Wind turbine health assessment method and system based on ai multi-source data fusion

Abstract

The application relates to the technical field of data analysis, and provides a wind turbine health evaluation method and system based on AI multi-source data fusion, which are used for realizing accurate and dynamic wind turbine health evaluation. The method comprises the following steps: collecting a first operation data set of a wind turbine through a set sensor network, performing multidimensional state feature mining on the first operation data set to obtain a multidimensional state feature set; combining the feature set, using a pre-constructed state deduction model to perform state deduction iteration, introducing different initial conditions and parameter disturbances to perform multi-path deduction, and generating a plurality of state deduction iteration results; performing fusion processing on the state deduction iteration results based on conflict resolution, combining a feature correlation weight matrix and a credibility evaluation mechanism to obtain a state deduction fusion result; and finally, performing state evaluation based on synchronous time sequence analysis on the state deduction fusion result and a second operation data set collected at a time later than the first operation data set, to obtain a health state evaluation label of the wind turbine.

Description

Technical Field

[0001] This application belongs to the field of data analysis technology, specifically relating to a method and system for health assessment of wind turbine units based on AI multi-source data fusion. Background Technology

[0002] In the field of wind turbine health assessment, early technologies relied solely on single-type sensor data, such as collecting vibration data to evaluate the turbine's health. This approach yielded limited data, failing to comprehensively reflect the actual operating conditions of the wind turbine. Later, multi-sensor data acquisition was adopted, but this simply involved summarizing and analyzing the data without delving into the underlying dynamic characteristics and potential anomaly patterns. Regarding state extrapolation, traditional methods often rely on fixed initial conditions and parameters for single-path extrapolation, failing to account for the uncertainties and diversity inherent in wind turbine operation. This leads to significant discrepancies between extrapolated results and actual conditions. Furthermore, existing technologies lack effective conflict resolution mechanisms, often resorting to simple discarding or averaging for conflicting data, failing to achieve effective information complementarity and impacting the accuracy and reliability of the fused results. Therefore, existing technologies lack flexibility and accuracy in wind turbine health assessment. Summary of the Invention

[0003] This application provides a method and system for wind turbine health assessment based on AI multi-source data fusion, which can achieve accurate and dynamic wind turbine health assessment.

[0004] In a first aspect, embodiments of this application provide a wind turbine health assessment method based on AI multi-source data fusion, applied to a wind turbine health assessment system, the method comprising:

[0005] The initial operational dataset of the wind turbine was collected through the established sensor network.

[0006] Multidimensional state feature mining is performed on the first running dataset to obtain a multidimensional state feature set characterizing the dynamic characteristics and potential anomaly modes of the wind turbine.

[0007] Combining the multidimensional state feature set, the wind turbine is subjected to state deduction iteration based on dynamic characteristics and operating laws using a pre-built state deduction model. Different initial conditions and parameter disturbances are introduced during the state deduction iteration process to carry out multi-path deduction, generating multiple state deduction iteration results that reflect the evolution trend of different operating states.

[0008] The multiple state inference iteration results are fused based on conflict resolution, and conflict coordination and information complementarity are carried out in combination with the established feature correlation weight matrix and credibility assessment mechanism to generate state inference fusion results. The state inference fusion results and the second operating dataset are subjected to state assessment based on synchronous time series analysis to obtain the health status assessment label of the wind turbine. The collection time of the second operating dataset is later than the collection time of the first operating dataset.

[0009] Secondly, embodiments of this application provide a wind turbine health assessment system, which includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the above-described method.

[0010] Thirdly, embodiments of this application provide a computer-readable storage medium including a computer program, which, when run on a wind turbine health assessment system, causes the wind turbine health assessment system to perform the steps of the above-described method.

[0011] This application embodiment, by setting up a sensor network to collect a first operating dataset and performing multi-dimensional state feature mining on the first operating dataset, can deeply analyze the dynamic characteristics and potential abnormal modes of wind turbine units, uncover key information that is difficult to discover using traditional methods, and improve the depth of understanding of the operating status of wind turbine units.

[0012] By using a pre-built state extrapolation model for state extrapolation iteration and introducing different initial conditions and parameter disturbances for multi-path extrapolation, the uncertainty and diversity of wind turbine operation process are fully considered, generating multiple results that reflect the evolution trend of different operating states. Compared with single-path extrapolation, it can more comprehensively predict the future state of wind turbine.

[0013] By performing conflict resolution-based fusion processing on multiple state inference iteration results, and combining feature correlation weight matrix and credibility assessment mechanism, the conflict problem in the multi-source data fusion process is effectively solved, conflict coordination and information complementarity are achieved, and the state inference fusion results are made more accurate and reliable.

[0014] Finally, the state simulation fusion results are combined with the second operating dataset to perform a state assessment based on synchronous time-series analysis, resulting in a health status assessment label for the wind turbine. This assessment method, which combines data from different time periods, can dynamically and accurately reflect the actual health status of the wind turbine, improving the accuracy and timeliness of the health assessment. It can also promptly detect potential faults in the wind turbine, reduce maintenance costs, and improve power generation efficiency. Attached Figure Description

[0015] Figure 1This is a flowchart illustrating a wind turbine health assessment method based on AI multi-source data fusion, provided in an embodiment of this application.

[0016] Figure 2 This is a schematic diagram of the structure of a wind turbine health assessment system provided in an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this application. Obviously, the described embodiments are only some embodiments of the technical solutions of this application, and not all embodiments. Based on the embodiments recorded in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the technical solutions of this application.

[0018] See Figure 1 This is a wind turbine health assessment method based on AI multi-source data fusion provided in the embodiments of this application. This method can be applied to the wind turbine health assessment system. The specific process is as follows: steps 110-140.

[0019] Step 110: Collect the first operational dataset of the wind turbine through the established sensor network.

[0020] In wind turbine health assessment scenarios, to comprehensively acquire wind turbine operating information, a sensor network is installed at multiple key locations within the wind turbine. These sensors are diverse in type and function, each responsible for collecting different types of data. For example, speed sensors and torque sensors are installed on the rotor. The speed sensor converts the rotor's rotational speed into an electrical signal for recording using the principle of electromagnetic induction, while the torque sensor uses devices such as strain gauges to measure the magnitude of the torque acting on the rotor. Voltage sensors and current sensors are installed at the generator. The voltage sensor measures the generator's output voltage based on the principle of electromagnetic induction or resistive voltage division, while the current sensor acquires current data through methods such as the Hall effect. Temperature sensors and vibration sensors are arranged at the gearbox. The temperature sensor uses elements such as thermistors or thermocouples to sense temperature changes in the gearbox, while the vibration sensor detects gearbox vibration through piezoelectric effects.

[0021] The sensor network continuously collects data at a certain sampling frequency. The sampling frequency setting needs to take into account factors such as data timeliness and storage costs. The collected data is transmitted to the data storage device in the form of digital signals, forming the first operational dataset, which contains various operating parameter information of the wind turbine over a period of time.

[0022] Step 120: Perform multidimensional state feature mining on the first running dataset to obtain a multidimensional state feature set characterizing the dynamic characteristics and potential anomaly modes of the wind turbine.

[0023] The first running dataset contains a large amount of raw data. In order to extract information that can reflect the dynamic characteristics and potential anomaly modes of wind turbines, multi-dimensional state feature mining is required.

[0024] Step 121: Perform time series decomposition processing on the first running dataset to separate a multi-component time series set containing periodic fluctuation features and non-periodic fluctuation features.

[0025] When performing time series decomposition on the first running dataset, a suitable time series decomposition algorithm is employed. Taking wind turbine power data as an example, this data exhibits certain regularities in its changes over time, but also contains some irregular fluctuations. Common time series decomposition methods include additive and multiplicative models; the embodiments of this application can select a suitable model based on the characteristics of the data.

[0026] In the additive model, the time series is decomposed into trend components, seasonal components, and residual components. The trend component reflects the long-term trend of the data, the seasonal component reflects the periodic fluctuations, and the residual component contains aperiodic fluctuation characteristics. By decomposing the power data, periodic fluctuation characteristics, such as daily or weekly periodic changes, and aperiodic fluctuation characteristics, which may be caused by sudden changes in weather conditions or temporary equipment failures, can be obtained. This yields a multi-component time series set containing both periodic and aperiodic fluctuation characteristics.

[0027] Step 122: Perform dynamic characteristic extraction processing on each component time series in the multi-component time series set to generate vibration mode features reflecting mechanical vibration characteristics and energy transfer features reflecting energy conversion efficiency.

[0028] Dynamic characteristics are extracted for each component time series in a multi-component time series dataset. For the vibration data components of wind turbines, signal processing methods are used to extract vibration mode features. First, the vibration data is filtered to remove noise interference. Then, the time-domain signal is converted into a frequency-domain signal using methods such as Fourier transform to extract vibration frequency, amplitude, and other features. These features can reflect the mechanical vibration characteristics of the wind turbine; for example, vibrations at different frequencies may correspond to failures in different components.

[0029] For energy-related data components, such as power and wind speed data, the relationships between them are analyzed to extract energy transfer characteristics that reflect energy conversion efficiency. For example, power output at different wind speeds is calculated to obtain the energy conversion efficiency coefficient. By statistically analyzing the efficiency coefficients at multiple time points, the distribution and variation patterns of energy transfer characteristics are obtained.

[0030] Step 123: Perform correlation analysis on the vibration mode features and the energy transfer features to identify the synergistic change patterns between the vibration mode features and the energy transfer features, and generate a feature correlation map.

[0031] When performing correlation analysis on the extracted vibration mode features and energy transfer features, statistical analysis and machine learning algorithms are employed. First, the correlation coefficient between the vibration mode features and energy transfer features is calculated to assess their linear relationship. Then, machine learning algorithms, such as decision trees and neural networks, are used to explore the nonlinear relationships between them.

[0032] Analysis revealed that in some cases, an increase in vibration frequency is accompanied by a decrease in energy conversion efficiency, potentially indicating a mechanical fault in the wind turbine that affects normal energy conversion. Based on these coordinated variation patterns, a feature correlation map was generated. In the map, nodes represent features, edges represent the correlations between features, and edge weights indicate the strength of the correlation. The feature correlation map provides a visual representation of the relationship between vibration mode features and energy transfer features.

[0033] Step 124: Based on the feature association map, perform anomaly-sensitive feature screening on the vibration mode features and the energy transfer features, and extract a subset of key features that are sensitive to potential anomalous modes.

[0034] Based on the generated feature correlation maps, anomaly-sensitive features are screened for vibration mode features and energy transfer features. The correlation strength and trends between various features are analyzed within the feature correlation maps. Features closely associated with potential anomalous modes and exhibiting high sensitivity to change are given special attention.

[0035] For example, analysis of historical data reveals that when a specific frequency vibration mode exhibits abnormal changes, energy conversion efficiency also decreases significantly, and this is often accompanied by wind turbine failures. Therefore, the vibration mode characteristics at this specific frequency and the corresponding energy conversion efficiency characteristics are identified as key features sensitive to potential abnormal modes. These key features are extracted using screening algorithms, such as threshold-based methods, forming a subset of key features.

[0036] Step 125: Perform dimensional reorganization on the key feature subset to generate a multidimensional state feature set containing time dimension features, frequency dimension features, and amplitude dimension features.

[0037] The key feature subset is restructured to generate a multidimensional state feature set containing time-dimensional, frequency-dimensional, and amplitude-dimensional features. First, the time dimension of each feature in the key feature subset is analyzed to extract the feature's variation over time, such as the rate of change and periodicity, thus forming the time-dimensional feature.

[0038] Then, frequency dimension analysis is performed on the features. Time-domain features are converted to frequency-domain features using methods such as Fourier transform, and the frequency components of the features are extracted to form frequency dimension features. Finally, amplitude information of the features, such as vibration amplitude and energy conversion efficiency, is extracted to form amplitude dimension features.

[0039] By combining time-dimensional features, frequency-dimensional features, and amplitude-dimensional features, a multi-dimensional state feature set is obtained, which can more comprehensively and accurately characterize the dynamic characteristics and potential anomaly modes of wind turbine units.

[0040] In the exemplary application, the time series decomposition process in step 121 uses an additive model. For the power data of wind turbines, the decomposition parameters are configured as follows: the trend component is calculated using the moving average method, with the moving window size set to 7 days, meaning a trend value is calculated for every 7 days of data; the seasonal component is determined by calculating the average value of the same period each year, for example, the average power value of the same month each year is used as the seasonal component for that month; the residual component is obtained by subtracting the trend component and the seasonal component from the original data. The above parameter configuration can effectively separate the periodic and non-periodic fluctuation characteristics in the power data.

[0041] In step 122, the Fourier transform algorithm is used to extract vibration mode features. For the vibration data of the wind turbine, the sampling frequency is set to 100Hz, that is, 100 data points are collected per second. When performing the Fourier transform, a data length of 1024 points is selected, which can obtain a relatively accurate frequency resolution in the frequency domain. In this way, the frequency and amplitude features of the vibration can be accurately extracted.

[0042] In step 123, a decision tree algorithm is used for association analysis. The maximum depth of the decision tree is set to 5, meaning the tree has a maximum of 5 levels; the minimum number of sample splits is set to 10, meaning each internal node requires at least 10 samples to continue splitting; and the minimum number of leaf nodes is set to 5, meaning each leaf node requires at least 5 samples. This approach improves the accuracy of association analysis while ensuring moderate model complexity.

[0043] Step 130: Combining the multidimensional state feature set, the wind turbine is subjected to state deduction iteration based on dynamic characteristics and operating laws using a pre-constructed state deduction model. Different initial conditions and parameter disturbances are introduced during the state deduction iteration process to conduct multi-path deduction and generate multiple state deduction iteration results that reflect the evolution trend of different operating states.

[0044] By combining the obtained multidimensional state feature set, the wind turbine is iterated through state inference using a pre-built state inference model. This model is based on the dynamic characteristics and operating laws of the wind turbine and can simulate the evolution of the wind turbine's operating state under different conditions.

[0045] Step 131: Standardize the multidimensional state feature set based on the historical operating state records of the wind turbine to generate a standardized multidimensional state feature set; input the standardized multidimensional state feature set into the initial state configuration layer of the state inference model, and generate multiple differentiated initial state vectors based on the historical operating state records of the wind turbine.

[0046] The multidimensional state feature set is standardized based on the historical operating status records of the wind turbine. Since different features may have different value ranges and units, standardization makes these features comparable. Exemplary standardization methods include Z-score standardization and Min-Max standardization.

[0047] Taking Z-score standardization as an example, the mean and standard deviation of each feature are calculated. The mean is subtracted from the feature value, and then the result is divided by the standard deviation to obtain the standardized feature value. In this way, a multidimensional state feature set is transformed into a standardized multidimensional state feature set.

[0048] A standardized multidimensional state feature set is input into the initial state configuration layer of the state extrapolation model. In this layer, multiple differentiated initial state vectors are generated based on the historical operating state records of the wind turbine. By analyzing historical data, different initial conditions, such as different wind speeds and rotor speeds, are determined. Combined with the standardized multidimensional state feature set, multiple vectors representing different initial states are generated. These initial state vectors provide different starting points for multi-path extrapolation.

[0049] Step 132: Perform parameter perturbation processing on each initial state vector, randomly adjust the dynamic characteristic parameters within the preset perturbation range, and generate a set of perturbation state vectors containing different parameter combinations.

[0050] Each initial state vector is subjected to parameter perturbation to simulate various uncertainties that the wind turbine may encounter during operation. Within a preset perturbation range, dynamic characteristic parameters, such as wind speed, rotor blade angle, and gearbox transmission efficiency, are randomly adjusted.

[0051] For example, within a preset wind speed disturbance range, a certain wind speed value can be randomly increased or decreased, while the angle of the wind turbine blades can be slightly adjusted. By randomly combining and adjusting multiple dynamic characteristic parameters, a set of disturbance state vectors containing different parameter combinations is generated. These disturbance state vectors provide more possibilities for multi-path simulation, making the simulation results more reflective of various situations in actual wind turbine operation.

[0052] Step 133: Input the set of disturbance state vectors into the deduction iteration layer of the state deduction model, perform multi-step state deduction calculation based on the wind turbine dynamic equation, and generate a state evolution path sequence corresponding to each disturbance state vector.

[0053] The set of disturbance state vectors is input into the derivation iteration layer of the state derivation model. In this layer, multi-step state derivation calculations are performed based on the wind turbine dynamic equations.

[0054] Step 1331: Convert each perturbation state vector in the set of perturbation state vectors into a standard state vector that matches the input format of the state deduction model.

[0055] Different models may have specific requirements for the format of the input data. Therefore, each perturbation state vector needs to be converted into a standard state vector that matches the input format of the state inference model. First, the dimensions of the perturbation state vector are checked to ensure that they are consistent with the model input requirements. If the dimensions are inconsistent, they are adjusted using methods such as interpolation or dimensionality reduction.

[0056] Then, the order of the elements in the disturbance state vector is adjusted to conform to the order requirements of the model input. For example, the model requires the input vector elements to be arranged in the order of wind speed, turbine speed, and power, but the order of the elements in the disturbance state vector may be different, requiring corresponding adjustments. Through these processes, the disturbance state vector is transformed into a standard state vector.

[0057] Step 1332: Call the dynamic calculation kernel of the state deduction model, and perform a single-step state transition calculation on the standard state vector based on the mechanical transmission system characteristics and aerodynamic characteristic equations of the wind turbine, to obtain the state prediction vector for the next moment.

[0058] The dynamic calculation kernel of the state derivation model is invoked, which contains the mechanical transmission system characteristics and aerodynamic characteristic equations of the wind turbine. Using the standard state vector as input, a single-step state transition calculation is performed based on these equations.

[0059] For example, based on the current rotational speed and blade angle of the wind turbine, and combining aerodynamic principles and mechanical transmission relationships, the state parameters of the wind turbine at the next moment, such as rotational speed and torque, are calculated. During the calculation process, factors such as the inertia and friction of the wind turbine are considered to ensure the accuracy of the results. The state prediction vector for the next moment is obtained through single-step state transition calculation.

[0060] Step 1333: Extract the key state parameters from the state prediction vector, normalize and compare them with similar parameters in the historical operating state database, and generate a parameter similarity score.

[0061] Key state parameters, such as power and rotational speed, are extracted from the state prediction vector. These key state parameters are then normalized to similar parameters in the historical operating state database to ensure comparability. The normalization process can employ the same standardization method as in step 131.

[0062] Next, a similarity comparison algorithm is used to calculate the similarity between the predicted parameters and historical parameters, generating a parameter similarity score. Common similarity comparison algorithms include Euclidean distance and cosine similarity. Taking Euclidean distance as an example, the distance between the predicted parameters and historical parameters is calculated; the smaller the distance, the higher the similarity. The distance value is then converted into a similarity score; a higher score indicates a greater similarity between the predicted and historical states.

[0063] Step 1334: Perform credibility weighting on the state prediction vector based on the parameter similarity score to generate a weighted state prediction vector.

[0064] Based on the generated parameter similarity scores, the state prediction vectors are weighted according to their credibility. If the parameter similarity score is high, it means that the predicted state is more similar to the historical state, and the prediction vector has high credibility and is given a higher weight; conversely, if the similarity score is low, it is given a lower weight.

[0065] For example, the parameter similarity score can be used as a weighting coefficient and multiplied by each element in the state prediction vector to obtain a weighted state prediction vector. In this way, the state prediction vector can be adjusted to better reflect the actual situation.

[0066] Step 1335: Using the weighted state prediction vector as the new input state vector, repeatedly perform single-step state transition calculation, similarity comparison and confidence weighting processing until the preset number of deduction steps are completed, generating a state evolution path sequence containing timestamps.

[0067] The weighted state prediction vector is used as the new input state vector, and the single-step state transition calculation, similarity comparison, and confidence weighting process described above are repeated. Each calculation result is timestamped, recording the time when each state occurred.

[0068] For example, after calculating the state prediction vector in the first step, similarity comparison and confidence weighting are performed to obtain a weighted state prediction vector. This vector is then used as the new input for the second step of single-step state transition calculation, and so on. This process is repeated until the preset number of steps are completed, ultimately generating a state evolution path sequence containing timestamps. This sequence records in detail the evolution of the wind turbine's operating state at different times.

[0069] Step 134: During the generation of the state evolution path sequence, the deviation between the deduced state and the historical operating state is monitored in real time. When the deviation exceeds the preset deviation, the path correction mechanism is triggered to dynamically adjust the state transition matrix of the subsequent deduced steps.

[0070] During the generation of the state evolution path sequence, it is necessary to monitor the deviation between the simulated state and the historical operating state in real time to ensure the accuracy of the simulation results.

[0071] Step 1341: Extract the deduced state vector at the current moment from the state evolution path sequence, and extract the actual operating state vector at the corresponding moment from the historical operating state database.

[0072] Extract the inferred state vector at the current moment from the state evolution path sequence. Since the state evolution path sequence contains state information at each moment, the inferred state vector at the current moment can be accurately extracted by using timestamp marking.

[0073] Simultaneously, the actual operating state vector for the corresponding moment is extracted from the historical operating state database. The historical operating state database records the actual operating state of the wind turbine at different times in the past. Based on the timestamp, the actual operating state vector corresponding to the current moment can be found for deviation calculation.

[0074] Step 1342: Normalize the inferred state vector and the actual operating state vector, calculate the cosine similarity or correlation coefficient between the normalized vectors, and convert the cosine similarity or correlation coefficient into a difference index as a deviation quantification index.

[0075] The projected state vector and the actual operating state vector are normalized to make them comparable. The normalization process can use the same standardization method as in step 131.

[0076] Next, the cosine similarity or correlation coefficient between the normalized vectors is calculated. Cosine similarity measures the cosine of the angle between two vectors; a value closer to 1 indicates greater similarity. The correlation coefficient reflects the strength of the linear relationship between the two vectors. The calculated cosine similarity or correlation coefficient is then converted into a difference index, for example, by subtracting the cosine similarity or correlation coefficient from 1. The result is used as a deviation metric to measure the degree of deviation between the projected state and the actual state.

[0077] Step 1343: Compare the deviation quantification index with a preset deviation threshold. When the deviation quantification index is greater than the deviation threshold, activate the path correction mechanism.

[0078] The calculated deviation metric is compared with a preset deviation threshold. The preset deviation threshold is determined based on historical data and experience of wind turbine units and is used to determine whether the deviation between the simulated state and the actual state is within an acceptable range.

[0079] If the deviation metric exceeds the deviation threshold, it indicates a significant discrepancy between the simulated state and the actual state, necessitating a correction of the simulation path. In this case, the path correction mechanism is activated to ensure more accurate simulation results.

[0080] Step 1344: Call the state transition matrix adjustment module in the path correction mechanism to calculate the matrix correction coefficient based on the magnitude of the deviation quantification index. The matrix correction coefficient is positively correlated with the deviation quantification index.

[0081] The state transition matrix adjustment module in the path correction mechanism is invoked. This module calculates the matrix correction coefficient based on the magnitude of the deviation metric. Since the matrix correction coefficient is positively correlated with the deviation metric, the larger the deviation, the larger the matrix correction coefficient.

[0082] For example, a mapping function can be used to map the deviation metric to the range of values ​​for the matrix correction coefficients. When the deviation metric increases, the matrix correction coefficients also increase accordingly, allowing the state transition matrix to be dynamically adjusted based on the magnitude of the deviation.

[0083] Step 1345: Use the matrix correction coefficients to perform element-wise weighted adjustment on the state transition matrix in the state deduction model to generate the corrected state transition matrix.

[0084] The calculated matrix correction coefficients are used to perform element-wise weighted adjustments on the state transition matrix in the state deduction model. Specifically, the matrix correction coefficients are multiplied by each element of the state transition matrix to obtain the corrected state transition matrix.

[0085] For example, each element in the state transition matrix represents the probability of transitioning from one state to another, and these probabilities are adjusted using matrix correction coefficients. If the deviation is large, the matrix correction coefficients will be large, resulting in significant adjustments to the elements of the state transition matrix, changing the probabilities of state transitions, and thus altering the deduction path.

[0086] Step 1346: Continue to execute subsequent state deduction calculation steps using the corrected state transition matrix until the deviation quantification index is less than or equal to the deviation threshold or all deduction steps are completed.

[0087] The subsequent state deduction calculation steps are continued using the corrected state transition matrix. During the deduction process, the deviation metric is continuously monitored.

[0088] When the deviation metric is less than or equal to the deviation threshold, it indicates that the deviation between the simulated state and the actual state is within an acceptable range, and path correction is stopped; or, when all simulation steps are completed, the entire simulation process ends. This method ensures the accuracy and reliability of the simulation results.

[0089] Step 135: Repeat the state deduction calculation and path correction mechanism until the deduction process within the preset time window is completed, and output multiple state deduction iteration results containing multiple independent evolution paths.

[0090] Repeat the state deduction calculation and path correction mechanism described above to continuously update the state prediction vector and state transition matrix. Continue the deduction process within a preset time window.

[0091] By introducing different initial conditions and parameter disturbances, multiple state deduction iteration results with multiple independent evolution paths are obtained. These results reflect the evolution trend of the wind turbine's operating state under different conditions, providing rich information for state assessment.

[0092] In the above embodiments, for some non-limiting application examples, in step 131, the Z-score standardization method is used to standardize the multidimensional state feature set. For features such as power and speed of wind turbine generators, their mean and standard deviation are calculated. Taking power features as an example, the mean of power data over a period of time is 500kW, and the standard deviation is 50kW. During standardization, the mean of 500kW is subtracted from each power data point, and then divided by the standard deviation of 50kW to obtain the standardized power value.

[0093] In step 1332, the dynamic calculation kernel of the state deduction model adopts a calculation method based on physical equations. For the mechanical transmission system of the wind turbine, the rotational inertia of the wind turbine is considered to be 10000 kg·m. 2The gearbox has a transmission efficiency of 0.95. During single-step state transition calculations, based on the aforementioned physical parameters and aerodynamic characteristic equations, the rotor speed and torque, among other state parameters, are accurately calculated for the next moment.

[0094] In step 1333, a cosine similarity algorithm is used for similarity comparison. For the state prediction vector and the vector in the historical operating state database, a similarity score is obtained by calculating the cosine value of these two vectors. The cosine similarity value ranges from -1 to 1, with a value closer to 1 indicating higher similarity. In step 1342, when calculating the deviation metric, cosine similarity is used to convert it into a difference metric.

[0095] Step 140: Perform conflict resolution-based fusion processing on the multiple state inference iteration results, and combine the established feature correlation weight matrix and credibility assessment mechanism to coordinate conflicts and complement information, generate state inference fusion results, and perform state assessment based on synchronous time series analysis on the state inference fusion results and the second operating dataset to obtain the health status assessment label of the wind turbine; the collection time of the second operating dataset is later than the collection time of the first operating dataset.

[0096] Multiple state simulation iterations are fused to obtain more accurate state assessment results. Simultaneously, synchronous time-series analysis is performed using the second operational dataset to determine the health status assessment labels for the wind turbine units.

[0097] Step 141: Perform time axis alignment processing on the multiple state deduction iteration results to keep all state deduction iteration results synchronized in the time dimension; extract the key state feature sequence in each state deduction iteration result, and calculate the feature similarity matrix between different state deduction iteration results.

[0098] Time axis alignment is performed on the results of multiple state extrapolation iterations. Since different extrapolation results may have different time starting points and sampling frequencies, time axis alignment is required to keep all results synchronized in the time dimension.

[0099] First, determine the maximum time range of all simulation results. Then, using the minimum sampling interval as a benchmark, interpolate each simulation result to ensure that it has a corresponding state value at the same time point.

[0100] Next, key state feature sequences, such as power and rotational speed, are extracted from each state extrapolation iteration result. A feature similarity matrix is ​​calculated between different state extrapolation iteration results. Using similarity calculation methods, such as Euclidean distance and cosine similarity, the similarity between every two key state feature sequences is calculated, and the similarity values ​​are combined into a matrix. This matrix reflects the degree of similarity between different extrapolation results.

[0101] Step 142: Identify conflicting feature sequence pairs based on the feature similarity matrix, classify the conflict type of the conflicting feature sequence pairs, and determine the feature dimension and time node where the conflict occurred.

[0102] Step 1421: Traverse the feature similarity matrix according to the preset feature similarity threshold, and mark feature sequence pairs with similarity values ​​lower than the feature similarity threshold as potential conflicting feature sequence pairs.

[0103] The feature similarity matrix is ​​iterated through according to a preset feature similarity threshold. The preset feature similarity threshold is determined based on experience and historical data and is used to determine whether there is a conflict between two feature sequences.

[0104] When the similarity value is below the threshold, the corresponding feature sequence pair is marked as a potentially conflicting feature sequence pair. For example, for power feature sequences, if the similarity between the power feature sequences in two inference results is below the threshold, then these two sequence pairs are marked as potentially conflicting feature sequence pairs.

[0105] Step 1422: Perform time series segmentation processing on the potential conflict feature sequence pairs, dividing each feature sequence into multiple time window segments of equal length.

[0106] The potential conflict feature sequence pairs are processed by time series segmentation. Each feature sequence is divided into multiple time window segments of equal length according to a certain time length.

[0107] For example, the power feature sequence can be divided into multiple time window segments, with each hour serving as a time window. This allows for a more detailed analysis of the distribution of conflicts at different points in time.

[0108] Step 1423: Normalize the feature sequence pairs within each time window segment, calculate the dynamic time bending distance of the normalized sequence pairs, and mark the time window segments with a dynamic time bending distance greater than a preset distance threshold as conflict time windows.

[0109] The feature sequence pairs within each time window segment are normalized to make them comparable. The normalization process can use the same standardization method as in step 131.

[0110] Next, the dynamic time warp distance of the normalized sequence pairs is calculated. Dynamic time warp distance is a method for measuring the similarity between two time series, and it can handle issues of different sequence lengths and time offsets. Time window segments with dynamic time warp distances greater than a preset distance threshold are marked as conflict time windows. The preset distance threshold is determined based on experience and historical data and is used to determine whether two time series conflict within this time window.

[0111] Step 1424: Extract the start and end timestamps of the conflict time window in the feature sequence to determine the target time node where the conflict occurred.

[0112] Extract the start and end timestamps from the segments marked as conflict time windows. These two timestamps determine the target time point at which the conflict occurred.

[0113] For example, if a conflict window starts at 9:00 AM and ends at 10:00 AM, then these two times are the target time points for the conflict to occur. By determining the target time points, the causes and impacts of the conflict can be further analyzed.

[0114] Step 1425: Perform feature dimension decomposition on the feature sequence pairs within the conflict time window, identify the target feature dimension where there is a difference jump, and mark the target feature dimension as the conflict feature dimension.

[0115] The feature sequence pairs within the conflict time window are decomposed into feature dimensions. Each feature sequence may contain multiple feature dimensions, such as power, speed, and temperature.

[0116] By analyzing the changes in feature sequences across different feature dimensions, target feature dimensions exhibiting discrepancies are identified. For example, if within a conflict time window, the power feature sequence shows a significant discrepancy, while the speed and temperature feature sequences show relatively small changes, then the power feature dimension is the conflicting feature dimension. These target feature dimensions are then labeled as conflicting feature dimensions.

[0117] Step 1426: Based on the number and type of the conflict feature dimensions and the distribution characteristics of the conflict time windows, classify the conflict feature sequence pairs into conflict types and generate a conflict analysis report containing a conflict type identifier, a set of conflict feature dimensions, and a list of conflict time nodes.

[0118] Based on the number and type of conflict feature dimensions, as well as the distribution characteristics of conflict time windows, conflict feature sequence pairs are classified into conflict types.

[0119] For example, if there is only one conflict feature dimension and the conflict time window is concentrated in one time period, it can be classified as a single-dimensional local conflict; if there are multiple conflict feature dimensions and the conflict time windows are more dispersed, it can be classified as a multi-dimensional global conflict.

[0120] Generate a conflict analysis report containing a conflict type identifier, a set of conflict feature dimensions, and a list of conflict time points. The report details the type of conflict, the feature dimensions involved, and the time point in time when the conflict occurred.

[0121] Step 143: Call the established feature correlation weight matrix, extract the correlation weight values ​​of conflict feature dimensions in historical data, and calculate the credibility score of each conflict feature sequence in combination with the credibility assessment mechanism.

[0122] Step 1431: Perform hierarchical deconstruction on the feature correlation weight matrix, locate the coordinate index of the conflicting feature dimension in the matrix, and extract the row vector corresponding to the coordinate index as the initial correlation weight vector.

[0123] The feature correlation weight matrix is ​​hierarchically deconstructed. This two-dimensional matrix contains elements representing the correlation weight between two feature dimensions. By locating the coordinate indices of conflicting feature dimensions within the matrix, the row vector corresponding to those indices is extracted as the initial correlation weight vector. For example, if the conflicting feature dimension is power, the row containing power is located in the matrix, and the vector from that row is extracted as the initial correlation weight vector. This vector reflects the degree of correlation between the power feature dimension and other feature dimensions.

[0124] Step 1432: Perform correlation mapping processing between the initial correlation weight vector and the historical performance features of the conflict feature dimension in the historical operation database to generate dynamic weight adjustment coefficients. The historical performance features include historical conflict occurrence rate and historical fusion contribution.

[0125] The initial association weight vector is mapped to the historical performance characteristics of the conflict feature dimension in the historical database. These historical performance characteristics include the historical conflict occurrence rate and the historical fusion contribution. The historical conflict occurrence rate reflects the frequency of conflicts occurring in the past for this conflict feature dimension, while the historical fusion contribution reflects the importance of this feature dimension in the fusion process. Through this mapping process, the initial association weight vector is comprehensively considered in conjunction with these historical performance characteristics to generate dynamic weight adjustment coefficients. For example, if one of the conflict feature dimensions has a high historical conflict occurrence rate and a low historical fusion contribution, the dynamic weight adjustment coefficients will correspondingly reduce the weight of that feature dimension.

[0126] Step 1433: The conflicting feature sequences are processed through a multi-source evidence fusion module of the credibility assessment mechanism. The multi-dimensional evidence includes internal consistency evidence of the feature sequences, external correlation evidence with other non-conflicting feature sequences, and reference evidence in similar historical scenarios.

[0127] The multi-source evidence fusion module, employing a credibility assessment mechanism, performs multi-dimensional evidence collection and processing on conflicting feature sequences. This multi-dimensional evidence includes internal consistency evidence of the feature sequences, external correlation evidence with other non-conflicting feature sequences, and reference evidence from historically similar scenarios.

[0128] Evidence of internal consistency of a characteristic sequence can be obtained by analyzing its fluctuations and trend changes, reflecting the stability and reliability of the sequence itself. Evidence of external correlation with other non-conflicting characteristic sequences can be obtained by calculating their correlations, reflecting the synergistic relationship between the current sequence and other characteristic sequences. Reference evidence from similar historical scenarios can be obtained from historical operational databases. By finding historical scenarios similar to the current conflict characteristic sequence and analyzing their processing results and impacts, a reference can be provided for assessing the credibility of the current conflict characteristic sequence.

[0129] Step 1434: Quantify the evidence strength of the collected multi-dimensional evidence to generate an evidence strength value and evidence reliability factor for each piece of evidence.

[0130] The collected multi-dimensional evidence undergoes quantification processing to assess its strength. By setting different quantification standards, the strength of each piece of evidence is quantified, generating an evidence strength value. Simultaneously, considering the source and reliability of the evidence, an evidence reliability factor is assigned to each piece of evidence. For example, evidence from high-precision sensors has a higher reliability factor, while evidence from experience-based judgment has a lower reliability factor.

[0131] Step 1435: Perform weighted correction processing on the initial correlation weight vector based on the weight dynamic adjustment coefficient to generate the corrected correlation weight vector.

[0132] The initial association weight vector is weighted and corrected based on a dynamic adjustment coefficient. The dynamic adjustment coefficient is multiplied by each element of the initial association weight vector to obtain the corrected association weight vector. For example, if the dynamic adjustment coefficient is 0.8 and the initial association weight vector is [0.2, 0.3, 0.5], then the corrected association weight vector is [0.16, 0.24, 0.4]. In this way, the association weights are adjusted based on the historical performance of the conflict feature dimension.

[0133] Step 1436: Normalize the modified association weight vector and the evidence strength values ​​of the multi-dimensional evidence, and then perform a weighted fusion operation to obtain a preliminary credibility score matrix.

[0134] The corrected correlation weight vector and the evidence strength values ​​of the multi-dimensional evidence are normalized to make them comparable. The normalization process can use the same standardization method as in step 131. Then, a weighted fusion operation is performed. The corrected correlation weight vector and the normalized evidence strength values ​​are multiplied by weight and then summed to obtain a preliminary credibility score matrix. This matrix reflects the preliminary credibility score of each conflicting feature sequence under different pieces of evidence.

[0135] Step 1437: Perform element-wise weighted averaging on the preliminary credibility score matrix based on the evidence reliability factor to generate credibility sub-scores for each conflict feature sequence in different conflict time windows.

[0136] The preliminary credibility score matrix is ​​then subjected to an element-wise weighted average based on the evidence reliability factor. Each element in the preliminary credibility score matrix is ​​multiplied by its corresponding evidence reliability factor, then summed and divided by the total number of pieces of evidence to obtain the credibility sub-score for each conflict feature sequence in different conflict time windows.

[0137] For example, if there are three pieces of evidence within a conflict time window for a conflict feature sequence, with initial confidence scores of 0.6, 0.7, and 0.8, and corresponding evidence reliability factors of 0.8, 0.9, and 0.7, then the confidence sub-score of the conflict feature sequence within the conflict time window is (0.6×0.8+0.7×0.9+0.8×0.7) / 3.

[0138] Step 1438: Perform time-series accumulation processing on the credibility sub-scores of all conflict time windows to generate credibility scores covering the entire conflict feature sequence.

[0139] The credibility sub-scores for all conflict time windows are accumulated over time. The credibility sub-scores for each conflict time window are accumulated in chronological order to obtain a credibility score covering the entire conflict feature sequence.

[0140] For example, for a conflict feature sequence with three conflict time windows, the confidence sub-scores are 0.7, 0.8, and 0.9 respectively. Then the confidence score of the conflict feature sequence is 0.7 + 0.8 + 0.9, which reflects the confidence level of the entire conflict feature sequence.

[0141] Step 144: Perform weighted fusion processing on the conflict feature sequences according to the credibility score, prioritize the retention of feature sequences with weight values ​​higher than the preset weight, and perform feature correction processing on feature sequences with weight values ​​lower than the preset weight.

[0142] Conflicting feature sequences are weighted and fused based on credibility scores. The credibility scores are used as weights to calculate a weighted average of the conflicting feature sequences. Feature sequences with weights higher than a preset weight are preferentially retained. The preset weights are determined based on experience and historical data and are used to determine the reliability of a feature sequence. Feature sequences with weights lower than the preset weight are subject to feature correction. For example, low-weight feature sequences can be corrected by interpolation or fitting with other reliable feature sequences to better reflect reality.

[0143] Step 145: Concatenate the fused feature sequence with the non-conflicting feature sequence to generate a state inference fusion result containing complete state feature information; standardize the second operation dataset based on the historical operation state records of the wind turbine to generate a standardized second operation dataset; synchronize the timestamps of the sampling points of the state inference fusion result and the standardized second operation dataset, extract the state feature values ​​of the corresponding time of the synchronized state inference fusion result and the standardized second operation dataset, and calculate the feature deviation sequence.

[0144] The fused feature sequence is concatenated with the non-conflicting feature sequence. The fused feature sequence after conflict resolution is then concatenated with the non-conflicting feature sequence in chronological order to generate a state inference fusion result containing complete state feature information. The second operating dataset is standardized based on the historical operating state records of the wind turbine. The standardization method is the same as the standardization of the multi-dimensional state feature set in step 131, making the second operating dataset comparable. The state inference fusion result and the standardized second operating dataset are synchronized using timestamps for sampling points. Since the sampling times of the two datasets may differ, timestamp synchronization is necessary to ensure they have corresponding state values ​​at the same time points. The state feature values ​​at corresponding times in the synchronized state inference fusion result and the standardized second operating dataset are extracted, and a feature deviation sequence is calculated. By calculating the difference between the state feature values ​​at each corresponding time point, the feature deviation sequence is obtained, which reflects the difference between the state inference fusion result and the actual operating state.

[0145] Step 146: Perform trend analysis on the characteristic deviation sequence to identify the type and rate of change of the deviation.

[0146] Step 1461: Input the feature deviation sequence into the sequence preprocessing module of the trend analysis model for sequence stabilization and outlier identification to generate a purified deviation sequence.

[0147] The characteristic deviation sequence is input into the sequence preprocessing module of the trend analysis model. This module first performs sequence stationarization processing, eliminating trend and seasonal components in the sequence through methods such as differencing and logarithmic transformation, making the sequence stationary. Then, outlier identification processing is performed, identifying outliers in the sequence by setting thresholds or using statistical methods. Outliers can be handled using interpolation or deletion methods. Through these processes, a purified deviation sequence is generated, providing more accurate data for trend analysis.

[0148] Step 1462: Perform multi-scale decomposition on the purified deviation sequence to decompose it into a multi-component sequence set containing fluctuation components and trend components.

[0149] The purified deviation sequence is then subjected to multi-scale decomposition. Common multi-scale decomposition methods include wavelet decomposition. Wavelet decomposition breaks down the deviation sequence into components of different scales, including fluctuation and trend components. The fluctuation component reflects the short-term fluctuations of the sequence, while the trend component reflects the long-term trend. This decomposition allows for a clearer analysis of the variation characteristics of the deviation sequence.

[0150] Step 1463: Extract the trend component from the multi-component sequence set and input it into the trend type recognizer for pattern matching processing. The pattern matching processing includes comparing the trend component with a preset standard trend pattern library. The standard trend pattern library includes continuous upward pattern, continuous downward pattern, fluctuating upward pattern, fluctuating downward pattern, and stable pattern.

[0151] Trend components are extracted from a multi-component sequence set and input into a trend type recognizer for pattern matching. The trend type recognizer contains a pre-defined standard trend pattern library, which includes continuous upward, continuous downward, fluctuating upward, fluctuating downward, and stable patterns. The trend components are compared morphologically with patterns in the standard trend pattern library, and the best-matching pattern is found by calculating similarity or matching degree. For example, methods such as dynamic time curvature distance are used to calculate the similarity between the trend component and each standard pattern, and the pattern with the highest similarity is selected as the matching result.

[0152] Step 1464: Determine the main trend type corresponding to the trend component based on the principle of highest similarity in pattern matching, and extract the fluctuation frequency and fluctuation amplitude features of the fluctuation component.

[0153] The primary trend type corresponding to the trend component is determined based on the principle of highest similarity in pattern matching. If the trend component has the highest similarity to a continuous upward pattern, then the primary trend type is continuous upward. Simultaneously, the fluctuation frequency and amplitude features of the fluctuation component are extracted. The fluctuation frequency can be obtained by calculating the period of the fluctuation component, and the fluctuation amplitude can be obtained by calculating the difference between the maximum and minimum values ​​of the fluctuation component. These features can further describe the changes in the deviation sequence.

[0154] Step 1465: Based on the main trend type, fluctuation frequency, and fluctuation amplitude characteristics, call the trend descriptor generator to generate a trend feature descriptor that includes trend direction identifier, fluctuation characteristic parameters, and trend stability indicators.

[0155] Based on the main trend type, fluctuation frequency, and fluctuation amplitude characteristics, a trend descriptor generator is invoked to generate trend feature descriptors. The trend feature descriptor includes a trend direction identifier, fluctuation characteristic parameters, and a trend stability index. The trend direction identifier is determined according to the main trend type, such as "upward," "downward," or "stable." Fluctuation characteristic parameters include fluctuation frequency and fluctuation amplitude. The trend stability index can be obtained by calculating the variance or standard deviation of the fluctuation components, reflecting the stability of the trend. Through trend feature descriptors, the changing trend of the deviation sequence can be comprehensively described.

[0156] Step 1466: Perform sliding window processing on the purified deviation sequence, set observation windows of different lengths, and calculate the sequence morphology change rate within each observation window.

[0157] A sliding window process was applied to the purified deviation sequence. Observation windows of different lengths were set, such as 10 time points, 20 time points, etc.

[0158] Within each observation window, the morphological change rate of the sequence is calculated. This can be obtained by calculating indicators such as the slope or curvature of the sequence within the window. For example, for an observation window with 10 time points, the morphological change rate of the sequence within that window is obtained by calculating the ratio of the difference between the first and last time points within the window to the window length.

[0159] Step 1467: Perform statistical fusion processing on the morphological change rates of different observation windows, and combine them with the fluctuation characteristic parameters in the trend feature descriptor to generate a comprehensive change trend assessment index.

[0160] The morphological change rates of different observation windows can be statistically fused using weighted average or other statistical methods to synthesize the morphological change rates of different windows.

[0161] By combining the fluctuation characteristic parameters in the trend feature descriptor, a comprehensive evaluation index for the changing trend is generated. For example, by weighting the rate of change of shape with the fluctuation frequency and fluctuation amplitude, a comprehensive evaluation index is obtained, which can more comprehensively reflect the changing trend of the deviation sequence.

[0162] Step 1468: The comprehensive change situation assessment index is semantically converted using the rate of change inference model, mapping the comprehensive change situation assessment index to the corresponding level in the preset rate of change level system, and generating a trend analysis result report containing trend type identifier and rate of change level.

[0163] The comprehensive change situation assessment indicators are semantically transformed using a rate of change inference model. The rate of change inference model includes a pre-defined rate of change level system, such as "rapid increase", "slow increase", "rapid decrease", and "slow decrease".

[0164] The comprehensive trend assessment indicators are mapped to corresponding levels in this grading system. For example, by setting different threshold ranges, the assessment indicators are divided into different levels. A trend analysis results report is generated, which includes a trend type identifier and a rate of change level. This report details the trend and rate of change of the deviation sequence.

[0165] Step 147: Based on the change trend type and change rate, query the preset health status assessment rule base to determine the health status assessment label corresponding to the wind turbine.

[0166] The system queries a pre-defined health status assessment rule base based on the type and rate of change of the trend. This rule base contains health status assessment tags corresponding to different trend types and rates of change, such as "healthy," "sub-healthy," "fault warning," and "fault."

[0167] Based on the trend type identifier and rate of change level in the trend analysis results report, the corresponding health status assessment label is searched in the rule base. For example, if the trend type is continuously rising and the rate of change is rapidly rising, the corresponding label in the rule base might be "fault warning". This method is used to determine the health status assessment label for the wind turbine.

[0168] For example, in step 141, the Euclidean distance algorithm is used when calculating the feature similarity matrix. For two key state feature sequences, such as sequence A=[1, 2, 3, 4, 5] and sequence B=[2, 3, 4, 5, 6], the Euclidean distance is obtained by calculating the square root of the sum of the squares of the differences between their corresponding elements. The calculated Euclidean distance is 2.24. This distance value is converted into a similarity value, for example, by using a mapping function to convert the distance value to the range of 0-1, thus obtaining a similarity score.

[0169] In step 1423, when calculating the dynamic time bending distance of the normalized sequence pairs, the length of the time window segment is set to 10 time points. For feature sequences within two time window segments, their dynamic time bending distance is calculated using a dynamic programming algorithm. For example, if two sequences are [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0] and [0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1], the calculated dynamic time bending distance is 0.5. This distance is compared with a preset distance threshold (e.g., 0.6) to determine whether it is a conflicting time window.

[0170] In step 1431, the feature correlation weight matrix is ​​a 10×10 matrix representing the correlation weights between the 10 feature dimensions. For example, if the conflict feature dimension is power, the row containing power is found in the matrix, and the row vector is [0.2, 0.3, 0.1, 0.15, 0.05, 0.1, 0.05, 0.03, 0.02, 0.0], which is used as the initial correlation weight vector.

[0171] In step 1432, the calculation of the weight dynamic adjustment coefficient takes into account the historical conflict occurrence rate and the historical fusion contribution. For example, the historical conflict occurrence rate of the power feature dimension is 0.2, and the historical fusion contribution is 0.3. The weight dynamic adjustment coefficient is calculated by a linear combination, such as coefficient = 0.8 × (1 - 0.2) + 0.2 × 0.3.

[0172] In step 1461, the sequence preprocessing module of the trend analysis model uses a first-order differencing method for sequence stationarization. For the characteristic deviation sequence, each data point is subtracted from the previous data point to obtain the differencing sequence. For example, the original sequence is [1, 2, 3, 4, 5], and after first-order differencing, it becomes [1, 1, 1, 1]. In outlier identification, a threshold of 3 times the standard deviation is set; that is, if the difference between any data point and the mean exceeds 3 times the standard deviation, it is considered an outlier.

[0173] In step 1463, the trend type identifier uses dynamic time bending distance for pattern matching. The standard trend pattern library contains a continuous upward trend sequence [1, 2, 3, 4, 5], and a fluctuating upward trend sequence [1, 1.5, 2, 2.5, 3], etc. For example, a trend component sequence [1.2, 2.1, 3.2, 4.1, 5.2] is used; by calculating its dynamic time bending distance with each standard pattern, the best-matching pattern is found.

[0174] In an alternative embodiment, the method further includes:

[0175] Step 210: Perform historical data tracing analysis on the generated wind turbine health status assessment labels and extract historical assessment records that are the same as or similar to the current health status assessment labels.

[0176] Historical data retrospective analysis is performed on the generated wind turbine health status assessment labels. Historical assessment records with the same or similar labels as the current one are searched in the historical data. Analysis of these historical records reveals the subsequent operation and handling measures of the wind turbines under the same or similar health conditions. For example, if the current health status assessment label is "fault warning," all records marked "fault warning" are searched in the historical data, and information such as the corresponding wind turbine operating parameters, fault occurrence time, and handling methods are analyzed.

[0177] Step 220: Collect the subsequent operating status data of the wind turbine corresponding to the historical assessment records, and establish a health status evolution case library.

[0178] Collect subsequent operational status data of wind turbines corresponding to historical assessment records. This data includes changes in operating parameters after a failure, maintenance records, and downtime. Organize and categorize this data to establish a health status evolution case library. Each case in the library includes a health status assessment label, corresponding operational status data, and processing results. This case library can provide a reference for the current health status assessment and handling of wind turbines.

[0179] Step 230: Perform feature extraction processing on the case data in the health status evolution case library to identify the status evolution path features corresponding to different health status assessment labels.

[0180] Feature extraction processing is performed on the case data in the health status evolution case library. The changing patterns of operational status data in each case are analyzed to extract features that reflect the status evolution. For example, for cases tagged "fault warning," the changing trends of parameters such as power and speed after the warning are analyzed to extract status evolution path features. By analyzing cases with different health status assessment tags, the status evolution path features corresponding to each tag are identified. These features can help predict the development trend of wind turbines under different health states.

[0181] Step 240: Construct a health status prediction model based on the state evolution path features, input the current multidimensional state feature set into the health status prediction model, and generate a health status prediction sequence within a preset time window.

[0182] A health status prediction model can be built based on state evolution path characteristics, using machine learning algorithms such as neural networks and decision trees. The current multidimensional state feature set is input into the health status prediction model, which learns and predicts based on the state evolution path characteristics. Through the model's calculations, a health status prediction sequence within a preset time window is generated. For example, predicting the health status changes of wind turbine units over the next week, the sequence includes health status assessment labels at different time points.

[0183] Step 250: Assess the risk level of the health status prediction sequence and determine the risk level corresponding to different time points.

[0184] A risk level assessment is performed on the health status prediction sequence. Based on the severity and probability of occurrence of the health status assessment labels, a risk level is assigned to the health status at each time point. For example, a "healthy" status corresponds to a lower risk level, while a "faulty" status corresponds to a higher risk level. By assessing the health status at each time point in the prediction sequence, the risk level corresponding to different time points is determined, and these risk levels can help in developing appropriate maintenance strategies.

[0185] Step 260: Develop a corresponding maintenance recommendation plan based on the risk level. The maintenance recommendation plan includes a recommended maintenance time window and a recommended maintenance measure type.

[0186] Develop corresponding maintenance recommendation plans based on risk levels. For time points with higher risk levels, formulate more timely and stringent maintenance measures. The maintenance recommendation plan includes a suggested maintenance time window and suggested maintenance measure types. For example, if the risk level of a time point is "high," the suggested maintenance time window can be set to the near future, and suggested maintenance measure types can include comprehensive inspection, replacement of vulnerable components, etc. By developing maintenance recommendation plans, wind turbine maintenance can be carried out in advance, reducing the probability of failure.

[0187] Step 270: Perform a coordination analysis between the maintenance recommendation scheme and the current operation plan of the wind turbine, adjust the time window of the maintenance recommendation scheme, and output the adjusted maintenance recommendation scheme.

[0188] A coordination analysis is conducted between the proposed maintenance plan and the current operating schedule of the wind turbines. Considering factors such as the wind turbines' power generation targets and maintenance resources, it is determined whether the timing of the proposed maintenance plan conflicts with the current operating schedule. If a conflict exists, the timing of the proposed maintenance plan is adjusted. For example, if the proposed maintenance timing conflicts with peak power generation periods, the maintenance time can be adjusted to off-peak periods. Through coordination analysis and adjustment, a revised maintenance plan is output to ensure that maintenance work can guarantee the safe operation of the wind turbines while minimizing the impact on power generation targets.

[0189] In step 240, the health status prediction model employs a neural network model. The neural network has 10 input layer nodes, corresponding to the 10 features of the multidimensional state feature set; two hidden layers are set, with the first hidden layer having 20 nodes and the second hidden layer having 15 nodes; and three output layer nodes, corresponding to the three types of health status assessment labels (e.g., healthy, sub-healthy, and faulty). The learning rate is set to 0.01, and the number of training epochs is set to 100.

[0190] In an alternative embodiment, the method further includes:

[0191] Step 310: Collect historical operation datasets and corresponding health status assessment results of multiple wind turbine units to construct a multi-unit health status analysis database.

[0192] Historical operational datasets and corresponding health status assessment results from multiple wind turbines were collected. These data and results originated from operational records of different wind turbines at different time periods. This data was then organized and integrated to construct a multi-turbine health status analysis database. The database contains information such as operating parameters, health status assessment tags, and fault records for each wind turbine. This database allows for comprehensive analysis and comparison of the health status of multiple wind turbines.

[0193] Step 320: Standardize the data in the multi-unit health status analysis database to eliminate the impact of equipment and environmental differences between different units.

[0194] Data in the multi-unit health status analysis database is standardized. Differences in equipment specifications and installation environments among different wind turbines can affect data comparability. Standardization methods, such as Z-score standardization or Min-Max standardization, are used to process the operating parameters of each wind turbine. This standardization process eliminates the influence of equipment and environmental differences between different units, ensuring data comparability.

[0195] Step 330: Extract the multidimensional state feature set of the standardized historical operation dataset, perform correlation analysis with the corresponding health status assessment results, and construct a feature-health status correlation model.

[0196] Extract a multidimensional state feature set from the standardized historical running dataset. The method for extracting the multidimensional state feature set is the same as the method used for processing the first running dataset in steps 120-125. Perform correlation analysis between the extracted multidimensional state feature set and the corresponding health status assessment results. Statistical analysis methods or machine learning algorithms can be used to analyze the relationship between the features and the health status.

[0197] For example, decision tree algorithms can be used to identify features that have a significant impact on health status. Based on the association analysis results, a feature-health status association model can be constructed. This model can be used to predict the health status of wind turbine units, outputting corresponding health status assessment labels based on the input multidimensional state feature set.

[0198] Step 340: Use the feature-health status association model to perform similarity analysis on the multidimensional state feature sets of different wind turbine units, and identify turbine clusters with similar health status evolution trends.

[0199] A feature-health status association model is used to perform similarity analysis on the multidimensional state feature sets of different wind turbines. The similarity between the multidimensional state feature sets of every two wind turbines is calculated. Methods such as Euclidean distance and cosine similarity can be used for calculation. Based on the similarity results, wind turbines with similar health status evolution trends are grouped into a single cluster. For example, if two wind turbines have high similarity in their multidimensional state feature sets, it indicates that their health status evolution trends are similar, and they are classified into the same cluster. By identifying clusters, wind turbines can be categorized and managed, improving maintenance efficiency.

[0200] Step 350: Extract common features from the multidimensional state feature set of each unit cluster to determine the key common features that affect the health status of the units within the cluster.

[0201] Common features are extracted from the multidimensional state feature sets of each wind turbine cluster. The multidimensional state feature sets of each wind turbine within the cluster are analyzed to identify common characteristics among them. Methods such as principal component analysis can be used to extract key common features that represent the cluster's characteristics. For example, for a wind turbine cluster, if features such as power and speed show similar trends in most wind turbines, these are key common features. By identifying these key common features, we can gain a deeper understanding of the factors affecting the health status of the turbines within the cluster.

[0202] Step 360: Compare the key common characteristics of different unit clusters, identify the main influencing factors that lead to differences in health status, construct a health status optimization model based on the main influencing factors, generate personalized operating parameter adjustment suggestions for different unit clusters, input the personalized operating parameter adjustment suggestions into the wind turbine control system, and adjust the operating parameters of the wind turbine in real time.

[0203] Compare the key common characteristics of different wind turbine clusters. Analyze the differences in the value range and trends of key common characteristics among different clusters. Through comparison, identify the main influencing factors leading to differences in health status. For example, if one cluster generally has lower power while another cluster has higher power, it may be due to factors such as wind speed and equipment efficiency. Construct a health status optimization model based on the main influencing factors, using optimization algorithms such as genetic algorithms and particle swarm optimization. The model takes the main influencing factors as input and aims to optimize health status to calculate the optimal operating parameters. Generate personalized operating parameter adjustment suggestions for different wind turbine clusters. Input the personalized operating parameter adjustment suggestions into the wind turbine control system, and the control system adjusts the wind turbine operating parameters in real time according to the suggestions. For example, adjusting parameters such as the rotor blade angle and generator output power can improve the health status and power generation efficiency of the wind turbine.

[0204] Step 370: Continuously monitor the changes in the health status of the adjusted wind turbine units, and iteratively update the health status optimization model based on the monitoring results.

[0205] Continuously monitor changes in the health status of adjusted wind turbines. Real-time operational data from the wind turbines is collected via a sensor network, and their health status is assessed using a health status evaluation method. The health status optimization model is iteratively updated based on the monitoring results. If the adjusted operating parameters do not achieve the expected health status optimization effect, the reasons are analyzed, and the model is adjusted. For example, parameters or optimization algorithms in the model are adjusted to enable it to more accurately predict and adjust the operating parameters of the wind turbines. Through iterative updates, the accuracy and effectiveness of the health status optimization model are continuously improved.

[0206] For example, in step 330, the feature-health status association model is constructed using a logistic regression algorithm. The regularization parameter is set to 0.1 to control the model's complexity and prevent overfitting. The model is trained using training data, and the model's weight parameters are adjusted so that the model can accurately predict the health status of the wind turbine.

[0207] This application embodiment, by setting up a sensor network to collect a first operating dataset and performing multi-dimensional state feature mining on the first operating dataset, can deeply analyze the dynamic characteristics and potential abnormal modes of wind turbine units, uncover key information that is difficult to discover using traditional methods, and improve the depth of understanding of the operating status of wind turbine units.

[0208] By using a pre-built state extrapolation model for state extrapolation iteration and introducing different initial conditions and parameter disturbances for multi-path extrapolation, the uncertainty and diversity of wind turbine operation process are fully considered, generating multiple results that reflect the evolution trend of different operating states. Compared with single-path extrapolation, it can more comprehensively predict the future state of wind turbine.

[0209] By performing conflict resolution-based fusion processing on multiple state inference iteration results, and combining feature correlation weight matrix and credibility assessment mechanism, the conflict problem in the multi-source data fusion process is effectively solved, conflict coordination and information complementarity are achieved, and the state inference fusion results are made more accurate and reliable.

[0210] Finally, the state simulation fusion results are combined with the second operating dataset to perform a state assessment based on synchronous time-series analysis, resulting in a health status assessment label for the wind turbine. This assessment method, which combines data from different time periods, can dynamically and accurately reflect the actual health status of the wind turbine, improving the accuracy and timeliness of the health assessment. It can also promptly detect potential faults in the wind turbine, reduce maintenance costs, and improve power generation efficiency.

[0211] Based on the same inventive concept, embodiments of this application also provide a wind turbine health assessment system. See also... Figure 2 As shown, it is a schematic diagram of a possible wind turbine health assessment system provided in an embodiment of this application. Figure 2 In the wind turbine health assessment system 200, there are a processor 210 and a memory 220. The memory 220 stores computer programs that can be executed by the processor 210. By executing the instructions stored in the memory 220, the processor 210 can perform the steps of the aforementioned wind turbine health assessment method based on AI multi-source data fusion.

[0212] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium including a computer program. When the computer program is run on a wind turbine health assessment system, it causes the system to perform the steps of the aforementioned AI-based multi-source data fusion-based wind turbine health assessment method. In some possible implementations, various aspects of the AI-based multi-source data fusion-based wind turbine health assessment method provided in this application can also be implemented as a program product, including a computer program. When the program product is run on a wind turbine health assessment system, it causes the system to perform the steps of the aforementioned AI-based multi-source data fusion-based wind turbine health assessment method. For example, the wind turbine health assessment system can perform actions such as... Figure 1 The steps are shown in the figure.

[0213] The above content is only a preferred exemplary embodiment of this application and is not intended to limit the implementation of this application. Those skilled in the art can easily make corresponding modifications or alterations based on the main concept and spirit of this application.

Claims

1. A method for health assessment of wind turbine units based on AI multi-source data fusion, characterized in that, include: The initial operational dataset of the wind turbine was collected through the established sensor network. Multidimensional state feature mining is performed on the first running dataset to obtain a multidimensional state feature set characterizing the dynamic characteristics and potential anomaly modes of the wind turbine. Combining the multidimensional state feature set, the wind turbine is subjected to state deduction iteration based on dynamic characteristics and operating laws using a pre-built state deduction model. Different initial conditions and parameter disturbances are introduced during the state deduction iteration process to carry out multi-path deduction, generating multiple state deduction iteration results that reflect the evolution trend of different operating states. The multiple state inference iteration results are fused based on conflict resolution, and conflict coordination and information complementarity are carried out by combining the established feature correlation weight matrix and credibility assessment mechanism to generate state inference fusion results. The state inference fusion results and the second operating dataset are subjected to state assessment based on synchronous time series analysis to obtain the health status assessment label of the wind turbine. The second running dataset was collected later than the first running dataset.

2. The method as described in claim 1, characterized in that, The step of performing multidimensional state feature mining on the first operational dataset to obtain a multidimensional state feature set characterizing the dynamic characteristics and potential anomaly modes of the wind turbine includes: The first running dataset is subjected to time series decomposition processing to separate a multi-component time series set containing periodic fluctuation features and non-periodic fluctuation features; The dynamic characteristics of each component time series in the multi-component time series set are extracted to generate vibration mode features reflecting mechanical vibration characteristics and energy transfer features reflecting energy conversion efficiency. Correlation analysis is performed on the vibration mode features and the energy transfer features to identify the synergistic change patterns between them and generate a feature correlation map. Based on the feature association map, the vibration mode features and energy transfer features are subjected to anomaly-sensitive feature screening to extract a subset of key features that are sensitive to potential anomalies. The key feature subset is reorganized to generate a multidimensional state feature set containing time dimension features, frequency dimension features, and amplitude dimension features.

3. The method as described in claim 1, characterized in that, The process involves combining the multidimensional state feature set and using a pre-built state deduction model to perform state deduction iterations on the wind turbine based on its dynamic characteristics and operating laws. Different initial conditions and parameter disturbances are introduced during the state deduction iteration process to perform multi-path deductions, generating multiple state deduction iteration results reflecting the evolution trends of different operating states, including: Based on the historical operating status records of the wind turbine, the multidimensional state feature set is standardized to generate a standardized multidimensional state feature set; the standardized multidimensional state feature set is input into the initial state configuration layer of the state inference model to generate multiple differentiated initial state vectors according to the historical operating status records of the wind turbine. For each initial state vector, parameter perturbation processing is performed, and the dynamic characteristic parameters are randomly adjusted within a preset perturbation range to generate a set of perturbed state vectors containing different parameter combinations; The set of disturbance state vectors is input into the deduction iteration layer of the state deduction model, and multi-step state deduction calculation is performed based on the wind turbine dynamic equation to generate a state evolution path sequence corresponding to each disturbance state vector. During the generation of the state evolution path sequence, the deviation between the deduced state and the historical operating state is monitored in real time. When the deviation exceeds the preset deviation, the path correction mechanism is triggered to dynamically adjust the state transition matrix of the subsequent deduced steps. Repeatedly execute state deduction calculations and path correction mechanisms until the deduction process within the preset time window is completed, and output multiple state deduction iteration results containing multiple independent evolution paths.

4. The method as described in claim 3, characterized in that, The step involves inputting the set of disturbance state vectors into the iterative layer of the state deduction model, performing multi-step state deduction calculations based on the wind turbine dynamics equations, and generating a state evolution path sequence corresponding to each disturbance state vector, including: Each perturbation state vector in the set of perturbation state vectors is converted into a standard state vector that matches the input format of the state deduction model; The dynamic calculation kernel of the state deduction model is invoked, and a single-step state transition calculation is performed on the standard state vector based on the mechanical transmission system characteristics and aerodynamic characteristic equations of the wind turbine to obtain the state prediction vector for the next moment. Key state parameters are extracted from the state prediction vector and normalized and compared with similar parameters in the historical operating state database to generate parameter similarity scores. The state prediction vector is weighted based on the parameter similarity score to generate a weighted state prediction vector. The weighted state prediction vector is used as the new input state vector, and the single-step state transition calculation, similarity comparison and confidence weighting are repeatedly performed until the preset number of deduction steps are completed, generating a state evolution path sequence containing timestamps.

5. The method as described in claim 3, characterized in that, During the generation of the state evolution path sequence, the deviation between the deduced state and the historical operating state is monitored in real time. When the deviation exceeds a preset deviation, a path correction mechanism is triggered to dynamically adjust the state transition matrix of subsequent deduced steps, including: Extract the inferred state vector at the current moment from the state evolution path sequence, and extract the actual operating state vector at the corresponding moment from the historical operating state database; The inferred state vector and the actual operating state vector are normalized, and the cosine similarity or correlation coefficient between the normalized vectors is calculated. The cosine similarity or correlation coefficient is then converted into a difference index, which is used as a deviation quantification index. The deviation metric is compared with a preset deviation threshold. When the deviation metric is greater than the deviation threshold, the path correction mechanism is activated. The state transition matrix adjustment module in the path correction mechanism is invoked to calculate the matrix correction coefficient based on the magnitude of the deviation metric index. The matrix correction coefficient is positively correlated with the deviation metric index. The state transition matrix in the state deduction model is adjusted element-wise using the matrix correction coefficients to generate the corrected state transition matrix. The modified state transition matrix is ​​used to continue the subsequent state deduction calculation steps until the deviation quantification index is less than or equal to the deviation threshold or all deduction steps are completed.

6. The method as described in claim 1, characterized in that, The process involves fusing the multiple state deduction iteration results based on conflict resolution, and combining the established feature correlation weight matrix and credibility assessment mechanism for conflict coordination and information complementarity to generate a state deduction fusion result. The state deduction fusion result is then compared with the second operational dataset to perform a state assessment based on synchronous time-series analysis to obtain the health status assessment label of the wind turbine, including: The multiple state deduction iteration results are aligned on the time axis to keep all state deduction iteration results synchronized in the time dimension; the key state feature sequence in each state deduction iteration result is extracted, and the feature similarity matrix between different state deduction iteration results is calculated; Based on the feature similarity matrix, conflicting feature sequence pairs are identified, and the conflicting feature sequence pairs are classified by conflict type to determine the feature dimension and time node where the conflict occurred. The established feature correlation weight matrix is ​​invoked to extract the correlation weight values ​​of conflict feature dimensions in historical data, and the credibility score of each conflict feature sequence is calculated in combination with the credibility assessment mechanism. Based on the credibility score, the conflict feature sequences are weighted and fused. Feature sequences with weight values ​​higher than the preset weight are retained first, and feature sequences with weight values ​​lower than the preset weight are modified. The fused feature sequence is concatenated with the non-conflicting feature sequence to generate a state inference fusion result containing complete state feature information; the second operation dataset is standardized based on the historical operation state record of the wind turbine to generate a standardized second operation dataset; the timestamp of the sampling point is synchronized between the state inference fusion result and the standardized second operation dataset, and the state feature values ​​at the corresponding time of the synchronized state inference fusion result and the standardized second operation dataset are extracted and the feature deviation sequence is calculated. The characteristic deviation sequence is subjected to trend analysis to identify the type and rate of change of the deviation. Based on the change trend type and change rate, a preset health status assessment rule base is queried to determine the health status assessment label corresponding to the wind turbine.

7. The method as described in claim 6, characterized in that, The process of identifying conflicting feature sequence pairs based on the feature similarity matrix, classifying the conflict type of the conflicting feature sequence pairs, and determining the feature dimension and time point at which the conflict occurred includes: The feature similarity matrix is ​​traversed according to a preset feature similarity threshold, and feature sequence pairs with similarity values ​​lower than the feature similarity threshold are marked as potentially conflicting feature sequence pairs. The potential conflict feature sequence pairs are subjected to time series segmentation processing, dividing each feature sequence into multiple time window segments of equal length; Normalize the feature sequence pairs within each time window segment, calculate the dynamic time bending distance of the normalized sequence pairs, and mark the time window segments with a dynamic time bending distance greater than a preset distance threshold as conflict time windows. Extract the start and end timestamps of the conflict time window from the feature sequence to determine the target time node where the conflict occurred; The feature sequence pairs within the conflict time window are decomposed into feature dimensions to identify the target feature dimension where there is a difference jump, and the target feature dimension is marked as the conflict feature dimension. Based on the number and type of the conflict feature dimensions and the distribution characteristics of the conflict time windows, the conflict feature sequence pairs are classified into conflict types, and a conflict analysis report containing a conflict type identifier, a set of conflict feature dimensions, and a list of conflict time nodes is generated.

8. The method as described in claim 6, characterized in that, The aforementioned call establishes a feature correlation weight matrix, extracts the correlation weight values ​​of conflict feature dimensions in historical data, and calculates the credibility score for each conflict feature sequence using a credibility assessment mechanism, including: The feature correlation weight matrix is ​​subjected to hierarchical deconstruction to locate the coordinate index of the conflicting feature dimension in the matrix, and the row vector corresponding to the coordinate index is extracted as the initial correlation weight vector. The initial correlation weight vector is correlated with the historical performance characteristics of the conflict feature dimension in the historical operation database to generate dynamic weight adjustment coefficients. The historical performance characteristics include historical conflict occurrence rate and historical fusion contribution. The multi-source evidence fusion module of the credibility assessment mechanism performs multi-dimensional evidence collection and processing on conflict feature sequences. The multi-dimensional evidence includes internal consistency evidence of the feature sequences, external correlation evidence with other non-conflicting feature sequences, and reference evidence in similar historical scenarios. The collected multi-dimensional evidence is subjected to evidence strength quantification processing to generate an evidence strength value and evidence reliability factor for each piece of evidence. The initial associated weight vector is weighted and corrected based on the aforementioned weight dynamic adjustment coefficient to generate the corrected associated weight vector. The modified correlation weight vector and the evidence strength values ​​of the multi-dimensional evidence are normalized, and then a weighted fusion operation is performed to obtain a preliminary credibility score matrix. The initial credibility score matrix is ​​weighted element-wise based on the evidence reliability factor to generate a credibility sub-score for each conflict feature sequence in different conflict time windows. The credibility sub-scores for all conflict time windows are accumulated over time to generate credibility scores covering the entire conflict feature sequence.

9. The method as described in claim 6, characterized in that, The step of performing trend analysis on the characteristic deviation sequence to identify the type and rate of change of the deviation trend includes: The characteristic deviation sequence is input into the sequence preprocessing module of the trend analysis model for sequence stabilization and outlier identification to generate a purified deviation sequence. The purified deviation sequence is decomposed into a multi-component sequence set containing fluctuation and trend components by multi-scale decomposition. Extract the trend component from the multi-component sequence set and input it into the trend type recognizer for pattern matching processing. The pattern matching processing includes comparing the trend component with a preset standard trend pattern library. The standard trend pattern library includes continuous upward pattern, continuous downward pattern, fluctuating upward pattern, fluctuating downward pattern, and stable pattern. The main trend type corresponding to the trend component is determined based on the principle of highest similarity in pattern matching, and the fluctuation frequency and fluctuation amplitude features of the fluctuation component are extracted. Based on the main trend type, fluctuation frequency, and fluctuation amplitude characteristics, the trend descriptor generator is invoked to generate a trend feature descriptor that includes trend direction identifier, fluctuation characteristic parameters, and trend stability indicators. The purified deviation sequence was processed by sliding window, and observation windows of different lengths were set to calculate the rate of change of sequence morphology within each observation window. The morphological change rates of different observation windows are statistically fused and combined with the fluctuation characteristic parameters in the trend feature descriptor to generate a comprehensive change trend assessment index. The comprehensive change situation assessment index is semantically transformed by the change rate inference model, and the comprehensive change situation assessment index is mapped to the corresponding level in the preset change rate level system, generating a trend analysis result report containing trend type identifier and change rate level.

10. A wind turbine health assessment system, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of any one of the methods described in claims 1 to 9.

Images