A method, system, storage medium, and program product for monitoring the connection status of towers.
By acquiring multi-source time-domain data and fusing cross-modal features, combined with a Bayesian inference network, the problem of delayed early identification of minor loosening in tower connection status monitoring was solved, achieving efficient monitoring and early defect identification of wind turbine tower connection status.
Patent Information
- Application Number
- CN202611175934.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-04
- Publication Date
- 2026-11-17
AI Technical Summary
In the monitoring of tower connection status, the existing technology has a delay in identifying the early stage of slight loosening, which leads to untimely identification of connection defects. Especially in the micro-slippage stage of high-strength bolt connections, the existing system cannot effectively capture the weak cross-modal coupling characteristics.
Multi-source time-domain data acquisition is adopted. Wavelet packet decomposition and cross-modal feature fusion techniques are used in combination with Bayesian inference network to integrate multi-source heterogeneous data, extract frequency domain feature subspace and generate cross-modal energy feature vector, and perform probabilistic inference to identify the tower connection status.
It improves the accuracy of early micro-defect identification, reduces identification delay, enhances sensitivity to micro-slip conditions, and ensures the safe operation of wind turbine generators.
Smart Images

Figure CN123117322A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data identification technology, and in particular to a tower connection status monitoring method, system, storage medium and program product. Background Technology
[0002] As the core load-bearing structure, the tower of a wind turbine generator typically relies on high-strength bolts for fastening its flange connections. Under long-term alternating loads and complex operating conditions, the structural integrity of the tower connection points directly affects the safe operation of the entire equipment. Continuous monitoring of the physical condition of concealed connection points is a fundamental aspect of maintaining the safe service of the equipment.
[0003] In related technologies, when monitoring the connection status of towers, various physical quantity signals, such as stress signals and vibration signals, are typically collected from the tower connection area. The system performs independent feature extraction and status determination on the single-mode signals collected by various sensors. At the decision-making level, the determination results of each single signal are fused through a voting mechanism or weighted algorithm to output the final tower connection status assessment result.
[0004] However, with the higher demand for accurate identification of connection defects in the early and slight loosening stage, the relevant technology filters out weak cross-modal coupling features in the independent judgment stage of each signal. This blocks the correlation information between high-frequency and low-frequency signals in the early stage of micro-slippage, causing the system to show obvious feature shift lag in the initial loosening stage of the anchor plate, resulting in a delay in the identification of early micro-defects. Summary of the Invention
[0005] This application provides a tower connection status monitoring method, system, storage medium, and program product, which are used to reduce the identification delay of early micro-defects in a multi-source heterogeneous data environment.
[0006] Firstly, this application provides a method for monitoring the connection status of a tower, applied to a tower monitoring system. The method includes: acquiring stress signals and ultrasonic signals from the tower connection area to obtain multi-source time-domain data; performing wavelet packet decomposition on the multi-source time-domain data to obtain multiple frequency-domain feature subspaces; calculating the sum of the squared amplitudes of the wavelet packet coefficients in each frequency-domain feature subspace to generate frequency band energy values for each frequency-domain feature subspace; performing cross-modal feature fusion on the frequency band energy values corresponding to the stress signals and the ultrasonic signals to generate cross-modal energy feature vectors; inputting the cross-modal energy feature vectors into a Bayesian inference network to perform probabilistic inference operations to obtain a joint probability distribution of the connection status; matching the corresponding status label based on the maximum probability value in the joint probability distribution to determine the tower connection status result; and generating corresponding wind turbine control commands based on the tower connection status result to perform status response operations on the wind turbine generator set.
[0007] In the above embodiments, the tower monitoring system collects multi-source time-domain data and performs wavelet packet decomposition to extract the frequency band energy values of each frequency domain feature subspace. Then, it performs cross-modal feature fusion to generate cross-modal energy feature vectors, which are input into a Bayesian inference network for probabilistic inference. This system can integrate the deep correlation of multi-source heterogeneous data, avoid the loss of weak coupling features caused by single-modal judgment, and reduce the early micro-defect identification delay.
[0008] In conjunction with some embodiments of the first aspect, in some embodiments, the step of acquiring stress signals and ultrasonic signals from the tower connection area to obtain multi-source time-domain data specifically includes: acquiring low-frequency stress signals from the tower connection area; extracting alternating load fluctuation characteristics of the low-frequency stress signals; when the alternating load fluctuation characteristics meet preset fatigue excitation conditions, sending a wake-up command to the ultrasonic acquisition probe in a dormant state to acquire high-frequency ultrasonic signals; and using the timestamp of the wake-up command as a reference, performing time window interception on the low-frequency stress signals and high-frequency ultrasonic signals to obtain multi-source time-domain data.
[0009] In the above embodiments, the tower monitoring system extracts the alternating load fluctuation characteristics of the low-frequency stress signal and wakes up the ultrasonic acquisition probe when the fatigue excitation conditions are met. The time window is intercepted based on the wake-up command time stamp. This ensures the synchronization of the low-frequency stress signal and the high-frequency ultrasonic signal in the time dimension, providing an accurate data foundation for subsequent cross-modal feature fusion.
[0010] In conjunction with some embodiments of the first aspect, in some embodiments, the step of performing time window interception on low-frequency stress signals and high-frequency ultrasonic signals based on the wake-up command issuance timestamp to obtain multi-source time-domain data specifically includes: extracting the hardware wake-up delay time of the ultrasonic acquisition probe from the sleep state to the data output state, and calculating the ultrasonic alignment timestamp in combination with the issuance timestamp; intercepting stress data segments in the low-frequency stress signals starting from the issuance timestamp; intercepting ultrasonic data segments in the high-frequency ultrasonic signals starting from the ultrasonic alignment timestamp; and combining the stress data segments and ultrasonic data segments to obtain multi-source time-domain data.
[0011] In the above embodiments, the tower monitoring system calculates the ultrasonic alignment timestamp by extracting the hardware wake-up delay time, and combines the stress data segment and the ultrasonic data segment separately; this compensates for the time deviation caused by the hardware wake-up, improves the time alignment accuracy of multi-source data, and reduces the feature fusion error caused by time misalignment.
[0012] In conjunction with some embodiments of the first aspect, in some embodiments, the step of performing wavelet packet decomposition on multi-source time-domain data to obtain multiple frequency-domain feature subspaces specifically includes: performing multi-scale wavelet packet transform on multi-source time-domain data to obtain an initial frequency-domain node sequence; calculating the energy entropy value of each node in the initial frequency-domain node sequence; selecting target frequency-domain nodes representing the dominant energy based on the energy entropy values; and performing a frequency band merging operation on the target frequency-domain nodes to generate multiple frequency-domain feature subspaces.
[0013] In the above embodiments, the tower monitoring system obtains the initial frequency domain node sequence through multi-scale wavelet packet transform, calculates the energy entropy value and filters the target frequency domain nodes, and performs frequency band merging; it can accurately extract the frequency domain features that characterize the dominant energy, eliminate redundant frequency bands, reduce the data processing dimension, and improve the computational efficiency of subsequent feature fusion.
[0014] In conjunction with some embodiments of the first aspect, in some embodiments, after performing a band merging operation on the target frequency domain node to generate multiple frequency domain feature subspaces, the method further includes: extracting a set of target subspaces corresponding to stress concentration frequency bands from the multiple frequency domain feature subspaces; dividing the set of target subspaces into a stable connection state subspace characterizing effective fastening characteristics and a loosening hazard state subspace characterizing micro-slippage characteristics; and calculating a fastening reliability index based on the ratio of the frequency band energy value of the stable connection state subspace to the total energy, as a state constraint condition input to the Bayesian inference network.
[0015] In the above embodiments, the tower monitoring system divides the stable connection state subspace and the loosening potential state subspace, and calculates the fastening reliability index as a state constraint condition; this provides additional physical state constraints for the Bayesian inference network, enhances the sensitivity of the inference model to early micro-slip states, and improves the accuracy of state determination.
[0016] In some embodiments of the first aspect, the step of inputting the cross-modal energy feature vector into a Bayesian inference network to perform probabilistic inference operations and obtain the joint probability distribution of the connection states specifically includes: obtaining the historical posterior probability distribution of the previous monitoring period and constructing the prior probability density function of the current period in combination with the tightness reliability index; calculating the likelihood function values of the cross-modal energy feature vector under different connection states based on a preset multivariate conditional probability model; performing Bayesian formula operations based on the prior probability density function and the likelihood function values to obtain the joint probability distribution characterizing the early microslip state and the stable tightness state; and performing feature filtering on the tightness features under single-source monitoring based on the joint probability distribution to reduce the proportion of tightness features to a preset early fusion feature threshold range.
[0017] In the above embodiments, the tower monitoring system constructs a prior probability density function by combining historical posterior probability distribution and fastening reliability index, calculates the likelihood function value and performs Bayesian formula calculation, and finally performs feature filtering; it effectively suppresses the fastening features that dominate under single-source monitoring, highlights early micro-slip features, and improves the ability to identify weak defects.
[0018] In conjunction with some embodiments of the first aspect, in some embodiments, the step of fusing the energy values of each frequency band corresponding to the stress signal and the energy values of each frequency band corresponding to the ultrasonic signal across modal features to generate a cross-modal energy feature vector specifically includes: normalizing the energy values of each frequency band corresponding to the stress signal and the energy values of each frequency band corresponding to the ultrasonic signal to obtain a stress energy feature sequence and an ultrasonic energy feature sequence; constructing a Kalman filter model with the real energy characteristics of the tower connection state as the state variable and the stress energy feature sequence and the ultrasonic energy feature sequence as the observation variables; calculating the optimal state estimate after eliminating random noise based on the Kalman filter model, and extracting the gain matrix of the Kalman filter model in the observation update process as the fusion gain coefficient; and performing weighted early fusion of the stress energy feature sequence and the ultrasonic energy feature sequence according to the fusion gain coefficient to generate a cross-modal energy feature vector.
[0019] In the above embodiments, the tower monitoring system calculates the optimal state estimate by constructing a Kalman filter model and extracts the gain matrix as the fusion gain coefficient to perform weighted early fusion; it can dynamically adjust the fusion weights of different modal features, eliminate random noise interference, generate high-quality cross-modal energy feature vectors, and ensure the reliability of state monitoring.
[0020] In a second aspect, embodiments of this application provide a tower monitoring system, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, which includes computer instructions, and the one or more processors call the computer instructions to cause the tower monitoring system to perform the method described in the first aspect and any possible implementation thereof.
[0021] Thirdly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a tower monitoring system, cause the tower monitoring system to perform the method described in the first aspect and any possible implementation thereof.
[0022] Fourthly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on a tower monitoring system, cause the tower monitoring system to perform the method described in the first aspect and any possible implementation thereof.
[0023] Understandably, the tower monitoring system provided in the second aspect, the computer storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the methods provided in the embodiments of this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.
[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:
[0025] 1. Due to the adoption of cross-modal feature fusion and Bayesian inference network, it can integrate the deep correlation of multi-source heterogeneous data, effectively solving the problem of weak coupling feature loss caused by single modality judgment in the existing technology, thereby reducing the early micro-defect recognition delay.
[0026] 2. Due to the adoption of energy entropy value screening and frequency band merging operations, the frequency domain features representing the dominant energy can be accurately extracted, effectively solving the problems of computational redundancy and noise interference caused by full-band processing in existing technologies, thereby achieving effective reduction of feature dimensions and preservation of key information.
[0027] 3. Because the Kalman filter model is used to extract the fusion gain coefficient, the fusion weights of different modal features can be dynamically adjusted, which effectively solves the problem that static weighting cannot adapt to changes in environmental noise in the existing technology, and thus achieves high-quality fusion of cross-modal energy features. Attached Figure Description
[0028] Figure 1 This is a flowchart illustrating a tower connection status monitoring method in an embodiment of this application;
[0029] Figure 2 This is another flowchart illustrating the tower connection status monitoring method in this application embodiment;
[0030] Figure 3 This is a schematic diagram of the physical device structure of a tower monitoring system in an embodiment of this application. Detailed Implementation
[0031] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.
[0032] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.
[0033] During operation, wind turbine generators experience complex alternating loads in the tower connection area. Alternating loads refer to loads whose magnitude and direction change periodically or non-periodically over time. Under long-term service, bolts at flange connections are prone to micro-slippage. Micro-slippage refers to the minute relative sliding that occurs at the local contact surface between connecting parts before macroscopic relative displacement occurs. To accurately monitor this physical state, stress signals and ultrasonic signals need to be acquired. Stress signals reflect the macroscopic stress state of the structure, while ultrasonic signals reflect the microscopic acoustic emission characteristics within the structure. By using cross-modal feature fusion technology, data from different physical modes are correlated and integrated at the feature layer, extracting weak coupling features that cannot be characterized by a single mode. This provides high-dimensional information input for subsequent state inference, ensuring the accuracy of early micro-defect monitoring.
[0034] The method provided in this embodiment is described in detail below. Please refer to [link / reference]. Figure 1 This is a flowchart illustrating a tower connection status monitoring method in an embodiment of this application.
[0035] S101. Collect stress signals and ultrasonic signals in the tower connection area to obtain multi-source time-domain data.
[0036] Among them, stress signal refers to the data sequence reflecting the stress deformation of the tower structure obtained by sensors such as strain gauges; ultrasonic signal refers to the high-frequency vibration data reflecting the acoustic emission phenomenon inside the material obtained by ultrasonic probe; multi-source time-domain data refers to the original time series set containing the above two different physical modes within the same time observation window.
[0037] Specifically, when the wind turbine is in grid-connected power generation operation, the tower monitoring system synchronously collects stress signals and ultrasonic signals through a sensor array deployed in the flange connection area. It is necessary to ensure that the data sampling rate of different sensors meets the Nyquist sampling theorem of their respective signals, and record the timestamp of each sampling point through a hardware clock synchronization mechanism. Finally, the data from different channels are encapsulated into a multi-source time-domain data matrix.
[0038] In some embodiments, the acquisition of multi-source time-domain data can be achieved in several ways: Optionally, the tower monitoring system is configured with independent data acquisition cards, with separate low-frequency and high-frequency channels, synchronously triggering sampling commands to convert the acquired analog signals into digital signals and store them in a local buffer; Optionally, the tower monitoring system employs distributed intelligent sensor nodes, with each node synchronizing its clock based on a precise time protocol, and after independently acquiring signals, sending time-stamped data packets to the central processing unit via industrial Ethernet for alignment and splicing. It is understood that other methods can also be used to acquire multi-source time-domain data, and these are not limited here.
[0039] In some embodiments, electromagnetic interference at the site may cause pulse noise in the acquired signals. To address this, the tower monitoring system introduces a median filtering algorithm before data encapsulation to smooth the original time-domain sequence, remove abnormal pulse extreme points, and ensure the quality of multi-source time-domain data.
[0040] S102. Perform wavelet packet decomposition on the multi-source time-domain data to obtain multiple frequency domain feature subspaces.
[0041] Among them, wavelet packet decomposition refers to a mathematical transformation method that uses wavelet packet functions to simultaneously perform multi-level refinement decomposition of the low-frequency and high-frequency parts of a signal; the frequency domain feature subspace refers to the independent data set covering a specific frequency range obtained after decomposition.
[0042] Specifically, after receiving multi-source time-domain data, the tower monitoring system selects a specific wavelet basis function to decompose the data sequence into a preset number of layers; the original wideband signal is divided into multiple frequency domain feature subspaces with equal bandwidth and no overlap, so that the local features of the signal in different frequency bands can be revealed, providing a fine frequency domain division basis for subsequent energy feature extraction.
[0043] In some embodiments, wavelet packet decomposition can be implemented in several ways: Optionally, the tower monitoring system sets the decomposition level to three levels, selects the Daubechies wavelet system as the basis function, performs layer-by-layer filtering and downsampling on the time-domain data, and generates eight frequency-domain node sequences; Optionally, the tower monitoring system adaptively determines the decomposition level based on the spectral distribution characteristics of the signal, selects the Symlets wavelet system to perform the decomposition, and performs orthogonality checks on the reconstructed signals of each node, outputting a frequency-domain feature subspace that meets the requirements. It is understood that other methods can also be used to implement wavelet packet decomposition, which are not limited here.
[0044] In some embodiments, excessively high decomposition levels can lead to excessive computational resource consumption. To address this, the tower monitoring system pre-calculates the frequency band energy concentration of the signal. When the concentration exceeds a set threshold, it dynamically reduces the number of wavelet packet decomposition levels, thereby reducing computational overhead while ensuring feature extraction accuracy.
[0045] S103. Calculate the sum of the squared amplitudes of the wavelet packet coefficients in each frequency domain feature subspace to generate the frequency band energy value of each frequency domain feature subspace.
[0046] Among them, wavelet packet coefficients refer to the numerical representation of a signal after projection onto a specific frequency domain feature subspace; amplitude squared cumulative value refers to the sum of the squares of all coefficients in the subspace; and frequency band energy value is used to represent the magnitude of physical energy contained in the signal within that frequency band.
[0047] Specifically, the tower monitoring system traverses all frequency domain feature subspaces and extracts wavelet packet coefficient sequences within each subspace. The system squares each coefficient in the sequence and accumulates all squared values within the same subspace to obtain the frequency band energy value for that subspace; this reduces the high-dimensional coefficient sequence to a single energy scalar, intuitively reflecting the energy distribution state of different frequency bands.
[0048] In some embodiments, the generation of the band energy value can be achieved in several ways: Optionally, the tower monitoring system reads the subspace coefficients sequentially, calls the hardware multiplier to calculate the square, stores the squares in the accumulator register, and outputs the register value as the band energy value after traversal; Optionally, the tower monitoring system divides the coefficient sequence into multiple data blocks, uses multi-threaded parallel calculation of the sum of squares of each data block, and finally summarizes and merges the results of each thread to generate the final band energy value. It is understood that other methods can also be used to generate the band energy value, and this is not limited here.
[0049] It should be noted that when performing the amplitude square accumulation operation of wavelet packet coefficients in each frequency domain feature subspace, the tower monitoring system can generate the three-layer wavelet packet decomposition. Using 1 feature subspace as the processing object, extract the discrete wavelet packet coefficient sequence within the j-th subspace (j ranges from 0 to 7). , where k represents the sampling point index in the sequence. The system calculates the frequency band energy value of the corresponding subspace by squaring the absolute value of each coefficient in the sequence and integrating along the time axis. Its underlying data processing logic follows mathematical expressions. Here, n represents the total number of coefficients within the subspace. After calculating the energy values for all eight independent subspaces, the system further linearly adds these eight frequency band energy values to generate a total energy value that characterizes the overall physical energy intensity of the multi-source signals within the current time window. Its calculation logic is as follows: For example, when the system acquires a stress signal containing minor loosening characteristics, it decomposes the signal into eight frequency bands through three levels. The system then calculates the sum of the squares of these eight frequency bands. If the energy values of the first four frequency bands are 50, 40, 15, and 5 respectively, and the energy of the last four frequency bands is negligible, the system will not only output the energy values of these eight independent frequency bands but also calculate the total energy value. The value is 110, thus transforming the intangible connection state evolution into a frequency band energy mathematical model that can be quantified and calculated.
[0050] S104. Perform cross-modal feature fusion on the energy values of each frequency band corresponding to the stress signal and the energy values of each frequency band corresponding to the ultrasonic signal to generate a cross-modal energy feature vector.
[0051] Cross-modal feature fusion refers to the process of mathematically combining data features from different physical sources according to specific rules; cross-modal energy feature vectors refer to the comprehensive data structure containing multi-source energy information formed after fusion.
[0052] Specifically, the tower monitoring system acquires the frequency band energy values of stress signals and ultrasonic signals respectively, and uses a preset fusion algorithm to perform correlation calculations on the two sets of values; it breaks the feature island of a single signal, maps the low-frequency macroscopic stress characteristics and high-frequency microscopic acoustic characteristics to the same feature space, and constructs a cross-modal energy feature vector that can comprehensively characterize the connection state of the tower.
[0053] In some embodiments, cross-modal feature fusion can be achieved in several ways: Optionally, the tower monitoring system concatenates the two sets of frequency band energy values in descending order of frequency to form a high-dimensional cascaded feature vector, and performs principal component analysis for dimensionality reduction; Optionally, the tower monitoring system calculates the cross-correlation coefficients of the two sets of frequency band energy values in corresponding frequency bands, and performs a weighted summation of the energy values using the cross-correlation coefficients as weights to generate the fused cross-modal energy feature vector. It is understood that other methods can also be used to achieve cross-modal feature fusion, and this is not limited here.
[0054] In some embodiments, there may be a situation where the energy magnitude difference between the two sets of signals is too large, causing a shift in the fusion characteristics. To address this, the tower monitoring system performs Z-score normalization on the energy values of the two frequency bands separately before fusion to eliminate differences in dimensions and magnitudes, ensuring the balance of the fusion process.
[0055] S105. Input the cross-modal energy feature vector into the Bayesian inference network to perform probability inference operations and obtain the joint probability distribution of the connection states.
[0056] Among them, Bayesian inference network refers to a probabilistic graphical model built on Bayes' theorem for handling uncertain inference; probabilistic inference operation refers to the process of calculating the probability of each target state based on input features; joint probability distribution refers to the set containing all possible connection states and their corresponding probability values.
[0057] Specifically, the tower monitoring system inputs cross-modal energy feature vectors as observational evidence into a pre-trained Bayesian inference network. Network nodes, based on conditional probability tables and prior probabilities, execute message passing and belief update algorithms to calculate the posterior probabilities of the tower connection being in various states—normal, slightly slipped, or loose—under the current feature vector, and output the joint probability distribution.
[0058] In some embodiments, probabilistic inference can be implemented in several ways: Optionally, the tower monitoring system employs the joint tree algorithm in exact inference algorithms to transform the Bayesian network into a clustered tree structure, and calculates the marginal probability distribution of each state node through local message passing; Optionally, the tower monitoring system employs the Markov chain Monte Carlo sampling method in approximate inference algorithms to generate a large number of sample sequences in the network, and approximates the true joint probability distribution by statistically analyzing sample frequencies. It is understood that other methods can also be used to implement probabilistic inference, and these are not limited here.
[0059] In some embodiments, the network conditional probability table may experience a decrease in inference accuracy due to a lack of updates over a long period. To address this, the tower monitoring system periodically incorporates historical confirmed case data and employs the expectation-maximization algorithm to learn and fine-tune the network parameters online, maintaining the accuracy of probability inference.
[0060] S106. Based on the maximum probability value in the joint probability distribution, match the corresponding status label to determine the tower connection status result.
[0061] Among them, the maximum probability value refers to the posterior probability with the largest value in the joint probability distribution; the state label refers to the predefined identifier that characterizes the physical state of the tower connection; and the tower connection state result refers to the final judgment conclusion output by the system.
[0062] Specifically, the tower monitoring system analyzes the joint probability distribution set and uses a sorting algorithm to extract the probability term with the highest value. The system then maps and matches the state node corresponding to the highest probability value with a preset state label library to extract the corresponding physical state description, which is used as the tower connection status result for the current monitoring period.
[0063] In some embodiments, the determination of the status result can be achieved in multiple ways: Optionally, the tower monitoring system traverses the probability distribution array, records the maximum value and its index, retrieves the corresponding status label by looking up the index, and outputs the status result; Optionally, the tower monitoring system sets a confidence threshold, and when the maximum probability value exceeds the threshold, directly outputs the corresponding status label; if it does not exceed the threshold, it outputs "status pending" and triggers a retest mechanism. It is understood that other methods can also be used to determine the status result, and these are not limited here.
[0064] In some embodiments, there may be situations where multiple states have similar probability values and none of them have an absolute advantage. To address this, the tower monitoring system extracts the state labels with the top two probability values and combines them with historical state results from the previous monitoring period to perform smoothing filtering decisions, thus avoiding frequent jumps in state determination.
[0065] S107. Generate corresponding wind turbine control commands based on the tower connection status results to perform status response operations on the wind turbine generator set.
[0066] Among them, wind turbine control commands refer to control codes used to adjust the operating parameters of wind turbine generator sets; status response operations refer to the specific physical actions performed by the unit after receiving the commands.
[0067] Specifically, the tower monitoring system inputs the determined tower connection status results to the main control logic module. Based on a preset safety response strategy, the main control logic module matches control commands appropriate to the current state and sends them via the industrial bus to the wind turbine's yaw, pitch, or shutdown actuators, driving the unit to perform corresponding status response operations to ensure equipment safety.
[0068] In some embodiments, command generation and response can be implemented in multiple ways: Optionally, when the status result is slight slip, the system generates a load reduction command to limit the output power and sends it to the pitch system to perform pitch angle adjustment; Optionally, when the status result is severe looseness, the system generates an emergency stop command, sends it to the braking system to execute mechanical braking, and simultaneously sends an alarm message to the operation and maintenance platform. It is understood that other methods can also be used to implement command generation and response, which are not limited here.
[0069] In some embodiments, network congestion may cause delays in the issuance of control commands. To address this, the tower monitoring system assigns the highest network transmission priority to control commands corresponding to high-risk states and employs redundant channels for dual-transmission to ensure timely status response operations.
[0070] To further improve the temporal alignment accuracy and feature fusion quality of multi-source data, this application proposes a more refined data extraction and weighted fusion mechanism based on the above embodiments. This mechanism provides specific solutions for practical engineering problems such as hardware wake-up latency and environmental noise interference.
[0071] The method provided in this embodiment will now be described in more detail. Please refer to [link / reference]. Figure 2 This is another flowchart illustrating the tower connection status monitoring method in this application embodiment.
[0072] S201. Collect stress signals and ultrasonic signals from the tower connection area to obtain multi-source time-domain data.
[0073] Refer to step S101, which will not be repeated here.
[0074] In some embodiments, the tower monitoring system performs a time window interception operation to ensure data synchronization. Specifically, the tower monitoring system acquires low-frequency stress signals in the tower connection area; extracts the alternating load fluctuation characteristics of the low-frequency stress signals; when the alternating load fluctuation characteristics meet the preset fatigue excitation conditions, it sends a wake-up command to the ultrasonic acquisition probe in a dormant state to acquire high-frequency ultrasonic signals; and, based on the timestamp of the wake-up command, performs time window interception on the low-frequency stress signals and high-frequency ultrasonic signals to obtain multi-source time-domain data.
[0075] Among them, low-frequency stress signal refers to a sequence with a low sampling rate that reflects macroscopic deformation; alternating load fluctuation characteristics refer to the periodic amplitude variation pattern presented in the signal; fatigue excitation condition refers to the set load threshold for triggering ultrasonic testing; wake-up command refers to the control signal that activates the hibernation hardware; issued timestamp refers to the precise system time at which the command is issued; and time window interception refers to the operation of extracting data for a specific time period from a continuous data stream.
[0076] Specifically, the tower monitoring system continuously monitors low-frequency stress signals and calculates their peak-to-valley difference and fluctuation frequency as characteristics of alternating load fluctuation. When this characteristic reaches a preset fatigue excitation condition, the system determines that it is currently in a stress state prone to micro-slippage and sends a wake-up command to the ultrasonic acquisition probe. The system records the timestamp of this command and uses it as the origin to extract data segments of fixed duration, forming multi-source time-domain data.
[0077] In some embodiments, time window interception can be implemented in several ways: Optionally, the tower monitoring system sets up a ring buffer to cache stress signals, locks the current buffer pointer when a wake-up command is issued, extracts data before and after the pointer, and synchronously records ultrasonic data; Optionally, the tower monitoring system uses a hardware trigger line to synchronously trigger direct memory access (DMA) transfer between the two channels when the hardware level flips when the wake-up command is issued, intercepting a data block of a specified length. It is understood that other methods can also be used to implement time window interception, which are not limited here.
[0078] In some embodiments, there may be instances where the ultrasonic probe fails to wake up, resulting in the loss of high-frequency signals. To address this, the tower monitoring system will set a response timeout detection after issuing the command. If no ultrasonic data is received within the timeout period, an error log will be recorded and the wake-up command will be resent.
[0079] In some embodiments, the tower monitoring system performs precise time alignment compensation. Specifically, the tower monitoring system extracts the hardware wake-up delay time of the ultrasonic acquisition probe from the sleep state to the data output state, and calculates the ultrasonic alignment timestamp by combining it with the sent timestamp. Starting from the sent timestamp, it extracts stress data segments from the low-frequency stress signal; starting from the ultrasonic alignment timestamp, it extracts ultrasonic data segments from the high-frequency ultrasonic sound signal; and combines the stress data segments and ultrasonic data segments to obtain multi-source time-domain data.
[0080] Among them, hardware wake-up delay time refers to the inherent physical time taken for the probe to go from receiving a command to outputting valid data; ultrasonic alignment timestamp refers to the actual data start time after compensation for the delay; stress data segment and ultrasonic data segment refer to the extracted independent data blocks.
[0081] Specifically, the tower monitoring system reads the factory configuration parameters or historical calibration records of the ultrasonic probe to obtain the hardware wake-up delay time. The system adds the sent timestamp to this delay time to obtain the ultrasonic alignment timestamp. The system uses both the sent timestamp and the ultrasonic alignment timestamp as starting points, extracts data segments of the same length from the data streams of the two channels, and combines them to eliminate time misalignment caused by differences in hardware response.
[0082] In some embodiments, time alignment compensation can be achieved in several ways: Optionally, the tower monitoring system calculates the time offset at the software level and extracts aligned data segments by moving the read index of the data array; alternatively, the tower monitoring system sets an adjustable delay line in the programmable logic device to set the transmission delay of the stress signal to be consistent with the ultrasonic hardware wake-up delay, thereby achieving hardware-level real-time data alignment. It is understood that other methods can also be used to achieve time alignment compensation, which are not limited here.
[0083] In some embodiments, changes in ambient temperature may cause the hardware wake-up latency to drift. To address this, the tower monitoring system introduces a temperature compensation coefficient to dynamically adjust the hardware wake-up latency parameter based on the current ambient temperature, ensuring the accuracy of the timestamp alignment.
[0084] S202. Perform multi-scale wavelet packet transform on the multi-source time-domain data to obtain the initial frequency domain node sequence.
[0085] Refer to step S102, which will not be repeated here.
[0086] S203. Calculate the energy entropy value of each node in the initial frequency domain node sequence.
[0087] Among them, the energy entropy value refers to a numerical index calculated based on the information entropy theory to measure the uniformity of signal energy distribution.
[0088] Specifically, the tower monitoring system extracts the energy values of each node in the initial frequency domain node sequence and calculates the proportion of each node's energy to the total energy. The system then substitutes this proportion into the Shannon entropy formula to calculate the energy entropy value of each node. The smaller this value, the more concentrated the energy distribution and the more obvious the characteristics.
[0089] It should be noted that when calculating the energy entropy value of each node in the initial frequency domain node sequence, the tower monitoring system first obtains the absolute energy value of the i-th node in the current node sequence. And calculate the sum of the energies of all nodes. Then, through ratio calculation The probability weight of the energy of the i-th node in the overall frequency band is obtained. Then, the system substitutes this probability weight sequence into the information entropy processing logic, and performs mathematical operations by calculating the negative sum of the product of each probability weight and its natural logarithm. This generates the energy entropy value for each node. For example, if the energy of a certain ultrasonic signal is highly concentrated on a specific high-frequency node, the probability proportion of that node... When the value is close to 1, the calculated energy entropy value H will approach 0. Based on this logic, the system can accurately identify and filter out target frequency domain nodes with extremely low entropy values that represent the dominant physical characteristics, thereby effectively eliminating redundant frequency bands with uniform energy distribution and containing a large amount of random noise.
[0090] In some embodiments, energy entropy can be calculated in several ways: Optionally, the tower monitoring system first calculates the total energy across the entire frequency band, then calculates the energy proportion of each node, and finally calls a logarithmic function library to calculate the entropy value; alternatively, the tower monitoring system constructs an energy probability density function, solves the energy entropy distribution curve within the continuous frequency band through integration, and extracts the entropy value of discrete nodes. It is understood that other methods can also be used to calculate energy entropy, and this is not limited here.
[0091] In some embodiments, the energy percentage calculation may result in a division-to-zero anomaly under weak signals. To address this, the tower monitoring system adds a very small bias constant to the total energy term in the denominator to prevent calculation overflow and ensure the stability of the energy entropy value calculation.
[0092] S204. Select target frequency domain nodes that characterize the dominant energy based on energy entropy values.
[0093] Among them, dominant energy refers to the energy component that occupies the main part of the overall signal energy and contains key physical characteristics; target frequency domain nodes refer to the specific set of nodes that have been selected and retained.
[0094] Specifically, the tower monitoring system sorts the calculated energy entropy values of each node in ascending order. Based on a preset retention ratio or a set entropy threshold, the system selects the nodes that rank highly or meet the threshold conditions, confirming them as target frequency domain nodes representing the dominant energy, and discards redundant nodes.
[0095] In some embodiments, the selection of target frequency domain nodes can be achieved in several ways: Optionally, the tower monitoring system sets a fixed entropy threshold, traverses the node sequence, and retains nodes with entropy values lower than the threshold; Optionally, the tower monitoring system uses a clustering algorithm to divide the nodes into high-entropy clusters and low-entropy clusters according to their entropy values, and directly extracts the nodes in the low-entropy clusters as target frequency domain nodes. It is understood that other methods can also be used to select target frequency domain nodes, which are not limited here.
[0096] In some embodiments, there may be situations where all node entropy values are too high, resulting in empty screening results. To address this, the tower monitoring system will forcibly retain the three nodes with the largest absolute energy values as target frequency domain nodes, ensuring data input for subsequent processing stages.
[0097] S205. Perform a frequency band merging operation on the target frequency domain node to generate multiple frequency domain feature subspaces.
[0098] Among them, the frequency band merging operation refers to the process of splicing and combining adjacent or similar frequency domain nodes.
[0099] Specifically, the tower monitoring system analyzes and selects the frequency range of the target frequency domain nodes, merges the nodes with continuous frequencies at their boundaries to form a wider frequency band; it reconstructs the frequency domain partitioning structure and generates a smaller number of frequency domain feature subspaces with more concentrated features.
[0100] In some embodiments, the band merging operation can be implemented in several ways: Optionally, the tower monitoring system traverses the target nodes in ascending frequency order; if adjacent nodes have continuous frequencies, they are merged into a subspace, and the boundary parameters are updated. Optionally, the tower monitoring system calculates the mutual information between each target node, and merges nodes with mutual information greater than a set threshold to generate a frequency domain feature subspace. It is understood that other methods can also be used to implement the band merging operation, which are not limited here.
[0101] S206. Calculate the sum of the squared amplitudes of the wavelet packet coefficients in each frequency domain feature subspace to generate the frequency band energy value of each frequency domain feature subspace.
[0102] Refer to step S103, which will not be repeated here.
[0103] In some embodiments, the tower monitoring system introduces state constraints to enhance inference accuracy. Specifically, the tower monitoring system extracts the target subspace set corresponding to the stress concentration frequency band from multiple frequency domain feature subspaces; divides the target subspace set into a stable connection state subspace representing effective fastening characteristics and a loosening hazard state subspace representing micro-slippage characteristics; and calculates the fastening reliability index based on the ratio of the frequency band energy value of the stable connection state subspace to the total energy, which is then used as a state constraint input to the Bayesian inference network.
[0104] Among them, the stress concentration frequency band refers to the frequency range in which the stress deformation characteristics are most significant; the target subspace set refers to the total set of subspaces within this range; the stable connection state subspace refers to the set containing energy characteristics under normal tightness; the loosening potential state subspace refers to the set containing minor loosening characteristics; the tightness reliability index refers to the numerical parameters that quantify and evaluate the current connection stability; and the state constraint conditions refer to the additional information used to limit the output range of the inference model.
[0105] Specifically, the tower monitoring system locates the stress concentration frequency band based on prior knowledge and extracts the corresponding target subspace set. The system then divides this subspace into a stable connected state subspace and a loosening potential state subspace based on frequency distribution patterns. The system calculates the frequency band energy value of the stable connected state subspace and divides it by the total energy of the target subspace set to obtain the tightness reliability index. This index is input as a state constraint into the Bayesian inference network to correct the prior probability.
[0106] In some embodiments, the fastening reliability index can be calculated in several ways: Optionally, the tower monitoring system directly reads the energy scalar of each subspace and performs a simple division operation to obtain the reliability index; alternatively, the tower monitoring system assigns a higher confidence weight to the energy value of the stable connection state subspace, performs a weighted summation of the total energy, and then performs a ratio calculation to generate a weighted fastening reliability index. It is understood that other methods can also be used to calculate the fastening reliability index, and this is not limited here.
[0107] In some embodiments, the total energy may be extremely low, causing drastic fluctuations in the ratio calculation results. To address this, the tower monitoring system will maintain the previous cycle's fastening reliability index unchanged when the total energy falls below a set lower limit, thus avoiding the introduction of invalid constraints.
[0108] It should be noted that when extracting stress concentration frequency bands from multiple frequency domain feature subspaces and calculating fastening reliability indices, the tower monitoring system can anchor the first four feature subspaces (i.e., frequency bands 1 to 4) with high energy concentration in the concealed engineering state as the target subspace set, and strictly divide them into two state spaces according to the physical response mechanism: frequency bands 1 and 2, which characterize effective fastening features, are classified as the stable connection state subspace, and frequency bands 3 and 4, which characterize early micro-slip features, are classified as the loosening hazard state subspace. Subsequently, the system extracts the frequency band energy values corresponding to the stable connection state subspace (denoted as...). and ), and call the total energy value calculated earlier. By calculating the percentage of the sum of steady-state energies in the total energy, a fastening reliability index is generated. Its core data processing logic strictly follows the formula ( ) For example, when the tower anchor plate is in the initial loosening stage, the system extracts the sum of the energy of frequency band 1 and frequency band 2 as 45, and the total energy is 100. The system will then calculate the fastening reliability index as 45% and input this quantitative index as a state constraint condition into the Bayesian inference network, thereby breaking the recognition lag barrier in traditional single-source monitoring where the fastening feature accounts for as much as 60% to 70%.
[0109] S207. Normalize the energy values of each frequency band corresponding to the stress signal and the energy values of each frequency band corresponding to the ultrasonic signal to obtain the stress energy characteristic sequence and the ultrasonic energy characteristic sequence.
[0110] Normalization refers to the mathematical transformation process that maps data of different dimensions and orders of magnitude to the same specific interval; stress energy characteristic sequence and ultrasonic energy characteristic sequence refer to the dimensionless data sets obtained after processing.
[0111] Specifically, the tower monitoring system extracts the frequency band energy value sets of stress signals and ultrasonic signals respectively, and finds the maximum and minimum values in each set. The system uses a maximum-minimum normalization formula to perform a linear transformation on each energy value in the set, outputting a standardized stress energy characteristic sequence and ultrasonic energy characteristic sequence.
[0112] In some embodiments, normalization can be implemented in several ways: Optionally, the tower monitoring system statistically analyzes global extreme values based on a historical sample database and uses these global extreme values to map the energy value of the current period; alternatively, the tower monitoring system employs a mean-variance normalization method to calculate the mean and standard deviation of the current sequence, converting the data into a characteristic sequence conforming to a standard normal distribution. It is understood that other methods can also be used to implement normalization, and this is not limited here.
[0113] In some embodiments, a single outlier may cause severe compression of the overall sequence after normalization. To address this, the tower monitoring system uses the Raida criterion to remove outliers before normalization and uses the second-largest value as the normalization benchmark to ensure the distribution range of valid data.
[0114] S208. Construct a Kalman filter model with the actual energy characteristics of the tower connection state as the state variable and the stress energy characteristic sequence and ultrasonic energy characteristic sequence as the observation variables.
[0115] Among them, the Kalman filter model refers to an algorithm model that uses the state equation of a linear system to make the optimal estimate of the system state through the system input and output observation data; the state variable refers to the real physical quantity inside the system that cannot be directly measured; the observation variable refers to the data sequence actually collected and processed by the sensor.
[0116] Specifically, the tower monitoring system sets the actual energy characteristics of the tower connection state as the state variables of the Kalman filter model and establishes a state transition matrix. The system uses the normalized stress energy characteristic sequence and ultrasonic energy characteristic sequence as observation variables to establish an observation matrix and initialize the process noise covariance and measurement noise covariance matrices. When initializing the process noise covariance and measurement noise covariance matrices, the tower monitoring system pre-collects the background stress noise and ultrasonic noise floor of the wind turbine generator in a stopped and stationary state, calculates the variance of the two as the initial diagonal element of the measurement noise covariance matrix; at the same time, based on the energy characteristic fluctuation rate under historical normal grid-connected power generation conditions, the process noise covariance matrix is set to ensure the convergence stability of the Kalman filter model in the initial iteration stage.
[0117] In some embodiments, the Kalman filter model can be constructed in several ways: Optionally, the tower monitoring system assumes the state transition process is a first-order Markov process, sets the state transition matrix as the identity matrix, and constructs a standard linear Kalman filter model; alternatively, the tower monitoring system introduces a nonlinear state transition function, uses Taylor expansion for linearization approximation, and constructs an extended Kalman filter model. It is understood that other methods can also be used to construct the Kalman filter model, and this is not limited here.
[0118] It should be noted that when constructing the Kalman filter model and performing weighted early fusion, the tower monitoring system sets the actual energy characteristics of the tower connection state as the system state variable. The normalized stress energy characteristic sequence and the ultrasonic energy characteristic sequence are concatenated as the observed variable. The system then performs state prediction based on the optimal estimate from the previous time step. During the observation update phase, the system calculates the ratio of prediction error covariance to observation noise covariance, and derives the Kalman gain matrix. This matrix essentially dynamically quantifies the reliability weights of the stress mode and ultrasonic mode data at the current moment. Subsequently, the system extracts the weight coefficients from this gain matrix and directly performs a weighted summation operation on the input stress energy feature sequence and ultrasonic energy feature sequence at the feature layer (rather than the final decision layer) to generate a cross-modal energy feature vector. For example, when on-site wind noise causes high-frequency random fluctuations in the stress sensor, the system calculates a surge in the stress observation noise covariance using Kalman filtering logic, and then automatically lowers the weight coefficients allocated to the stress features in the gain matrix while increasing the weight coefficients of the ultrasonic features for weighted fusion. This early fusion mechanism, completed during the feature extraction stage, effectively avoids the convergence effect caused by misjudgment of a single signal in late fusion strategies, resulting in a significant increase of approximately 40% in the expression weights of early micro-slip and other potential hazards in the fused feature vector.
[0119] S209. Based on the Kalman filter model, calculate the optimal state estimate after eliminating random noise, and extract the gain matrix of the Kalman filter model in the observation update process as the fusion gain coefficient.
[0120] Among them, the optimal state estimate refers to the numerical value that approximates the true state output by the model; the gain matrix refers to the weight matrix used in the update step to weigh the confidence of the predicted value and the observed value; and the fusion gain coefficient refers to the extracted parameters used for subsequent weighted calculations.
[0121] Specifically, the tower monitoring system executes the Kalman filter time update equation to calculate the predicted state value and the prior error covariance. Subsequently, the system executes the observation update equation, calculates the Kalman gain matrix, and uses this matrix in conjunction with the current observed variables to calculate the optimal state estimate. The system extracts the elements from this gain matrix as fusion gain coefficients representing the weights of the two types of signals.
[0122] In some embodiments, the optimal state estimate and coefficients can be calculated and extracted in several ways: Optionally, the tower monitoring system performs matrix multiplication and addition operations according to the standard Kalman filter iterative formula, and writes the gain matrix to a specified memory address for subsequent calls at the end of each iteration; Optionally, the tower monitoring system uses the square root Kalman filter algorithm to update the square root of the covariance matrix through orthogonal transformation, thereby improving the stability of numerical calculation and extracting the fused gain coefficients. It is understood that other methods can also be used to calculate the optimal state estimate and extract coefficients, and this is not limited here.
[0123] In some embodiments, matrix inversion may fail due to matrix singularity. To address this, the tower monitoring system performs a condition number check on the matrix before inversion. If the condition number is too large, a small regularization perturbation term is added to the diagonal to ensure matrix invertibility.
[0124] S210. Perform weighted early fusion of the stress energy feature sequence and the ultrasonic energy feature sequence according to the fusion gain coefficient to generate a cross-modal energy feature vector.
[0125] Weighted early fusion refers to the process of linearly combining different modal data at the feature layer (rather than the decision layer) according to specific weights.
[0126] Specifically, the tower monitoring system reads the extracted fusion gain coefficients and uses them as multiplication weights for the stress energy feature sequence and the ultrasonic energy feature sequence, respectively. The system performs element-wise addition on the two weighted sequences to generate a fused cross-modal energy feature vector; this effectively suppresses interference from high-noise modes and highlights the characteristics of high-quality modes.
[0127] In some embodiments, weighted early fusion can be performed in several ways: Optionally, the tower monitoring system utilizes a hardware multiply-adder array to simultaneously perform multiplication and addition instructions on corresponding elements of the two sets of sequences, outputting a fusion vector; alternatively, the tower monitoring system concatenates the two sets of sequences into a matrix form and performs matrix multiplication with a weight vector containing fusion gain coefficients to generate a cross-modal energy feature vector. It is understood that other methods can also be used to perform weighted early fusion, which are not limited here.
[0128] In some embodiments, the numerical range of the feature vectors may shift after weighted summation. To address this, the tower monitoring system adds a scaling and translation transformation after the fusion operation to remap the cross-modal energy feature vectors to the standard range, adapting them to the input requirements of the subsequent inference network.
[0129] S211. Input the cross-modal energy feature vector into the Bayesian inference network to perform probability inference operations and obtain the joint probability distribution of the connection states.
[0130] Refer to step S105, which will not be repeated here.
[0131] In some embodiments, the tower monitoring system performs feature filtering to highlight early defect features. Specifically, the tower monitoring system acquires the historical posterior probability distribution of the previous monitoring cycle and constructs the prior probability density function of the current cycle by combining it with the fastening reliability index. Based on a preset multivariate conditional probability model, it calculates the likelihood function values of the cross-modal energy feature vector under different connection states. Based on the prior probability density function and the likelihood function values, it performs Bayesian formula operations to obtain the joint probability distribution characterizing the early microslip state and the stable fastening state. Based on the joint probability distribution, it performs feature filtering on the fastening features under single-source monitoring to reduce the proportion of fastening features to a preset early fusion feature threshold range.
[0132] Among them, the historical posterior probability distribution refers to the probability result derived from the inference at the previous time step; the prior probability density function refers to the initial probability distribution assumption before the inference at the current time step; the multivariate conditional probability model refers to the mathematical model describing the mapping relationship between features and states; the likelihood function value refers to the probability of observing the current feature under a specific state; feature filtering refers to the process of suppressing useless features and retaining key features; and the feature threshold range refers to the set reasonable proportion range.
[0133] Specifically, the tower monitoring system reads the historical posterior probability distribution, smooths it using a tightness reliability index, and generates the prior probability density function for the current period. The system substitutes the cross-modal energy feature vector into a multivariate conditional probability model to calculate the likelihood function value for each state. The system performs Bayesian formula calculations to obtain the joint probability distribution and uses this distribution as a mask to attenuate and filter strong tightness features in the original features, thus highlighting micro-slip features.
[0134] In some embodiments, feature filtering can be performed in several ways: Optionally, the tower monitoring system generates an attenuation coefficient based on the probability value of a stable fastening state in the joint probability distribution, multiplies it by the original fastening feature vector, and reduces the feature proportion; Optionally, the tower monitoring system sets a hard threshold function, and when the probability of determining a micro-slip state exceeds the threshold, directly sets the value of the corresponding dimension of the fastening feature to zero, thus completing feature filtering. It is understood that other methods can also be used to perform feature filtering, which are not limited here.
[0135] It should be noted that when filtering the fastened features under single-source monitoring based on the joint probability distribution, the tower monitoring system uses the joint probability distribution output by the Bayesian inference network as a dynamic adjustment mask to extract the posterior probability value representing the "stable fastened state" and calculate its deviation coefficient from the preset standard probability. The system converts this deviation coefficient into a decay weight with a value between 0 and 1, which is directly applied to the feature components belonging to the stable connected state subspace (band 1 and band 2) in the cross-modal energy feature vector. Through multiplication operations, these dominant fastened features are numerically suppressed. For example, when the Bayesian network detects an increase in the probability of an "early micro-slip state," the system simultaneously generates a decay weight of 0.5, multiplies it by the value of the tight feature in the feature vector, thereby forcibly weakening the absolute amplitude of the tight feature at the underlying data structure. This causes the proportion of the tight feature in the overall energy to rapidly and smoothly decrease from 60% to 70% in single-source monitoring to the early fusion feature threshold range of 30% to 45%, fundamentally amplifying the subtle loosening potential features and ensuring high sensitivity for subsequent state label matching.
[0136] In some embodiments, the prior probability density function may deviate from the true state due to the accumulation of historical errors. To address this, the tower monitoring system introduces a forgetting factor, gradually reducing the weight of historical posterior probabilities and increasing the weight of current fastening reliability indicators when constructing prior probabilities, thus preventing error divergence.
[0137] S212. Based on the maximum probability value in the joint probability distribution, match the corresponding status label to determine the tower connection status result.
[0138] Refer to step S106, which will not be repeated here.
[0139] S213. Generate corresponding wind turbine control commands based on the tower connection status results to perform status response operations on the wind turbine generator set.
[0140] Refer to step S107, which will not be repeated here.
[0141] In this embodiment, cross-modal feature fusion and Bayesian inference network are used, which can integrate the deep correlation of multi-source heterogeneous data, effectively solve the problem of weak coupling feature loss caused by single modality judgment in the prior art, and thus realize the reduction of early micro-defect recognition delay.
[0142] The tower monitoring system in the embodiments of this invention is described below from the perspective of hardware processing. Please refer to [link / reference]. Figure 3 This is a schematic diagram of the physical device structure of a tower monitoring system in an embodiment of this application.
[0143] It should be noted that, Figure 3 The structure of the tower monitoring system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.
[0144] like Figure 3 As shown, the tower monitoring system includes a CPU 301, which can perform various appropriate actions and processes according to a program stored in ROM 302 or a program loaded from storage section 308 into RAM 303, such as executing the methods described in the above embodiments. RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via bus 304. I / O interface 305 is also connected to bus 304.
[0145] The following components are connected to I / O interface 305: input section 306 including audio input devices, push-button switches, etc.; output section 307 including liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 308 including hard disks, etc.; and communication section 309 including network interface cards such as LAN (Local Area Network) cards, modems, etc. Communication section 309 performs communication processing via a network such as the Internet. Drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as disks, optical disks, magneto-optical disks, semiconductor memories, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.
[0146] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by CPU 301, it performs the various functions defined in the present invention.
[0147] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, program segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.
[0148] Specifically, the tower monitoring system of this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the tower connection status monitoring method provided in the above embodiment.
[0149] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the tower monitoring system described in the above embodiments; or it may exist independently and not assembled into the tower monitoring system. The storage medium carries one or more computer programs, which, when executed by a processor of the tower monitoring system, cause the tower monitoring system to implement the tower connection status monitoring method provided in the above embodiments.
[0150] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.
[0151] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".
Claims
1. A method for monitoring the connection status of a tower, characterized in that, The method, applied to a tower monitoring system, includes: Stress signals and ultrasonic signals from the tower connection area were collected to obtain multi-source time-domain data. Perform wavelet packet decomposition on the multi-source time-domain data to obtain multiple frequency-domain feature subspaces; Calculate the sum of the squared amplitudes of the wavelet packet coefficients in each frequency domain feature subspace to generate the frequency band energy value of each frequency domain feature subspace; The energy values of each frequency band corresponding to the stress signal and the energy values of each frequency band corresponding to the ultrasonic sound signal are fused across modal features to generate a cross-modal energy feature vector. The cross-modal energy feature vector is input into the Bayesian inference network to perform probability inference operations, and the joint probability distribution of the connection states is obtained. The tower connection status result is determined by matching the corresponding status label with the maximum probability value in the joint probability distribution. Based on the tower connection status results, corresponding wind turbine control commands are generated to perform status response operations on the wind turbine generator set.
2. The method according to claim 1, characterized in that, The step of acquiring stress signals and ultrasonic signals from the tower connection area to obtain multi-source time-domain data specifically includes: Acquire low-frequency stress signals in the tower connection area; Extract the alternating load fluctuation characteristics of the low-frequency stress signal; When the alternating load fluctuation characteristics meet the preset fatigue excitation conditions, a wake-up command is sent to the ultrasonic acquisition probe in a dormant state to acquire high-frequency ultrasonic signals. Based on the timestamp of the wake-up command, time windows are extracted from the low-frequency stress signal and the high-frequency ultrasonic signal to obtain multi-source time-domain data.
3. The method according to claim 2, characterized in that, The step of performing time window extraction on the low-frequency stress signal and the high-frequency ultrasonic signal based on the timestamp of the wake-up command to obtain multi-source time-domain data specifically includes: Extract the hardware wake-up delay time of the ultrasonic acquisition probe from the sleep state to the data output state, and calculate the ultrasonic alignment timestamp in combination with the sent timestamp; Starting from the issued timestamp, extract stress data segments from the low-frequency stress signal; Starting from the ultrasonic alignment timestamp, an ultrasonic data segment is extracted from the high-frequency ultrasonic signal; The stress data segment and the ultrasonic data segment are combined to obtain multi-source time-domain data.
4. The method according to claim 1, characterized in that, The step of performing wavelet packet decomposition on the multi-source time-domain data to obtain multiple frequency-domain feature subspaces specifically includes: Perform multi-scale wavelet packet transform on the multi-source time-domain data to obtain an initial frequency-domain node sequence; Calculate the energy entropy value of each node in the initial frequency domain node sequence; Based on the energy entropy value, target frequency domain nodes representing the dominant energy are selected; A frequency band merging operation is performed on the target frequency domain node to generate multiple frequency domain feature subspaces.
5. The method according to claim 4, characterized in that, After the step of performing a band merging operation on the target frequency domain node to generate multiple frequency domain feature subspaces, the method further includes: Extract the target subspace set corresponding to the stress concentration frequency band from the multiple frequency domain feature subspaces; The target subspace set is divided into a stable connection state subspace that characterizes effective fastening features and a loosening potential state subspace that characterizes micro-slippage features; Based on the ratio of the frequency band energy value to the total energy of the stable connected state subspace, a tight reliability index is calculated and used as a state constraint input to the Bayesian inference network.
6. The method according to claim 5, characterized in that, The step of inputting the cross-modal energy feature vector into a Bayesian inference network to perform probability inference operations and obtain the joint probability distribution of the connection states specifically includes: Obtain the historical posterior probability distribution of the previous monitoring period, and construct the prior probability density function of the current period in combination with the fastening reliability index; Based on a pre-defined multivariate conditional probability model, the likelihood function values of the cross-modal energy feature vector under different connectivity states are calculated. Based on the prior probability density function and the likelihood function value, Bayes' theorem is performed to obtain the joint probability distribution characterizing the early microslip state and the stable tight state. Based on the joint probability distribution, feature filtering is performed on the fastening features under single-source monitoring so that the proportion of the fastening features is reduced to a preset early fusion feature threshold range.
7. The method according to claim 1, characterized in that, The step of fusing the energy values of each frequency band corresponding to the stress signal and the energy values of each frequency band corresponding to the ultrasonic signal across modes to generate a cross-modal energy feature vector specifically includes: The energy values of each frequency band corresponding to the stress signal and the energy values of each frequency band corresponding to the ultrasonic signal are normalized to obtain the stress energy feature sequence and the ultrasonic energy feature sequence. A Kalman filter model is constructed with the actual energy characteristics of the tower connection state as the state variable and the stress energy characteristic sequence and the ultrasonic energy characteristic sequence as the observation variables. Based on the Kalman filter model, the optimal state estimate after eliminating random noise is calculated, and the gain matrix of the Kalman filter model in the observation update process is extracted as the fusion gain coefficient. Based on the fusion gain coefficient, a weighted early fusion is performed on the stress energy feature sequence and the ultrasonic energy feature sequence to generate a cross-modal energy feature vector.
8. A tower monitoring system, characterized in that, The tower monitoring system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the tower monitoring system to perform the method as described in any one of claims 1-7.
9. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on the tower monitoring system, the tower monitoring system performs the method as described in any one of claims 1-7.
10. A computer program product, characterized in that, When the computer program product is run on the tower monitoring system, it causes the tower monitoring system to perform the method as described in any one of claims 1-7.
