Wind turbine health diagnosis method and system fusing multi-sensor and intelligent algorithm

Abstract

The application provides a kind of fusion multisensor and intelligent algorithm's wind turbine health diagnosis method and system, it is related to wind power generation technical field, including through acquisition multisensor signal, extract envelope spectrum peak frequency sequence mapping to high-dimensional feature space, utilize recurrent neural unit learning frequency migration law to generate conditional probability density field and mark abnormal frequency, based on abnormal frequency retrieval historical fault knowledge base, through dynamic time warping algorithm calculation form matching degree to obtain remaining evolution time, identify the three measuring points of maximum fluctuation amplitude to construct dynamic phase space, when forming stable attractor structure, extract topological feature and fault mode feature library to isomorphism matching to determine evolution terminal state, output fault diagnosis conclusion and intervention strategy.The application improves the accuracy and early warning capability of wind turbine health diagnosis.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, and in particular to a method and system for health diagnosis of wind turbine generators that integrates multiple sensors and intelligent algorithms. Background Technology

[0002] In the wind power sector, turbine health diagnostics are crucial for reducing operation and maintenance costs and preventing catastrophic accidents. Current conventional practices often rely on signal acquisition and processing from a single sensor (such as a vibration accelerometer or temperature sensor), using fixed thresholds or statistical characteristics to trigger alarms for abnormalities. These methods depend on field experience to calibrate thresholds, making it difficult to adapt to signal fluctuations across different turbine models, operating conditions, and environments. This results in early, minor faults being masked by noise or generating numerous false alarms.

[0003] In addition, traditional machine learning models are used to classify the extracted frequency domain features. However, these models typically assume that fault features are statically invariant over time, ignoring the dynamic migration patterns of faults from initiation to evolution. In actual operation, the frequency of fault features will undergo nonlinear drift with the degradation process, and static classifiers cannot capture this trend, resulting in low recognition rates in the early stages of faults.

[0004] Existing methods generally lack the ability to quantitatively predict fault evolution time. Most systems only output a binary judgment of "normal" or "fault," failing to inform maintenance personnel how much time remains between the occurrence of an anomaly and actual failure. This deficiency keeps maintenance decisions at a passive response level, preventing proactive condition-based maintenance. Furthermore, multi-sensor information fusion often relies on simple weighted averaging or voting mechanisms, failing to fully utilize the complementary evolutionary information of different signals in space, resulting in crude diagnosis and poor fault tolerance for complex fault modes.

[0005] In summary, existing conventional methods have significant shortcomings in capturing frequency migration patterns, predicting evolution time, and deeply integrating multi-source information, making it difficult to achieve high-precision and predictive health diagnosis of wind turbine units. Summary of the Invention

[0006] This invention provides a method and system for health diagnosis of wind turbine units that integrates multiple sensors and intelligent algorithms, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides a method for health diagnosis of wind turbine generators that integrates multiple sensors and intelligent algorithms, comprising:

[0008] Collect signals from multiple sensors in the wind turbine generator;

[0009] The envelope spectrum peak frequency sequence is extracted from multi-sensor signals and mapped to a high-dimensional feature space. Frequency migration rules are learned through recursive neural units and a conditional probability density field is generated. If the measured peak frequency deviates from the confidence interval of the conditional probability density field, the measured peak frequency is marked as an abnormal frequency.

[0010] Based on the abnormal frequency, historical cases with frequency feature matching are retrieved from the historical fault knowledge base, and the evolution trajectory of the corresponding multi-sensor signals is extracted. The morphological matching degree between the current evolution trajectory and the historical evolution trajectory is calculated by the dynamic time warping algorithm. The time interval from the occurrence of the abnormality to the confirmation of the fault is obtained from the historical case with the highest matching degree, and the remaining evolution time is obtained.

[0011] During the remaining evolution time, the three measurement points with the largest fluctuation amplitude are identified and sampled at high frequency. The synchronous measurement values ​​of the three measurement points are used as coordinates to construct a dynamic phase space and draw the state trajectory. When the state trajectory forms a stable attraction structure, the topological features of the stable attraction structure are extracted and isomorphically matched with the topological signature in the fault mode feature library to determine the final evolution state.

[0012] Based on the final state of evolution, output fault diagnosis conclusions and intervention strategies.

[0013] In one optional embodiment, the envelope spectrum peak frequency sequence is extracted from multi-sensor signals and mapped to a high-dimensional feature space. Frequency migration patterns are learned through recursive neural units, and a conditional probability density field is generated. If the measured peak frequency deviates from the confidence interval of the conditional probability density field, the measured peak frequency is marked as an abnormal frequency, including:

[0014] The multi-sensor signals are subjected to envelope demodulation to obtain an envelope signal. The envelope signal is then subjected to spectral analysis to determine the envelope spectrum. The frequencies corresponding to the amplitude peaks are extracted from the envelope spectrum to obtain peak frequencies. The peak frequencies are then arranged in chronological order to form a peak frequency sequence.

[0015] Each frequency value in the peak frequency sequence is converted into a high-dimensional vector. The high-dimensional vector is then mapped to a high-dimensional feature space using a kernel function. The manifold distance between frequency vectors at adjacent time points is calculated in the high-dimensional feature space to construct the topological manifold structure for frequency migration.

[0016] The topological manifold structure and manifold distance are input into the recurrent neural unit to capture the migration trajectory of the frequency on the topological manifold, and a conditional probability density field is generated based on the migration trajectory and manifold distance.

[0017] Calculate the confidence interval of the conditional probability density field, and determine whether the measured peak frequency falls outside the confidence interval. If it falls outside the confidence interval, mark the measured peak frequency as an abnormal frequency.

[0018] In one optional embodiment, the topological manifold structure and manifold distance are input into a recurrent neural unit to capture the migration trajectory of frequencies on the topological manifold, and the conditional probability density field is generated based on the migration trajectory and manifold distance, including:

[0019] The representations of frequencies at each time step in the high-dimensional feature space of the topological manifold are organized into an input sequence in chronological order. The manifold distance between adjacent time step frequency vectors is extracted from the topological manifold to construct a manifold distance sequence.

[0020] The input sequence and the manifold distance sequence are simultaneously input into the recurrent neural unit. The forget gate, input gate and output gate in the recurrent neural unit use the manifold distance at the current time as a geometric control factor to generate forget weight, input weight and output weight respectively.

[0021] Candidate memory content is generated based on the frequency vector at the current time and the hidden state at the previous time. The memory unit at the previous time is filtered by forgetting weight to obtain the retained memory information. The candidate memory content is controlled by input weight to obtain the new memory information. The retained memory information and the new memory information are fused to form the memory unit at the current time. The memory unit at the current time is activated to obtain the activated memory information. The hidden state at the current time is extracted from the activated memory information by output weight. The hidden states at each time are organized to form the migration trajectory of the frequency on the topological manifold.

[0022] The hidden state and manifold distance at the current time are simultaneously input into the probability distribution parameters to generate the network. The mean vector is determined based on the hidden state at the current time, and the covariance matrix is ​​determined by using the manifold distance at the current time as a dispersion control factor. The conditional probability density field is constructed based on the mean vector and the covariance matrix.

[0023] In one optional embodiment, historical cases with frequency feature matching are retrieved from a historical fault knowledge base based on the abnormal frequency, and the evolution trajectory of the corresponding multi-sensor signals is extracted. The morphological matching degree between the current evolution trajectory and the historical evolution trajectory is calculated using a dynamic time warping algorithm. The time interval from the occurrence of the anomaly to the confirmation of the fault is obtained from the historical case with the highest matching degree, and the remaining evolution time is obtained, including:

[0024] The abnormal frequency is converted into a frequency feature vector. The vector distance between the frequency feature vector of each historical case and the frequency feature vector of the abnormal frequency is calculated in the historical fault knowledge base. Candidate historical cases are selected based on the vector distance and the corresponding multi-sensor signal evolution trajectory is extracted.

[0025] The evolution trajectory of the current multi-sensor signal and the historical evolution trajectory are input into the dynamic time warping algorithm. A nonlinear time mapping relationship is established through the time alignment path and the morphological matching degree is calculated.

[0026] The historical case with the highest morphological matching degree is selected and the time interval between the time of anomaly frequency occurrence and the time of fault confirmation is extracted to obtain the historical evolution time. The time scale factor is determined according to the ratio of the number of time points of the current evolution trajectory and the historical evolution trajectory in the time alignment path. The historical evolution time is multiplied by the time scale factor to obtain the remaining evolution time.

[0027] In one optional embodiment, the evolution trajectory of the current multi-sensor signal and the historical evolution trajectory are input into a dynamic time warping algorithm. A nonlinear time mapping relationship is established through a time alignment path, and the morphological matching degree is calculated, including:

[0028] The evolution trajectory of the current multi-sensor signal and the historical evolution trajectory are divided into the current time series and the historical time series, respectively. The point-to-point distance between each time point in the current time series and the historical time series is calculated.

[0029] Initialize the cumulative distance matrix, traverse the time points of the current time series and the historical time series in ascending order of time, select the minimum cumulative distance from the adjacent previous position of the current time point in the cumulative distance matrix for the current time point, add the minimum cumulative distance to the point-to-point distance corresponding to the current time point to update the elements of the cumulative distance matrix, and complete the construction of the cumulative distance matrix.

[0030] Backtracking from the end position of the cumulative distance matrix to the start position, the preceding position with the smallest cumulative distance is selected as the path node to form a time alignment path. The path nodes in the time alignment path establish a non-linear time mapping relationship between the current time series and the historical time series. The point-to-point distances corresponding to each path node on the time alignment path are extracted and accumulated and normalized to obtain the morphological matching degree.

[0031] In one optional embodiment, the three measurement points with the largest fluctuation amplitude are identified and sampled at high frequency during the remaining evolution time. The synchronous measurement values ​​of the three measurement points are used as coordinates to construct a dynamic phase space and draw a state trajectory. When the state trajectory forms a stable attraction structure, the topological features of the stable attraction structure are extracted and isomorphically matched with the topological signatures in the fault mode feature library to determine the final evolution state, including:

[0032] During the remaining evolution time, the multi-sensor signals at each measurement point are statistically analyzed using a sliding window. The range of the signals within the sliding window is calculated as the fluctuation amplitude. The three measurement points with the largest fluctuation amplitude are selected, and high-frequency sampling is performed on the three measurement points to obtain synchronous measurement values.

[0033] The synchronous measurement values ​​of the three measuring points are mapped to the coordinate axis values ​​of the three-dimensional spatial coordinate system, and the spatial position points corresponding to each sampling time are calibrated in the three-dimensional spatial coordinate system to construct a dynamic phase space;

[0034] Connect the spatial locations in chronological order to form a state trajectory, calculate the envelope size of the state trajectory, calculate the spatial volume occupied by the state trajectory as the envelope size increases, and form a stable attraction structure when the variance of the spatial volume within the continuous monitoring window is less than the preset stability judgment threshold.

[0035] The topological boundary is determined by fitting the spatial location points in the stable attraction structure. The geometric shape parameters of the topological boundary and the distribution density features of the spatial location points are extracted as topological features. The topological signatures corresponding to each fault mode are extracted from the fault mode feature library. The shape similarity and distribution similarity between the topological features and each topological signature are calculated. The shape similarity and distribution similarity are fused to obtain the topological isomorphism. The fault mode corresponding to the topological signature with the highest topological isomorphism is selected as the final evolution state.

[0036] In one optional embodiment, extracting the geometric shape parameters and spatial location point distribution density features of the topological boundary as topological features includes:

[0037] Calculate the principal axis direction of the topological boundary, measure the maximum span of the topological boundary along the principal axis direction as the major axis dimension, measure the maximum span of the topological boundary perpendicular to the principal axis direction as the minor axis dimension, and calculate the ratio of the major axis dimension to the minor axis dimension to determine the flatness parameter;

[0038] The topological boundary is divided into multiple sections along the principal axis, the area of ​​each section is calculated, the change curve of the section area along the principal axis is extracted as the morphological evolution feature, and the degree of aggregation of spatial location points in each section in the stable attraction structure is statistically analyzed as the distribution density feature. The flatness parameter, morphological evolution feature and distribution density feature together constitute the topological feature.

[0039] A second aspect of this invention provides a wind turbine health diagnosis system integrating multiple sensors and intelligent algorithms, comprising:

[0040] The signal acquisition unit is used to acquire signals from multiple sensors in the wind turbine.

[0041] An anomaly detection unit is used to extract the peak frequency sequence of the envelope spectrum from multi-sensor signals and map it to a high-dimensional feature space. It learns the frequency migration law through recursive neural units and generates a conditional probability density field. If the measured peak frequency deviates from the confidence interval of the conditional probability density field, the measured peak frequency is marked as an anomaly frequency.

[0042] The fault matching unit is used to retrieve historical cases with frequency feature matching in the historical fault knowledge base based on the abnormal frequency, extract the evolution trajectory of the corresponding multi-sensor signals, calculate the morphological matching degree between the current evolution trajectory and the historical evolution trajectory through the dynamic time warping algorithm, obtain the time interval from the occurrence of the abnormality to the fault confirmation from the historical case with the highest matching degree, and obtain the remaining evolution time.

[0043] The state trajectory unit is used to identify the three measurement points with the largest fluctuation amplitude in the remaining evolution time and sample them at high frequency. The synchronous measurement values ​​of the three measurement points are used as coordinates to construct a dynamic phase space and draw the state trajectory. When the state trajectory forms a stable attraction structure, the topological features of the stable attraction structure are extracted and isomorphically matched with the topological signature in the fault mode feature library to determine the final evolution state.

[0044] The diagnostic decision unit is used to output fault diagnosis conclusions and intervention strategies based on the final evolution state.

[0045] A third aspect of the present invention provides an electronic device, comprising:

[0046] processor;

[0047] Memory used to store processor-executable instructions;

[0048] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0049] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0050] In this embodiment of the invention, by mapping the peak frequency sequence of the envelope spectrum to a high-dimensional feature space and utilizing recursive neural units to learn frequency migration patterns, weak abnormal frequency fluctuations can be accurately identified, significantly reducing the false negative rate. The confidence interval of the conditional probability density field provides a dynamic statistical benchmark for anomaly marking, effectively filtering out noise interference and making the diagnostic system sensitive to early degradation and stable and reliable. Based on the retrieval of abnormal frequencies in the historical fault knowledge base and the morphological matching of evolution trajectories, a quantitative estimation of the remaining evolution time is achieved. The dynamic time warping algorithm analyzes the morphological similarity between the current and historical trajectories, removes time scale differences, and significantly improves the generalization ability across units and operating conditions. This estimation result provides a clear time window for operation and maintenance decisions, avoiding economic losses and safety risks caused by sudden outages. High-frequency sampling and dynamic phase space reconstruction of the three measurement points with the largest fluctuation amplitudes within the remaining evolution time can capture the evolution process of the system from chaos to order. The extraction of topological features of stable attracting structures and isomorphic matching of fault mode feature libraries transform qualitative fault diagnosis into quantitative topological comparison, greatly enhancing the ability to identify complex faults. Attached Figure Description

[0051] Figure 1 A flowchart illustrating a method for health diagnosis of wind turbine units that integrates multiple sensors and intelligent algorithms;

[0052] Figure 2 The flowchart shows the logic for determining the final state of evolution. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0055] Figure 1 This is a flowchart illustrating the wind turbine health diagnosis method integrating multiple sensors and intelligent algorithms according to an embodiment of the present invention.

[0056] Wind turbine health diagnosis methods that integrate multiple sensors and intelligent algorithms include:

[0057] Collect signals from multiple sensors in the wind turbine generator;

[0058] The envelope spectrum peak frequency sequence is extracted from multi-sensor signals and mapped to a high-dimensional feature space. Frequency migration rules are learned through recursive neural units and a conditional probability density field is generated. If the measured peak frequency deviates from the confidence interval of the conditional probability density field, the measured peak frequency is marked as an abnormal frequency.

[0059] Based on the abnormal frequency, historical cases with frequency feature matching are retrieved from the historical fault knowledge base, and the evolution trajectory of the corresponding multi-sensor signals is extracted. The morphological matching degree between the current evolution trajectory and the historical evolution trajectory is calculated by the dynamic time warping algorithm. The time interval from the occurrence of the abnormality to the confirmation of the fault is obtained from the historical case with the highest matching degree, and the remaining evolution time is obtained.

[0060] During the remaining evolution time, the three measurement points with the largest fluctuation amplitude are identified and sampled at high frequency. The synchronous measurement values ​​of the three measurement points are used as coordinates to construct a dynamic phase space and draw the state trajectory. When the state trajectory forms a stable attraction structure, the topological features of the stable attraction structure are extracted and isomorphically matched with the topological signature in the fault mode feature library to determine the final evolution state.

[0061] Based on the final state of evolution, output fault diagnosis conclusions and intervention strategies.

[0062] In one optional embodiment, the envelope spectrum peak frequency sequence is extracted from multi-sensor signals and mapped to a high-dimensional feature space. Frequency migration patterns are learned through recursive neural units, and a conditional probability density field is generated. If the measured peak frequency deviates from the confidence interval of the conditional probability density field, the measured peak frequency is marked as an abnormal frequency, including:

[0063] The multi-sensor signals are subjected to envelope demodulation to obtain an envelope signal. The envelope signal is then subjected to spectral analysis to determine the envelope spectrum. The frequencies corresponding to the amplitude peaks are extracted from the envelope spectrum to obtain peak frequencies. The peak frequencies are then arranged in chronological order to form a peak frequency sequence.

[0064] Each frequency value in the peak frequency sequence is converted into a high-dimensional vector. The high-dimensional vector is then mapped to a high-dimensional feature space using a kernel function. The manifold distance between frequency vectors at adjacent time points is calculated in the high-dimensional feature space to construct the topological manifold structure for frequency migration.

[0065] The topological manifold structure and manifold distance are input into the recurrent neural unit to capture the migration trajectory of the frequency on the topological manifold, and a conditional probability density field is generated based on the migration trajectory and manifold distance.

[0066] Calculate the confidence interval of the conditional probability density field, and determine whether the measured peak frequency falls outside the confidence interval. If it falls outside the confidence interval, mark the measured peak frequency as an abnormal frequency.

[0067] In one specific implementation, envelope demodulation of multi-sensor signals is a fundamental step in extracting the peak frequency sequence. The acquired vibration signals typically contain amplitude modulation components, making direct spectral analysis of the original signal insufficient for effectively identifying fault characteristic frequencies. Therefore, a Hilbert transform is first performed on the original signal to obtain its analytical form, and then its instantaneous amplitude is extracted as the envelope signal. The envelope signal reflects the amplitude modulation pattern of the original vibration signal, effectively separating low-frequency fault characteristics from the high-frequency carrier. A fast Fourier transform is then performed on the envelope signal to obtain the envelope spectrum. In the envelope spectrum, frequency points with amplitudes significantly higher than the background noise level correspond to periodic impact characteristics during unit operation. Extracting the corresponding frequency values ​​from these amplitude peaks yields the peak frequency at the current moment. Following the temporal order of signal acquisition, the peak frequencies extracted from multiple consecutive moments are arranged sequentially to form a peak frequency sequence. This sequence records the dynamic changes in the unit's frequency characteristics within a time window, providing the initial input for subsequent high-dimensional mapping and anomaly detection.

[0068] Each frequency value in a peak frequency sequence is itself a one-dimensional scalar. Analyzing the evolution of frequencies directly in one-dimensional space makes it difficult to capture the nonlinear correlations between frequencies. Therefore, each frequency value... Convert to high-dimensional vector ,in This serves as the time step index. The high-dimensional vector is constructed as follows: using... Centered on a given point, a multidimensional feature vector is formed by concatenating the frequency values ​​and corresponding amplitude, phase, and other auxiliary features from several time points before and after it. Subsequently, a kernel function maps this high-dimensional vector to the reproducing kernel Hilbert space, i.e., the high-dimensional feature space. The radial basis function is chosen as the kernel function, and its expression is: ,in and The first and the A high-dimensional vector at each moment, This is the bandwidth parameter of the kernel function. This represents the Euclidean distance. Through kernel function mapping, the frequency feature relationships that were originally linearly inseparable in the low-dimensional space are effectively expanded in the high-dimensional feature space, making the frequency transfer patterns easier to capture by subsequent models.

[0069] In a high-dimensional feature space, the manifold distance between frequency vectors at adjacent time points is calculated to characterize the topological evolution of frequency features over time. The manifold distance is calculated using the geodesic distance approximation method: [The text then describes a process involving constructing...] Nearest neighbor graph, which associates each vector with its nearest neighbor. The Euclidean distance between the nearest neighbor vectors is used as the edge weights of the graph, and then the geodesic distance between any two vectors is estimated using the shortest path algorithm. Based on the manifold distances between all adjacent time pairs, a topological manifold structure for frequency migration is constructed. This topological manifold structure is stored in the form of a graph, where nodes correspond to the frequency vectors at each time point, and edge weights correspond to the manifold distances between adjacent time points. This encodes the temporal evolution of the frequency sequence into a geometric topological representation, preserving the nonlinear structural information during the frequency migration process.

[0070] The topological manifold structure and manifold distance sequence are input into the recurrent neural unit (RNU) to learn the frequency transfer patterns along the manifold. The RNU employs a gated recurrent unit (GRU) structure, which effectively handles long-term temporal dependencies and avoids the vanishing gradient problem. At each time step, the GRU receives the frequency vector at the current time step. and its manifold distance from the previous time step. Combined with hidden state By working together to update and reset the door, the current hidden state is output. After complete training on historical frequency sequences, the recurrent neural unit can capture the typical migration trajectories of frequencies on the topological manifold, i.e., the regular paths of frequency feature evolution over time under normal operating conditions. Based on the captured migration trajectories and the corresponding manifold distance distribution, a conditional probability density field is generated using a kernel density estimation method. Conditional probability density field Describes the migration state in a given history Under the condition of, the peak frequency at the next moment The probability distribution, where The frequency value to be evaluated. For the recurrent neural unit in the first The hidden state vector output by the step. The conditional probability density field expresses the normal fluctuation range of frequency in the form of a continuous probability distribution, and can adaptively reflect the dynamic changes of frequency characteristics under different operating conditions.

[0071] Based on the conditional probability density field, confidence intervals are calculated to determine a reasonable range for normal frequencies. The confidence intervals are calculated using a two-sided quantile method: given a confidence level... (Usually taken as 0.95 or 0.99), determine the conditional probability density field respectively. lower quantile and upper quantile ,satisfy Confidence interval It covers frequency values ​​in terms of probability under normal migration patterns. The range of occurrence. The measured peak frequency for each new acquisition moment. Determine whether it falls outside the confidence interval: If or If the frequency deviates from the normal migration pattern, it is considered to be an abnormal frequency. The marking of abnormal frequencies is not based on a static judgment of a single threshold, but rather combines the dynamic context of historical migration states. Therefore, it has a strong adaptive ability to changes in operating conditions and can effectively distinguish between frequency changes caused by fluctuations in normal operating conditions and frequency anomalies caused by fault initiation.

[0072] In practical engineering applications, the operating conditions of wind turbines are frequently affected by external conditions such as wind speed and load, resulting in fluctuations in the peak frequency of the envelope spectrum even under normal conditions. By combining conditional probability density fields with topological manifold structures, the influence of operating condition changes can be considered simultaneously when learning frequency migration patterns, thereby reducing the false alarm rate. When the measured peak frequencies at multiple consecutive times are marked as anomalous frequencies, a subsequent historical case retrieval and evolution analysis process is triggered, providing reliable anomalous frequency input for fault diagnosis.

[0073] In one optional embodiment, the topological manifold structure and manifold distance are input into a recurrent neural unit to capture the migration trajectory of frequencies on the topological manifold, and the conditional probability density field is generated based on the migration trajectory and manifold distance, including:

[0074] The representations of frequencies at each time step in the high-dimensional feature space of the topological manifold are organized into an input sequence in chronological order. The manifold distance between adjacent time step frequency vectors is extracted from the topological manifold to construct a manifold distance sequence.

[0075] The input sequence and the manifold distance sequence are simultaneously input into the recurrent neural unit. The forget gate, input gate and output gate in the recurrent neural unit use the manifold distance at the current time as a geometric control factor to generate forget weight, input weight and output weight respectively.

[0076] Candidate memory content is generated based on the frequency vector at the current time and the hidden state at the previous time. The memory unit at the previous time is filtered by forgetting weight to obtain the retained memory information. The candidate memory content is controlled by input weight to obtain the new memory information. The retained memory information and the new memory information are fused to form the memory unit at the current time. The memory unit at the current time is activated to obtain the activated memory information. The hidden state at the current time is extracted from the activated memory information by output weight. The hidden states at each time are organized to form the migration trajectory of the frequency on the topological manifold.

[0077] The hidden state and manifold distance at the current time are simultaneously input into the probability distribution parameters to generate the network. The mean vector is determined based on the hidden state at the current time, and the covariance matrix is ​​determined by using the manifold distance at the current time as a dispersion control factor. The conditional probability density field is constructed based on the mean vector and the covariance matrix.

[0078] In one specific implementation, after obtaining the peak frequency sequence of the envelope spectrum and completing the high-dimensional feature space mapping, the topological manifold structure and manifold distance information need to be jointly input into the recurrent neural unit to drive subsequent frequency migration law capture and conditional probability density field generation. The topological manifold structure stores the embedded representations of the peak frequencies at each time step in the high-dimensional feature space. These embedded representations are arranged sequentially according to the acquisition time order to form the input sequence. Simultaneously, the manifold distances between frequency vectors at adjacent time steps are extracted from the topological manifold structure and arranged according to the corresponding time steps to form the manifold distance sequence. The manifold distance reflects the true geometric span of the frequency state on the nonlinear manifold, rather than the straight-line distance in Euclidean space, thus more accurately characterizing the amplitude and direction changes of frequency evolution between adjacent time steps. The input sequence and the manifold distance sequence are strictly aligned on the time axis to ensure that the frequency vector and the corresponding manifold distance at each time step can be synchronously fed into the recurrent neural unit for joint processing.

[0079] The recurrent neural unit employs a long short-term memory (LSTM) structure, internally containing three gating mechanisms: a forget gate, an input gate, and an output gate. Unlike traditional LSM networks, this approach uses the manifold distance at the current time step as a geometric control factor, injecting it into each of the three gating processes. This ensures that the gating weights depend not only on the current frequency vector and the hidden state at the previous time step, but also on the explicit adjustment of the degree of local geometric change in the frequency along the manifold. Let the current time step number be... The frequency vector at the current time is The hidden state vector at the previous time step is The manifold distance at the current moment is Forgotten goalkeeper , and After concatenation, a linear transformation and sigmoid activation are performed to obtain the forgetting weights. The input gate obtains the input weights in the same way. The output gate obtains the output weights in the same way. When manifold distance When the value is large, it indicates a significant geometric jump in frequency state between adjacent time steps. The forget gate tends to output a smaller forget weight, prompting the network to forget the historical memory accumulated in the previous time step more significantly, thus quickly adapting to the abrupt change in frequency state; when... When the frequency state is small, the frequency state changes gradually on the manifold, and the forget gate tends to retain more historical memory, maintaining the continuity and stability of the frequency evolution trajectory.

[0080] After the gating weights are determined, based on the frequency vector at the current time... Hidden state from the previous moment Candidate memory content is generated through linear transformation and tanh activation. Candidate memory content represents new information that may be written to the memory unit at the current moment. (Previous moment's memory unit) With forgetting weight Element-wise multiplication yields the retained memory information, which is the historical state information that remains after filtering. Candidate memory content. With input weights Element-wise multiplication yields the newly added memory information, which is the new content actually written to the memory unit at the current moment. The retained memory information and the newly added memory information are then added element-wise to form the memory unit at the current moment. ,Right now ,in This represents element-wise multiplication. It refers to the memory unit at the current time step. Applying the tanh activation transformation yields the activation memory information. Then, with the output weights Element-wise multiplication yields the hidden state at the current time. ,Right now Process each time step of the input sequence sequentially along the time axis, and output the hidden state at each time step. Organized chronologically, the frequencies are traced on the topological manifold. This trace encodes the historical path of the frequency states as they move across the manifold over time, as well as the information on local geometric changes at each node, regulated by manifold distance.

[0081] After establishing the migration trajectory, a conditional probability density field needs to be constructed to quantify the distribution range of frequency values ​​at future time points, providing a probabilistic basis for subsequent anomaly detection. The current state is then hidden. Distance from the manifold at the current moment A synchronous input probability distribution parameter generation network is proposed. This network is a lightweight feedforward structure with two parallel branches: the mean branch only uses the hidden states. As input, the mean vector is output after linear transformation. The mean vector represents the expected frequency position of the conditional distribution at the current time, and its dimension is consistent with the dimension of the frequency feature space; the covariance branch uses the hidden state The fundamental covariance parameters are obtained through linear transformation, and the manifold distance is also included. The basic covariance parameter is scaled and adjusted as a dispersion control factor, ultimately outputting the covariance matrix. When manifold distance When the value is large, the frequency states span a larger geometric distance on the manifold, implying higher uncertainty in frequency evolution. Consequently, the diagonal elements of the covariance matrix increase, the dispersion of the conditional distribution expands, and the confidence interval widens accordingly. When... When the value is small, the frequency evolution stabilizes, the covariance matrix shrinks, the confidence interval tightens, and the sensitivity to identify anomalous frequencies improves. (Based on mean vector) With covariance matrix We construct a multivariate Gaussian conditional probability density field, which describes the probability distribution of the peak frequency at the next moment given the current migration trajectory and local geometry.

[0082] For each measured peak frequency, it is mapped to a high-dimensional feature space, and its probability density value under the current conditional probability density field is calculated and compared with the quantile boundary corresponding to the preset confidence level. If the measured peak frequency falls outside the confidence interval jointly determined by the mean vector and covariance matrix, the frequency is determined to be an anomalous frequency, triggering the subsequent historical case retrieval process. The entire construction process of the conditional probability density field deeply integrates manifold geometric information and recursive memory mechanism, which not only preserves the long-range dependence of frequency evolution, but also realizes dynamic response to local geometric mutations through real-time injection of manifold distance. This allows the generated probability density field to adaptively track the frequency drift pattern of wind turbines under different operating conditions, thereby significantly improving the accuracy and robustness of anomalous frequency identification.

[0083] In one optional embodiment, historical cases with frequency feature matching are retrieved from a historical fault knowledge base based on the abnormal frequency, and the evolution trajectory of the corresponding multi-sensor signals is extracted. The morphological matching degree between the current evolution trajectory and the historical evolution trajectory is calculated using a dynamic time warping algorithm. The time interval from the occurrence of the anomaly to the confirmation of the fault is obtained from the historical case with the highest matching degree, and the remaining evolution time is obtained, including:

[0084] The abnormal frequency is converted into a frequency feature vector. The vector distance between the frequency feature vector of each historical case and the frequency feature vector of the abnormal frequency is calculated in the historical fault knowledge base. Candidate historical cases are selected based on the vector distance and the corresponding multi-sensor signal evolution trajectory is extracted.

[0085] The evolution trajectory of the current multi-sensor signal and the historical evolution trajectory are input into the dynamic time warping algorithm. A nonlinear time mapping relationship is established through the time alignment path and the morphological matching degree is calculated.

[0086] The historical case with the highest morphological matching degree is selected and the time interval between the time of anomaly frequency occurrence and the time of fault confirmation is extracted to obtain the historical evolution time. The time scale factor is determined according to the ratio of the number of time points of the current evolution trajectory and the historical evolution trajectory in the time alignment path. The historical evolution time is multiplied by the time scale factor to obtain the remaining evolution time.

[0087] In one specific implementation, after detecting an abnormal frequency, it needs to be converted into a structured representation that can be used for knowledge base retrieval. The abnormal frequency itself carries frequency domain feature information of the fault, but a single frequency value is insufficient to describe the complete feature pattern of the fault. Therefore, the abnormal frequency is converted into a frequency feature vector. This vector not only contains the absolute value of the abnormal frequency but also integrates its relative position in the spectrum, its ratio to the rotational frequency and its harmonics, its deviation from the bearing's characteristic frequencies (inner race fault frequency, outer race fault frequency, rolling element fault frequency), and the statistical characteristics of the abnormal frequency over a time window (mean, variance, drift trend slope, etc.). This multi-dimensional information collectively constitutes the frequency feature vector, enabling subsequent knowledge base retrieval to achieve accurate matching of frequency features at the semantic level, rather than relying solely on numerical similarity.

[0088] Let the frequency feature vector corresponding to the current abnormal frequency be... The first in the historical fault knowledge base The frequency feature vector of each historical case is Calculate the vector distance between the two. A weighted Euclidean distance metric is used, assigning different weights to each dimension of the frequency feature vector. Dimensions related to the harmonic relationship of the rotational frequency have higher weights to highlight the characteristic sensitivity of mechanical periodic faults. All historical cases are analyzed according to... Sort in ascending order and select the first one with the smallest distance. These cases form a candidate historical case set, among which The value of is configured during actual deployment based on the size of the knowledge base and the required retrieval accuracy, and is usually an integer between 10 and 30. The multi-sensor signal evolution trajectory corresponding to each candidate case is extracted. This trajectory is formed by splicing the time series of signals from multiple sensors, such as vibration sensors, temperature sensors, and current sensors, before and after the anomaly occurs, thus forming a multi-dimensional time series matrix.

[0089] The current evolution trajectory of the multi-sensor signal and each candidate historical evolution trajectory are input into the dynamic time warping algorithm for morphological matching calculation. The core of the dynamic time warping algorithm lies in constructing a cumulative distance matrix to find an optimal path that non-linearly aligns two time series on the time axis, thereby overcoming the time axis misalignment problem caused by differences in operating conditions, inconsistent signal sampling frequencies, or different fault development rates. Let the current evolution trajectory be of length... A multidimensional time series, with a historical evolution trajectory of length Multidimensional time series, construct Cumulative distance matrix Each element in the matrix Indicates the current trajectory number 1 The first point in time and historical trajectory The local distance between each time point is added to the minimum cumulative cost to reach that location. The local distance is calculated using Euclidean distance from multiple sensor signals, comprehensively considering the differences in multi-dimensional signal components such as vibration amplitude, temperature readings, and current harmonics. The optimal time alignment path is obtained by recursively applying dynamic programming from the upper left corner to the lower right corner of the matrix. The cumulative distance at the end of the path, after normalization, is the shape matching degree. , The smaller the value, the more similar the shapes, and the higher the matching degree.

[0090] After obtaining the morphological matching degree of all candidate cases, select The smallest historical case is selected as the optimal matching case. The time interval between the first marked time of the anomaly frequency and the final confirmed time of the fault is extracted from the case's record information and denoted as the historical evolution time. The unit is hours. Historical evolution time reflects the complete time span from the appearance of early abnormal signals to the evolution into a confirmed fault state in historical cases of this type of fault, and is an important reference benchmark for predicting the remaining lifetime of the current fault.

[0091] However, due to differences in current equipment operating conditions, load conditions, and ambient temperature compared to historical cases, directly using historical evolution time as the remaining evolution time will introduce systematic bias. To eliminate this bias, the optimal time alignment path obtained during dynamic time warping is utilized. Extract the time scale factor. Time alignment path. A series of coordinate pairs Composition, in which Index of the current trajectory's time point. This is an index of historical trajectory time points. The number of time points covered by the current evolutionary trajectory in the path is... The number of time points covered by the historical evolution trajectory is Time scale factor Defined as the ratio of the two, that is This ratio reflects the relative relationship between the current failure evolution rate and the historical case evolution rate: if This indicates that the current fault evolution is slower than in historical cases, and the remaining time is correspondingly longer; if This indicates that the current fault is evolving faster, and the remaining time is correspondingly shorter.

[0092] Time of historical evolution Multiply by time scale factor To obtain the remaining evolution time ,Right now Remaining evolution time This represents the estimated time remaining from the current moment until the final confirmation of the fault state, given the current evolution rate. In practical engineering applications, to improve the robustness of the prediction, the time required to predict the future timeframe can be adjusted. The remaining evolution time is calculated for each historical case with the highest matching degree, and then a weighted average is performed using the matching degree as the weight to obtain a more robust estimate of the remaining evolution time. The number of cases participating in the weighted average is typically 3 to 5. This multi-case weighted fusion strategy can effectively reduce the impact of random errors from a single historical case on the prediction results, especially when there are multiple cases with similar matching degrees in the historical case library, it can comprehensively reflect the probability of different evolutionary paths.

[0093] In actual deployment, the historical fault knowledge base needs continuous maintenance and updates. Whenever a new fault case is manually confirmed, its frequency feature vector, multi-sensor signal evolution trajectory, and the time interval from the anomaly's occurrence to fault confirmation are all entered into the knowledge base. The feature vector index structure in the knowledge base is also incrementally updated to ensure that retrieval efficiency does not significantly decrease as the database size increases. The knowledge base also needs to periodically perform clustering and redundancy removal on historical cases, eliminating highly similar duplicate cases and retaining typical representative cases of various fault modes, thereby controlling the computational load while ensuring coverage. Through these mechanisms, the retrieval quality of the knowledge base continuously improves with accumulated runtime, and the prediction accuracy of remaining evolution time also gradually improves.

[0094] In one optional embodiment, the evolution trajectory of the current multi-sensor signal and the historical evolution trajectory are input into a dynamic time warping algorithm. A nonlinear time mapping relationship is established through a time alignment path, and the morphological matching degree is calculated, including:

[0095] The evolution trajectory of the current multi-sensor signal and the historical evolution trajectory are divided into the current time series and the historical time series, respectively. The point-to-point distance between each time point in the current time series and the historical time series is calculated.

[0096] Initialize the cumulative distance matrix, traverse the time points of the current time series and the historical time series in ascending order of time, select the minimum cumulative distance from the adjacent previous position of the current time point in the cumulative distance matrix for the current time point, add the minimum cumulative distance to the point-to-point distance corresponding to the current time point to update the elements of the cumulative distance matrix, and complete the construction of the cumulative distance matrix.

[0097] Backtracking from the end position of the cumulative distance matrix to the start position, the preceding position with the smallest cumulative distance is selected as the path node to form a time alignment path. The path nodes in the time alignment path establish a non-linear time mapping relationship between the current time series and the historical time series. The point-to-point distances corresponding to each path node on the time alignment path are extracted and accumulated and normalized to obtain the morphological matching degree.

[0098] In one specific implementation, after acquiring candidate historical cases, it is necessary to perform an accurate morphological similarity measurement between the evolution trajectory of the current multi-sensor signal and the historical evolution trajectory. Since the fault development rate of wind turbines varies under different operating conditions, the same fault evolution process may exhibit different expansion or compression patterns on the time axis. Therefore, it is not feasible to simply align the time axis and directly compare the signal amplitudes. Instead, a dynamic time warping algorithm is needed to establish a nonlinear time mapping relationship, thereby calculating the morphological matching degree between the two trajectories while allowing for elastic deformation of the time axis.

[0099] The evolution trajectory of the current multi-sensor signals is recorded as the current time series, and the historical evolution trajectories are recorded as the historical time series. The current time series includes... The historical time series contains [number] time points, each recording synchronous sampling vectors from multiple sensors; The nth time point also carries the corresponding multi-sensor sampling vector. For the nth time series in the current time series... Multi-sensor vector at each time point In the historical time series, the first Multi-sensor vector at each time point Calculate the Euclidean distance between the two as the point-to-point distance. ,Right now Arrange the point-to-point distances corresponding to all time points into a single array. Point-to-point distance matrix , of which Line number The elements of the column are Point-to-point distance matrix It reflects the local morphological differences between the two time series at various time point combinations and serves as the basic input for the subsequent construction of the cumulative distance matrix.

[0100] Construct the cumulative distance matrix Its dimension is the same as the point-to-point distance matrix. They are the same, both are During initialization, The Line number Column elements are set to That is, the point-to-point distance between the starting points of the two sequences; the first... The rest of the positions Initialize to prefix accumulation value , will the The remaining positions of the column Initialize to prefix accumulation value This initialization operation ensures that the path can only originate from... Starting from the beginning, extend in ascending order of time, without allowing time to reverse. Then traverse the matrix in ascending order of time. from arrive , from arrive All positions, for the current traversed position From its three legal preorder positions , , Select the minimum cumulative distance. ,according to Update the cumulative distance matrix elements at the current position. The three preceding positions correspond to three time alignment methods: single-step advancement of the current time series, single-step advancement of the historical time series, and simultaneous advancement of both. By selecting the minimum cumulative cost at each step, the final path is guaranteed to have global optimality. After traversal, Stored from the starting position To the end position The total morphological difference cost accumulated along the optimal time alignment path.

[0101] Complete the cumulative distance matrix After construction, from the termination position The process begins by executing a backtracking operation, progressively tracing the optimal time alignment path. At each current position... Check its three valid preorder positions. , , In the cumulative distance matrix, the position with the smallest cumulative distance is selected as the previous node of the path, and this node is added to the time-aligned path. Then update the current position to the previous node, and repeat the above process until the starting position is reached. The final time alignment path Consists of a series of path nodes Composition, in which For path node indexing, from Total number of nodes on the path Path nodes Indicates the current time series number The first point in time and the historical time series A nonlinear temporal correspondence was established between the time points. This nonlinear mapping allows a certain time period in the current sequence to be matched with corresponding segments of different time spans in the historical sequence, thereby eliminating the interference of differences in fault evolution rates on morphological similarity assessment.

[0102] Extract time alignment path Each path node Corresponding point-to-point distance The total path distance is obtained by summing the point-to-point distances of all nodes along the path. To eliminate path length To address the differences in units caused by variations, the total path distance is normalized to the total number of path nodes. As a normalization factor, calculate the normalized path distance. Normalized path distance The smaller the value, the closer the current evolutionary trajectory is to the historical evolutionary trajectory in terms of morphology. This morphological matching degree... Defined as a monotonically decreasing transformation of the normalized path distance, specifically, , making The range of values ​​falls within Within the interval, The closer The higher the degree of morphological matching, The closer The greater the morphological difference, the better. The dynamic time warping process described above is applied to all candidate historical cases to obtain their respective morphological matching degrees. ,according to Sort the data from largest to smallest, select the historical cases with the highest morphological matching degree as a reference, extract the time interval between the occurrence of the anomaly and the confirmation of the fault, and combine the ratio of the current evolution rate to the historical evolution rate to further estimate the remaining evolution time, providing a time window constraint for subsequent phase space construction and final state identification.

[0103] In practical applications, to improve the computational efficiency of the dynamic time warping algorithm, Sakoe-Chiba band constraints can be applied to the time alignment path, that is, to limit the time offset of the path nodes. Not exceeding the preset bandwidth parameters This reduces the matrix traversal range from the entire matrix to a diagonal band, significantly reducing computational complexity while ensuring shape matching is completed within a reasonable time scaling range, avoiding physically unreasonable extreme time alignment situations. Bandwidth parameter The constraint range can be set based on the maximum change factor of the fault evolution rate in historical data to ensure that the constraint range covers the actual possible time alignment scenarios.

[0104] like Figure 2The diagram shown illustrates the logic flowchart for determining the final state of evolution.

[0105] In one optional embodiment, the three measurement points with the largest fluctuation amplitude are identified and sampled at high frequency during the remaining evolution time. The synchronous measurement values ​​of the three measurement points are used as coordinates to construct a dynamic phase space and draw a state trajectory. When the state trajectory forms a stable attraction structure, the topological features of the stable attraction structure are extracted and isomorphically matched with the topological signatures in the fault mode feature library to determine the final evolution state, including:

[0106] During the remaining evolution time, the multi-sensor signals at each measurement point are statistically analyzed using a sliding window. The range of the signals within the sliding window is calculated as the fluctuation amplitude. The three measurement points with the largest fluctuation amplitude are selected, and high-frequency sampling is performed on the three measurement points to obtain synchronous measurement values.

[0107] The synchronous measurement values ​​of the three measuring points are mapped to the coordinate axis values ​​of the three-dimensional spatial coordinate system, and the spatial position points corresponding to each sampling time are calibrated in the three-dimensional spatial coordinate system to construct a dynamic phase space;

[0108] Connect the spatial locations in chronological order to form a state trajectory, calculate the envelope size of the state trajectory, calculate the spatial volume occupied by the state trajectory as the envelope size increases, and form a stable attraction structure when the variance of the spatial volume within the continuous monitoring window is less than the preset stability judgment threshold.

[0109] The topological boundary is determined by fitting the spatial location points in the stable attraction structure. The geometric shape parameters of the topological boundary and the distribution density features of the spatial location points are extracted as topological features. The topological signatures corresponding to each fault mode are extracted from the fault mode feature library. The shape similarity and distribution similarity between the topological features and each topological signature are calculated. The shape similarity and distribution similarity are fused to obtain the topological isomorphism. The fault mode corresponding to the topological signature with the highest topological isomorphism is selected as the final evolution state.

[0110] In one specific implementation, within the remaining evolution time window, multi-sensor signals from each measuring point of the wind turbine are continuously acquired, and the time series of each measuring point is statistically analyzed segment by segment using a sliding window of fixed length. The range of the signal within the sliding window is defined as the difference between the maximum and minimum signal values ​​within that window, used to quantify the fluctuation amplitude of the measuring point in the current time period. The range values ​​of all measuring points are sorted, and the three measuring points with the largest range values ​​are selected as target measuring points for subsequent high-frequency sampling. The calculation method of the range value intuitively reflects the instantaneous dynamic range of the signal, and can effectively capture the areas of drastic changes in physical quantities such as vibration, temperature, or electrical quantities during the abnormal evolution of the unit, thereby ensuring that the selected measuring points have the strongest fault information carrying capacity. After determining the three target measuring points, a high-frequency sampling mode is initiated for them. The sampling frequency is set according to the signal bandwidth requirements to ensure that the sampling times of the three measuring points are strictly synchronized, avoiding phase errors introduced due to time deviations.

[0111] The synchronous measurement values ​​of the three measuring points at each sampling time are mapped to the three coordinate axes of a three-dimensional spatial coordinate system, that is, the measurement value of the first measuring point corresponds to... The coordinates of the axes correspond to the measured values ​​of the second measuring point. The coordinates of the axes correspond to the measured values ​​of the third measuring point. Axial coordinates. Each sampling moment corresponds to a spatial location point in the three-dimensional coordinate system. As sampling continues, these spatial location points accumulate, forming a dynamic phase space. The physical meaning of this construction method is that the joint state of the three measuring points can be represented by a unique point in three-dimensional space at any given moment. Therefore, the dynamic evolution process of the unit is transformed into a time-series trajectory problem of a set of points in three-dimensional space. The dynamism of the phase space is reflected in the continuous updating of the set of spatial location points as new sampling data is added, and the trajectory shape evolves over time.

[0112] The spatial locations are connected sequentially according to the sampling time to form a continuous state trajectory. To determine whether the state trajectory has converged to a stable attracting structure, two quantitative indicators are introduced: the envelope size and the spatial volume. The envelope size is defined as the maximum span of all current spatial locations along each axis of the three-dimensional coordinate system, reflecting the overall distribution range of the trajectory in space. The spatial volume is defined as the volume of the three-dimensional convex hull occupied by the state trajectory. The convex hull algorithm is used to calculate the envelope of all current spatial locations to obtain the volume value of the minimum convex polyhedron, denoted as [missing value]. Within the continuous monitoring window, the spatial volume value corresponding to each window is... Perform statistical analysis and calculate the relationship between adjacent monitoring windows. The variance is denoted as .when Less than the preset stability threshold At that time, it is determined that the state trajectory has formed a stable attraction structure. The basis of this determination logic is that if the dynamic evolution of the unit has entered a certain stable fault mode, its phase space trajectory will tend to traverse repeatedly within a finite volume, and the volume fluctuation will approach zero; conversely, if the evolution is still in the transition stage, the trajectory will continue to expand, and the volume variance will be large.

[0113] After the stable attraction structure is formed, boundary fitting is performed on all spatial points within the structure to determine the topological boundary. Boundary fitting employs... The alpha-shape algorithm controls the compactness of the boundary by adjusting the contraction parameter, ensuring that the fitted boundary both encompasses all spatial points and reflects the local concavity and convexity of the trajectory. After fitting, the geometric shape parameters of the topological boundary are extracted, including the mean principal curvature, surface area, major-minor axis ratio, and fractal dimension of the boundary surface. These parameters collectively describe the macroscopic geometry of the attraction structure. Simultaneously, a density distribution analysis is performed on the spatial points within the stable attraction structure. The three-dimensional space is divided into a uniform grid, the number of points in each grid cell is counted, and after normalization, a point density distribution field is obtained. The peak position, density gradient direction, and entropy value of the density distribution are extracted as distribution density features. The geometric shape parameters and distribution density features together constitute the topological feature vector of the current state, denoted as . .

[0114] The fault mode feature library pre-stores topological signatures corresponding to various typical fault modes. Each topological signature also consists of two parts: geometric shape parameters and distribution density features. Let the first one be the topological signature. The topological signature of each failure mode is Calculate the current topological feature vector. With each topology signature Shape similarity between Similarity to distribution Shape similarity is calculated based on the cosine similarity between geometric shape parameter sub-vectors, while distribution similarity is calculated based on the Bhattacharyya coefficient between density feature sub-vectors. Both are normalized to [value missing]. The interval is defined by a higher numerical value, indicating a greater degree of similarity. A weighted fusion of shape similarity and distribution similarity yields the topological isomorphism. The fusion weights are set based on historical diagnostic experience, and the shape similarity weight is denoted as... The distribution similarity weight is denoted as ,satisfy ,but .

[0115] Traverse all fault modes in the fault mode feature library and calculate the corresponding topological isomorphism. Select The failure mode corresponding to the maximum value is taken as the final evolutionary state. If the maximum topological isomorphism is lower than the preset confidence threshold... If the current attraction structure does not match any known fault modes, an "unknown fault mode" flag is output, prompting maintenance personnel to conduct manual verification. The current topology feature vector and evolution process data are then archived into the fault knowledge base to support the expansion and updating of subsequent fault modes. After the final evolution state is determined, it is passed to the subsequent diagnostic conclusion output stage. Combined with intervention strategies from historical cases, maintenance suggestions and handling plans for the current fault mode are generated, completing the closed loop of the entire health diagnosis process.

[0116] In one optional embodiment, extracting the geometric shape parameters and spatial location point distribution density features of the topological boundary as topological features includes:

[0117] Calculate the principal axis direction of the topological boundary, measure the maximum span of the topological boundary along the principal axis direction as the major axis dimension, measure the maximum span of the topological boundary perpendicular to the principal axis direction as the minor axis dimension, and calculate the ratio of the major axis dimension to the minor axis dimension to determine the flatness parameter;

[0118] The topological boundary is divided into multiple sections along the principal axis, the area of ​​each section is calculated, the change curve of the section area along the principal axis is extracted as the morphological evolution feature, and the degree of aggregation of spatial location points in each section in the stable attraction structure is statistically analyzed as the distribution density feature. The flatness parameter, morphological evolution feature and distribution density feature together constitute the topological feature.

[0119] In one specific implementation, after the stable attraction structure is formed, it is necessary to extract topological features from its geometry that can effectively characterize the fault modes. The topological boundary is formed by the outer envelope of the state trajectory in the dynamic phase space, reflecting the overall geometric contour of the system's motion in the phase space. To quantify this geometric contour into a feature vector that can be used for matching, two dimensions are needed: first, the geometric shape parameters of the topological boundary; and second, the distribution density characteristics of spatial points within the boundary. Together, these constitute a topological feature vector that can distinguish different fault modes. .

[0120] When calculating the principal axis direction of the topological boundary, principal component analysis is performed on the entire set of spatial locations covered by the stable attracting structure, and the direction with the largest variance is extracted as the principal axis direction. Specifically, a covariance matrix is ​​constructed for the set of spatial locations, and eigenvalue decomposition is performed on the covariance matrix. The eigenvector corresponding to the largest eigenvalue is taken as the unit vector of the principal axis direction, denoted as . The principal axis direction represents the direction in which the state trajectory extends most significantly in phase space, and usually corresponds to the main dynamic degrees of freedom of the fault excitation. Along... The direction projects all spatial points onto the principal axis; the difference between the maximum and minimum projected values ​​is the major axis dimension, denoted as . Perpendicular to The plane is formed by the direction of the secondary principal components. and Zhang Cheng, after projecting all spatial points onto this plane, takes the projected coordinates at... and The combined maximum span in both directions is taken as the minor axis dimension, denoted as Flatness parameter Defined as the ratio of the major axis dimension to the minor axis dimension, i.e. .when When the value is close to 1, the topological boundary is approximately spherical, corresponding to a random vibration-type fault; when... When the value is significantly greater than 1, the topological boundary is an elongated ellipsoid, corresponding to a single-degree-of-freedom periodic excitation type fault; when When the value is in the intermediate range, it corresponds to a multi-degree-of-freedom coupled fault. The flatness parameter is one of the core scalar features that distinguish the topological signatures of different fault modes.

[0121] Obtaining the principal axis direction Then, the topological boundary is divided along the principal axis. A number of equally spaced sections. The division method is as follows: the principal axis projection range... Evenly divided There are 10 intervals, each corresponding to a cross-sectional slice. For slices falling into the 10th interval... For each spatial point within a cross-sectional slice, calculate its two-dimensional convex hull area in a plane perpendicular to the principal axis, denoted as . ,in This yields the cross-sectional area sequence. This sequence describes the variation of the cross-sectional area of ​​the topological boundary along the principal axis, and is called the morphological evolution feature, denoted as a vector. Morphological evolution characteristics carry information about the dynamic evolution of fault modes in phase space: for bearing outer ring spalling faults, the morphological evolution characteristics typically exhibit a spindle-shaped distribution that is wide in the middle and narrow at both ends; for gear meshing impact faults, they exhibit multi-peak fluctuation characteristics; and for unbalanced mass faults, they exhibit an approximately uniform distribution. By normalizing the morphological evolution characteristics, amplitude deviations caused by differences in sensor ranges are eliminated, making the morphological evolution characteristics under different operating conditions comparable. The normalization method is to... Dividing each element by the maximum value in the sequence yields the normalized morphological evolution feature vector. .

[0122] The extraction of distribution density features is based on the quantification of the degree of clustering of spatial locations within each cross-section. For the first... For each cross-sectional slice, count the number of spatial locations falling within that slice. and divide it by the convex hull area of ​​the slice. , obtained the surface density of each cross section Surface density sequence This reflects the frequency distribution of the state trajectory in phase space, i.e., the relative length of time the system spends in different phase space regions. Sections with high areal density correspond to slowly changing regions of system motion, while sections with low areal density correspond to rapidly changing regions. The areal density sequence, after normalization, forms the distribution density feature vector, denoted as... The normalization method is to divide by the sum of the sequences so that the sum of all components is 1, forming a probability distribution. Distribution density features play an important role in distinguishing fault modes with similar geometries but different dynamic properties. For example, two faults may produce attractive structures with similar shapes, but their access frequency distributions of state trajectories are significantly different. Distribution density features can effectively capture this difference.

[0123] Flatness parameter Normalized morphological evolution feature vector With normalized distribution density eigenvector Concatenate them sequentially to form a complete topological feature vector. ,Right now Where the flatness parameter is a scalar, and the morphological evolution feature vector and the distribution density feature vector are respectively... Therefore, the total dimension of the topological feature vectors is . In practical applications, Typically, an integer between 16 and 32 is chosen to strike a balance between feature resolution and computational cost. Too few cross sections will prevent morphological evolution features from capturing detailed changes, while too many cross sections will introduce noise and increase computational cost.

[0124] During the establishment phase of the fault mode feature library, a topological signature vector is extracted for each known fault mode following the same procedure. Its dimensions and Maintain consistency. During the isomorphic matching phase, the current topological feature vector... Topological signature vectors for each failure mode Perform comparisons and calculate shape similarity separately. Similarity to distribution The shape similarity is calculated based on the cosine similarity of the difference in flatness parameters and morphological evolution characteristics, while the distribution similarity is calculated based on the reciprocal form of the Kullback-Leibler divergence of the distribution density feature vector. The final result is the fusion of these factors to obtain the topological isomorphism. and credibility threshold By comparison, the failure mode corresponding to the final evolutionary state is determined.

[0125] A second aspect of this invention provides a wind turbine health diagnosis system integrating multiple sensors and intelligent algorithms, comprising:

[0126] The signal acquisition unit is used to acquire signals from multiple sensors in the wind turbine.

[0127] An anomaly detection unit is used to extract the peak frequency sequence of the envelope spectrum from multi-sensor signals and map it to a high-dimensional feature space. It learns the frequency migration law through recursive neural units and generates a conditional probability density field. If the measured peak frequency deviates from the confidence interval of the conditional probability density field, the measured peak frequency is marked as an anomaly frequency.

[0128] The fault matching unit is used to retrieve historical cases with frequency feature matching in the historical fault knowledge base based on the abnormal frequency, extract the evolution trajectory of the corresponding multi-sensor signals, calculate the morphological matching degree between the current evolution trajectory and the historical evolution trajectory through the dynamic time warping algorithm, obtain the time interval from the occurrence of the abnormality to the fault confirmation from the historical case with the highest matching degree, and obtain the remaining evolution time.

[0129] The state trajectory unit is used to identify the three measurement points with the largest fluctuation amplitude in the remaining evolution time and sample them at high frequency. The synchronous measurement values ​​of the three measurement points are used as coordinates to construct a dynamic phase space and draw the state trajectory. When the state trajectory forms a stable attraction structure, the topological features of the stable attraction structure are extracted and isomorphically matched with the topological signature in the fault mode feature library to determine the final evolution state.

[0130] The diagnostic decision unit is used to output fault diagnosis conclusions and intervention strategies based on the final evolution state.

[0131] A third aspect of the present invention provides an electronic device, comprising:

[0132] processor;

[0133] Memory used to store processor-executable instructions;

[0134] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0135] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0136] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for health diagnosis of wind turbine units integrating multiple sensors and intelligent algorithms, characterized in that, include: Collect signals from multiple sensors in the wind turbine generator; The envelope spectrum peak frequency sequence is extracted from multi-sensor signals and mapped to a high-dimensional feature space. Frequency migration rules are learned through recursive neural units and a conditional probability density field is generated. If the measured peak frequency deviates from the confidence interval of the conditional probability density field, the measured peak frequency is marked as an abnormal frequency. Based on the abnormal frequency, historical cases with frequency feature matching are retrieved from the historical fault knowledge base, and the evolution trajectory of the corresponding multi-sensor signals is extracted. The morphological matching degree between the current evolution trajectory and the historical evolution trajectory is calculated by the dynamic time warping algorithm. The time interval from the occurrence of the abnormality to the confirmation of the fault is obtained from the historical case with the highest matching degree, and the remaining evolution time is obtained. During the remaining evolution time, the three measurement points with the largest fluctuation amplitude are identified and sampled at high frequency. The synchronous measurement values ​​of the three measurement points are used as coordinates to construct a dynamic phase space and draw the state trajectory. When the state trajectory forms a stable attraction structure, the topological features of the stable attraction structure are extracted and isomorphically matched with the topological signature in the fault mode feature library to determine the final evolution state. Based on the final state of evolution, output fault diagnosis conclusions and intervention strategies.

2. The method according to claim 1, characterized in that, The envelope spectrum peak frequency sequence is extracted from multi-sensor signals and mapped to a high-dimensional feature space. Frequency migration patterns are learned through recursive neural units, and a conditional probability density field is generated. If the measured peak frequency deviates from the confidence interval of the conditional probability density field, the measured peak frequency is marked as an anomalous frequency, including: The multi-sensor signals are subjected to envelope demodulation to obtain an envelope signal. The envelope signal is then subjected to spectral analysis to determine the envelope spectrum. The frequencies corresponding to the amplitude peaks are extracted from the envelope spectrum to obtain peak frequencies. The peak frequencies are then arranged in chronological order to form a peak frequency sequence. Each frequency value in the peak frequency sequence is converted into a high-dimensional vector. The high-dimensional vector is then mapped to a high-dimensional feature space using a kernel function. The manifold distance between frequency vectors at adjacent time points is calculated in the high-dimensional feature space to construct the topological manifold structure for frequency migration. The topological manifold structure and manifold distance are input into the recurrent neural unit to capture the migration trajectory of the frequency on the topological manifold, and a conditional probability density field is generated based on the migration trajectory and manifold distance. Calculate the confidence interval of the conditional probability density field, and determine whether the measured peak frequency falls outside the confidence interval. If it falls outside the confidence interval, mark the measured peak frequency as an abnormal frequency.

3. The method according to claim 2, characterized in that, The topological manifold structure and manifold distance are input into the recurrent neural unit to capture the migration trajectory of frequencies on the topological manifold. The conditional probability density field is generated based on the migration trajectory and manifold distance, including: The representations of frequencies at each time step in the high-dimensional feature space of the topological manifold are organized into an input sequence in chronological order. The manifold distance between adjacent time step frequency vectors is extracted from the topological manifold to construct a manifold distance sequence. The input sequence and the manifold distance sequence are simultaneously input into the recurrent neural unit. The forget gate, input gate and output gate in the recurrent neural unit use the manifold distance at the current time as a geometric control factor to generate forget weight, input weight and output weight respectively. Candidate memory content is generated based on the frequency vector at the current time and the hidden state at the previous time. The memory unit at the previous time is filtered by forgetting weight to obtain the retained memory information. The candidate memory content is controlled by input weight to obtain the new memory information. The retained memory information and the new memory information are fused to form the memory unit at the current time. The memory unit at the current time is activated to obtain the activated memory information. The hidden state at the current time is extracted from the activated memory information by output weight. The hidden states at each time are organized to form the migration trajectory of the frequency on the topological manifold. The hidden state and manifold distance at the current time are simultaneously input into the probability distribution parameters to generate the network. The mean vector is determined based on the hidden state at the current time, and the covariance matrix is ​​determined by using the manifold distance at the current time as a dispersion control factor. The conditional probability density field is constructed based on the mean vector and the covariance matrix.

4. The method according to claim 1, characterized in that, Based on the abnormal frequency, historical cases with frequency feature matching are retrieved from the historical fault knowledge base, and the evolution trajectory of the corresponding multi-sensor signals is extracted. The morphological matching degree between the current evolution trajectory and the historical evolution trajectory is calculated using a dynamic time warping algorithm. The time interval from the occurrence of the anomaly to the confirmation of the fault is obtained from the historical case with the highest matching degree, and the remaining evolution time is obtained, including: The abnormal frequency is converted into a frequency feature vector. The vector distance between the frequency feature vector of each historical case and the frequency feature vector of the abnormal frequency is calculated in the historical fault knowledge base. Candidate historical cases are selected based on the vector distance and the corresponding multi-sensor signal evolution trajectory is extracted. The evolution trajectory of the current multi-sensor signal and the historical evolution trajectory are input into the dynamic time warping algorithm. A nonlinear time mapping relationship is established through the time alignment path and the morphological matching degree is calculated. The historical case with the highest morphological matching degree is selected and the time interval between the time of anomaly frequency occurrence and the time of fault confirmation is extracted to obtain the historical evolution time. The time scale factor is determined according to the ratio of the number of time points of the current evolution trajectory and the historical evolution trajectory in the time alignment path. The historical evolution time is multiplied by the time scale factor to obtain the remaining evolution time.

5. The method according to claim 4, characterized in that, The evolution trajectories of current multi-sensor signals and historical evolution trajectories are input into a dynamic time warping algorithm. A nonlinear time mapping relationship is established through a time alignment path, and the morphological matching degree is calculated, including: The evolution trajectory of the current multi-sensor signal and the historical evolution trajectory are divided into the current time series and the historical time series, respectively. The point-to-point distance between each time point in the current time series and the historical time series is calculated. Initialize the cumulative distance matrix, traverse the time points of the current time series and the historical time series in ascending order of time, select the minimum cumulative distance from the adjacent previous position of the current time point in the cumulative distance matrix for the current time point, add the minimum cumulative distance to the point-to-point distance corresponding to the current time point to update the elements of the cumulative distance matrix, and complete the construction of the cumulative distance matrix. Backtracking from the end position of the cumulative distance matrix to the start position, the preceding position with the smallest cumulative distance is selected as the path node to form a time alignment path. The path nodes in the time alignment path establish a non-linear time mapping relationship between the current time series and the historical time series. The point-to-point distances corresponding to each path node on the time alignment path are extracted and accumulated and normalized to obtain the morphological matching degree.

6. The method according to claim 1, characterized in that, During the remaining evolution time, the three measurement points with the largest fluctuation amplitude are identified and sampled at high frequency. The synchronous measurement values ​​of the three measurement points are used as coordinates to construct a dynamic phase space and draw the state trajectory. When the state trajectory forms a stable attraction structure, the topological features of the stable attraction structure are extracted and isomorphically matched with the topological signatures in the fault mode feature library to determine the final evolution state, including: During the remaining evolution time, the multi-sensor signals at each measurement point are statistically analyzed using a sliding window. The range of the signals within the sliding window is calculated as the fluctuation amplitude. The three measurement points with the largest fluctuation amplitude are selected, and high-frequency sampling is performed on the three measurement points to obtain synchronous measurement values. The synchronous measurement values ​​of the three measuring points are mapped to the coordinate axis values ​​of the three-dimensional spatial coordinate system, and the spatial position points corresponding to each sampling time are calibrated in the three-dimensional spatial coordinate system to construct a dynamic phase space; Connect the spatial locations in chronological order to form a state trajectory, calculate the envelope size of the state trajectory, calculate the spatial volume occupied by the state trajectory as the envelope size increases, and form a stable attraction structure when the variance of the spatial volume within the continuous monitoring window is less than the preset stability judgment threshold. The topological boundary is determined by fitting the spatial location points in the stable attraction structure. The geometric shape parameters of the topological boundary and the distribution density features of the spatial location points are extracted as topological features. The topological signatures corresponding to each fault mode are extracted from the fault mode feature library. The shape similarity and distribution similarity between the topological features and each topological signature are calculated. The shape similarity and distribution similarity are fused to obtain the topological isomorphism. The fault mode corresponding to the topological signature with the highest topological isomorphism is selected as the final evolution state.

7. The method according to claim 6, characterized in that, Extracting the geometric shape parameters and spatial location point distribution density features of the topological boundary as topological features includes: Calculate the principal axis direction of the topological boundary, measure the maximum span of the topological boundary along the principal axis direction as the major axis dimension, measure the maximum span of the topological boundary perpendicular to the principal axis direction as the minor axis dimension, and calculate the ratio of the major axis dimension to the minor axis dimension to determine the flatness parameter; The topological boundary is divided into multiple sections along the principal axis, the area of ​​each section is calculated, the change curve of the section area along the principal axis is extracted as the morphological evolution feature, and the degree of aggregation of spatial location points in each section in the stable attraction structure is statistically analyzed as the distribution density feature. The flatness parameter, morphological evolution feature and distribution density feature together constitute the topological feature.

8. A wind turbine health diagnosis system integrating multiple sensors and intelligent algorithms, used to implement the method as described in any one of claims 1-7, characterized in that, include: The signal acquisition unit is used to acquire signals from multiple sensors in the wind turbine generator. An anomaly detection unit is used to extract the peak frequency sequence of the envelope spectrum from multi-sensor signals and map it to a high-dimensional feature space. It learns the frequency migration law through recursive neural units and generates a conditional probability density field. If the measured peak frequency deviates from the confidence interval of the conditional probability density field, the measured peak frequency is marked as an anomaly frequency. The fault matching unit is used to retrieve historical cases with frequency feature matching in the historical fault knowledge base based on the abnormal frequency, extract the evolution trajectory of the corresponding multi-sensor signals, calculate the morphological matching degree between the current evolution trajectory and the historical evolution trajectory through the dynamic time warping algorithm, obtain the time interval from the occurrence of the abnormality to the fault confirmation from the historical case with the highest matching degree, and obtain the remaining evolution time. The state trajectory unit is used to identify the three measurement points with the largest fluctuation amplitude in the remaining evolution time and sample them at high frequency. The synchronous measurement values ​​of the three measurement points are used as coordinates to construct a dynamic phase space and draw the state trajectory. When the state trajectory forms a stable attraction structure, the topological features of the stable attraction structure are extracted and isomorphically matched with the topological signature in the fault mode feature library to determine the final evolution state. The diagnostic decision unit is used to output fault diagnosis conclusions and intervention strategies based on the final evolution state.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images