Method and System for Identifying Corrosion Status of Wind Turbine Bolt Materials

Abstract

This invention discloses a method and system for identifying the corrosion state of wind turbine bolt materials. The method includes: acquiring raw signals using a ring ultrasonic transducer array in full-matrix acquisition mode; obtaining focused echo signals through full-focusing post-processing; performing wavelet packet decomposition and soft-threshold denoising on the focused echo signals to extract the energy entropy of each sub-band; concatenating the raw time-domain signal, sub-band reconstructed signal, and energy entropy into a multi-channel feature sequence; inputting the sequence into a deep convolutional neural network combined with a channel attention mechanism to extract corrosion feature semantic vectors; acquiring load and environmental data and mapping them to external feature vectors, fusing them with the corrosion feature semantic vectors through tensor product fusion to obtain a dynamic embedding vector; constructing the dynamic embedding vectors of the same bolt at different times as an input sequence, and outputting the corrosion level, residual strength ratio, and remaining life in parallel through a bidirectional long short-term memory network. This invention can improve the accuracy and multi-dimensional quantitative assessment capability of identifying the corrosion state of wind turbine bolts under complex and harsh working conditions.

Description

Technical Field

[0001] This invention belongs to the field of wind power equipment testing technology, specifically a method and system for identifying the corrosion status of wind power bolt materials. Background Technology

[0002] As a critical connecting component of wind turbine generators, the reliability of wind turbine bolts directly affects the overall operational safety and service life of the turbine. Statistics show that a typical wind turbine uses over a thousand high-strength bolts, covering key load-bearing components such as tower flange connections, blade root connections, and nacelle base fixation. During service, these bolts endure combined alternating tensile, bending, and torsional loads, while simultaneously being exposed to harsh environments such as salt spray, humidity, and drastic temperature changes, making them highly susceptible to stress corrosion, corrosion fatigue, and intergranular corrosion. In recent years, with the development of offshore wind power towards deeper waters and larger capacities, bolt corrosion has become increasingly prominent. Offshore wind turbines are subjected to the combined effects of high salt spray, high humidity, and wave impact, resulting in corrosion rates several times higher than onshore turbines. Bolt corrosion has become one of the main causes of unplanned downtime and safety accidents in wind turbine generators.

[0003] Existing corrosion detection technologies for wind turbine bolts generally suffer from insufficient feature extraction capabilities, inability to reflect corrosion evolution mechanisms, and a lack of quantitative assessment methods. Regarding feature extraction, conventional ultrasonic testing methods primarily rely on limited parameters such as echo amplitude and propagation time to determine the corrosion state. However, early-stage corrosion, such as intergranular corrosion and pitting corrosion, is extremely weak in ultrasonic echoes. Single feature quantities are insufficient to effectively separate corrosion-sensitive information from noise and structural reflection interference, resulting in low detection sensitivity and difficulty in issuing early warnings in the early stages of corrosion development. In terms of corrosion mechanism modeling, existing methods typically rely solely on ultrasonic testing results for independent judgment, failing to incorporate load data and environmental condition data into the analysis. In reality, the corrosion process of wind turbine bolts is not simply chemical corrosion, but rather the result of the synergistic effect of the corrosive medium and alternating stress. Ignoring the coupling effect of load and environmental factors leads to a distorted description of the corrosion evolution law, making it difficult to accurately reflect the degradation trend of bolts under actual service conditions. In terms of assessment dimensions, most current detection technologies are still at the stage of qualitative judgment, and can only output a binary conclusion of "corrosion present" or "no corrosion present". They lack the ability to quantitatively predict key performance parameters such as residual strength and remaining life. This qualitative judgment method cannot provide a quantitative basis for replacement for operation and maintenance decisions, often leading to two extreme situations: over-maintenance or untimely maintenance.

[0004] To address the aforementioned issues, scholars have already conducted relevant research. For example, Brandon Mills et al. published a research paper entitled "Inspection of wind turbine bolted connections using the ultrasonic phased array system" in the journal *Heliyon* in 2024 (DOI: 10.1016 / j.heliyon.2024.e34579). This study used ultrasonic phased array detection technology to inspect the bolted connections of wind turbine units, achieving simultaneous monitoring of bolt stress and defect detection through multiple imaging modes. However, this technical solution still has significant shortcomings: First, the method relies solely on ultrasonic echo signals for defect identification, without performing multi-scale decomposition and weak feature enhancement processing on the ultrasonic signals, thus limiting its ability to identify microscale damage such as early pitting corrosion and intergranular corrosion. Second, the method does not incorporate load data and environmental condition data, failing to reflect the synergistic mechanism between the corrosive medium and alternating stress, making it difficult to guarantee the accuracy of assessment under complex service conditions. Third, the method's output still relies on the presence or absence of defects as the basis for judgment, failing to provide quantitative assessment indicators such as corrosion level, residual strength, and remaining life, thus unable to provide direct decision support for the preventive maintenance of wind turbine bolts. Therefore, developing a corrosion identification method for wind turbine bolts capable of enhancing and extracting weak corrosion features, fusing multi-source data, and quantitatively predicting corrosion status and remaining life has become an urgent technical challenge to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for identifying the corrosion state of wind turbine bolt materials, in order to solve the following technical problems mentioned in the background art: Existing technologies rely on single ultrasonic features to extract weak corrosion information; they do not integrate load and environmental data, failing to reflect the coupling effect of corrosion and stress; they can only qualitatively determine the presence or absence of corrosion, lacking the ability to quantitatively predict residual strength and remaining life.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: on the one hand: A method for identifying the corrosion state of wind turbine bolt materials includes the following steps: S1: Acquire the raw ultrasonic signal of the wind turbine bolt in full matrix capture mode using a ring ultrasonic transducer array, and post-process the received signal using a full focusing method to obtain the focused echo signal; S2: Perform wavelet packet decomposition and soft thresholding denoising on the focused echo signal to obtain the reconstructed signal of each sub-band, and extract the energy entropy of each sub-band. Then, concatenate the focused echo signal as the original time domain signal with the reconstructed signal and energy entropy of the sub-band to form a multi-channel feature sequence. S3: Input the multi-channel feature sequence into a deep convolutional neural network and extract the erosion feature semantic vector by combining the channel attention mechanism; S4: Obtain the load data and environmental condition data of the bolts, normalize them and map them into external feature vectors. Then, fuse the corrosion feature semantic vector and the external feature vectors by tensor product and reduce the dimensionality to obtain a dynamic embedding vector. S5: The dynamic embedding vectors of the same bolt at different times are constructed into an input sequence in chronological order, and processed by a bidirectional long short-term memory network to output corrosion level, residual strength ratio and remaining life in parallel.

[0007] Furthermore, in step S1, the spatial coordinates of the array elements of the annular ultrasonic transducer array are determined according to the annular radius and equally spaced angles; the imaging grid step size of the full focusing method is 0.1-0.15 mm, the axial imaging range is 200-300 mm, and the Hilbert envelope is extracted from the A-scan signal before superposition.

[0008] Furthermore, in step S2, the wavelet packet decomposition uses the db8 wavelet basis function, and the decomposition level is 4; the threshold calculation formula for soft threshold denoising is: , in, The noise standard deviation is defined as the noise level within the last third of the bottom echo interval, where N is the signal length; energy entropy. ,in, Let i be the energy of the i-th subband signal. Let i be the energy of the j-th sub-band signal, where i and j are both sub-band indices.

[0009] Furthermore, in step S3, the deep convolutional neural network includes three convolutional layers, a global average pooling layer, two fully connected layers, and a dropout layer, and a channel attention mechanism is introduced before the global average pooling; the output erosion feature semantic vector has a dimension of 512.

[0010] Furthermore, in step S4, the load data includes tensile stress, bending stress, and torsional stress, and the environmental condition data includes salt spray concentration, relative humidity, and temperature; the time synchronization rule is to take the average value within 10 minutes before the ultrasonic detection time, and missing values ​​are interpolated linearly; normalization adopts the min-max method, and the normalization parameter is fixed to the statistical value of the training set; after tensor product fusion, it is sequentially passed through batch normalization layer and three fully connected network for dimensionality reduction, and gradient pruning is used during training.

[0011] Furthermore, in step S5, the length of the input sequence is the result of the last 5 detections, and if it is insufficient, it is padded with zeros; the number of hidden layer units of the bidirectional long short-term memory network is 128; the corrosion level is divided into 0-4 levels according to the maximum corrosion depth and pitting density, where the pitting density is defined as the number of effective connected regions divided by the area of ​​the detection area, and the detection area is a circular area with a center radius of 15 mm on the bolt end face.

[0012] Furthermore, the training data construction steps include: taking bolts from the same batch for accelerated corrosion-fatigue coupling tests, periodically taking out some bolts for ultrasonic testing and destructive tensile tests to obtain real labels, calibrating the residual strength ratio of undamaged bolts through finite element simulation, mixing real label samples and simulation calibration samples at a ratio of 1:2, and assigning loss weights of 1.0 and 0.5 respectively.

[0013] Furthermore, it also includes a 3D reconstruction step: performing a circumferential B-scan on the identified corrosion area, converting the polar coordinates to Cartesian coordinates to construct volume data, extracting corrosion pits through connected component filtering and artifact removal, and calculating the maximum depth, volume, and projected area; the connected component filtering retains a voxel count ≥10 and a volume ≥ For regions with a signal-to-noise ratio ≥ 6dB, the artifact removal rules include an aspect ratio ≤ 3, an axial distance ≥ 2mm and not greater than the effective length of the bolt.

[0014] Two aspects: A corrosion status identification system for wind turbine bolt materials, used to perform the method described in any of the preceding aspects, comprising: Ultrasonic acquisition module: It adopts a ring ultrasonic transducer array, supports full matrix acquisition mode, and has a built-in full-focus imaging unit; Signal preprocessing module: used for wavelet packet decomposition, soft thresholding denoising, and energy entropy extraction; Deep feature extraction module: Based on convolutional neural network and channel attention mechanism, it outputs erosion feature semantic vector; Multi-source data fusion module: used to acquire payload and environmental data, and perform time synchronization, normalization and tensor product fusion; Timing prediction module: Based on a bidirectional long short-term memory network, it outputs corrosion level, residual strength ratio and remaining lifetime.

[0015] Furthermore, it also includes a handheld self-diagnostic probe, which integrates a pulse generator, a low-noise amplifier, an analog-to-digital converter, a wireless communication module, and an inertial measurement unit; the system also includes a sparse deconvolution unit for measuring the transducer impulse response and performing iterative soft thresholding algorithm processing to improve axial resolution.

[0016] Furthermore, it also includes an online monitoring unit, which adopts a PVDF piezoelectric thin film array. The array contains 8×8 electrodes, which are attached to the surface of the pad and cured with acoustic impedance matching silicone adhesive. It is equipped with an independent miniature piezoelectric ceramic transmitter for synchronous triggering and acquisition. The wireless communication adopts a LoRa module with the following parameters: frequency 470MHz, transmit power 10dBm, spreading factor SF=9, bandwidth 125kHz, coding rate 4 / 5, preamble length 12 symbols, and antenna gain 2dBi.

[0017] Furthermore, it also includes a self-diagnostic unit, which performs a self-test every 24 hours, measures the transducer impedance spectrum and extracts the resonant frequency and bandwidth, compares the percentage deviation of the resonant frequency from the nominal value and the bandwidth change, measures the attenuation of the bottom echo amplitude, calibrates the initial value under the conditions of room temperature 20±5℃ and coupling pressure 20N, and corrects the resonant frequency using a temperature compensation formula.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This invention obtains high-quality echo signals through full-matrix acquisition and full-focusing technology, extracts weak corrosion features by combining wavelet packet energy entropy, and then performs tensor product deep fusion of ultrasonic features with load and environmental data. Finally, it uses a bidirectional long short-term memory network to simultaneously output corrosion level, residual strength ratio and remaining life, which significantly improves the accuracy of wind turbine bolt corrosion status identification and multi-dimensional quantitative evaluation capability. It is especially suitable for complex and harsh working conditions such as salt spray and alternating loads. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of the wind turbine bolt material corrosion state identification method of the present invention; Figure 2 This is a schematic diagram of the components of the wind turbine bolt material corrosion state identification system module of the present invention; Figure 3 This is a schematic diagram of the self-diagnosis process of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: This embodiment provides a method for identifying the corrosion state of wind turbine bolts under harsh operating conditions, such as... Figure 1As shown. This method achieves quantitative assessment of bolt health status through deep fusion of multidimensional information and nonlinear mapping. Specifically, it includes the following steps: For ultrasonic signal acquisition, this embodiment employs a 16-channel ultrasonic transducer array in a ring topology. The ring array has a diameter of 30 mm (radius 15 mm), each element is 1.5 mm wide and 10 mm high, and the center-to-center angle between adjacent elements is 22.5 degrees. Let the angle of the i-th element be... Where i = 0, 1, ..., 15, then the spatial coordinates of the array element are... The coordinate unit is millimeters. The transducer center frequency is 5MHz, which balances penetration depth and resolution for M30 and larger bolts with a length of 200mm or more; for M24 and smaller bolts, a center frequency of 10MHz is used. The operating mode is full matrix acquisition mode: each array element is excited sequentially, and all array elements receive simultaneously during each excitation, acquiring 256 groups of A-scan signals in a 16×16 format. The sampling frequency is fixed at 50MHz, which is 10 times the center frequency. The pulse transmission voltage is 200V, the pulse width is 100 nanoseconds, and the receiving gain is 40 dB.

[0022] Since the received raw signal cannot be directly used to form a focused acoustic beam, post-processing using a full-focusing method is required. Specifically, in this embodiment, the imaging area is divided into a two-dimensional grid with an axial depth range of 0 to 200 mm and a radial range of 0 to 25 mm, with a grid step size of 0.1 mm in both the axial and radial directions. The focused signal at each grid point is obtained by superimposing the delayed signals of all transmit-receive pairs. Before superposition, a Hilbert transform is performed on each A-scan signal to obtain the envelope, in order to avoid phase cancellation of the radio frequency signals. For grid point P, its signal amplitude I(P) is calculated using the following formula: in, The A-scan signal is transmitted by the i-th array element and received by the j-th array element. Extraction of Hilbert envelope, The sound wave propagation time from transmitting element i to grid point P and then to receiving element j is calculated using the following formula: In the formula, c is the longitudinal wave velocity in the bolt material, which is taken as 5900 meters per second for 42CrMo steel; , , These represent the spatial coordinates of the transmitting element, receiving element, and grid points, respectively. This algorithm can obtain well-focused echo signals at each depth location.

[0023] In this embodiment, the transducer is coupled and scanned by contacting the bolt end face via a 2mm thick flexible silicone rubber gasket with a Shore hardness of A60. The coupling agent is 500cSt dimethyl silicone oil, with a kinematic viscosity variation of no more than 15% within a temperature range of -40°C to 80°C. The scanning trajectory is controlled by a high-precision two-axis encoder: axial step 1mm and circumferential step 1.5 degrees. The received analog signal is converted into a digital signal by a 16-bit analog-to-digital converter, and the polar coordinates of each acquisition point relative to the center of the bolt end face are recorded, including radius r, angle θ, and axial position z.

[0024] Preferably, after obtaining the raw ultrasound signal, the digitized signal needs to be preprocessed. First, the effective analysis interval is extracted for each A-scan signal. The start point of the effective interval is set 2 microseconds before the initial wave front, and the end point is set 20 microseconds after the bottom echo. The arrival time of the bottom echo is calculated based on the bolt length L and the longitudinal wave velocity c. End time For bolts of unknown length, the position of the bottom echo peak can be automatically detected as a reference. The length of the truncated signal is denoted as... .

[0025] The truncated signal is subjected to 4-level wavelet packet decomposition, with the db8 wavelet basis function selected. The decomposition yields 16 frequency band sub-signals, each with a bandwidth of approximately 3.125MHz. For the high-frequency sub-bands of the 4th level, i.e., those with a center frequency higher than 3.5MHz, a soft thresholding function is used for denoising. The threshold T is calculated using the following formula: in, Let N be the signal length and σ be the standard deviation of the noise in this sub-band. The method for estimating the noise standard deviation is as follows: Take the time interval after the bottom echo in each A-scan signal, i.e., the last third of the interval when the total length is L. This interval contains no valid echoes and can be considered pure noise. Calculate its standard deviation as σ. This method can effectively avoid the initial echo and defect echo regions.

[0026] After denoising, the energy of each sub-band signal is calculated. And calculate the energy entropy H according to the following formula: Energy entropy H, as a primary damage indicator, is directly concatenated into the subsequent input feature vector.

[0027] Then, a semantic vector of erosion features is constructed. Specifically, the following three parts of data are concatenated into a multi-channel one-dimensional sequence: the denoised original time-domain signal, the reconstructed signals from 16 sub-bands, and the energy entropy H, which is copied and expanded to a vector of the same length as the original signal. The total number of channels is 1 + 16 + 1 = 18 channels, and the sequence length is... .

[0028] The 18-channel feature sequence is then input into a deep convolutional neural network. The network structure is as follows: First convolutional layer: kernel size 64×1, stride 2, output channels 32, ReLU activation; Second convolutional layer: kernel size 8×1, stride 1, output channels 64, ReLU activation; The third convolutional layer has a kernel size of 3×1, a stride of 1, 128 output channels, and ReLU activation. Global average pooling: outputs 128-dimensional features; Fully connected layer 1: 128 is mapped to 256, ReLU is activated, and the dropout rate is 0.5; Fully connected layer 2: 256 mapped to 512, ReLU activated, dropout rate 0.5.

[0029] Furthermore, a channel attention mechanism is introduced before global average pooling, employing the SENet structure: the 128-channel feature map output from layer 3 is subjected to global average pooling to obtain a 128-dimensional vector, which is then compressed to 16 dimensions by two fully connected layers and expanded back to 128 dimensions to generate channel weights, which are then multiplied with the original feature map. The final output is a 512-dimensional erosion feature semantic vector. .

[0030] Furthermore, multi-source data dynamic embedding was performed to acquire historical operating load data and environmental condition data of the wind turbine bolts. Load data was collected using a triaxial 45-degree strain gauge attached to the bolt head, with a sampling period of 10 minutes, and tensile stress was extracted. Bending stress Torsional stress The original values ​​ranged from 0-1000 MPa, 0-500 MPa, and 0-300 MPa, respectively. Environmental data included salt spray concentration, relative humidity, and temperature, which were acquired through electrochemical sensors, humidity sensors, and thermocouples, respectively, with original values ​​ranging from 0-5 mg / m³, 0-100%, and -40 to 80 degrees Celsius.

[0031] The rules for time synchronization of multi-source data are as follows: In this embodiment, the time of each ultrasound detection is recorded as _____. For the load and environmental data at that moment, a time interval is taken. The arithmetic mean of all sampled values ​​within a given interval is used as the input data for that detection. If there are fewer than 3 valid sampling points within the interval, the process is extended forward until at least 3 sampling points are obtained. If a sensor has no data within an interval, linear interpolation is used to fill in the gaps using values ​​from adjacent detection times. Normalization is performed using the min-max formula. Parameters and As described above, after normalization, it forms a 6-dimensional vector. The vector is then processed by a 3-layer fully connected network to map its 6-dimensional dimensions to 32-dimensional dimensions, and then to 128-dimensional dimensions, resulting in a 128-dimensional external feature vector. .

[0032] The fusion process in this embodiment uses tensor product plus dimensionality reduction. To avoid gradient instability caused by ultra-high dimensional tensor products, the following measures are taken: First, the tensor product is calculated. The dimension of T is 512 × 128, which equals 65536 dimensions. After flattening T, it is input into a three-layer fully connected network. Batch normalization layers are inserted between each layer, and the output of each layer is batch normalized before being input into the activation function. Gradient clipping is used during training, and the maximum gradient norm is set to 1.0. The optimizer is Adam, with a learning rate of 0.001. The three-layer fully connected network structure is as follows: the first layer maps 65536 to 2048, the second layer maps 2048 to 512, and the third layer maps 512 to 256, with ReLU as the activation function for all layers. The final output is a 256-dimensional dynamic embedding vector Z. As an alternative, a low-rank decomposition method can be used to approximate the tensor product as the Kronecker product of two low-rank matrices to reduce the number of parameters; however, this embodiment directly uses a fully connected approach with gradient stability.

[0033] Furthermore, to quantitatively analyze the correlation between corrosion state and mechanical property degradation, in this embodiment, for each bolt, the dynamic embedding vector obtained from each detection is stored in chronological order. , where t = 1, 2, …, T, and T is the number of historical detections. The input sequence is constructed with a length of 5, representing the 5 most recent detection results. When the number of historical detections is less than 5, leading zeros are padded to ensure the most recent detection result is always at the end of the sequence. For example, with only 3 historical detections, the sequence is: ,in This is the latest test result.

[0034] First, tensor product fusion can fully capture the nonlinear synergistic coupling relationship between corrosion feature semantic vectors and external feature vectors, thus characterizing the synergistic mechanism of corrosion damage, load, and environment. This overcomes the limitation of simple feature splicing, which can only linearly superimpose and cannot reflect the true law of corrosion evolution. Dimensionality reduction after fusion can filter out high-dimensional redundant information generated by tensor product, avoiding parameter explosion and model overfitting, reducing computational burden, and ensuring the stability of subsequent time-series modeling and real-time performance of on-site detection while retaining core interactive features. By fusing before inputting into the time-series network, the bidirectional long short-term memory network models the complete health state of a single bolt detection, rather than fragmented independent feature time sequences. This fully aligns with the bidirectional coupled evolution process of "accelerated corrosion due to changing working conditions and deterioration of bearing capacity due to corrosion damage," avoiding the fragmentation of mechanisms caused by separate time-series modeling and significantly improving the prediction accuracy of time-related indicators. The bidirectional long short-term memory network was chosen because its bidirectional gating structure can simultaneously learn the forward temporal dependence of corrosion evolution from history to the present, as well as the reverse dependence of inferring historical evolution from the present. This is suitable for the field conditions of wind turbine bolts, which have long service cycles, low inspection frequency, and limited historical data. Its unique gating mechanism can selectively retain key corrosion damage features and filter redundant noise, effectively solving the gradient vanishing problem of conventional recurrent neural networks in long-term modeling. It can accurately capture the long-term cumulative effect of early micro-corrosion, and support multi-task parallel end-to-end output of corrosion level classification, residual strength ratio regression, and remaining life prediction. This ensures the internal logical consistency of multi-dimensional evaluation results, simplifies the field deployment process, and improves the overall efficiency of detection and evaluation.

[0035] The sequence is input into a bidirectional long short-term memory network (BiLSTM), with the following structure: BiLSTM layer: 128 hidden units, including 64 units in the forward direction and 64 units in the backward direction, with a sequence input dimension of 256; Fully connected layer 1: 256 is mapped to 128, ReLU is activated, and the dropout rate is 0.3; Three parallel output heads: Classification header: 128 is mapped to 5, softmax activation is used, and the erosion level is output; The regression head 1:128 is mapped to 1, sigmoid activation is used, and the residual intensity ratio is output; The regression head 2:128 is mapped to 1, ReLU activation is used, and the remaining lifetime is output in months.

[0036] The quantitative criteria for determining the corrosion level are as follows: Grade 0 (No Corrosion): Maximum corrosion depth less than 0.05 mm; Grade 1 (Surface Pitting): Maximum corrosion depth is between 0.05 mm and 0.2 mm, and the pitting density is less than 10 pits per square centimeter. The pitting density is calculated as follows: In the 3D reconstructed volume data, this embodiment uses a six-neighbor connectivity labeling algorithm to identify independent pitting connected regions, retaining only connected regions with a voxel count greater than or equal to 3 and a signal-to-noise ratio greater than 6 dB as valid pitting. The detection area is a circular region with a center radius of 15 mm on the bolt end face. Pitting density = number of valid pitting connected regions / area of ​​the detection area, which is 7.07 square centimeters.

[0037] Level 2: The maximum corrosion depth is between 0.2 mm and 0.5 mm, or the depth of a single corrosion pit is greater than 0.5 mm but the projected area is less than 5 square millimeters; Level 3: Metallographic examination shows that the proportion of grain boundary cracks is greater than 20%, or it can be determined according to the following ultrasonic backscattering method. The formula for calculating the ultrasonic backscattering enhancement rate is as follows: Take the time window in each A-scan signal that is before the bottom echo and after the initial wave. The starting point of the time window is 2 microseconds after the end of the initial wave. The end point of the initial wave is defined as follows: Perform Hilbert envelope extraction on the A-scan signal, search forward from the peak position of the initial wave to the starting point of the rising edge of the envelope, and then search backward to the position where the envelope drops to 10% of the peak amplitude. This position is the end point of the initial wave. The time window length is 10 microseconds. Calculate the root mean square value of the signal within this window. The root mean square value within the same time window of the uncorroded area of ​​the same bolt or the same batch of uncorroded bolts was used as the baseline. The backscattering enhancement rate is calculated using the following formula: when When the corrosion rate is ≥30%, it is classified as intergranular corrosion level 3.

[0038] Level 4: The corrosion depth is greater than 10% of the bolt diameter, or a through-hole defect has been confirmed by 3D reconstruction.

[0039] The model training employs an end-to-end joint optimization strategy. The total loss function... The weighted sum of the classification loss and the two regression losses is shown in the following formula: in, =0.5, =0.1. During backpropagation, the total loss is gradient-wise with respect to all trainable parameters of the CNN, external feature extraction network, fusion network, and BiLSTM, and updated simultaneously. The optimizer is Adam, with an initial learning rate of 0.001, weights initialized using a He normal distribution, and a batch size of 32. The learning rate uses an exponential decay strategy, multiplying by 0.95 every 10 epochs. The training consists of 150 epochs, with an early stopping strategy where training stops and the system rolls back to the optimal weights if the validation set loss does not decrease for 15 consecutive epochs. Gradient clipping is enabled during training, and the maximum gradient norm is set to 1.0.

[0040] Regarding the construction of training data, this embodiment adopts the following hybrid approach. All accelerated corrosion tests were performed in accordance with ASTM B117 salt spray test standard and ASTM E466 constant amplitude fatigue test standard.

[0041] Take 120 wind turbine bolts from the same batch, made of 42CrMo with a strength grade of 10.9, numbered 1 to 120.

[0042] Furthermore, all bolts were placed in an accelerated corrosion-fatigue coupled testing device, which employed a servo fatigue loading module built into the salt spray chamber. Salt spray conditions: 5% NaCl solution, pH 6.5 to 7.2, temperature 35°C, continuous spraying, with the solution replaced every 72 hours. Alternating load: tensile stress amplitude 300 MPa, stress ratio R=0.1, waveform sine wave, frequency 1 Hz, 8 hours per day, total test period 56 days. Bolts were mounted inside the salt spray chamber using specialized clamps, exposing their heads to the spray while simultaneously bearing axial tensile fatigue load. Loading was paused every 7 days, and 10 bolts were removed for ultrasonic testing. After removal, they were cleaned with deionized water and dried at room temperature for 2 hours to obtain ultrasonic signals and the dynamic embedding vector Z.

[0043] Ten bolts were subjected to destructive tensile tests to determine their true residual strength ratio and remaining life. Remaining life is defined as the number of remaining fatigue cycles from the current corrosion state to fracture, determined through constant amplitude fatigue tests under the same conditions as the aforementioned alternating load. To convert the number of cycles to months, the Miner linear accumulation criterion was used. First, based on the historical SCADA data of the wind farm, the annual average equivalent fatigue cycle count was calculated to be 2.5 × 10^6 cycles, corresponding to the DLC1.2 load condition in the standard load spectrum IEC61400-1. This value has been converted according to the Miner criterion to an equivalent stress amplitude of 300 MPa for a single-stage constant amplitude load. The conversion formula is: ,That This represents the remaining number of cycles measured during constant amplitude fatigue testing. This conversion is valid only if the equivalent stress amplitude of the field load spectrum matches the test stress amplitude; if applied to other wind fields, recalibration is required.

[0044] Preferably, for the remaining bolts that are not damaged, their residual strength ratio is calibrated through finite element simulation. Specifically, a linear elastic fracture mechanics model is adopted, and the corrosion pits obtained by ultrasonic 3D reconstruction are equivalent to semi-elliptical surface cracks. The equivalence rules are as follows: In the mesh of the corrosion pit surface in 3D reconstruction, connected component screening and artifact removal are performed first. Connected component screening adopts the six-neighbor connected component labeling algorithm, and only connected components with more than 10 voxels and a volume greater than 0.5 cubic millimeters are retained. The artifact removal rules are as follows: Calculate the aspect ratio of each connected component. If the aspect ratio is greater than 3, it is regarded as a strip artifact and removed; calculate the axial distance from the geometric center of the connected component to the bolt end face. If the distance is less than 2 mm or greater than the effective length of the bolt, it is regarded as near-field interference or end face reflection artifact and removed; calculate the maximum signal-to-noise ratio of voxels in the connected component. The signal-to-noise ratio is defined as the ratio of the voxel amplitude to the root mean square of the background noise. Only voxels with a signal-to-noise ratio greater than 6 dB are retained. After screening, find the voxel with the largest depth in the retained connected components; its depth value is 'a'. Moving along the bolt axis from the point of maximum depth to both sides until the corrosion pit depth drops to 10% of 'a', the distance between these two boundary points is 'c'. If multiple scattered pits exist, take the point of maximum depth and its corresponding 'c'. If the distance between multiple pits is less than 5 times the average depth, merge them into an equivalent crack; 'a' is the maximum depth of each pit, and 'c' is the half-width of the merged envelope. The length-to-depth ratio defaults to a / c = 0.5; this value is used when the measured 'c' is unavailable. Residual strength ratio. Calculate using the following formula: in, The original tensile strength of the material is taken as 1080 MPa for 42CrMo; The strength reduction caused by corrosion is calculated using the following formula: In the formula, For fracture toughness, take 60. Y is the geometric correction factor; SF is the safety factor, taken as 1.5. The model is then corrected against subsequent destructive testing results: the theoretical attenuation calculated in the above formula is linearly regressed with the measured attenuation to obtain the correction coefficient. It equals the measured value divided by the theoretical value. Based on data from 120 bolts... The mean was 0.92 and the standard deviation was 0.07. Subsequent simulation calibrations were multiplied by k_s to keep the calibration accuracy error within 5%.

[0045] Approximately 600 samples with real labels were obtained from damaged bolts, and approximately 1400 samples with simulation calibration labels were obtained from undamaged bolts, totaling 2000 samples. During training, different loss weights were assigned to the two classes of samples: real-label samples had a weight of 1.0, and simulation calibration samples had a weight of 0.5. Sampling was performed by class balancing, with the ratio of real samples to simulation samples in each batch being 1:2 to match the natural proportion of the two classes in the dataset. The data was then divided into training, validation, and test sets in an 8:1:1 ratio.

[0046] This construction method has been validated in the laboratory, with a mean residual strength error of 4.2% on the test set. Preferably, in one model output and application, the prediction results are transmitted to a remote monitoring center via an industrial gateway using the Modbus TCP protocol for the purpose of developing a maintenance plan.

[0047] Furthermore, the method in this embodiment also includes preprocessing rules for model deployment and inference. When the model training is completed and deployed to the field inference system, the parameters for preprocessing the real-time acquired ultrasonic signals and environmental data must be completely consistent with those in the training phase and must not be recalculated. Specifically, this includes: effective interval truncation rules, wavelet packet decomposition parameters, soft threshold denoising calculation formula, energy entropy calculation method, and 18-channel feature splicing order. Normalization parameters use fixed min and max values ​​statistically derived from the training set: tensile stress 0 and 1000 MPa, bending stress 0 and 500 MPa, torsional stress 0 and 300 MPa, salt spray concentration 0 and 5 mg / m³, relative humidity 0 and 100%, and temperature -40 and 80 degrees Celsius. All inference inputs are normalized using these fixed values ​​and are not recalculated online.

[0048] This embodiment also provides a three-dimensional reconstruction function. For the identified critical corrosion areas, a circumferential B-scan method is used: the transducer steps 1 degree circumferentially along the bolt, and an axial line scan is collected at each circumferential position. From the center of the end face, the radial step is 0.5 mm outward, and 512 sampling points are collected at each position. The collected data is stored in polar coordinates: radial distance r, circumferential angle θ, axial depth z, and signal amplitude A.

[0049] The coordinate transformation uses the following formula to convert polar coordinates to Cartesian coordinates in order to construct volume data: Furthermore, all acquisition points are transformed to form a sparse point cloud. Three-dimensional volumetric data is reconstructed on a regular grid using cubic spline interpolation. The voxel size of the regular grid is 0.5 mm × 0.5 mm × 0.2 mm. During volumetric data stitching, overlapping areas between adjacent circumferential scans are merged using a weighted average. The weight is inversely proportional to the confidence level of the scan position, which is obtained by normalizing the bottom echo amplitude. The surface mesh of the corrosion pits is extracted using the Marching Cubes algorithm, with the isosurface threshold set to 40% of the bottom echo amplitude. This threshold is calibrated as follows: for 10 bolts with different corrosion levels, 3D reconstructions are performed using different thresholds and compared with the measured corrosion pit boundaries using a metallographic microscope. The threshold that minimizes the average boundary error is selected and ultimately determined to be 40% of the bottom echo amplitude. After extraction, the maximum depth, volume, and projected area of ​​the corrosion pits are calculated. The connected component filtering and artifact removal rules have been detailed in point 4 of the training data construction method and will not be repeated here.

[0050] In this embodiment, before on-site testing, the bolt end face is cleaned with anhydrous ethanol to remove oil and rust. Then, 0.5 ml of dimethyl silicone coupling agent is evenly applied to a circular area with a diameter of 30 mm at the center of the bolt end face. The transducer gasket is then attached to the end face, and an axial pressure of 15 to 20 Newtons is applied to compress and deform the gasket by approximately 0.5 mm. During testing, if the ambient temperature is below -20 degrees Celsius, the transducer must be preheated to above 0 degrees Celsius before use; in rainy or snowy weather, the testing area must be covered with a waterproof cover, and the end face must be dried with a heat gun before testing. After each test, residual oil is wiped off with a non-woven cloth, and the gasket surface is checked for scratches or aging cracks.

[0051] Example 2: This embodiment is based on Embodiment 1 and makes adaptive improvements for large-capacity offshore wind turbines. Large-capacity offshore wind turbines refer to single-unit capacities of 10 MW and above, with bolt specifications of M42 and above.

[0052] For signal acquisition, a 32-channel parallel acquisition system is used, with each channel having an independent analog-to-digital converter, 14-bit accuracy, and a sampling rate of 100 MHz. The transducer housing is made of 316L stainless steel and encapsulated with epoxy resin. The flexible gasket is made of hydrogenated nitrile butadiene rubber (HNBR) with a thickness of 3 mm. Scan step: 3 mm axially, 15 degrees radially.

[0053] The 32 elements of the ring array are evenly distributed, and the ring radius R = 20 mm. The spatial coordinates of the i-th element are calculated using the following formula: in, The TFM imaging grid step size is adjusted to 0.15 mm, and the imaging axial depth range is extended to 300 mm. To compensate for the attenuation of the sound beam during long-path propagation, distance gain compensation is applied to each A-scan signal before TFM superposition. The compensation amount is calculated to increase by 2 dB for every 10 mm of depth, with a maximum compensation not exceeding 40 dB.

[0054] Furthermore, in the signal preprocessing stage, a 50 Hz notch filter with a quality factor Q=10 is added before wavelet packet decomposition. The dynamic thresholding uses the local variance method: a sliding window of 256 points, with the window moving 64 points at a time. The local standard deviation is calculated within each window. The threshold is set to 3 times. The window starts at the first 200 points of the signal, i.e., the initial wave region, to ensure that noise estimation is not affected by the initial wave. After extracting the energy entropy of each sub-band, the mutual information entropy between sub-bands is calculated as a supplementary feature.

[0055] Preferably, for the feature extraction network, the CNN input channels are expanded to 32, containing the original signal and 31 sub-band reconstructed signals. The size of the first convolutional kernel is changed to 128×1, and the rest is the same as in Example 1. The output semantic vector dimension is 1024.

[0056] Preferably, in terms of multi-source data fusion, the output power, speed, and pitch angle of the wind turbine are obtained from the SCADA system, and the composite stress is calculated by combining the strain gauge data. Environmental data includes seawater temperature, flow velocity, and dissolved oxygen concentration, with dissolved oxygen concentration in milligrams per liter (mg / L), and original ranges of 0-30 degrees Celsius, 0-3 meters per second, and 0-15 mg / L, respectively. The external feature vector dimension is increased to 256. The fusion network is adjusted as follows: after tensor product, the dimension is 1024 × 256, equal to 262144 dimensions, with four fully connected layers. Batch normalization layers are also inserted between each layer, and the gradient clipping norm is 1.0. The network structure is as follows: the first layer maps 262144 to 4096, the second layer maps 4096 to 1024, the third layer maps 1024 to 512, and the fourth layer maps 512 to 256, outputting a dynamic embedding vector.

[0057] Preferably, the BiLSTM has 256 hidden layer elements for time-series prediction and fracture mechanics assessment. The output layer increases the range of stress intensity factors. The estimated value, in units of Used to assist in assessing fracture risk Calculate using the following formula: in, For equivalent stress, the root mean square values ​​of tensile and bending stresses are used; 'a' represents the pit depth, obtained from 3D reconstruction; and 'Y' is the geometric correction factor. For surface cracks at the root of cylindrical threads, the Newman-Raju formula is used: In the formula, c is the half-width of the corrosion pit, defined and extracted using the same equivalent rule as in the training data construction of Example 1; h is the radius of the thread root fillet, which is 1.2 mm for M42 bolts, and can be found in mechanical design manuals or measured for other specifications. Exceeding the material's fracture toughness For 42CrMo, take 60 Immediately issue a replacement alert.

[0058] In this embodiment, the transfer learning strategy is to lock the first three layers of the CNN and the BiLSTM layer, and only retrain the fusion network and output head. The specific hyperparameter settings for transfer learning are as follows: initial learning rate 0.0001, which is one-tenth of the source domain training rate; batch size 16; 50 training epochs; early stopping strategy: stop when the validation set loss does not decrease for 10 consecutive epochs; gradient clipping norm is kept at 1.0. The required target domain sample size does not exceed 50 sets, i.e., a small number of measured bolt samples from the new wind farm.

[0059] Optionally, regarding vibration compensation, a triaxial accelerometer with a range of ±5g is integrated into the transducer. When the effective value of vibration acceleration is detected... When the amplitude exceeds 0.1g, delayed sampling is triggered: the system monitors the vibration acceleration signal and selects a time window near its zero-crossing point with an amplitude less than 0.05g to trigger acquisition. If a time window meeting the condition cannot be found for 3 consecutive seconds, the threshold is lowered to 0.15g and a vibration interference marker is recorded. For acquired signals with vibration interference markers, a phase-weighted suppression method is used for post-processing. Let the A-scan signal acquired in this instance be s(t), and the effective vibration acceleration value be... The weighting coefficient w is defined as follows: Where k is the sensitivity coefficient, which has been experimentally calibrated to be... Multiply the phase spectrum of the original signal by this weighting coefficient, i.e. , where S(f) is the Fourier transform of s(t). This operation can suppress the influence of random phase jitter caused by vibration on coherent superposition. The processed signal is then used for subsequent feature extraction.

[0060] Example 3: This embodiment provides an integrated handheld self-diagnostic system, such as... Figure 3 As shown.

[0061] In this embodiment, the handheld probe incorporates a pulse generator, a low-noise amplifier, a 16-bit analog-to-digital converter, a Wi-Fi module, a three-axis gyroscope, and an STM32H7 main controller. The pulse generator outputs 200-volt pulses with a 100-nanosecond width. The low-noise amplifier has a gain of 60 dB and a bandwidth of 1 to 20 MHz. The Wi-Fi module supports the 802.11ac protocol. The three-axis gyroscope is an MPU6050. The probe measures 45 mm in diameter and 100 mm in length, weighing 0.8 kg. It is powered by two 18650 lithium batteries connected in series, with a voltage of 7.4 volts and a capacity of 2600 mAh, providing 4 hours of continuous operation and 72 hours of standby time.

[0062] In terms of signal processing, a sparse deconvolution algorithm is introduced to improve axial resolution. First, the impulse response of the transducer is measured. During measurement, a steel block with acoustic properties consistent with the bolt material being measured should be used, specifically 42CrMo alloy steel or other steel with a sound velocity error of less than 1%. The steel block should be 50 mm thick, with a bottom surface parallelism of less than 0.01 mm and a surface roughness Ra ≤ 1.6 μm. The bottom surface echo is collected from this steel block, and this echo signal is used as the initial impulse response. .right The following preprocessing is performed: DC component is removed, its mean is calculated and subtracted from the signal; signal energy is normalized to 1; peak position is detected and aligned to zero time; a Hamming window is applied to reduce sidelobes, with a window length of 80% of the signal length. The preprocessed signal is denoted as h(t), serving as the impulse response. The measured signal y(t) is considered as the convolution of the defect reflection sequence x(t) and h(t) with noise, i.e. The iterative soft thresholding algorithm ISTA is used to solve the problem. The convergence condition during iteration is set to the relative change of the solution between two adjacent iterations being less than 0.001, and the maximum number of iterations is set to 50. Regularization coefficient. This value is determined by performing a mesh search on signals from 10 bolts with different corrosion levels, minimizing the mean square error of the reconstructed signal. The final sparse solution is output when the iteration stops. This algorithm can improve the axial resolution from 2 mm to 0.3 mm, making it possible to distinguish continuous pitting pits with a spacing of less than 0.3 mm.

[0063] Preferably, this embodiment also includes a time-frequency graph branch, which performs continuous wavelet transform on the original signal using Morlet wavelets with a center frequency of 5 MHz and a scale of 32, generating a 128×128 pixel time-frequency graph. This time-frequency graph is input into a simplified CNN containing two convolutional layers: the first layer is a 3×3 convolution with 16 output channels; the second layer is a 3×3 convolution with 32 output channels; followed by global average pooling, outputting 128-dimensional features. This branch is jointly trained end-to-end with the main network: the output features of the time-frequency graph branch are concatenated with the 512-dimensional features of the main path along the channel dimension, and the resulting 640-dimensional features are fed into the subsequent fusion network. The total loss function calculates the gradient of all parameters of both branches simultaneously and updates them. Before training, the output features of both branches are L2 normalized to normalize the magnitude of each feature to 1, balancing the feature scales of the two branches. The same optimizer and learning rate settings as in Embodiment 1 are used during training.

[0064] Optionally, this embodiment also includes a geographic information system (GIS) auxiliary function, automatically matching the regional corrosion database based on GPS coordinates. The coastal wind field is divided into four zones, with salt spray deposition rates of 0.2, 0.5, 0.8, and 1.2 mg / cm² / day for each zone. The baseline corrosion rate for this region is used as prior knowledge and embedded additively into the intermediate layer of the fusion network. Specifically, a baseline corrosion rate encoding vector is added to the 512-dimensional output of the second fully connected layer of the fusion network. This encoding vector is obtained by mapping from the salt spray deposition rate through an independent fully connected layer, also with a dimension of 512.

[0065] Maintenance decision output: When the residual strength ratio is greater than 0.95, it is judged to be in a green normal state, and the inspection cycle is 12 months. When the residual strength ratio is between 0.90 and 0.95, it is judged as a yellow alert status, and the inspection cycle is shortened to 6 months; When the residual strength ratio is between 0.85 and 0.90, it is considered an orange alert status, and replacement is recommended within 3 months; If the residual strength ratio is less than 0.85, it is considered a red danger state, and the machine should be stopped and replaced immediately.

[0066] This embodiment employs a PVDF piezoelectric film array with a thickness of 200 micrometers, containing 64 electrodes in an 8×8 pattern. Each electrode is a square with a side length of 2 millimeters, and the electrode spacing is 0.5 millimeters. The piezoelectric constant d33 of the PVDF material is 25 picocoulombs per Newton, which ensures the generation of approximately 2.5 picocoulombs of charge at a typical ultrasonic echo pressure of 0.1 MPa, resulting in a millivolt-level signal after charge amplification. The array is adhered to the gasket surface using an acoustic impedance-matching silicone adhesive, specifically Dow Corning Q3-6575. This adhesive has an acoustic impedance of approximately 3 MRayle after curing, providing good matching with the PVDF and HNBR gasket. The adhesion process involves uniformly coating the gasket surface with an adhesive thickness of approximately 0.1 millimeters, applying a pressure of 0.1 MPa after attaching the PVDF array, and curing at room temperature for 24 hours. The gasket material is HNBR, 1 millimeter thick, and the array transmits acoustic pressure to the bolt end face via the gasket. The transmitter is an independent miniature piezoelectric ceramic sheet with a center frequency of 5 MHz and a diameter of 3 mm, attached to the center of the array. The transmit pulse voltage is 100 volts, the pulse width is 100 nanoseconds, and the transmit frequency is matched with the center frequency of the receive array. Synchronous triggering method: After the system wakes up every 6 hours, the main control outputs a 5-volt TTL trigger signal, simultaneously activating the transmit circuit and 64-channel parallel acquisition, with a trigger delay of less than 1 microsecond. During reception, each electrode of the PVDF film independently outputs a charge signal, which is converted into a voltage by a charge amplifier. The charge amplifier uses a negative feedback circuit based on an operational amplifier, with a feedback capacitor of 10 picofarads, an input impedance of 10 gigahertz, and a magnification factor determined by the feedback capacitor, approximately 100 times. The filtering bandwidth is 1 to 10 MHz. The same dimethyl silicone oil coupling agent as in Example 1 is applied between the transmitter and the PVDF array, and between the PVDF array and the bolt end face, with a thickness of approximately 0.2 mm.

[0067] The core parameters of the LoRa wireless communication module in this embodiment are set as follows: operating frequency 470 MHz, transmit power 10 dBmW, spreading factor SF=9, bandwidth 125 kHz, coding rate 4 / 5, and preamble length 12 symbols. The data packet format is: 2-byte CRC16 checksum followed by a variable-length payload, with no retransmission mechanism. The antenna is a spring antenna with a gain of 2 dB. With these parameters, the communication distance in open environments can reach 1.5 km, meeting the online monitoring requirements inside wind farm towers.

[0068] Preferably, the system wakes up every 6 hours, transmits a pulse, and then performs parallel data acquisition across 64 channels, taking 2 seconds. The acquired data is compressed and transmitted via a LoRa module, immediately entering deep sleep mode after transmission. Power is supplied by two CR3032 lithium-manganese batteries connected in parallel, with a total capacity of 1000 mAh and a self-discharge rate of less than 1% per year. The system is powered by a low-power power management chip, TPS62740, with a sleep current of less than 5 microamps. Calculations show an average daily power consumption of approximately 0.25 mAh, and the two CR3032 batteries can support approximately 3.5 years of battery life.

[0069] This embodiment also provides a marine correction coefficient lookup table, which is pre-calculated based on computational fluid dynamics (CFD) simulation results. Flow velocity v, salinity S, and dissolved oxygen DO are used as independent variables, and the correction coefficient k is the dependent variable. The complete value range is: flow velocity v from 0 to 5 m / s, with a step size of 0.5 m / s; salinity S from 25‰ to 40‰, with a step size of 2.5‰; and dissolved oxygen DO from 0 to 15 mg / L, with a step size of 1 mg / L. A total of 11 × 7 × 16 = 1232 grid points are generated. The correction coefficient for each grid point is obtained through CFD simulation. The simulation boundary condition is a seawater temperature of 15 degrees Celsius. The k-ε turbulence model is used, and the Butler-Volmer equation is adopted for the electrochemical corrosion model. Example data is shown in Table 1: Table 1 In practical applications, the system acquires v, S, and DO in real time. When the input value exceeds the minimum or maximum value defined in the table, truncation is performed: the input value is restricted to the range of the table boundary values, and a warning log is logged. For values ​​within the boundary, three-dimensional linear interpolation is used to obtain k. The predicted corrosion depth is multiplied by k to obtain the compensated predicted depth. This lookup table method has low computational cost and is suitable for embedded real-time implementation.

[0070] This embodiment also has a self-diagnostic function, specifically, as follows: Figure 3 As shown, a self-test program is automatically executed every 24 hours. The self-test should be performed at room temperature (20±5 degrees Celsius) and a coupling pressure of 20 Newtons between the transducer and the end face of a defect-free bolt to calibrate the initial values. The transducer impedance spectrum is measured, with a frequency sweep range of 1 to 10 MHz and a step size of 0.1 MHz, to extract the resonant frequency. And a bandwidth of -3 dB Δf. If If the deviation from the nominal value by more than ±5% within 5 MHz, or if Δf is greater than 1.5 times the nominal value, the piezoelectric ceramic sheet is considered to have cracks or aging. The effect of temperature on the resonant frequency is compensated by the following formula: Where T represents the current temperature. Simultaneously, the bottom surface echo amplitude is measured. The initial amplitude was measured under the same temperature and pressure conditions during the initial installation. In comparison, if Divide by If the value is less than 0.5, the flexible coupling gasket is considered to have failed or the coupling agent has dried out. If the self-test fails, the system generates a report containing fault codes and suggested repair solutions. Fault code E01 indicates a transducer failure, and E02 indicates a coupling failure. The report is sent to the remote maintenance team via LoRa.

[0071] In practice, an M36 bolt of grade 10.9, which had been in service for 6 years, was inspected at an offshore wind farm. The system output a maximum corrosion depth of 1.21 mm, with a residual strength ratio of 0.876, representing a decrease of 12.4%. After disassembling the bolt, the maximum corrosion depth was measured using a metallographic microscope to be 1.17 mm, with an error of 3.4%; the laboratory tensile test showed a residual strength ratio of 0.869, representing a decrease of 13.1%. The two measurements are consistent, verifying the effectiveness of the proposed method.

[0072] Example 4: A corrosion status identification system for wind turbine bolt materials, such as Figure 2 As shown, it includes: Ultrasonic acquisition module: It adopts a ring ultrasonic transducer array, supports full matrix acquisition mode, and has a built-in full-focus imaging unit; Signal preprocessing module: used for wavelet packet decomposition, soft thresholding denoising, and energy entropy extraction; Deep feature extraction module: Based on convolutional neural network and channel attention mechanism, it outputs erosion feature semantic vector; Multi-source data fusion module: used to acquire payload and environmental data, and perform time synchronization, normalization and tensor product fusion; Timing prediction module: Based on a bidirectional long short-term memory network, it outputs corrosion level, residual strength ratio and remaining lifetime.

[0073] In a further optimized embodiment, this embodiment also includes a handheld self-diagnostic probe, which integrates a pulse generator, a low-noise amplifier, an analog-to-digital converter, a wireless communication module, and an inertial measurement unit; the system also includes a sparse deconvolution unit for measuring the transducer impulse response and performing iterative soft thresholding algorithm processing to improve axial resolution.

[0074] Further optimized, this embodiment also includes an online monitoring unit, which adopts a PVDF piezoelectric thin film array. The array contains 8×8 electrodes, which are attached to the surface of the pad and cured with acoustic impedance matching silicone adhesive. It is equipped with an independent miniature piezoelectric ceramic transmitter for synchronous triggering and acquisition. The wireless communication adopts a LoRa module with the following parameters: frequency 470MHz, transmit power 10dBm, spreading factor SF=9, bandwidth 125kHz, coding rate 4 / 5, preamble length 12 symbols, and antenna gain 2dBi.

[0075] In a further optimized embodiment, this embodiment also includes a self-diagnostic unit, which performs a self-test every 24 hours, measures the transducer impedance spectrum and extracts the resonant frequency and bandwidth, compares the percentage deviation of the resonant frequency from the nominal value and the bandwidth change, measures the attenuation of the bottom echo amplitude, calibrates the initial value under the conditions of room temperature 20±5℃ and coupling pressure 20N, and corrects the resonant frequency using a temperature compensation formula.

[0076] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying the corrosion state of wind turbine bolt materials, characterized in that, Includes the following steps: S1: Acquire the raw ultrasonic signal of the wind turbine bolt in full matrix capture mode using a ring ultrasonic transducer array, and post-process the received signal using a full focusing method to obtain the focused echo signal; S2: Perform wavelet packet decomposition and soft thresholding denoising on the focused echo signal to obtain the reconstructed signal of each sub-band, and extract the energy entropy of each sub-band. Then, concatenate the focused echo signal as the original time domain signal with the reconstructed signal and energy entropy of the sub-band to form a multi-channel feature sequence. S3: Input the multi-channel feature sequence into a deep convolutional neural network and extract the erosion feature semantic vector by combining the channel attention mechanism; S4: Obtain the load data and environmental condition data of the bolts, normalize them and map them into external feature vectors. Then, fuse the corrosion feature semantic vector and the external feature vectors by tensor product and reduce the dimensionality to obtain a dynamic embedding vector. S5: The dynamic embedding vectors of the same bolt at different times are constructed into an input sequence in chronological order, and processed by a bidirectional long short-term memory network to output corrosion level, residual strength ratio and remaining life in parallel.

2. The method for identifying the corrosion state of wind turbine bolt materials according to claim 1, characterized in that, In step S1, the spatial coordinates of the array elements of the annular ultrasonic transducer array are determined according to the annular radius and the equally spaced angles; the imaging grid step size of the full focusing method is 0.1-0.15 mm, the axial imaging range is 200-300 mm, and the Hilbert envelope is extracted from the A-scan signal before superposition.

3. The method for identifying the corrosion state of wind turbine bolt materials according to claim 1, characterized in that, In step S2, wavelet packet decomposition uses the db8 wavelet basis function, with a decomposition level of 4 layers; the threshold calculation formula for soft threshold denoising is: ， in, The noise standard deviation is defined as the noise level within the last third of the bottom echo interval, where N is the signal length; energy entropy. ,in, Let i be the energy of the i-th subband signal. Let i be the energy of the j-th sub-band signal, where i and j are both sub-band indices.

4. The method for identifying the corrosion state of wind turbine bolt materials according to claim 1, characterized in that, In step S3, the deep convolutional neural network includes three convolutional layers, a global average pooling layer, two fully connected layers, and a dropout layer, and a channel attention mechanism is introduced before the global average pooling; the output erosion feature semantic vector has a dimension of 512.

5. The method for identifying the corrosion state of wind turbine bolt materials according to claim 1, characterized in that, In step S4, the load data includes tensile stress, bending stress, and torsional stress, and the environmental condition data includes salt spray concentration, relative humidity, and temperature. The time synchronization rule is to take the average value within 10 minutes before the ultrasonic detection time, and to use linear interpolation for missing values. Normalization uses the min-max method, and the normalization parameter is fixed to the statistical value of the training set. After tensor product fusion, the data is successively passed through a batch normalization layer and a three-layer fully connected network for dimensionality reduction. Gradient pruning is used during training.

6. The method for identifying the corrosion state of wind turbine bolt materials according to claim 1, characterized in that, In step S5, the length of the input sequence is the result of the last 5 detections, and if it is insufficient, it is padded with zeros; the number of hidden layer units of the bidirectional long short-term memory network is 128; the corrosion level is divided into 0-4 levels according to the maximum corrosion depth and pitting density, where the pitting density is defined as the number of effective connected regions divided by the area of ​​the detection area, and the detection area is a circular area with a center radius of 15 mm on the bolt end face.

7. The method for identifying the corrosion state of wind turbine bolt materials according to claim 1, characterized in that, It also includes the training data construction steps: take bolts from the same batch for accelerated corrosion-fatigue coupling test, periodically take out bolts for ultrasonic testing and destructive tensile test to obtain real labels, calibrate the residual strength ratio of undamaged bolts through finite element simulation, mix real label samples and simulation calibration samples at a ratio of 1:2, and assign loss weights of 1.0 and 0.5 respectively.

8. A system for identifying the corrosion state of wind turbine bolt materials, used to execute the method for identifying the corrosion state of wind turbine bolt materials according to any one of claims 1 to 7, characterized in that, include: Ultrasonic acquisition module: used to acquire raw ultrasonic signals of wind turbine bolts in full matrix capture mode through a ring ultrasonic transducer array, and to post-process the received signals using a full focusing method to obtain focused echo signals; Signal preprocessing module: used to perform wavelet packet decomposition and soft thresholding denoising on the focused echo signal to obtain the reconstructed signal of each sub-band, and extract the energy entropy of each sub-band. The focused echo signal is used as the original time domain signal and is concatenated with the reconstructed signal of the sub-band and the energy entropy to form a multi-channel feature sequence. Deep feature extraction module: used to input multi-channel feature sequences into a deep convolutional neural network and extract erosion feature semantic vectors by combining channel attention mechanism; Multi-source data fusion module: used to acquire bolt load data and environmental condition data, which are normalized and mapped to external feature vectors. The corrosion feature semantic vector and the external feature vector are fused by tensor product and dimensionality reduced to obtain a dynamic embedding vector. The time-series prediction module is used to construct the dynamic embedding vectors of the same bolt at different times into an input sequence in chronological order. It is processed using a bidirectional long short-term memory network and outputs corrosion level, residual strength ratio and remaining life in parallel.

9. A corrosion status identification system for wind turbine bolt materials according to claim 8, characterized in that, It also includes a handheld self-diagnostic probe, which includes a pulse generator, a low-noise amplifier, an analog-to-digital converter, a wireless communication module, and an inertial measurement unit; the system also includes a sparse deconvolution unit for measuring the transducer impulse response and performing iterative soft thresholding algorithm processing.

10. A corrosion state identification system for wind turbine bolt materials according to claim 8, characterized in that, It also includes a self-diagnostic unit, which performs a self-test every 24 hours, measures the transducer impedance spectrum and extracts the resonant frequency and bandwidth, compares the percentage deviation of the resonant frequency from the nominal value and the bandwidth change, measures the attenuation of the bottom echo amplitude, calibrates the initial value under the conditions of room temperature 20±5℃ and coupling pressure 20N, and corrects the resonant frequency using a temperature compensation formula.

Images