corrosion prediction result, T represents the theoretical corrosion value, and the corrosion prediction result includes an average value or a weighted average value determined by the first corrosion prediction result, the second corrosion prediction result, and the third corrosion prediction result; S75: Obtaining a comprehensive corrosion risk index by weighted calculation according to the root mean square error, the quantified corrosion risk and the degree of agreement; S76: Outputting a pipeline inner wall corrosion detection report of the water conservancy project according to the comprehensive corrosion risk index.

2. A pipeline inner wall corrosion detection method suitable for water conservancy projects according to claim 1, characterized in that: Step S20 specifically includes: S21: performing data cleaning on a plurality of sensor data to obtain a plurality of cleaned data; S22: Classifying the plurality of cleaning data according to sensor types to obtain a plurality of types of cleaning data; S23: According to the several types of cleaned data, feature extraction is performed on each type of cleaned data to obtain several detection data features.

3. A pipeline inner wall corrosion detection method suitable for water conservancy projects according to claim 1, characterized in that: Step S30 specifically includes: S31: dynamically adjusting the time of each detection data feature to align the time axis of each detection data feature to obtain a plurality of detection data features with aligned time axes; S32: performing coordinate transformation and geometric calibration on each detection data feature to obtain a number of detection data features under the same reference system; S33: according to the plurality of detection data features aligned with the time axis and the plurality of detection data features under the same reference system, a dynamic Bayesian probability fusion algorithm is used to obtain a time series confidence and a confidence weight of each detection data feature; S34: Perform weighted calculation according to each detection data feature and its corresponding time series confidence and confidence weight to obtain a fusion feature.

4. A pipeline inner wall corrosion detection method suitable for water conservancy projects according to claim 1, characterized in that: Step S40 specifically includes: S41: stacking several LSTM network units to form a LSTM network model; S42: Obtaining the time series of historical corrosion data of the inner wall of the pipeline; S43: Inputting the time series of the historical corrosion data of the inner wall of the pipeline and the fusion features into the long short-term memory network model to obtain a first corrosion prediction result.

5. A pipeline inner wall corrosion detection method suitable for water conservancy projects according to claim 1, characterized in that: Step S50 specifically includes: S51: establishing a number of random forest training sets according to the fusion features; S52: establishing a plurality of random forest decision trees using a random forest algorithm according to a plurality of random forest training sets, wherein the number of the random forest decision trees is the same as the number of the random forest training sets; S53: generating a plurality of decision results according to the plurality of random forest decision trees; S54: Voting on a number of decision results using a voting algorithm to obtain a second corrosion prediction result.

6. A pipeline inner wall corrosion detection system suitable for water conservancy projects, characterized in that: include: A sensor component, wherein the sensor component is disposed in the pipeline to be tested and is used to obtain a number of sensor data; A feature extraction module, used to extract features from the plurality of sensor data to obtain a plurality of detection data features; A feature fusion module, used for performing spatiotemporal synchronous fusion of the plurality of detection data features to obtain fused features; A first prediction module is used to obtain a time series of historical corrosion data of the inner wall of the pipeline, input the time series of historical corrosion data of the inner wall of the pipeline and the fusion feature into a long short-term memory network model, and obtain a first corrosion prediction result; A second prediction module, used for obtaining a second corrosion prediction result by using a random forest algorithm according to the fusion features; A third prediction module, used for obtaining a third corrosion prediction result by using a gradient boosting tree algorithm according to the fusion feature; An output module, used for outputting a pipeline inner wall corrosion detection report of a water conservancy project according to the first corrosion prediction result, the second corrosion prediction result and the third corrosion prediction result; Wherein, the third prediction module includes: A gradient boosting tree training set establishing unit, used for preprocessing the fusion features and establishing a gradient boosting tree training set according to the preprocessed fusion features; Gradient boosting decision tree unit, used to build gradient boosting decision trees; A prediction value generating unit, used for inputting the gradient boosting tree training set into a gradient boosting decision tree to generate a prediction value; A numerical residual calculation unit, which calculates a numerical residual according to the gradient boosting tree training set and the predicted value; A loss function calculation unit, used for presetting a corrosion risk weight coefficient and calculating a loss function according to the corrosion risk weight coefficient and the numerical residual; A multi-residual calculation unit, used for obtaining environmental data in the pipeline to be tested, and calculating physical residuals and structural residuals according to the environmental data, the gradient boosting tree training set and the predicted value; A fusion residual calculation unit, used for obtaining a fusion residual by weighted calculation according to the numerical residual, the physical residual and the structural residual; A third prediction unit is used to update the gradient boosting decision tree according to the fused residual and the predicted value, repeatedly execute the predicted value generating unit to the fused residual calculating unit, calculate the loss function using the corrosion risk weight coefficient and the numerical residual in each iteration until the loss function meets a preset cutoff condition, and output the predicted value at this time as a third corrosion prediction result; Wherein, the multi-residual calculation unit includes: A physical residual calculation subunit, used for calculating the physical residual according to the environmental data and the Arrhenius formula; A structural residual calculation subunit, used to calculate feature interaction entropy and conditional mutual information according to the gradient boosting tree training set and the predicted value, and then calculate the structural residual according to the feature interaction entropy and the conditional mutual information; Wherein, the output module includes: A data acquisition unit, used to acquire the material type, environmental factors and historical corrosion data of the pipeline to be tested; a root mean square error calculation unit, configured to calculate a root mean square error according to the first corrosion prediction result, the second corrosion prediction result and the third corrosion prediction result; A quantitative corrosion risk calculation unit, used for calculating the quantitative corrosion risk of the pipeline to be tested according to the material type of the pipeline to be tested, the environmental factors and the historical corrosion data; A unit for calculating a theoretical corrosion value according to the historical corrosion data, and calculating a degree of agreement according to the first corrosion prediction result, the second corrosion prediction result, the third corrosion prediction result and the theoretical corrosion value; The formula for the degree of fit is as follows: ; In the formula, G represents the degree of fit, P represents the corrosion prediction result, T represents the theoretical corrosion value, and the corrosion prediction result includes an average value or a weighted average value determined by the first corrosion prediction result, the second corrosion prediction result, and the third corrosion prediction result; A corrosion risk comprehensive index calculation unit, used for obtaining a corrosion risk comprehensive index by weighted calculation according to the root mean square error, the quantified corrosion risk and the degree of agreement; The report generating unit is used to output a pipeline inner wall corrosion detection report of the water conservancy project according to the comprehensive corrosion risk index.

7. An electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement a pipeline inner wall corrosion detection method applicable to water conservancy projects as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, a pipeline inner wall corrosion detection method applicable to water conservancy projects as described in any one of claims 1 to 5 is implemented.

Images