Method and system for evaluating service life of converter of wind turbine generator based on multiple scales

Abstract

The invention relates to the field of converter service life evaluation, and particularly discloses a multi-scale-based wind turbine generator converter service life evaluation method and system. High-frequency instantaneous junction temperature data and long-period average junction temperature data of a wind turbine generator converter are synchronously processed, and a dual-channel depth time sequence coding architecture is introduced; and carrying out deep deconstruction on the junction temperature sequences of two different time scales. And through a cross-scale interactive fusion mechanism, a nonlinear modulation effect of long-scale average junction temperature on short-scale damage accumulation is captured and quantified to obtain a deep fusion vector. And finally, based on the deep fusion vector driving life prediction model, precise evaluation of the life consumption rate under the current working condition is realized, dynamic iteration is carried out in combination with historical damage, and online prediction of the residual service life of the converter is completed. Therefore, the limitation of a traditional linear accumulation criterion can be effectively overcome, and the accuracy and reliability of wind turbine generator converter life prediction are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of converter life evaluation, and more particularly, to a wind turbine converter life evaluation method and system based on multi-scale. BACKGROUND

[0002] With the continuous advancement of the dual-carbon target, the new energy industry represented by wind power is entering a period of rapid development. Wind turbine, as the core equipment of wind farm, its reliability and stability of operation is directly related to the safety and economic benefits of the entire power grid. In the wind turbine, the converter is the key hub to realize the efficient conversion of wind energy to electric energy, but the internal power semiconductor devices (such as IGBT modules) bear complex electro-thermal stress in long-term operation, and are one of the weakest links prone to failure in the entire system. The main reason for device failure is the fluctuation of junction temperature, and the randomness of wind speed, the disturbance of power grid and the dynamic adjustment of control strategy make the converter junction temperature present complex fluctuation characteristics on different time scales: both the high-frequency, small-amplitude temperature rise fluctuations caused by millisecond switching frequency, and the low-frequency, large-amplitude average temperature drift caused by wind conditions and scheduling instruction changes. Therefore, in order to realize the accurate prediction of the remaining useful life of the converter and avoid the huge loss caused by unplanned downtime, a life evaluation model that can fully consider the influence of multi-scale thermal stress must be built.

[0003] However, in the prior art, the traditional life evaluation method usually decouples the thermal stress cycles of different scales and uses linear damage accumulation criteria (such as Miner's rule) for simple superposition calculation. The inherent assumption of this method is that the damage caused by each thermal stress cycle is independent of each other. But in physical reality, the damage accumulation of power devices is a highly complex nonlinear process. The average junction temperature level caused by the change of working conditions will significantly affect the fatigue characteristics of the material, thus modulating and accelerating the damage accumulation rate caused by the short time scale junction temperature fluctuation caused by high-frequency switching action. That is, there is a strong coupling effect between thermal stresses at different time scales, and simple linear superposition cannot accurately describe the nonlinear acceleration phenomenon of this damage.

[0004] Therefore, an optimized wind turbine converter life evaluation method and system are expected. SUMMARY

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a multi-scale wind turbine converter life assessment method and system. It synchronously processes high-frequency instantaneous junction temperature data and long-period average junction temperature data of the wind turbine converter, and introduces a dual-channel deep time-series coding architecture to deeply deconstruct the junction temperature sequences of the two different time scales. Furthermore, through a cross-scale interactive fusion mechanism, it captures and quantifies the nonlinear modulation effect of the long-scale average junction temperature on the accumulation of short-scale damage to obtain a deep fusion vector. Finally, based on this deep fusion vector, it drives a life prediction model to achieve an accurate assessment of the lifespan consumption rate under current operating conditions, and combines historical damage for dynamic iteration to complete the online prediction of the remaining service life of the converter. This effectively overcomes the limitations of traditional linear accumulation criteria and significantly improves the accuracy and reliability of wind turbine converter life prediction.

[0006] According to one aspect of this application, a multi-scale wind turbine converter life assessment method is provided, comprising: Obtain high-frequency junction temperature sequences and average junction temperature sequences based on SCADA cycles; Dual-channel parallel multi-scale feature encoding is performed on the high-frequency junction temperature sequence and the average junction temperature sequence based on SCADA cycle to obtain the encoding representation vector of the short-scale sequence and the encoding representation vector of the long-scale sequence. Cross-scale fusion of the encoding representation vectors of short-scale sequences and long-scale sequences is performed to obtain a multi-scale junction temperature fused encoding representation vector; The lifetime attrition rate is predicted based on the multi-scale junction temperature fusion coding representation vector to obtain the lifetime attrition rate predicted for the current window. The remaining useful life at the current moment is determined based on the cumulative damage at the previous moment and the life consumption rate predicted in the current window.

[0007] According to another aspect of this application, a multi-scale wind turbine converter life assessment system is provided, comprising: The junction temperature sequence acquisition module is used to acquire high-frequency junction temperature sequences and average junction temperature sequences based on SCADA cycles. A dual-channel parallel feature encoding module is used to perform dual-channel parallel multi-scale feature encoding on high-frequency junction temperature sequences and average junction temperature sequences based on SCADA cycles to obtain the encoding representation vectors of short-scale sequences and long-scale sequences. The cross-scale fusion module is used to perform cross-scale fusion of the encoding representation vectors of short-scale sequences and long-scale sequences to obtain a multi-scale junction temperature fused encoding representation vector. The lifetime attrition rate prediction module is used to predict the lifetime attrition rate based on the multi-scale junction temperature fusion encoded representation vector to obtain the lifetime attrition rate predicted for the current window. The remaining useful life calculation module is used to determine the remaining useful life at the current moment based on the cumulative damage at the previous moment and the life consumption rate predicted in the current window.

[0008] Compared with existing technologies, the multi-scale wind turbine converter life assessment method and system provided in this application simultaneously processes high-frequency instantaneous junction temperature data and long-period average junction temperature data of the wind turbine converter, and introduces a dual-channel deep time-series coding architecture to deeply deconstruct the junction temperature sequences of the two different time scales. Then, through a cross-scale interactive fusion mechanism, it captures and quantifies the nonlinear modulation effect of long-scale average junction temperature on short-scale damage accumulation to obtain a deep fusion vector. Finally, based on this deep fusion vector, a life prediction model is driven to achieve accurate assessment of the life consumption rate under current operating conditions, and dynamic iteration is performed by combining historical damage data to complete the online prediction of the remaining service life of the converter. This effectively overcomes the limitations of traditional linear accumulation criteria and significantly improves the accuracy and reliability of wind turbine converter life prediction. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 A flowchart of a multi-scale wind turbine converter life assessment method according to an embodiment of this application; Figure 2 This is a data flow diagram of the multi-scale wind turbine converter life assessment method according to an embodiment of this application; Figure 3 This is a flowchart of sub-step S2 of the multi-scale wind turbine converter life assessment method according to an embodiment of this application; Figure 4 This is a flowchart of sub-step S3 of the multi-scale wind turbine converter life assessment method according to an embodiment of this application; Figure 5 This is a flowchart of sub-step S34 of the multi-scale wind turbine converter life assessment method according to an embodiment of this application; Figure 6 This is a block diagram of a multi-scale wind turbine converter life assessment system according to an embodiment of this application. Detailed Implementation

[0011] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0012] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0013] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0014] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0015] It is worth noting that all data acquisition actions in this application were carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization granted by the owner of the relevant device.

[0016] To address the technical problems described in the background, this application proposes a multi-scale wind turbine converter life assessment method. This method simultaneously processes high-frequency instantaneous junction temperature data and long-period average junction temperature data of the wind turbine converter, and introduces a dual-channel deep time-series coding architecture to deeply deconstruct the junction temperature sequences of the two different time scales. Furthermore, through a cross-scale interactive fusion mechanism, it captures and quantifies the nonlinear modulation effect of the long-scale average junction temperature on the accumulation of short-scale damage to obtain a deep fusion vector. Finally, based on this deep fusion vector, a life prediction model is driven to achieve an accurate assessment of the lifespan consumption rate under current operating conditions. Combined with historical damage data for dynamic iteration, it completes the online prediction of the remaining service life of the converter. This effectively overcomes the limitations of traditional linear accumulation criteria and significantly improves the accuracy and reliability of wind turbine converter life prediction.

[0017] Figure 1 This is a flowchart of a multi-scale wind turbine converter life assessment method according to an embodiment of this application.Figure 2 This is a data flow diagram of a multi-scale wind turbine converter life assessment method according to an embodiment of this application. Figure 1 and Figure 2 As shown, the multi-scale wind turbine converter life assessment method includes the following steps: S1, acquiring a high-frequency junction temperature sequence and an average junction temperature sequence based on the SCADA cycle; S2, performing dual-channel parallel multi-scale feature encoding on the high-frequency junction temperature sequence and the average junction temperature sequence based on the SCADA cycle to obtain the encoding representation vector of the short-scale sequence and the encoding representation vector of the long-scale sequence; S3, performing cross-scale fusion on the encoding representation vector of the short-scale sequence and the encoding representation vector of the long-scale sequence to obtain a multi-scale junction temperature fusion encoding representation vector; S4, predicting the life consumption rate based on the multi-scale junction temperature fusion encoding representation vector to obtain the life consumption rate predicted for the current window; S5, determining the remaining service life at the current moment based on the cumulative damage at the previous moment and the life consumption rate predicted for the current window.

[0018] In the aforementioned multi-scale wind turbine converter life assessment method, step S1 involves acquiring a high-frequency junction temperature sequence and an average junction temperature sequence based on the SCADA cycle. It should be understood that the core driving factor for the failure of power semiconductor devices (such as IGBT modules) in wind turbine converters is multi-scale thermal stress cycling. High-frequency instantaneous junction temperature fluctuations (milliseconds to seconds) are directly related to short-scale power cycle damage, while long-cycle average junction temperature (minutes to hours) determines the modulation effect of medium- to long-scale operating conditions on damage accumulation. Therefore, this application simultaneously collects and acquires both the high-frequency junction temperature sequence and the average junction temperature sequence based on the SCADA cycle, providing complete raw data input for subsequent dual-channel multi-scale feature encoding, ensuring that thermal stress information at both time scales is included in the assessment system. This avoids the loss of key information caused by single-scale data, accurately covers the multi-scale characteristics of thermal stress, lays a data foundation for subsequently capturing cross-scale nonlinear coupling effects, and thus improves the accuracy and reliability of converter life assessment.

[0019] In the specific implementation process, the first step is to acquire the high-frequency junction temperature sequence. A high-precision thermocouple temperature sensor is integrated inside the power semiconductor device package of the converter. The sensor sampling frequency is set to 5kHz to collect instantaneous junction temperature data during device operation in real time. The collected instantaneous junction temperature data is then arranged continuously in timestamp order to form a high-frequency junction temperature sequence. The second step is to acquire the average junction temperature sequence based on the SCADA cycle. Temperature monitoring data corresponding to the converter is extracted from the wind turbine SCADA system, with the SCADA system sampling cycle set to 10 minutes. The arithmetic mean of all temperature monitoring data within each 10-minute sampling cycle is calculated to obtain the average junction temperature value for that cycle. The average junction temperature values ​​from each 10-minute cycle are then concatenated in chronological order to form the average junction temperature sequence based on the SCADA cycle.

[0020] In the aforementioned multi-scale wind turbine converter life assessment method, step S2 involves performing dual-channel parallel multi-scale feature encoding on the high-frequency junction temperature sequence and the average junction temperature sequence based on the SCADA cycle to obtain the encoding representation vectors for the short-scale sequence and the long-scale sequence. It should be understood that since the high-frequency junction temperature sequence and the average junction temperature sequence based on the SCADA cycle belong to different time scales, a single encoding method cannot adapt to the structural differences between the two types of data, easily leading to scale feature confusion or loss of key information. Therefore, this application further employs a dual-channel parallel architecture to perform multi-scale feature encoding on the two types of sequences separately, thereby specifically extracting short-scale instantaneous fluctuation features and long-scale trend features, obtaining the encoding representation vectors for the short-scale sequence and the long-scale sequence respectively. This avoids the limitations of single-channel encoding and fully preserves the core information of the junction temperature data at different scales. Figure 3 This is a flowchart of sub-step S2 of the multi-scale wind turbine converter life assessment method according to an embodiment of this application. Figure 3 As shown, step S2 includes the following steps: S21, inputting the high-frequency junction temperature sequence into the short-scale channel to obtain the encoded representation vector of the short-scale sequence; S22, inputting the average junction temperature sequence based on the SCADA cycle into the long-scale channel to obtain the encoded representation vector of the long-scale sequence.

[0021] Specifically, in step S21, the high-frequency junction temperature sequence is input into the short-scale channel to obtain the encoded representation vector of the short-scale sequence. It should be understood that since the high-frequency junction temperature sequence contains instantaneous thermal stress fluctuations at the millisecond to second level, these fluctuations directly correspond to the short-scale power cycling of power semiconductor devices and are a major cause of fatigue damage to the device's bond wires. Therefore, this application further inputs the high-frequency junction temperature sequence into the short-scale channel. In a specific example of this application, the short-scale channel uses a Transformer encoder. Utilizing the multi-head self-attention mechanism of the Transformer encoder, the dependency relationship of different instantaneous junction temperature fluctuations within the high-frequency sequence is accurately mined, and features directly related to the accumulation of short-scale damage are extracted to obtain the encoded representation vector of the short-scale sequence. This fully preserves the fine-grained fluctuation information and instantaneous correlation features of the high-frequency junction temperature sequence, accurately characterizing the direct impact of short-scale thermal stress on device damage. This provides accurate feature support for quantifying the short-scale damage effect during subsequent cross-scale fusion, avoiding damage assessment bias caused by insufficient extraction of short-scale features.

[0022] In the specific implementation process, the high-frequency junction temperature sequence is first preprocessed. An outlier detection algorithm is used to remove abnormal data points caused by instantaneous sensor interference. Then, the sequence is reordered according to timestamps to ensure temporal continuity. The preprocessed high-frequency junction temperature sequence is then input into a Transformer encoder in a short-scale channel. An embedding layer maps the junction temperature values ​​at each time step in the sequence into a high-dimensional feature vector, while positional encoding is added to preserve temporal order. Next, two multi-head self-attention layers, each with four attention heads, are used to mine the correlation features of junction temperature fluctuations at different time steps within the sequence. Each self-attention layer is followed by a layer normalization module and a residual connection module to stabilize gradient propagation during the encoding process. Finally, a global average pooling module aggregates the encoded features of all time steps to generate a fixed-dimensional short-scale sequence encoding representation vector, which is then transmitted to the cross-scale fusion module.

[0023] Specifically, in step S22, the average junction temperature sequence based on the SCADA cycle is input into the long-scale channel to obtain the encoded representation vector of the long-scale sequence. It should be understood that since the average junction temperature sequence based on the SCADA cycle exhibits a long-term trend ranging from minutes to hours, this trend reflects the medium-to-long-scale thermal stress background under changes in wind turbine operating conditions (such as wind speed and load). The magnitude of this trend directly affects the fatigue characteristics of power semiconductor device materials, thereby modulating the accumulation rate of short-scale damage. Therefore, this application further inputs the average junction temperature sequence into the long-scale channel. In a specific example of this application, the long-scale channel uses a Long Short-Term Memory (LSTM) network. Utilizing the collaborative mechanism of the LSM network's input gate, forget gate, and output gate, the long-term trend dependency of the average junction temperature sequence is accurately captured, and features related to the long-scale operating condition background are extracted to obtain the encoded representation vector of the long-scale sequence. This fully preserves the long-term trend information of the average junction temperature, accurately reflects the modulation effect of medium-to-long-scale thermal stress on device damage, and provides reliable feature support for quantifying the nonlinear influence of long-scale on short-scale during subsequent cross-scale fusion, ensuring that the lifetime assessment model can accurately identify damage acceleration factors at the long-scale level.

[0024] In the specific implementation process, the average junction temperature sequence based on the SCADA cycle is first preprocessed. A linear interpolation algorithm is used to fill in data gaps caused by SCADA system communication interruptions, and a sliding window smoothing algorithm is used to reduce the interference of short-term random fluctuations on trend features. The preprocessed average junction temperature sequence is then input into a long short-term memory network (LSTM) for long-scale channels. This network is configured with a single hidden layer, the number of hidden units is set to adapt to the scale of long-cycle data processing, the input feature dimension is set to 1 (corresponding only to the average junction temperature value), and the tanh function is used as the activation function. The network processes the sequence data sequentially according to the SCADA cycle time steps, filtering and retaining key long-term trend information through a forget gate, updating the trend features of the current time step through an input gate, and outputting the effective features of the current time step through an output gate. After all sequences have been processed, the hidden state of the last time step is extracted. This hidden state is the encoded representation vector of the long-scale sequence and is transmitted to the cross-scale fusion module.

[0025] In the aforementioned multi-scale-based wind turbine converter life assessment method, step S3 involves cross-scale fusion of the encoding representation vectors of the short-scale sequence and the long-scale sequence to obtain a multi-scale junction temperature fusion encoding representation vector. It should be understood that since the encoding representation vector of the short-scale sequence only carries the characteristics of short-scale thermal stress fluctuations, and the encoding representation vector of the long-scale sequence only reflects the characteristics of long-scale thermal stress trends, their independent existence cannot reflect the nonlinear modulation effect of the long-scale average junction temperature on the accumulation of short-scale damage. Therefore, this application further performs cross-scale fusion processing on the encoding representation vectors of the short-scale sequence and the long-scale sequence. By integrating multi-scale features and capturing their coupling relationships, a multi-scale junction temperature fusion encoding representation vector containing multi-scale collaborative information is generated. This effectively avoids the limitations of single-scale features, fully characterizes the modulation mechanism of long-scale background on short-scale damage, provides comprehensive and accurate feature input for subsequent lifespan rate prediction, and significantly improves the accuracy and reliability of converter remaining service life assessment. Figure 4 This is a flowchart of sub-step S3 of the multi-scale wind turbine converter life assessment method according to an embodiment of this application. Figure 4 As shown, step S3 includes the following steps: S31, performing linear projection of multi-scale features onto the encoded representation vectors of the short-scale sequence and the long-scale sequence to obtain a long-scale junction temperature query vector, a short-scale junction temperature key vector, and a short-scale junction temperature value vector; S32, calculating the correlation between the long-scale junction temperature query vector and the short-scale junction temperature key vector to obtain an attention weight matrix; S33, performing matrix multiplication of the attention weight matrix and the short-scale junction temperature value vector to obtain a cross-scale scenario correlation vector; S34, fusing the cross-scale scenario correlation vector and the encoded representation vector of the long-scale sequence to obtain the multi-scale junction temperature fusion encoded representation vector.

[0026] Specifically, in step S31, a linear projection of multi-scale features is performed on the encoded representation vectors of the short-scale sequence and the long-scale sequence to obtain a long-scale junction temperature query vector, a short-scale junction temperature key vector, and a short-scale junction temperature value vector. It should be understood that since the encoded representation vectors of the short-scale sequence and the long-scale sequence may reside in different feature spaces, their dimensions and semantic attributes differ, and directly calculating the correlation degree can easily lead to bias. Therefore, this application first performs a linear projection of multi-scale features on the encoded representation vectors of the short-scale sequence and the long-scale sequence, mapping the long-scale features to a query vector and the short-scale features to key and value vectors. This eliminates the spatial differences between features of different scales, ensuring that the subsequent attention mechanism can effectively mine the coupling relationship of multi-scale features and improve the effectiveness of cross-scale fusion.

[0027] In the specific implementation, three independent linear projection layers are first constructed, corresponding to the generation of query vector, key vector, and value vector, respectively. Each projection layer is configured with a learnable weight matrix and bias term. Then, the encoded representation vector of the long-scale sequence is input into the first linear projection layer, and through matrix multiplication and bias addition, it is mapped to a long-scale junction temperature query vector. Simultaneously, the encoded representation vector of the short-scale sequence is input into the other two linear projection layers, one outputting a short-scale junction temperature key vector and the other outputting a short-scale junction temperature value vector. During the projection process, layer normalization is used to stabilize the numerical distribution of the output vectors, ensuring that the three vectors are in the same feature space. Finally, all three are transmitted to the correlation calculation module.

[0028] Specifically, step S32 involves calculating the correlation between the long-scale junction temperature query vector and the short-scale junction temperature bond vector to obtain an attention weight matrix. It should be understood that, due to differences in the contribution of short-scale thermal stress fluctuations to damage accumulation under different long-scale thermal stress backgrounds, not all short-scale features have the same impact on current lifetime consumption. Ignoring this difference could lead to the obscuring of crucial damage information. Therefore, in a specific example of this application, step S32 includes: calculating the correlation between the long-scale junction temperature query vector and the short-scale junction temperature bond vector using the following formula:

[0029]

[0030] in, This is a long-scale junction temperature query vector. For short-scale junction temperature bond vectors, This is the transpose of the vector. Let be the dimension of the short-scale junction temperature bond vector. The original attention score. For normalization function, This is the attention weight matrix. It quantifies the correlation strength between the long-scale background and each short-scale feature, generating an attention weight matrix that reflects the importance of short-scale features. This allows for the accurate identification of short-scale features that play a dominant role in damage accumulation against the long-scale background. By assigning weights, key information is highlighted and irrelevant noise is suppressed, providing a basis for subsequent weighted extraction of effective short-scale features and improving the targeting of cross-scale fusion.

[0031] Specifically, in step S33, the attention weight matrix is ​​multiplied by the short-scale junction temperature vector to obtain a cross-scale scenario association vector. It should be understood that since the attention weight matrix quantifies the importance of each short-scale feature, and the short-scale junction temperature vector contains the core information of the short-scale features, the importance difference cannot be translated into actual feature selection solely through the weight matrix. Therefore, this application further multiplies the attention weight matrix by the short-scale junction temperature vector to weight and filter the information of the short-scale value vector according to the weights, retaining important short-scale feature information and weakening secondary information. In this way, a cross-scale scenario association vector that integrates long-scale background attention preferences can be generated, allowing short-scale feature information to be initially combined with long-scale modulation effects, providing high-quality short-scale side feature input for the subsequent final fusion with the long-scale encoded vector.

[0032] Specifically, step S34 involves fusing the cross-scale scenario correlation vector and the encoded representation vector of the long-scale sequence to obtain the multi-scale junction temperature fused encoded representation vector. It should be understood that since the cross-scale scenario correlation vector only contains weighted short-scale feature information, and the encoded representation vector of the long-scale sequence only retains long-scale trend information, neither vector alone can fully encompass the coupling effect of multi-scale thermal stress, which is the core basis for accurately assessing the converter's lifespan consumption rate. Therefore, this application further fuses the cross-scale scenario correlation vector and the encoded representation vector of the long-scale sequence to integrate short-scale weighted information and long-scale trend information, forming a fused encoded vector containing multi-scale interactive coupling features. This allows the fused vector to simultaneously carry the modulation effect of the long-scale on the short-scale and the trend characteristics of the long-scale itself, providing comprehensive and accurate feature support for subsequent lifespan consumption rate prediction models and effectively improving the accuracy of converter remaining service life prediction. Figure 5 This is a flowchart of sub-step S34 of the multi-scale wind turbine converter life assessment method according to an embodiment of this application. Figure 5As shown, step S34 includes the following steps: S341, extracting local feature domains based on one-dimensional convolutional coding from the cross-scale scenario association vector and the long-scale sequence encoding representation vector to obtain a set of local feature vectors for cross-scale scenario association and a set of local granular encoding vectors for long-scale sequences; S342, calculating the cost matrix between the set of local feature vectors for cross-scale scenario association and the set of local granular encoding vectors for long-scale sequences to obtain a scenario association cost matrix; S343, determining the optimal transmission planning matrix for scenario association based on the scenario association cost matrix; S344, performing latent spatial distribution matching on the set of local granular encoding vectors for long-scale sequences based on the optimal transmission planning matrix for scenario association to obtain a set of local granular encoding vectors after manifold matching for long-scale sequences; S345, performing local feature unit co-modulation on the set of local feature vectors for cross-scale scenario association and the set of local granular encoding vectors after manifold matching for long-scale sequences to obtain a multi-scale junction-temperature fusion encoding representation vector.

[0033] More specifically, step S341 involves extracting local feature domains based on one-dimensional convolutional coding from the cross-scale scenario association vector and the encoded representation vector of the long-scale sequence to obtain a set of local feature vectors for cross-scale scenario association and a set of local granular encoded vectors for the long-scale sequence, expressed by the formula:

[0034]

[0035] in, and These are the encoded representation vectors of the cross-scale scenario association vector and the long-scale sequence, respectively. This is a one-dimensional convolution operation. The radius of the convolution kernel, i.e., the local neighborhood range. Represented by feature points Centered Neighborhood, and These are, respectively, a set of local feature vectors for cross-scale context association and a set of local granular encoding vectors for long-scale sequences. and These are the first local feature vectors in the set of cross-scale scenario associations. The set of local feature vectors associated with cross-scale scenarios and the local granular encoding vectors of long-scale sequences, and the first... A long-scale sequence local granularity encoding vector The number of vectors in the set of local feature vectors associated across scale scenarios. The number of vectors in the set of local granularity encoded vectors for long-scale sequences.

[0036] In other words, this application uses a one-dimensional convolution operation to determine the local neighborhood based on the convolution kernel radius. and Local feature extraction is performed on feature points in the cross-scale scenario association vector and feature points in the long-scale sequence encoding representation vector, respectively, to generate a set of local feature vectors for cross-scale scenario association and a set of local granular encoding vectors for long-scale sequences. This decouples the global features of cross-scale scenario association semantics and long-scale sequence encoding semantics into physically meaningful local feature units, providing a foundation for quantifying the association strength of cross-scale local features. This ensures that multi-scale coupling effects are accurately captured at the local level, improving the accuracy of feature representation of damage accumulation patterns. For example, when performing latent spatial distribution matching on long-scale sequence local granular encoding vectors, structured local vectors allow the matching process to more accurately reflect the actual coupling patterns of multi-scale thermal stress in converters, thereby improving the accuracy of the multi-scale junction temperature fusion encoding representation vector in representing the multi-scale thermal stress coupling effect of converters and providing more reliable feature support for subsequent lifetime loss rate prediction.

[0037] More specifically, step S342 involves calculating the cost matrix between the set of local feature vectors for cross-scale scenario association and the set of local granular encoding vectors for long-scale sequences to obtain the scenario association cost matrix, expressed by the formula:

[0038] in, Let the first norm of the vector be 1. for function, for and Hyperbolic space distance between them In the context-related cost matrix The element at a given position.

[0039] In other words, hyperbolic spatial distance is further used to calculate the cost matrix between the set of local feature vectors associated with cross-scale scenarios and the set of local granular encoding vectors of long-scale sequences. This quantifies the similarity difference between any two local feature vectors from different sets, generating a scenario association cost matrix. This clearly presents the matching cost between local features at different scales. For example, when a long-scale sequence local granular encoding vector reflects the stable trend of the average junction temperature of a wind turbine converter due to stable wind speed over a certain period, the scenario association cost matrix can accurately determine the matching cost between this vector and different high-frequency junction temperature fluctuation features (such as instantaneous junction temperature changes caused by IGBT module switching) in the set of local feature vectors associated with cross-scale scenarios. Simultaneously, the scenario association cost matrix also provides crucial quantitative basis for subsequently determining the optimal transmission planning matrix for scenario association, ensuring that the potential spatial distribution matching of subsequent long-scale features accurately matches the actual coupling law of multi-scale thermal stress. This improves the accuracy of the multi-scale junction temperature fusion encoding representation vector in representing the factors affecting converter lifespan consumption, providing more reliable feature support for subsequent lifespan consumption rate prediction.

[0040] More specifically, step S343, based on the scenario association cost matrix, determines the scenario association optimal transmission planning matrix, expressed by the formula:

[0041]

[0042] in, For entropy regularization, For trainable weights, for middle The element at position, For context-aware transport mapping matrix, To obtain the minimum value , The optimal transmission planning matrix is ​​associated with the scenario.

[0043] That is, further optimizing the objectives, under constraints Under conditions where ≥0 and the sum of each row is 1, the scenario-related optimal transmission planning matrix that minimizes the objective is solved to determine the scenario-related optimal transmission planning matrix. This enables globally optimal matching of multi-scale local features. For example, when the local granularity encoding vector of a long-scale sequence shows a slow upward trend in the average junction temperature of the converter under a certain wind condition, the scenario-related optimal transmission planning matrix can accurately describe how to match the distribution of short-scale high-frequency junction temperature fluctuation features with the distribution of this long-scale trend feature at the lowest cost. This allows the multi-scale junction temperature fusion encoding representation vector to more accurately reflect the nonlinear modulation effect of long-scale thermal stress on short-scale damage accumulation, providing more realistic feature inputs for subsequent lifetime consumption rate prediction.

[0044] More specifically, step S344 involves performing latent spatial distribution matching on the set of local granularity coding vectors for the long-scale sequence based on the scenario-related optimal transmission planning matrix, to obtain the set of local granularity coding vectors after manifold matching for the long-scale sequence, expressed by the formula:

[0045]

[0046] in, for and Parallel transmission operators between them for middle The element at position, The first in the set of local granularity encoded vectors after manifold matching of long-scale sequences Local granularity encoded vectors after matching a long-scale sequence manifold.

[0047] In other words, based on the scenario-related optimal transmission planning matrix, the set of local granularity coding vectors for long-scale sequences is adjusted using parallel transmission operators to make the latent spatial distribution of long-scale features consistent with that of short-scale features. This allows long-scale features to contain contextual information aligned with the deep structure of short-scale features. For example, if the local granularity coding vector of a long-scale sequence originally reflects the slow decreasing trend of the converter's average junction temperature under a certain wind condition, and the set of local feature vectors related to cross-scale scenario correlation corresponds to high-frequency junction temperature pulses generated by grid disturbances during the same period, after latent spatial distribution matching, the long-scale vectors will incorporate the correlation information of the high-frequency pulses, and their distribution will be aligned with the short-scale feature space. Simultaneously, the local granularity coding vectors after long-scale sequence manifold matching can more accurately reflect the modulation effect on the accumulation of short-scale damage. For example, in subsequent local feature co-modulation, the mitigation effect of the average junction temperature decrease on high-frequency pulse damage can be more accurately quantified, thereby improving the representation accuracy of the multi-scale junction temperature fusion coding representation vector.

[0048] More specifically, step S345 involves performing local feature unit co-modulation on the set of local feature vectors associated with the cross-scale scenario and the set of local granularity coding vectors after long-scale sequence manifold matching to obtain a multi-scale junction-temperature fusion coding representation vector, expressed by the formula:

[0049]

[0050] in, For vector concatenation, The first in the set of local granularity encoded vectors after manifold matching of long-scale sequences Local granular encoding vector after matching a long-scale sequence manifold. For trainable weight matrix, for function, These are the 1st, 2nd, and 3rd elements in the set of multi-scale junction temperature fusion feature interactive encoding vectors. The and the first A multi-scale junction temperature fusion feature interactive encoding vector, Forward Encoder, This is a multi-scale junction temperature fusion encoding representation vector.

[0051] In other words, the matched local feature pairs are further concatenated, and then nonlinear modulation is performed using tanh activation and weight matrices to obtain a set of multi-scale junction temperature fusion feature interactive coding vectors. Finally, this set of multi-scale junction temperature fusion feature interactive coding vectors is input into a forward LSTM encoder to obtain a multi-scale junction temperature fusion coding representation vector. This achieves the collaborative modulation of local feature units and generates a multi-scale junction temperature fusion coding representation vector. This allows for deep fusion of multi-scale features at the local unit level, enabling the multi-scale junction temperature fusion coding representation vector to simultaneously carry short-scale damage features and long-scale modulation features. This provides crucial feature support for subsequent lifetime attrition rate prediction. For example, when input into a regression head, it can more accurately output the lifetime attrition rate under the current operating conditions, thereby improving the accuracy of converter remaining service life prediction and providing a reliable basis for predictive maintenance of wind turbine converters.

[0052] In the aforementioned multi-scale wind turbine converter life assessment method, step S4 involves predicting the lifespan attrition rate based on the multi-scale junction temperature fusion encoding representation vector to obtain the predicted lifespan attrition rate for the current window. In a specific example of this application, step S4 includes: inputting the multi-scale junction temperature fusion encoding representation vector into a regression head composed of one or more fully connected layers to obtain the predicted lifespan attrition rate for the current window. It should be understood that although the multi-scale junction temperature fusion encoding representation vector integrates key features of junction temperature at different time dimensions, these features are still in an abstract vector space and cannot be directly correlated with the specific equipment state quantification indicator, lifespan attrition rate. Therefore, this application further inputs the multi-scale junction temperature fusion encoding representation vector into a regression head composed of one or more fully connected layers to leverage the parameter learning capability of the fully connected layers, thereby mining the potential correlation between each feature dimension in the fusion vector and the lifespan attrition rate, completing the transformation from an abstract feature vector to a specific prediction indicator, and thus generating the lifespan attrition rate prediction result corresponding to the current window. This approach fully activates the predictive value of multi-scale junction temperature characteristics, avoids prediction bias caused by the disconnect between characteristics and target indicators, improves the accuracy and stability of lifespan consumption rate prediction results, and provides reliable data support for equipment health status assessment, remaining lifespan estimation, and maintenance strategy formulation.

[0053] In the specific implementation process, the network structure of the regression head is first constructed, which consists of three fully connected layers. The input multi-scale junction temperature fusion encoding representation vector first enters a fully connected layer containing 128 neurons and uses the ReLU activation function. Its output is then fed into a hidden layer containing 64 neurons, also using the ReLU activation function. Finally, through a single-neuron output layer, the predicted lifetime attrition rate for the current window is obtained. To enable the entire evaluation model to have predictive capabilities, it needs to be trained end-to-end. The training data can come from accelerated aging experiments of power semiconductor devices, in which high-frequency junction temperature and average junction temperature are recorded simultaneously, and the true lifetime attrition rate corresponding to each operating condition window is determined as the label through failure analysis. During training, the collected junction temperature sequence is input into the model to obtain the predicted lifetime attrition rate. Then, the mean squared error is used as the loss function to calculate the difference between the predicted value and the true label. Finally, the Adam optimizer is used to perform backpropagation and gradient update on all trainable parameters of the model (including the dual-channel encoder, the cross-scale fusion module, and all weights and biases in the regression head). For example, an initial learning rate of 0.001 can be set, and a learning rate decay strategy can be adopted, such as reducing the learning rate by 10% every 10 training cycles. By performing multiple rounds of iterative training on the entire training dataset until the model's loss converges on the validation set, an optimized model capable of accurately predicting lifetime attrition rate is obtained. Then, the prediction process is initiated, inputting the preprocessed multi-scale junction temperature fusion encoding representation vector into the constructed regression head, which sequentially passes through three fully connected layers for feature processing and mapping, finally outputting the predicted lifetime attrition rate value corresponding to the current window.

[0054] In the aforementioned multi-scale wind turbine converter life assessment method, step S5 determines the remaining service life at the current moment based on the accumulated damage at the previous moment and the life consumption rate predicted in the current window. It should be understood that since converter life consumption is a long-term cumulative process, relying solely on the life consumption rate predicted in the current window cannot reflect the damage generated during historical operation, and the accumulated damage at the previous moment alone cannot reflect the life consumption trend under current operating conditions. Without considering the correlation between the two, the actual remaining service life cannot be accurately obtained. Therefore, in a specific example of this application, step S5 includes: determining the remaining service life at the current moment based on the accumulated damage at the previous moment and the life consumption rate predicted in the current window using the following formula:

[0055]

[0056] in, This is the cumulative damage from the previous moment. The predicted lifetime attrition rate for the current window. Step size, This represents the remaining service life at the current moment. This represents the average lifetime attrition rate predicted for the current window. This involves calculating the cumulative damage at the current moment. In other words, by dynamically updating the cumulative damage and relating it to consumption trends under future operating conditions, the remaining service life at the current moment can be calculated. This fully integrates historical damage and real-time consumption information, avoiding prediction biases caused by focusing only on data from a single time point. It ensures that the remaining service life assessment results dynamically match the actual health status of the converter, providing a reliable basis for developing accurate predictive maintenance strategies for wind farms, and effectively reducing the risk of unplanned downtime and operation and maintenance costs.

[0057] In the specific implementation process, firstly, cumulative damage updates are performed. Based on a preset time step, the predicted lifetime attrition rate for the current window is calculated with the step size to obtain the damage increment within the current window. This increment is then added to the cumulative damage from the previous time step to obtain the cumulative damage at the current time step. Next, the future average lifetime attrition rate is determined. Based on the predicted lifetime attrition rates from the past 30 time windows, a weighted average algorithm (with recent data having higher weight than older data) is used to calculate the future average lifetime attrition rate. Then, the current remaining lifetime is calculated. The cumulative damage at the current time step is subtracted from the damage failure threshold (set to 1) to obtain the remaining tolerable damage amount. Finally, the remaining tolerable damage amount is divided by the future average lifetime attrition rate to obtain the final remaining lifetime at the current time step.

[0058] In summary, the multi-scale wind turbine converter life assessment method based on the embodiments of this application is elucidated. It synchronously processes high-frequency instantaneous junction temperature data and long-period average junction temperature data of the wind turbine converter, and introduces a dual-channel deep time-series coding architecture to deeply deconstruct the junction temperature sequences of the two different time scales. Furthermore, through a cross-scale interactive fusion mechanism, it captures and quantifies the nonlinear modulation effect of the long-scale average junction temperature on the accumulation of short-scale damage to obtain a deep fusion vector. Finally, based on this deep fusion vector, it drives a life prediction model to achieve an accurate assessment of the life consumption rate under current operating conditions, and combines historical damage for dynamic iteration to complete the online prediction of the remaining service life of the converter. This effectively overcomes the limitations of traditional linear accumulation criteria and significantly improves the accuracy and reliability of wind turbine converter life prediction.

[0059] Furthermore, a multi-scale wind turbine converter life assessment system is also provided.

[0060] Figure 6 This is a block diagram of a multi-scale wind turbine converter life assessment system according to an embodiment of this application. Figure 6As shown, the multi-scale wind turbine converter life assessment system 100 according to an embodiment of this application includes: a junction temperature sequence acquisition module 110, used to acquire a high-frequency junction temperature sequence and an average junction temperature sequence based on the SCADA cycle; a dual-channel parallel feature encoding module 120, used to perform dual-channel parallel multi-scale feature encoding on the high-frequency junction temperature sequence and the average junction temperature sequence based on the SCADA cycle to obtain an encoding representation vector for a short-scale sequence and an encoding representation vector for a long-scale sequence; a cross-scale fusion module 130, used to perform cross-scale fusion on the encoding representation vectors of the short-scale sequence and the long-scale sequence to obtain a multi-scale junction temperature fusion encoding representation vector; a lifespan attrition rate prediction module 140, used to predict the lifespan attrition rate based on the multi-scale junction temperature fusion encoding representation vector to obtain the lifespan attrition rate predicted in the current window; and a remaining service life calculation module 150, used to determine the remaining service life at the current moment based on the cumulative damage at the previous moment and the lifespan attrition rate predicted in the current window.

[0061] Here, those skilled in the art will understand that the specific operation of each module in the above-mentioned multi-scale wind turbine converter life assessment system has been referenced above. Figures 1 to 5 The description of the multi-scale-based wind turbine converter life assessment method is detailed here, and therefore, its repeated description will be omitted.

[0062] The basic principles of the present invention have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in the present invention are merely examples and not limitations, and should not be considered as essential features of each embodiment of the present invention. Furthermore, the specific details of the above embodiments are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the present invention to the necessity of employing the specific details described above.

[0063] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the unit division is only a logical functional division, and other division methods may exist in actual implementation. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0064] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the present invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.

[0065] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units stated in a system claim may also be implemented by a single unit through software or hardware.

[0066] Finally, it should be noted that the above description has been given for illustrative and descriptive purposes. Furthermore, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although modifications or equivalent substitutions may be made to the technical solutions with reference to preferred embodiments, they will not depart from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-scale wind turbine converter life assessment method, characterized in that, include: Obtain high-frequency junction temperature sequences and average junction temperature sequences based on SCADA cycles; Dual-channel parallel multi-scale feature encoding is performed on the high-frequency junction temperature sequence and the average junction temperature sequence based on SCADA cycle to obtain the encoding representation vector of the short-scale sequence and the encoding representation vector of the long-scale sequence. Cross-scale fusion of the encoding representation vectors of short-scale sequences and long-scale sequences is performed to obtain a multi-scale junction temperature fused encoding representation vector; The lifetime attrition rate is predicted based on the multi-scale junction temperature fusion coding representation vector to obtain the lifetime attrition rate predicted for the current window. The remaining useful life at the current moment is determined based on the cumulative damage at the previous moment and the life consumption rate predicted in the current window.

2. The method for assessing the lifespan of wind turbine converters based on multiple scales according to claim 1, characterized in that, Dual-channel parallel multi-scale feature encoding is performed on the high-frequency junction temperature sequence and the average junction temperature sequence based on SCADA cycles to obtain the encoded representation vectors of the short-scale sequence and the long-scale sequence, including: The high-frequency junction temperature sequence is input into the short-scale channel to obtain the encoded representation vector of the short-scale sequence; The average junction temperature sequence based on the SCADA cycle is input into the long-scale channel to obtain the encoded representation vector of the long-scale sequence.

3. The method for assessing the lifespan of wind turbine converters based on multiple scales according to claim 2, characterized in that, The short-scale channel uses a Transformer encoder, while the long-scale channel uses a Long Short-Term Memory network.

4. The method for assessing the lifespan of wind turbine converters based on multiple scales according to claim 1, characterized in that, Cross-scale fusion of the encoding representation vectors of short-scale sequences and long-scale sequences yields a multi-scale junction-temperature fused encoding representation vector, including: Linear projection of multi-scale features onto the encoding representation vectors of short-scale sequences and long-scale sequences yields long-scale junction temperature query vectors, short-scale junction temperature key vectors, and short-scale junction temperature value vectors. The correlation degree of the long-scale junction temperature query vector and the short-scale junction temperature key vector is calculated to obtain the attention weight matrix; The attention weight matrix is ​​multiplied by the short-scale junction temperature vector to obtain the cross-scale scenario association vector; The multi-scale junction temperature fusion encoding representation vector is obtained by fusing the cross-scale scenario association vector and the long-scale sequence encoding representation vector.

5. The method for assessing the lifespan of wind turbine converters based on multiple scales according to claim 4, characterized in that, The correlation between the long-scale junction temperature query vector and the short-scale junction temperature bond vector is calculated to obtain the attention weight matrix. This includes calculating the correlation between the long-scale junction temperature query vector and the short-scale junction temperature bond vector using the following formula: in, This is a long-scale junction temperature query vector. For short-scale junction temperature bond vectors, This is the transpose of the vector. Let be the dimension of the short-scale junction temperature bond vector. The original attention score. For normalization function, This is the attention weight matrix.

6. The method for assessing the lifespan of wind turbine converters based on multiple scales according to claim 4, characterized in that, The multi-scale junction temperature fusion encoded representation vector is obtained by fusing cross-scale scenario correlation vectors and long-scale sequence encoded representation vectors, including: Local feature domain extraction based on one-dimensional convolutional coding is performed on cross-scale scenario association vectors and long-scale sequence encoded representation vectors to obtain a set of local feature vectors for cross-scale scenario association and a set of local granular encoded vectors for long-scale sequences. The cost matrix between the set of local feature vectors of cross-scale scenario association and the set of local granular encoding vectors of long-scale sequence is calculated to obtain the scenario association cost matrix. Based on the scenario association cost matrix, determine the scenario association optimal transmission planning matrix; Based on the scenario-related optimal transmission planning matrix, the set of local granular coding vectors for long-scale sequences is matched for potential spatial distribution to obtain the set of local granular coding vectors after manifold matching for long-scale sequences. The set of local feature vectors associated with the cross-scale scenario and the set of local granularity coding vectors after long-scale sequence manifold matching are subjected to local feature unit co-modulation to obtain a multi-scale junction temperature fusion coding representation vector.

7. The method for assessing the lifespan of wind turbine converters based on multiple scales according to claim 1, characterized in that, The method for predicting lifetime attrition rate based on multi-scale junction temperature fusion encoded representation vector to obtain the lifetime attrition rate predicted for the current window includes: inputting the multi-scale junction temperature fusion encoded representation vector into a regression head composed of one or more fully connected layers to obtain the lifetime attrition rate predicted for the current window.

8. The method for assessing the lifespan of wind turbine converters based on multiple scales according to claim 7, characterized in that, The remaining useful life at the current moment is determined based on the cumulative damage at the previous moment and the lifetime attrition rate predicted for the current window. This includes determining the remaining useful life at the current moment using the following formula: in, This is the cumulative damage from the previous moment. The predicted lifetime attrition rate for the current window. Step size, This represents the remaining service life at the current moment. This represents the average lifetime attrition rate predicted for the current window. This represents the cumulative damage at the current moment.

9. A multi-scale wind turbine converter life assessment system, characterized in that, include: The junction temperature sequence acquisition module is used to acquire high-frequency junction temperature sequences and average junction temperature sequences based on SCADA cycles. A dual-channel parallel feature encoding module is used to perform dual-channel parallel multi-scale feature encoding on high-frequency junction temperature sequences and average junction temperature sequences based on SCADA cycles to obtain the encoding representation vectors of short-scale sequences and long-scale sequences. The cross-scale fusion module is used to perform cross-scale fusion of the encoding representation vectors of short-scale sequences and long-scale sequences to obtain a multi-scale junction temperature fused encoding representation vector. The lifetime attrition rate prediction module is used to predict the lifetime attrition rate based on the multi-scale junction temperature fusion encoded representation vector to obtain the lifetime attrition rate predicted for the current window. The remaining useful life calculation module is used to determine the remaining useful life at the current moment based on the cumulative damage at the previous moment and the life consumption rate predicted in the current window.

Images