Intelligent fan diagnosis method, system and equipment and medium

Abstract

The invention discloses an intelligent fan diagnosis method, system and equipment and a medium, and relates to the technical field of wind power equipment diagnos.The method comprises the steps that variable-rate collection is conducted on a multi-source sensor signal at a fan end, and original data are obtained; extracting and physically aligning multi-scale features based on the original data to obtain a physical feature vector; training an auto-encoder with physical constraints based on the physical feature vectors to obtain local model parameters; performing weighted aggregation and domain adaptation on local model parameters to obtain a global diagnosis model, and fusing an abnormal score, a classification probability and a group deviation degree to obtain fault probability distribution and an uncertainty index; and based on the fault probability distribution and the uncertainty index, reasoning through a sequential logic rule base to obtain a diagnosis conclusion. According to the method, high sensitivity, cross-domain strong generalization, high diagnosis confidence and explainability of early faults of the wind turbine generator can be realized.

Description

Technical Field

[0001] This invention relates to the field of wind power equipment diagnostic technology, and in particular to a wind turbine intelligent diagnostic method, system, equipment and medium. Background Technology

[0002] As the global installed capacity of wind power continues to expand, wind turbines operate in harsh environments with variable speeds and loads for extended periods, leading to a year-on-year increase in the failure rate of core components such as gearboxes, bearings, and generators. Sudden failures not only incur high maintenance costs but can also trigger a chain reaction of shutdowns, resulting in a sharp drop in overall power generation. Therefore, achieving high-precision, interpretable, and cross-unit generalization intelligent diagnosis in the early stages of a fault has become a critical technical challenge for wind power operation and maintenance.

[0003] To address these needs, existing technologies primarily employ two approaches: one is expert systems based on vibration signal threshold alarms or empirical rules, which determine anomalies by setting fixed thresholds; the other utilizes purely data-driven models, such as autoencoders, convolutional neural networks, or transfer learning, to perform end-to-end training on multi-source sensor data and output fault category probabilities. Some solutions further introduce federated learning frameworks to aggregate parameters from various wind field models in the cloud to alleviate the problem of data silos.

[0004] However, existing technologies still have significant shortcomings: threshold rules rely on human experience and have low sensitivity to complex coupled faults; pure data-driven models lack physical constraints, making early weak fault characteristics easily submerged by noise and resulting in poor interpretability of diagnostic results; the federated aggregation process ignores the differences in the distribution of different wind farm operating conditions, resulting in insufficient cross-domain generalization ability of the model; at the same time, existing solutions lack quantitative means for diagnostic uncertainty, cannot automatically trigger verification in high-risk scenarios, and are difficult to meet the high-confidence decision-making needs of unattended wind farms.

[0005] In summary, there is an urgent need for a novel diagnostic method that integrates physical mechanisms and data-driven approaches, supports uncertainty quantification and adaptive verification, in order to improve the accuracy, interpretability, and cross-domain robustness of early fault identification in wind turbines, and provide reliable technical support for smart wind power operation and maintenance. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the shortcomings of existing technologies, specifically the problems of insufficient early sensitivity in wind turbine fault diagnosis, poor cross-domain generalization ability, and lack of uncertainty quantification and adaptive verification mechanisms. Specifically, this invention provides a wind turbine intelligent diagnosis method, system, equipment, and medium, as detailed below: 1) In a first aspect, the present invention provides a method for intelligent diagnosis of wind turbines, the specific technical solution of which is as follows: S1, at the wind turbine end, acquires raw data by performing variable-rate acquisition of signals from multiple sources of sensors; S2, Based on the original data, extract and physically align multi-scale features to obtain a physical feature vector; S3, Train an autoencoder with physical constraints based on the physical feature vectors to obtain local model parameters; Perform weighted aggregation and domain adaptation on the local model parameters to obtain a global diagnostic model; S4. Based on the obtained global diagnostic model, the abnormal score, classification probability, and group deviation are integrated to obtain the fault probability distribution and uncertainty index; based on the fault probability distribution and the uncertainty index, the diagnostic conclusion is obtained through reasoning using the temporal logic rule base. S5. When the uncertainty index does not meet the preset threshold, increase the sampling frequency and repeat S2-S4 to obtain the final diagnostic conclusion after verification.

[0007] The beneficial effects of the intelligent diagnostic method for wind turbines provided by this invention are as follows: By employing variable-rate data acquisition, low power consumption is ensured under normal conditions while high-frequency capture is achieved during anomalies, reducing the rate of missed detections of early, minor faults. Raw data is converted into physically aligned feature vectors, enabling direct comparison of values ​​for the same fault across different units and wind farms, improving cross-domain consistency. Physically constrained autoencoder training and parameter uploading are performed at the edge, with the center only aggregating statistics, reducing raw data transmission and alleviating bandwidth and privacy pressures. The globally diagnostic model, after weighted aggregation and domain adaptation, allows newly connected wind farms to benefit directly without requiring extensive re-collection of samples, shortening deployment cycles. By fusing anomaly scores, classification probabilities, and group deviations to output uncertainty indicators, maintenance personnel can decide whether to immediately repair or postpone processing based on quantified confidence levels, reducing unnecessary downtime. When uncertainty exceeds limits, the sampling frequency is automatically increased and cyclical diagnosis is performed. If two conclusions are consistent, the confidence level is increased; otherwise, it is marked for further review, avoiding economic losses from single misjudgments.

[0008] Based on the above solution, the present invention can be further improved as follows.

[0009] Furthermore, the variable rate acquisition includes: The multi-source sensor signals at the wind turbine end under normal conditions are sampled and acquired at low frequency; High-frequency sampling is used to collect multi-source sensor signals at the wind turbine end under abnormal conditions.

[0010] Furthermore, S2 specifically refers to: The raw data includes at least vibration signals, acoustic signals, and rotational speed signals; The vibration signal is subjected to wavelet packet decomposition, envelope demodulation, and empirical mode decomposition to extract vibration features including sub-band energy, kurtosis, and power spectral entropy. Perform a short-time Fourier transform on the acoustic signal to obtain acoustic features that include time-frequency energy distribution; The spectral features are converted into order features using the rotational speed signal, and the frequency consistency measure and gear meshing frequency consistency measure are calculated. The frequency consistency measure and the gear meshing frequency consistency measure are used as rotational speed features. The vibration features, acoustic features, and rotational speed features are aligned according to a time window, normalized, and then concatenated to obtain the physical feature vector.

[0011] Furthermore, the physically constrained self-encoder specifically comprises: A physical consistency loss is introduced into the loss function of the autoencoder, and the physical consistency loss consists of the weighted sum of squared residuals of the frequency estimation variance, so as to force the reconstruction result to conform to the wind turbine dynamics constraints.

[0012] 2) In a second aspect, the present invention also provides a wind turbine intelligent diagnostic system, the specific technical solution of which is as follows: a variable speed acquisition module, a feature extraction module, a model parameter module, a fusion inference module, and a verification and diagnostic module; The variable speed acquisition module is used to acquire signals from multiple sources at a variable speed at the wind turbine end to obtain raw data; The feature extraction module is used to extract and physically align multi-scale features based on the original data to obtain a physical feature vector; The model parameter module is used to train a physically constrained autoencoder based on the physical feature vector to obtain local model parameters; and to obtain a global diagnostic model by weighted aggregation and domain adaptation of the local model parameters. The fusion reasoning module is used to fuse anomaly scores, classification probabilities, and group deviations based on the obtained global diagnostic model to obtain a fault probability distribution and uncertainty index; and to obtain a diagnostic conclusion based on the fault probability distribution and the uncertainty index through reasoning using a time-series logic rule base. The verification and diagnosis module is used to increase the sampling frequency and repeat the feature extraction module to the fusion inference module when the uncertainty index does not meet the preset threshold, so as to obtain the final diagnosis conclusion after verification.

[0013] Based on the above solution, the present invention can be further improved as follows.

[0014] Furthermore, the variable rate acquisition includes: The multi-source sensor signals at the wind turbine end under normal conditions are sampled and acquired at low frequency; High-frequency sampling is used to collect multi-source sensor signals at the wind turbine end under abnormal conditions.

[0015] Furthermore, the feature extraction module specifically comprises: The raw data includes at least vibration signals, acoustic signals, and rotational speed signals; The vibration signal is subjected to wavelet packet decomposition, envelope demodulation, and empirical mode decomposition to extract vibration features including sub-band energy, kurtosis, and power spectral entropy. Perform a short-time Fourier transform on the acoustic signal to obtain acoustic features that include time-frequency energy distribution; The spectral features are converted into order features using the rotational speed signal, and the frequency consistency measure and gear meshing frequency consistency measure are calculated. The frequency consistency measure and the gear meshing frequency consistency measure are used as rotational speed features. The vibration features, acoustic features, and rotational speed features are aligned according to a time window, normalized, and then concatenated to obtain the physical feature vector.

[0016] Furthermore, the physically constrained self-encoder specifically comprises: A physical consistency loss is introduced into the loss function of the autoencoder, and the physical consistency loss consists of the weighted sum of squared residuals of the frequency estimation variance, so as to force the reconstruction result to conform to the wind turbine dynamics constraints.

[0017] 3) In a third aspect, the present invention also provides a computer device, the computer device including a processor coupled to a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to enable the computer device to implement any of the above methods.

[0018] 4) In a fourth aspect, the present invention also provides a computer-readable storage medium storing at least one computer program, which is loaded and executed by a processor to enable a computer to implement any of the above methods.

[0019] It should be noted that the beneficial effects of the technical solutions of the second to fourth aspects of the present invention and their corresponding possible implementations can be found in the above description of the technical effects of the first aspect and its corresponding possible implementations, and will not be repeated here. Attached Figure Description

[0020] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating the steps of a wind turbine intelligent diagnostic method according to an embodiment of the present invention; Figure 2 This is a structural block diagram of a computer device according to the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0022] like Figure 1 As shown in the figure, an intelligent diagnostic method for wind turbines according to an embodiment of the present invention includes the following steps: S1, at the wind turbine end, acquires raw data by performing variable-rate acquisition of signals from multiple sources of sensors; S2, based on the original data, extracts and physically aligns multi-scale features to obtain physical feature vectors; S3: Train an autoencoder with physical constraints based on physical feature vectors to obtain local model parameters; then, through weighted aggregation and domain adaptation, obtain a global diagnostic model. S4. Based on the obtained global diagnostic model, the abnormal score, classification probability, and group deviation are integrated to obtain the fault probability distribution and uncertainty index; based on the fault probability distribution and uncertainty index, the diagnostic conclusion is obtained through reasoning using the temporal logic rule base. S5. When the uncertainty index does not meet the preset threshold, increase the sampling frequency and repeat S2-S4 to obtain the final diagnostic conclusion after verification.

[0023] The beneficial effects of the intelligent diagnostic method for wind turbines provided by this invention are as follows: By employing variable-rate data acquisition, low power consumption is ensured under normal conditions while high-frequency capture is achieved during anomalies, reducing the rate of missed detections of early, minor faults. Raw data is converted into physically aligned feature vectors, enabling direct comparison of values ​​for the same fault across different units and wind farms, improving cross-domain consistency. Physically constrained autoencoder training and parameter uploading are performed at the edge, with the center only aggregating statistics, reducing raw data transmission and alleviating bandwidth and privacy pressures. The globally diagnostic model, after weighted aggregation and domain adaptation, allows newly connected wind farms to benefit directly without requiring extensive re-collection of samples, shortening deployment cycles. By fusing anomaly scores, classification probabilities, and group deviations to output uncertainty indicators, maintenance personnel can decide whether to immediately repair or postpone processing based on quantified confidence levels, reducing unnecessary downtime. When uncertainty exceeds limits, the sampling frequency is automatically increased and cyclical diagnosis is performed. If two conclusions are consistent, the confidence level is increased; otherwise, it is marked for further review, avoiding economic losses from single misjudgments.

[0024] It should be noted that, for ease of understanding, the technical terms used in this solution will be explained one by one, and will not be repeated hereafter: Wind turbine end: refers to the data acquisition and computing unit installed on-site by the wind turbine, i.e., "edge computing node".

[0025] Multi-source sensor signals: These refer to output signals from various sensors, including vibration, acoustic, speed, temperature, current, and voltage sensors. In this embodiment, the sensors are positioned as follows: vibration sensors are placed on the main shaft of the wind turbine, the input shaft of the gearbox, the output shaft of the gearbox, and the generator bearings; acoustic sensors are placed in the nacelle; speed sensors are installed on the main shaft; and temperature sensors are placed on the main shaft bearings, the gearbox oil sump, and the generator windings. Simultaneously, the three-phase current and voltage signals of the generator are collected. Furthermore, timestamps are uniformly added to the multi-source sensor signals, and precise clock synchronization and clock drift compensation are employed. The speed sensor signal is used as a reference for synchronous resampling of the vibration signal. It should be noted that, in addition to the vibration, acoustic, and speed signals, the physical feature vectors used for subsequent calculations can also be derived from the normalized operating condition signals of temperature, torque, current, and voltage, which are then further used to concatenate the operating condition features to obtain the physical feature vectors.

[0026] Variable rate acquisition refers to a mechanism that uses low-frequency sampling to acquire multi-source sensor signals at the wind turbine end under normal conditions, and high-frequency sampling to acquire multi-source sensor signals at the wind turbine end under abnormal conditions. In this embodiment, only the feature vector and low-frequency data are uploaded under normal conditions, and the frequency range of the low-frequency data is 1Hz to 10Hz; when a local anomaly is detected or a secondary review trigger command is received, the original high-frequency waveform data is uploaded.

[0027] Raw data: refers to the sampled values ​​of the aforementioned multi-source sensor signals, which are timestamped and clock-synchronized. In this embodiment, it includes at least vibration signals, acoustic signals, and rotational speed signals, and may also include operating condition signals.

[0028] Physically aligned multi-scale features refer to aligning various types of raw data acquired from multi-source sensor signals, so that the output can be used for subsequent calculations as a physical feature vector. In this embodiment, physically aligned multi-scale features refer to performing wavelet packet decomposition, envelope demodulation, and empirical mode decomposition on the vibration signal to extract vibration features including sub-band energy, kurtosis, and power spectral entropy; performing short-time Fourier transform on the acoustic signal to obtain acoustic features including time-frequency energy distribution; using the rotational speed signal to convert the spectral features into order features, and calculating the frequency consistency measure and gear meshing frequency consistency measure, using the frequency consistency measure and gear meshing frequency consistency measure as rotational speed features; aligning the vibration features, acoustic features, and rotational speed features according to a time window, normalizing them, and concatenating them to obtain the physical feature vector.

[0029] The process of processing vibration signals to obtain vibration characteristics is as follows: The vibration signal is bandpass filtered and subjected to Hilbert transform to obtain the envelope signal. : ; in, It is a vibration signal. Perform Hilbert transform on the vibration signal; perform Fast Fourier Transform on the envelope signal to extract the fault characteristic frequencies of the bearing or gear; Empirical mode decomposition is performed on the vibration signal to obtain several intrinsic mode functions. False modes are screened out based on spectral energy and kurtosis index, and the screened intrinsic mode functions are reconstructed to obtain the denoised vibration characteristics.

[0030] The process of processing acoustic signals to obtain acoustic features is as follows: Perform a short-time Fourier transform on the acoustic signal to extract the time-frequency energy distribution and obtain the acoustic signal; The process of processing the speed signal to obtain the speed characteristics is as follows: The spectral characteristics are normalized to the order based on the rotational speed signal, and the calculation formula is as follows: ; in, For the characteristic frequency of the spectrum, This is the rotation speed signal. This is a characteristic of order.

[0031] Specifically, within each time window, based on the speed signal Calculate the spindle rotation frequency And the gear meshing frequencies (GMF) for each stage are as follows: ; in, For reference gear tooth count, This is the transmission ratio correction factor; Convert the spectral feature f into an order feature. And map the peak values ​​of the spectral features to order domain features: ; For each order feature Calculate its relative deviation from the nearest integer order k. and define a frequency consistency metric: ; in, This is the adjustment constant; For envelope spectrum peak Calculate the peak value of the envelope spectrum relative gear meshing frequency set minimum normalization bias and define GMF consistency metrics. : ; Where γ is the adjustment constant; For time domain signals According to the speed signal Perform equal-angle resampling to obtain the signal The order spectrum is obtained through Fast Fourier Transform. Frequency consistency is then measured. and gear meshing frequency consistency measurement As a characteristic of rotational speed.

[0032] The process of processing operating condition signals to obtain ergonomic characteristics is as follows: Sliding window statistics are performed on low-frequency signals in operating condition signals, including temperature signals, current signals, etc., to obtain the mean, variance, and gradient. Then, normalization processing is performed, and uniform operating condition characteristics are generated by aligning according to the time window.

[0033] Physical feature vector The physical feature vector is composed of vibration characteristics, acoustic characteristics, and rotational speed characteristics. In this embodiment, in addition to the above three features, operating condition characteristics can also be combined to obtain the physical feature vector.

[0034] Physically Constrained Autoencoders: These are physical information autoencoders whose loss function includes a physical consistency loss, which is composed of a weighted sum of squared residuals based on the frequency estimation variance, to force the reconstructed results to conform to wind turbine dynamic constraints. Feature representation learning and fault detection are performed using physical information autoencoders. These autoencoders consist of an encoder and a decoder structure, with the input physical feature vector... The output is the reconstructed vector. Its training objective function is as follows: ; in, To reconstruct the loss, Weights for physical consistency loss. For regularization weights, Regularization loss; Physical consistency loss weights An adaptive scheduling strategy is adopted, and the scheduling is set in the early stage of training. =0.01, and gradually increased to the target value during subsequent training. ∈[0.1,0.5], to balance the reconstruction loss. loss of physical consistency . contributions.

[0035] Physical consistency loss: The calculation formula is as follows: ; ; in, To estimate the variance for frequency, Let i be the i-th spectral characteristic frequency. The uncertainty of the physical eigenvector. As the confidence level weight, Let be the residual of the i-th feature under the physical consistency constraint.

[0036] Local model parameters: refer to the autoencoder weights obtained after training. And characteristic statistics, namely the local characteristic mean. With covariance .

[0037] Weighted aggregation: refers to the aggregation by which the central server aggregates data based on the number of samples from each client (i.e., the wind turbine). The weighted aggregation process is as follows: ; ; in, The calculated global model parameters are similar to the local model parameters. This includes the global feature mean and covariance, which are used in the following domain adaptation calculation process.

[0038] Domain adaptation involves minimizing the first and second-order statistical differences between the global distribution and the client distribution. Its domain adaptation loss is as follows: ; in, These are the global feature mean and covariance, respectively. This represents the Frobenius norm value. In this embodiment, during the domain adaptation process, adversarial training can also be performed between the domain discriminator D(z) and the gradient inversion layer.

[0039] Global diagnostic model: refers to the global model parameters issued after weighted aggregation and domain adaptation. .

[0040] Anomaly scoring: Anomaly scores are compared by setting an anomaly scoring threshold. The threshold is set using the quantile threshold method, which is based on the percentile of the training normal samples. The reconstruction error of the reconstructed vector is then calculated. The anomaly scoring calculation formula is as follows: ; Where α + β = 1, α and β are the reconstruction losses, respectively. loss of physical consistency The weighting coefficients, This is a function that normalizes based on the statistics of the training set.

[0041] Classification probability: refers to the class probability vector output by the supervised learning classifier. The supervised learning classifier receives the physical feature vector aligned with the physical features and calculates the class probability vector. ;in Number of fault categories Let be the probability that the input feature vector belongs to the i-th type of fault. It should be noted that... For physical feature vectors, Let be the probability that the input feature vector belongs to the i-th type of fault. This is the category probability vector.

[0042] Group Deviation: Calculated by comparing the input physical feature vector p with historical normal characteristic statistics of similar units. and The results are obtained through comparison, and the specific formula is as follows: ; in, This represents the physical feature vector of the current wind turbine unit relative to the historical normal state of similar units. The Mahalanobis distance deviation of the physical feature vector is the group deviation of the physical feature vector.

[0043] Failure probability distribution ( The calculation formula is as follows: ; in, It is obtained by normalizing the outlier scores using a nonlinear mapping function.

[0044] Uncertainty indicators ( The calculation formula is as follows: .

[0045] A temporal logic rule base refers to a combination of atomic predicates and logical formulas, where atomic predicates include, but are not limited to: Rotational speed change Compare with a threshold; Envelope spectrum harmonic amplitude increment Compare with a threshold; Temperature change Compare with a threshold; Anomaly scoring of physical information autoencoder Compare with a threshold; Class probability distribution of supervised learning classifiers Compare with category thresholds; Group deviation Compare with the threshold.

[0046] Specifically, atomic predicates are mapped to trigger strengths via the Sigmoid function. The contribution of rule evidence is obtained through rule logic combination. ; Rule-based evidence contribution The posterior probability distribution of the fault category (i.e., the fault probability distribution) output by the integrated inferrer. The uncertainty index σ is jointly fused to generate the final diagnostic posterior probability distribution. With the final uncertainty index .

[0047] Preset threshold: refers to the set value in "Uncertainty index exceeds preset threshold". It can be set according to actual needs.

[0048] The final diagnostic conclusion after review: refers to increasing the sampling frequency and repeating S2-S4 to obtain the final diagnostic conclusion after review. Specifically, in this embodiment, when the final uncertainty index... When the preset threshold is exceeded or a high-priority rule is triggered by the time-series logic rule base, the edge computing node (i.e., the wind turbine end for multi-source sensor signals) automatically switches to high-frequency sampling and re-runs steps S2-S4 to generate a secondary diagnostic result. The consistency between the initial and secondary diagnostic results is compared with the secondary diagnostic results at the edge computing nodes. If they are consistent, the diagnostic confidence is increased. If they are inconsistent, output a diagnostic conclusion pending review. It also records the evidence chain structure, including feature contribution, physical residual, rule trigger strength, and verification consistency.

[0049] In another embodiment of this solution, S1 is specifically implemented as follows: Vibration sensors are installed on the wind turbine main shaft, gearbox input shaft, gearbox output shaft, and generator bearings. Acoustic sensors are installed in the nacelle. Speed ​​sensors are installed on the main shaft, and temperature sensors are installed on the main shaft bearings, gearbox oil sump, and generator windings. Simultaneously, the generator's three-phase current and voltage signals are collected. All sensor signals are timestamped and synchronized with a precise clock and clock drift compensation are applied. The speed sensor signal is used as a reference to perform synchronous resampling of the vibration signal. Edge computing nodes default to sampling feature vectors and low-frequency data at low frequencies of 1Hz to 10Hz. When a local anomaly is detected or a secondary verification trigger command is received, the sampling frequency is automatically switched to ≥1kHz and the original waveform data is uploaded. At the same time, the health status of the acquisition link is monitored and reported, including sensor disconnection and analog-to-digital conversion saturation information, thus obtaining the original data with timestamps and clock synchronization.

[0050] In another embodiment of this solution, S2 is specifically implemented as follows: After performing multi-level wavelet packet decomposition on the vibration signal to obtain sub-bands, the energy, kurtosis, and power spectral entropy are calculated. Then, bandpass filtering and Hilbert transform are performed to obtain the envelope signal. The system performs Fast Fourier Transform (FFT) on the envelope signal to extract bearing or gear fault characteristic frequencies. Simultaneously, it performs Empirical Mode Decomposition (EMD) to obtain intrinsic mode functions (EMFs), and after filtering out pseudo-modes based on spectral energy and kurtosis, reconstructs a denoised sequence to obtain vibration characteristics. It also performs Short-Time Fourier Transform (SFT) on the acoustic signal to extract time-frequency energy distribution, obtaining acoustic characteristics. Furthermore, it performs sliding window statistics on low-frequency temperature, current, and voltage signals to obtain mean, variance, and gradient, thus obtaining operating condition characteristics. Finally, it calculates the spindle rotation frequency using the rotational speed signal n(t). and the meshing frequency of each gear Convert spectral features into order features The relative deviation between the order of calculation and the nearest integer order And obtain frequency consistency measure For the peak value of the envelope spectrum Calculate the minimum normalized deviation from the GMF set. And obtain the GMF consistency measure The rotational speed characteristics are obtained by resampling the time-domain signal at equal angles according to the rotational speed signal, obtaining x(θ), and then using a fast Fourier transform to obtain the order spectrum. Vibration characteristics, acoustic characteristics, and operating condition characteristics are then combined. , The physical feature alignment vector is formed by concatenating these features, normalizing all features and aligning them according to the time window, thus obtaining the physical feature vector.

[0051] In another embodiment of this solution, S3 is specifically implemented as follows: At the edge, the physical feature vector is input into an autoencoder consisting of an encoder and a decoder, and the total loss is set. Reconstruction loss Mean square error of input and output, physical consistency loss , Let i be the frequency-order residual of the i-th feature under the wind turbine dynamics model. Its relative uncertainty, The value was gradually increased from 0.01 to 0.1-0.5 with each training epoch; the local model parameters were obtained after training with normal samples. and local characteristic mean Covariance And upload to the central server, which will then sort the samples according to the number of wind turbines. Weighted aggregation of global parameters Subsequently, by minimizing the first and second-order statistical differences between the global distribution and the client distribution... Perform domain adaptation, and introduce a domain discriminator and gradient inversion layer for adversarial training. After distribution alignment is completed, the updated global model parameters are distributed, thus obtaining a global diagnostic model that can be loaded and used by each wind turbine.

[0052] In another embodiment of this solution, S4 is specifically implemented as follows: After the global diagnostic model is distributed, it is loaded onto the edge side: first, the reconstruction error and physical consistency residual are calculated using an autoencoder on the current physical feature vector, and then the anomaly score s(x) is obtained after statistical normalization according to the training set; the same vector is fed into the supervised learning classifier in parallel to obtain the class probability vector. Simultaneously, the Mahalanobis distance between the vector and the historical normal characteristic statistics of similar units (i.e., characteristic statistics under normal conditions) is calculated to obtain the group deviation. After normalizing the three types of information, they are fused using a Bayesian approach to obtain the posterior probability distribution of the fault, p(c|data)∝p(c|clf)·p(c|AE)·p(c|stat), and the uncertainty index σ=−Σp(c|data)log p(c|data). Subsequently, p(c|data), σ, along with the changes in rotational speed, the increment of the envelope spectrum amplitude, the change in temperature, the anomaly score, and the deviation atomic predicates, are mapped to trigger strengths via a Sigmoid function. These are then combined into rule evidence contributions using a bounded linear temporal logic rule base. It is then fused with p(c|data) and σ again to output the final diagnostic conclusion and the final uncertainty index, thus completing the reasoning.

[0053] In another embodiment of this solution, S5 is specifically implemented as follows: When the final uncertainty indicator When the preset threshold is exceeded or a high-priority rule is triggered by the rule base, the edge computing node immediately increases the sampling frequency from 1Hz–10Hz to ≥1kHz, re-acquires the complete original waveforms of vibration, acoustics, rotational speed, temperature, current, and voltage, and performs wavelet packet decomposition, envelope demodulation, empirical mode decomposition, and time-frequency analysis again to extract and physically align multi-scale features to obtain new physical feature vectors. Subsequently, the global diagnostic model is used to recalculate the anomaly score, classification probability, and group deviation. Bayesian fusion is used to obtain the secondary fault probability distribution and secondary uncertainty index. Then, secondary inference is performed through the temporal logic rule base. If the two diagnostic conclusions are consistent, the confidence level is increased and the final conclusion is output. If they are inconsistent, they are marked as pending review and an evidence chain structure containing feature contribution, physical residual, rule triggering strength, and consistency is generated.

[0054] Furthermore, variable rate acquisition includes: The multi-source sensor signals at the wind turbine end under normal conditions are sampled and acquired at low frequency; High-frequency sampling is used to collect multi-source sensor signals at the wind turbine end under abnormal conditions.

[0055] Furthermore, S2 specifically refers to: The raw data includes at least vibration signals, acoustic signals, and rotational speed signals; The vibration signal is subjected to wavelet packet decomposition, envelope demodulation and empirical mode decomposition to extract vibration features including sub-band energy, kurtosis and power spectral entropy; The acoustic signal is subjected to a short-time Fourier transform to obtain acoustic features that include time-frequency energy distribution; The spectral features are converted into order features using the rotational speed signal, and the frequency consistency measure and gear meshing frequency consistency measure are calculated. The frequency consistency measure and gear meshing frequency consistency measure are used as rotational speed features. After aligning the vibration features, acoustic features, and rotational speed features according to the time window, normalize them, and then splice them together to obtain the physical feature vector.

[0056] Furthermore, the autoencoder with physical constraints is specifically as follows: A physical consistency loss is introduced into the loss function of the autoencoder, and the physical consistency loss consists of the weighted sum of squared residuals of the frequency estimation variance, so as to force the reconstruction result to conform to the wind turbine dynamics constraints.

[0057] This invention constructs a physical information autoencoder model with integrated physical constraints, enabling feature extraction and anomaly sensitivity scoring of key wind turbine components such as gearboxes, bearings, and rotors, thereby improving the accuracy and interpretability of early fault detection. By introducing a supervised learning classifier and group reference statistics, probabilistic diagnosis and deviation assessment are performed on multi-source sensor data of the unit. Based on this, a rule base based on bounded linear temporal logic is used for temporal logic reasoning, and evidence fusion is achieved by combining rule weights and priorities, realizing joint judgment and uncertainty quantification of multi-source information. Through a dynamic sampling strategy of edge computing nodes and a secondary diagnostic closed-loop mechanism, the diagnostic system can automatically adjust the sampling frequency and verify the diagnostic results when the uncertainty index exceeds a threshold, thereby improving diagnostic confidence and reliability. This effectively improves the accuracy, real-time performance, and interpretability of the wind turbine intelligent diagnostic system, enhances its adaptability to complex fault modes, and provides intelligent and highly reliable decision support for wind farm operation and maintenance.

[0058] In the above embodiments, although the steps are numbered S1, S2, etc., they are only specific embodiments given by the present invention. Those skilled in the art can adjust the execution order of S1, S2, etc. according to the actual situation, and these situations are also within the protection scope of the present invention. It can be understood that in some embodiments, some or all of the above embodiments may be included.

[0059] The present invention also provides a wind turbine intelligent diagnostic system, the specific technical solution of which is as follows: a variable speed acquisition module, a feature extraction module, a model parameter module, a fusion inference module, and a verification and diagnostic module; The variable speed acquisition module is used to acquire signals from multiple sensors at a variable speed at the wind turbine end to obtain raw data; The feature extraction module is used to extract and physically align multi-scale features based on the original data to obtain physical feature vectors; The model parameter module is used to train a physically constrained autoencoder based on physical feature vectors to obtain local model parameters; the local model parameters are then weighted, aggregated, and adapted to the domain to obtain a global diagnostic model. The fusion reasoning module is used to integrate anomaly scores, classification probabilities, and group deviations based on the obtained global diagnostic model to obtain fault probability distribution and uncertainty indicators; based on the fault probability distribution and uncertainty indicators, it uses a time-series logic rule base to reason and obtain diagnostic conclusions. The verification and diagnosis module is used to increase the sampling frequency and repeat the feature extraction module to the fusion inference module when the uncertainty index does not meet the preset threshold, so as to obtain the final diagnosis conclusion after verification.

[0060] Based on the above solution, the present invention can be further improved as follows.

[0061] Furthermore, variable rate acquisition includes: The multi-source sensor signals at the wind turbine end under normal conditions are sampled and acquired at low frequency; High-frequency sampling is used to collect multi-source sensor signals at the wind turbine end under abnormal conditions.

[0062] Furthermore, the feature extraction module specifically includes: The raw data includes at least vibration signals, acoustic signals, and rotational speed signals; The vibration signal is subjected to wavelet packet decomposition, envelope demodulation and empirical mode decomposition to extract vibration features including sub-band energy, kurtosis and power spectral entropy; The acoustic signal is subjected to a short-time Fourier transform to obtain acoustic features that include time-frequency energy distribution; The spectral features are converted into order features using the rotational speed signal, and the frequency consistency measure and gear meshing frequency consistency measure are calculated. The frequency consistency measure and gear meshing frequency consistency measure are used as rotational speed features. After aligning the vibration features, acoustic features, and rotational speed features according to the time window, normalize them, and then splice them together to obtain the physical feature vector.

[0063] Furthermore, the autoencoder with physical constraints is specifically as follows: A physical consistency loss is introduced into the loss function of the autoencoder, and the physical consistency loss consists of the weighted sum of squared residuals of the frequency estimation variance, so as to force the reconstruction result to conform to the wind turbine dynamics constraints.

[0064] It should be noted that the beneficial effects of the intelligent wind turbine diagnostic system provided in the above embodiments are the same as those of the intelligent wind turbine diagnostic method described above, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.

[0065] like Figure 2 As shown, an embodiment of the present invention provides a computer device 300, which includes a processor 320 coupled to a memory 310. The memory 310 stores at least one computer program 330, which is loaded and executed by the processor 320 to enable the computer device 300 to implement any of the above-described methods. Specifically: The computer device 300 can vary considerably due to differences in configuration or performance. It may include one or more processors 320 (Central Processing Units, CPUs) and one or more memories 310. The one or more memories 310 store at least one computer program 330, which is loaded and executed by the one or more processors 320 to enable the computer device 300 to implement the wind turbine intelligent diagnostic method provided in the above embodiments. Of course, the computer device 300 may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input and output. The computer device 300 may also include other components for implementing device functions, which will not be elaborated here.

[0066] An embodiment of the present invention provides a computer-readable storage medium storing at least one computer program, which is loaded and executed by a processor to enable a computer to implement any of the above-described methods.

[0067] Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, a floppy disk, and an optical data storage device, etc.

[0068] In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform any of the aforementioned wind turbine intelligent diagnostic methods.

[0069] It should be noted that the terms "first," "second," etc., used in the specification of this application are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown in the figures or description.

[0070] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this disclosure can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, the invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0071] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0072] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for intelligent diagnosis of wind turbines, characterized in that, include: S1, at the wind turbine end, acquires raw data by performing variable-rate acquisition of signals from multiple sources of sensors; S2, Based on the original data, extract and physically align multi-scale features to obtain a physical feature vector; S3, Train an autoencoder with physical constraints based on the physical feature vectors to obtain local model parameters; Perform weighted aggregation and domain adaptation on the local model parameters to obtain a global diagnostic model; S4. Based on the obtained global diagnostic model, the abnormal score, classification probability, and group deviation are integrated to obtain the fault probability distribution and uncertainty index; based on the fault probability distribution and the uncertainty index, the diagnostic conclusion is obtained through reasoning using the temporal logic rule base. S5. When the uncertainty index does not meet the preset threshold, increase the sampling frequency and repeat S2-S4 to obtain the final diagnostic conclusion after verification.

2. The intelligent diagnostic method for wind turbines according to claim 1, characterized in that, The variable rate acquisition includes: The multi-source sensor signals at the wind turbine end under normal conditions are sampled and acquired at low frequency; High-frequency sampling is used to collect multi-source sensor signals at the wind turbine end under abnormal conditions.

3. The intelligent diagnostic method for wind turbines according to claim 1, characterized in that, Specifically, S2 is: The raw data includes at least vibration signals, acoustic signals, and rotational speed signals; The vibration signal is subjected to wavelet packet decomposition, envelope demodulation, and empirical mode decomposition to extract vibration features including sub-band energy, kurtosis, and power spectral entropy. Perform a short-time Fourier transform on the acoustic signal to obtain acoustic features that include time-frequency energy distribution; The spectral features are converted into order features using the rotational speed signal, and the frequency consistency measure and gear meshing frequency consistency measure are calculated. The frequency consistency measure and the gear meshing frequency consistency measure are used as rotational speed features. The vibration features, acoustic features, and rotational speed features are aligned according to a time window, normalized, and then concatenated to obtain the physical feature vector.

4. The intelligent diagnostic method for wind turbines according to claim 1, characterized in that, The physically constrained autoencoder specifically refers to: A physical consistency loss is introduced into the loss function of the autoencoder, and the physical consistency loss consists of the weighted sum of squared residuals of the frequency estimation variance, so as to force the reconstruction result to conform to the wind turbine dynamics constraints.

5. A wind turbine intelligent diagnostic system, characterized in that, include: The system includes a variable-speed acquisition module, a feature extraction module, a model parameter module, a fusion inference module, and a verification and diagnosis module. The variable speed acquisition module is used to acquire signals from multiple sources at a variable speed at the wind turbine end to obtain raw data; The feature extraction module is used to extract and physically align multi-scale features based on the original data to obtain a physical feature vector; The model parameter module is used to train a physically constrained autoencoder based on the physical feature vector to obtain local model parameters; and to obtain a global diagnostic model by weighted aggregation and domain adaptation of the local model parameters. The fusion reasoning module is used to fuse anomaly scores, classification probabilities, and group deviations based on the obtained global diagnostic model to obtain a fault probability distribution and uncertainty index; and to obtain a diagnostic conclusion based on the fault probability distribution and the uncertainty index through reasoning using a time-series logic rule base. The verification and diagnosis module is used to increase the sampling frequency and repeat the feature extraction module to the fusion inference module when the uncertainty index does not meet the preset threshold, so as to obtain the final diagnosis conclusion after verification.

6. The intelligent diagnostic system for wind turbines according to claim 5, characterized in that, The variable rate acquisition includes: The multi-source sensor signals at the wind turbine end under normal conditions are sampled and acquired at low frequency; High-frequency sampling is used to collect multi-source sensor signals at the wind turbine end under abnormal conditions.

7. The intelligent diagnostic system for wind turbines according to claim 5, characterized in that, The feature extraction module is specifically as follows: The raw data includes at least vibration signals, acoustic signals, and rotational speed signals; The vibration signal is subjected to wavelet packet decomposition, envelope demodulation, and empirical mode decomposition to extract vibration features including sub-band energy, kurtosis, and power spectral entropy. Perform a short-time Fourier transform on the acoustic signal to obtain acoustic features that include time-frequency energy distribution; The spectral features are converted into order features using the rotational speed signal, and the frequency consistency measure and gear meshing frequency consistency measure are calculated. The frequency consistency measure and the gear meshing frequency consistency measure are used as rotational speed features. The vibration features, acoustic features, and rotational speed features are aligned according to a time window, normalized, and then concatenated to obtain the physical feature vector.

8. The intelligent diagnostic system for wind turbines according to claim 5, characterized in that, The physically constrained autoencoder specifically refers to: A physical consistency loss is introduced into the loss function of the autoencoder, and the physical consistency loss consists of the weighted sum of squared residuals of the frequency estimation variance, so as to force the reconstruction result to conform to the wind turbine dynamics constraints.

9. A computer device, characterized in that, The computer device includes a processor coupled to a memory storing at least one computer program, which is loaded and executed by the processor to enable the computer device to perform the method as claimed in any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to enable the computer to perform the method as described in any one of claims 1 to 4.

Images