Scene information reverse transmission method and system based on digital twinning

Abstract

The application provides a scene information reverse transmission method and system based on digital twinning, and relates to the technical field of artificial intelligence. The method comprises the following steps: performing space-time alignment and outlier processing on collected multi-source data to form a standardized data sequence, selecting a reference data segment in the standardized data sequence, generating two characteristic subsequences based on the reference data segment, calculating the dynamic correlation of the two characteristic subsequences to form a characteristic interval; selecting one key data unit in the characteristic interval and two key data units outside the characteristic interval; constructing a nonlinear data track according to the time continuity of the three key data units; generating a correction factor according to the nonlinear data track; inputting the correction factor and the standardized data sequence into a generative adversarial network to generate extreme fault time series data under multi-source interference. The application improves the adaptability and safety of reverse transmission.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of artificial intelligence, in particular to a scene information reverse transmission method and system based on digital twinning. BACKGROUND

[0002] In the process of industrial intelligentization, digital twinning technology has become the core support for device fault diagnosis and operation and maintenance scheduling due to its virtual-real mapping and real-time interaction characteristics. However, the current scene information reverse transmission application based on digital twinning still faces many technical bottlenecks.

[0003] On the one hand, the data collected by the multi-source sensors in the industrial field generally has problems of space-time dislocation, noise interference and abnormal value pollution. The existing data processing methods mostly rely on static threshold or single feature compensation, and are difficult to adapt to nonlinear error correction in complex dynamic environments, resulting in large mapping consistency deviation between the virtual scene and the physical entity, which directly affects the accuracy of the reverse transmission information.

[0004] On the other hand, the device fault features are usually high-dimensional and contain a large amount of redundant information. Traditional dimension reduction and lightweight methods mostly use fixed parameters or static pruning strategies, which cannot dynamically match the real-time running state of the device, resulting in high occupation of computing resources and difficulty in meeting the stringent requirements of large devices such as fans for real-time scheduling. In addition, the adaptability of the scheduling strategy generated in the virtual scene to the physical execution end is insufficient, and there is a lack of effective multi-scene verification mechanism, which leads to problems such as low reliability and unbalanced cost control of the strategy in actual execution, thereby restricting the full play of the value of digital twinning technology in industrial operation and maintenance. SUMMARY

[0005] The technical problem to be solved by the application is to provide a scene information reverse transmission method based on digital twinning, which can improve the adaptability and security of reverse transmission.

[0006] To solve the above technical problems, the technical scheme of the application is as follows:

[0007] In a first aspect, a scene information reverse transmission method based on digital twinning, the method comprising:

[0008] Step 1: performing space-time alignment and abnormal value processing on the collected multi-source data to form a standardized data sequence, selecting a reference data segment in the standardized data sequence, generating two feature subsequences based on the reference data segment, calculating the dynamic correlation of the two feature subsequences to form a feature interval; selecting one key data unit in the feature interval and two key data units outside the feature interval; constructing a nonlinear data trajectory according to the time continuity of the three key data units; generating a correction factor according to the nonlinear data trajectory;

[0009] Step 2, input the correction factor into the generative adversarial network together with the standardized data sequence to generate extreme fault time series data under multi-source interference;

[0010] Step 3, according to the fault time series data, perform feature selection and dimension compression, combine the improved particle swarm algorithm for dynamic parameter adjustment, and use the second derivative pruning to realize lightweight calculation to obtain lightweight features;

[0011] Step 4, in the digital twin, fuse the lightweight features with the fan topology structure and real-time running state data to establish a spatio-temporal correlation matrix of the fault scene; minimize the power generation loss and repair cost as the double objective, calculate the interactive influence of the fault propagation path and the resource scheduling scheme through multi-physical field coupling simulation; based on the interactive influence, iteratively optimize the scheduling parameters to generate an adaptive scheduling strategy; in the virtual mapping environment, perform multi-scenario deduction verification, and output the verified scheduling strategy when the reliability index of the strategy reaches the preset threshold;

[0012] Step 5, issue the verified scheduling strategy to the physical execution end for execution.

[0013] Further, the collected multi-source data is spatio-temporally aligned and outlier processed to form a standardized data sequence, a reference data segment is selected in the standardized data sequence, two feature subsequences are generated based on the reference data segment, and the dynamic correlation of the two feature subsequences is calculated to form a feature interval, including:

[0014] Step 1.1, spatio-temporally align the received heterogeneous sensor data, align the data stream based on a unified timestamp, and fill in the missing position data through spatial interpolation;

[0015] Step 1.2, use the quantile outlier detection algorithm to process outliers, and perform Gaussian smoothing correction on data points exceeding three times the standard deviation to generate a standardized data sequence;

[0016] Step 1.3, in the standardized data sequence, select a preset length of continuous and smooth running data as a reference data segment;

[0017] Step 1.4, based on the reference data segment, perform double feature extraction, the first feature subsequence extracts time domain statistical features through a sliding window, and the second feature subsequence extracts frequency energy distribution features through fast Fourier transform;

[0018] Step 1.5, calculate the dynamic time-varying correlation of the two feature subsequences, use the windowed mutual information algorithm, and calculate the local mutual information value with a preset step size, mark the time period with a mutual information value continuously higher than the threshold as a feature interval.

[0019] Further, a key data unit is selected within the feature interval, and two key data units are selected outside the feature interval; a nonlinear data track is constructed according to the time continuity of the three key data units; a correction factor is generated according to the nonlinear data track, including:

[0020] Step 1.6, within the feature interval, the data unit corresponding to the mutual information peak time is selected as the first key data unit; within a preset time period before the start time of the feature interval, the data unit with the lowest statistical feature fluctuation rate is selected as the second key data unit; within a preset time period after the end time of the feature interval, the extreme point of the frequency domain energy change gradient is selected as the third key data unit;

[0021] Step 1.7, based on the time stamp continuity of the three key data units, taking the second key data unit as the starting point, the first key data unit as the relay point, and the third key data unit as the terminal point, a nonlinear data track spanning the feature interval boundary is generated by fitting a third-order Bezier curve;

[0022] Step 1.8, for the nonlinear data track, the Euclidean distance of adjacent points on the track is calculated and continuously accumulated to obtain the total length of the track; the absolute value of the curvature of each point on the track is also calculated; the product of the total length of the track and the absolute value of the curvature is normalized to generate an environmental parameter correction factor.

[0023] Further, the correction factor and the standardized data sequence are jointly input into the generative adversarial network to generate extreme fault time series data under multi-source interference, including:

[0024] Step 2.1, align the environmental parameter correction factor with the standardized data sequence according to the time stamp to construct a conditional input vector;

[0025] Step 2.2, input the conditional input vector into the generator of the generative adversarial network, map it to the latent space modulation signal through the fully connected layer, and concatenate the latent space modulation signal with the random noise vector in the channel dimension to form the combined input data of the generator;

[0026] Step 2.3, perform time convolution decoding operation on the combined input data to generate preliminary fault time series data, and optimize the generator parameters based on the adversarial loss feedback of the discriminator to adjust the distribution of the preliminary fault time series data;

[0027] Step 2.4, input the preliminary fault time series data into the discriminator, verify the data validity through the preset physical constraint rule, first verify whether the fluid dynamics coupling relationship of the standardized data sequence conforms to the physical law, and secondly verify the positive correlation between the fault amplitude and the correction factor value;

[0028] Step 2.5, when the comprehensive confidence output by the discriminator on the verified preliminary fault time series data reaches the preset threshold, it is determined that the data output by the current generator meets the requirements, and is marked as extreme fault time series data under multi-source interference.

[0029] Further, according to the feature selection and dimension compression of the fault time series data, the improved particle swarm algorithm is used for dynamic parameter adjustment, and the second derivative pruning is used for lightweight calculation to obtain lightweight features, including:

[0030] Step 3.1, receiving the extreme fault time series data under multi-source interference, and inputting it into the initial multi-layer perception network;

[0031] Step 3.2, performing structured feature selection on the full connection layer of the initial multi-layer perception network, and removing redundant neuron connections to generate a primary optimization network with compressed feature dimensions;

[0032] Step 3.3, performing dynamic optimization of weight parameters on the primary optimization network, taking the reconstruction error of the feature and the original fault data as the fitness function, updating the network parameters through the inertia weight nonlinear decay mechanism and the global final disturbance strategy of the improved particle swarm algorithm, and generating a parameter optimization network;

[0033] Step 3.4, performing lightweight processing on the parameter optimization network, calculating the second derivative absolute value of the activation function as the neuron importance indicator, removing neurons and connection edges with importance lower than the preset threshold, and generating a pruned lightweight network. The output data of the lightweight network is marked as lightweight features.

[0034] Further, in the digital twin, the lightweight features are fused with the fan topology structure and real-time running state data to establish a spatio-temporal correlation matrix of the fault scene; the interactive influence of the fault propagation path and the resource scheduling scheme is calculated through multi-physical field coupling simulation with the dual objectives of minimizing power generation loss and minimizing repair cost; the scheduling parameters are iteratively optimized based on the interactive influence to generate an adaptive scheduling strategy; and the multi-scenario deduction verification is performed in the virtual mapping environment. When the reliability index of the strategy reaches the preset threshold, the scheduling strategy that passes the verification is output, including:

[0035] Step 4.1, in the digital twin, the lightweight features are fused with the real-time acquired fan topology structure data and fan running state data according to the spatial position mapping relationship to construct a spatio-temporal correlation matrix of the fault scene;

[0036] Step 4.2, inputting the spatio-temporal correlation matrix into the multi-physical field coupling simulator, taking the minimization of power generation loss and the minimization of repair cost as dual optimization objectives, and outputting the interactive influence value of the fault propagation path and the resource scheduling scheme through fluid-structure-electric field coupling calculation;

[0037] Step 4.3, iteratively optimize the scheduling parameters according to the interaction influence value, increase the maintenance resource weight parameter when the interaction influence value shows that the fault propagation risk is higher than the threshold value, adjust the power redistribution parameter when the interaction influence value shows that the power generation loss exceeds the threshold value, and re-execute the coupling simulation after dynamically updating the parameters to generate an adaptive scheduling strategy;

[0038] Step 4.4, load the adaptive scheduling strategy in the digital twin virtual mapping environment, inject historical extreme scenario data and random disturbance variables for multiple rounds of deduction, and output the scheduling strategy that passes the verification when the reliability index of the strategy meets the fault suppression success rate greater than or equal to the preset threshold and the cost control deviation rate less than or equal to the preset threshold.

[0039] Further, the scheduling strategy that passes the verification is issued to the physical execution end for execution, including:

[0040] Step 5.1, receive the adaptive scheduling strategy and encapsulate it into an executable instruction set according to the physical execution end protocol;

[0041] Step 5.2, issue the instruction set to the target wind turbine's physical execution end through a preset encryption channel, and start the end-to-end time delay monitoring clock synchronously;

[0042] Step 5.3, the physical execution end parses the instruction set and performs device start-stop, power adjustment and maintenance resource scheduling operations, and returns the status code in real time;

[0043] Step 5.4, judge the execution integrity according to the status code, if the success feedback code of all operations is received and the time delay is not exceeded, mark the strategy execution success; if a failure feedback code is detected or the time delay exceeds the timeliness threshold, trigger the strategy rollback mechanism and generate a fault alarm;

[0044] Step 5.5, collect the actual power generation data and maintenance cost data of the physical execution end, compare and verify with the decision accuracy index, and generate an execution efficiency evaluation report.

[0045] The second aspect is a scene information reverse transmission system based on digital twinning, including:

[0046] The acquisition module is configured to perform spatio-temporal alignment and outlier processing on the collected multi-source data to form a standardized data sequence, select a reference data segment in the standardized data sequence, generate two feature sub-sequences based on the reference data segment, calculate the dynamic correlation of the two feature sub-sequences to form a feature interval, select a key data unit within the feature interval and two key data units outside the feature interval, construct a nonlinear data trajectory according to the time continuity of the three key data units, and generate a correction factor according to the nonlinear data trajectory;

[0047] The input module is used to input the correction factor and the standardized data sequence into the generative adversarial network to generate extreme fault time series data under multi-source interference.

[0048] The computation module performs feature selection and dimensionality compression based on fault time-series data, dynamically tunes parameters using an improved particle swarm optimization algorithm, and employs second-derivative pruning to achieve computational lightweighting, resulting in lightweight features. In the digital twin, these lightweight features are fused with wind turbine topology and real-time operating status data to establish a spatiotemporal correlation matrix for the fault scenario. With the dual objectives of minimizing power generation loss and repair costs, the module uses multi-physics coupling simulation to calculate the interaction between fault propagation paths and resource scheduling schemes. Based on this interaction, scheduling parameters are iteratively optimized to generate an adaptive scheduling strategy. Multi-scenario simulations are performed in a virtual mapping environment to verify the strategy. When the strategy's reliability index reaches a preset threshold, the verified scheduling strategy is output.

[0049] The processing module is used to send the verified scheduling policy to the physical execution end for execution.

[0050] Thirdly, a computing device, comprising:

[0051] One or more processors;

[0052] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0053] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0054] The above-described solution of the present invention has at least the following beneficial effects:

[0055] By performing spatiotemporal alignment and dynamic correction on multi-source data, the problems of spatiotemporal misalignment and noise interference in industrial field data are solved, improving the consistency between virtual mapping and physical entities. Through feature selection, dynamic parameter tuning, and lightweight processing, the computational resource consumption caused by redundancy in fault feature dimensions is reduced, meeting the real-time scheduling requirements of equipment. Through multi-physics coupling simulation and dual-objective optimization in digital twins, combined with multi-scenario simulation verification, the generated adaptive scheduling strategy can accurately balance power generation losses and repair costs, ensuring the reliability and applicability of the strategy. Finally, through encrypted transmission, execution monitoring, and performance evaluation, secure and efficient reverse transmission of scene information from virtual to physical is achieved, providing accurate and reliable technical support for fault diagnosis and operation and maintenance scheduling of industrial equipment (such as wind turbines), and improving the intelligent operation and maintenance level and operational economy of industrial systems. Attached Figure Description

[0056] Figure 1 is a flowchart of a scene information reverse transmission method based on digital twinning provided by an embodiment of the present application.

[0057] Figure 2 is a schematic diagram of a scene information reverse transmission system based on digital twinning provided by an embodiment of the present application. DETAILED DESCRIPTION

[0058] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0059] As shown in Figure 1 , an embodiment of the present application proposes a scene information reverse transmission method based on digital twinning, which comprises the following steps:

[0060] Step 1, performing spatio-temporal alignment and outlier processing on the collected multi-source data to form a standardized data sequence, selecting a reference data segment in the standardized data sequence, generating two feature subsequences based on the reference data segment, calculating the dynamic correlation of the two feature subsequences to form a feature interval; selecting one key data unit within the feature interval and two key data units outside the feature interval; constructing a nonlinear data trajectory according to the time continuity of the three key data units; generating a correction factor according to the nonlinear data trajectory;

[0061] Step 2, inputting the correction factor and the standardized data sequence into a generative adversarial network to generate extreme fault time series data under multi-source interference;

[0062] Step 3, performing feature selection and dimension compression according to the fault time series data, combining an improved particle swarm algorithm for dynamic parameter adjustment, and using second derivative pruning to realize lightweight calculation to obtain lightweight features;

[0063] Step 4, in the digital twinning, fusing the lightweight features with the fan topology structure and real-time running state data to establish a spatio-temporal correlation matrix of the fault scene; minimizing power generation loss and repair cost as dual objectives, calculating the interactive influence of fault propagation path and resource scheduling scheme through multi-physical field coupling simulation; iteratively optimizing the scheduling parameters based on the interactive influence to generate an adaptive scheduling strategy; performing multi-scenario deduction verification in a virtual mapping environment, and outputting the scheduling strategy that passes the verification when the reliability index of the strategy reaches a preset threshold;

[0064] Step 5, issuing the scheduling strategy that passes the verification to the physical execution end for execution.

[0065] In the embodiment of the present application, by spatio-temporal alignment and outlier processing of multi-source data, combined with dynamic correlation analysis to generate correction factors, the data standardization degree is improved, the problems of spatio-temporal misplacement, noise interference and the like of industrial data are solved, and the consistency of virtual mapping and physical entity is enhanced; the extreme fault time series data is generated by using the generative adversarial network, combined with feature selection, dynamic parameter adjustment and lightweight processing, the feature dimension redundancy and the calculation resource occupation are reduced, and the real-time scheduling demand is met; the spatio-temporal correlation matrix is constructed by fusing multi-source information in the digital twin, the adaptive scheduling strategy is generated by multi-physical field coupling simulation and double-objective optimization, and the reliability is ensured by multi-scenario deduction verification, finally the strategy is safely issued to the physical execution end, realizing the precise and efficient information reverse transmission from virtual to physical, and providing decision support for equipment fault operation and maintenance considering the minimization of power generation loss and the optimization of repair cost.

[0066] In a preferred embodiment of the present application, the above step 1 can include:

[0067] Step 1.1, spatio-temporal alignment of received heterogeneous sensor data, aligning data streams based on a unified timestamp, and filling in missing data by spatial interpolation;

[0068] Step 1.2, processing outliers by using quantile outlier detection algorithm, Gaussian smoothing correction is performed on data points exceeding three times the standard deviation, and a standardized data sequence is generated;

[0069] Step 1.3, selecting a preset length of data continuously and smoothly running as a reference data segment in the standardized data sequence;

[0070] Step 1.4, performing double feature extraction based on the reference data segment, the first feature subsequence is extracted by sliding window to extract time domain statistical features and the second feature subsequence is extracted by fast Fourier transform to extract frequency energy distribution features;

[0071] Step 1.5, calculating the dynamic time-varying correlation of the two feature subsequences, using windowed mutual information algorithm to calculate the local mutual information value by sliding with a preset step, and marking the time period with mutual information value continuously higher than the threshold as a feature interval.

[0072] In the embodiment of the present application, the spatio-temporal consistency of multi-source heterogeneous data is ensured by spatio-temporal alignment and spatial interpolation, the data noise is effectively purified by quantile outlier detection and Gaussian smoothing correction, the stable reference data segment is selected to provide a reliable reference for feature extraction, the double feature extraction realizes multi-dimensional characterization of data features by combining time domain statistics and frequency energy distribution, and the windowed mutual information algorithm accurately locks the feature interval with high correlation.

[0073] In the embodiment of the present application, when specifically applied, the above steps can be realized by the following steps, for example:

[0074] Step 1.1, resample the data streams collected by different sensors according to a unified time reference; for data points that do not match in time, use linear interpolation to estimate the value between two known data points; for data with inconsistent time intervals, use a fixed-size sliding window with a width of Δt to process the data, and take the average or median value within the window as the unified value corresponding to the time window; for missing data points in space, estimate the value from the surrounding known points. Common methods include Kriging interpolation or inverse distance weighting interpolation; the calculation method of the filled value is to multiply the value of each adjacent known point by a weight and sum it up, where the weight is inversely proportional to the distance, for example, the weight in IDW is the reciprocal of the distance to the p-th power, and p is an adjustable power parameter.

[0075] Step 1.2, calculate the quartiles of the data: Q1 is the 25th percentile and Q3 is the 75th percentile, the difference between the two is called the interquartile range, IQR = Q3 - Q1; define the normal data range as [Q1 - k x IQR, Q3 + k x IQR], usually k = 1.5, data points outside this range are identified as outliers. For data points outside the range of μ ± 3σ, μ is the mean and σ is the standard deviation, use Gaussian kernel function for smoothing; the calculation method of the corrected value is to take the abnormal point as the center and perform weighted average on the values of the surrounding adjacent points, the weight is determined by the Gaussian function, and the weight decays exponentially with the square of the distance.

[0076] Step 1.3, use a fixed-length sliding window to detect the stationarity of the data in the standardized data sequence; the stationarity judgment standard is that the variance of the data in the window should be less than a preset threshold δ, and the mean change rate of the data in the window should be less than another preset threshold ε, find the longest window segment that continuously satisfies the stationarity condition from the data sequence.

[0077] Step 1.4, use a sliding window to slide on the data, calculate the following statistics for the data in each window, including mean, standard deviation, skewness and kurtosis, perform fast Fourier transform on the data in the sliding window to convert the time-domain data to frequency-domain representation, calculate the energy of each frequency component, i.e. the square of the amplitude of the frequency component, extract features from the frequency energy distribution, such as the main frequency component and the energy proportion in a certain frequency band.

[0078] The step 1.5 uses a fixed-length sliding window to synchronously slide on the time-domain and frequency-domain feature subsequences at a specific step, and calculates mutual information for two subsequences in each window. The mutual information is an index for measuring the mutual dependence between two variables, and the calculation formula is: the logarithm of the ratio of the joint probability distribution of the two variables to the product of the respective marginal probability distributions, and then the expected value is calculated. In actual calculation, the discrete method or kernel density estimation method is usually used to approximate the probability distribution. A mutual information threshold θ is set, and when the mutual information values of the continuous M sliding windows are all higher than the threshold, the corresponding time period is marked as a feature interval. The boundaries of the feature interval are further optimized, the derivative, i.e. the change rate of the mutual information, is calculated, and the zero-crossing points of the derivative are found. The points where the change rate changes from positive to negative or from negative to positive can more accurately determine the start and end positions of the feature interval.

[0079] In a preferred embodiment of the present application, the step 1 can include:

[0080] Step 1.6: In the feature interval, the data unit at the time corresponding to the mutual information peak value is selected as the first key data unit; in a preset time period before the start time of the feature interval, the data unit with the lowest statistical feature fluctuation rate is selected as the second key data unit; in a preset time period after the end time of the feature interval, the extreme point of the frequency domain energy change gradient is selected as the third key data unit;

[0081] Step 1.7: Based on the time stamp continuity of the three key data units, a non-linear data trajectory spanning the boundaries of the feature interval is generated by a third-order Bezier curve fitting starting from the second key data unit, using the first key data unit as a relay point, and ending at the third key data unit.

[0082] Step 1.8: For the non-linear data trajectory, the Euclidean distance between adjacent points on the trajectory is calculated and continuously accumulated to obtain the total length of the trajectory; the absolute value of the curvature of each point on the trajectory is also calculated; and the product of the total length of the trajectory and the absolute value of the curvature is normalized to generate an environmental parameter correction factor.

[0083] In the embodiment of the present application, the core data in the feature interval and the key reference points outside the interval are accurately selected, taking into account the representativeness and dynamic correlation of data features; the third-order Bezier curve fitting generates a non-linear trajectory spanning the boundaries of the interval, which can smoothly capture the continuous change rule of the data inside and outside the feature interval, avoiding the distortion of linear fitting; the correction factor is generated based on the product of the total length of the trajectory and the absolute value of the curvature, which can comprehensively quantify the influence of the dynamic characteristics of the data on the environment, and after normalization, the accuracy of the quantification of the influence of the data features on the environmental parameters is improved.

[0084] In the embodiment of the present application, when specifically applied, the above steps can be implemented by the following steps, for example:

[0085] The first key data unit is found in the feature interval, and the time point at which the mutual information value reaches the maximum value in the marked feature interval is the first key data unit, denoted as P1, and the action is to capture the core state with the strongest correlation between the two feature subsequences in the feature interval; the second key data unit is calculated in a preset time period before the start time of the feature interval, and the statistical feature fluctuation rate of each data unit is calculated; the fluctuation rate is calculated for the extracted time domain feature subsequence, the change rate of the statistical quantity such as mean and standard deviation in each window is selected, and the data unit corresponding to the time point with the minimum change rate is selected as the second key data unit, denoted as P0, which can represent the stable state before the feature interval and serve as the starting point of trajectory fitting; the third key data unit is calculated in a preset time period after the end time of the feature interval, and the gradient of the extracted frequency energy distribution sequence, i.e., the energy change rate, is calculated; the difference between the frequency energy values of adjacent time points is calculated, and the data unit corresponding to the time point with the maximum absolute difference value is selected as the third key data unit, denoted as P2, which can capture the point with the most drastic energy change after the end of the feature interval and serve as the end point of trajectory fitting.

[0086] The third-order Bezier curve is determined by four control points: the starting point P0, the end point P2, and two intermediate control points C0 and C1; the position of any point B(t) on the curve can be represented as the weighted sum of the four control points, and the weight is determined by the parameter t, which ranges from 0 to 1.

[0087] The intermediate control points are determined as follows:

[0088] C0 is located in the direction from P0 to P1, and the distance is α times the distance from P0 to P1, where α is a proportionality coefficient, usually taking a value of 0.3-0.5; C1 is located in the direction from P2 to P1, and the distance is β times the distance from P2 to P1, where β is a proportionality coefficient, usually taking a value of 0.3-0.5; the time stamps t0, t1, and t2 corresponding to the three points P0, P1, and P2 on the actual time axis are mapped to the [0, 1] interval of the parameter t, and a linear mapping method is usually used, i.e., t = (current time - t0) ÷ (t2 - t0).

[0089] The Bezier curve is discretized into N points, the straight-line distance between adjacent points is calculated, and then all the distances are accumulated to obtain the total length of the trajectory, where the total length of the trajectory = the sum of the Euclidean distances between adjacent points; for each point on the curve, the first derivative is calculated to represent the tangent direction, and the second derivative is calculated to represent the change rate of the tangent direction.

[0090] The curvature formula is: where, the first derivative of the direction coordinate with respect to the parameter is the instantaneous rate of change of the direction coordinate with respect to the parameter, the first derivative of the direction coordinate with respect to the parameter is the instantaneous rate of change of the direction coordinate with respect to the parameter, the first derivative of the direction coordinate with respect to the parameter is the instantaneous rate of change of the direction coordinate with respect to the parameter, the second derivative of the direction coordinate with respect to the parameter is the instantaneous rate of change of the first derivative of the direction coordinate with respect to the parameter, the second derivative of the direction coordinate with respect to the parameter is the instantaneous rate of change of the first derivative of the direction coordinate with respect to the parameter, the numerator is the absolute value of the cross product of the first derivative and the second derivative, and the denominator is the third power of the modulus of the first derivative, which normalizes the bending degree and eliminates the influence of the trajectory scale, and the absolute value of the curvature is used to measure the bending degree of the trajectory at the point, and the greater the value, the more abrupt the change of the trajectory at the point; the average curvature is multiplied by the total length of the trajectory to obtain an original correction value, and the original correction value is mapped to the interval [0, 1], and the specific method is: correction factor = (original value-historical minimum value) ÷ (historical maximum value-historical minimum value), wherein the minimum value and the maximum value are obtained by analyzing historical trajectory data.

[0091] In a preferred embodiment of the present application, the above step 2 can comprise:

[0092] Step 2.1, align the environmental parameter correction factor with the standardized data sequence according to the timestamp to construct a conditional input vector;

[0093] Step 2.2, input the conditional input vector into the generator of the generative adversarial network, map it to the latent space modulation signal through the full connection layer, and splice the latent space modulation signal with the random noise vector in the channel dimension to form the combined input data of the generator;

[0094] Step 2.3, perform time convolution decoding operation on the combined input data to generate preliminary fault time series data, and optimize the generator parameters based on the adversarial loss feedback of the discriminator to adjust the distribution of the preliminary fault time series data;

[0095] Step 2.4, input the preliminary fault time series data into the discriminator to verify the data validity through the preset physical constraint rule, first verify whether the fluid dynamics coupling relationship of the standardized data sequence conforms to the physical law, and secondly verify the positive correlation between the fault amplitude and the correction factor value;

[0096] Step 2.5, when the comprehensive confidence of the preliminary fault time series data output by the discriminator reaches the preset threshold, it is determined that the data output by the current generator meets the requirements, and is marked as the extreme fault time series data under multi-source interference.

[0097] In the embodiment of the present application, the spatiotemporal correlation between the generated data and the original data and the environmental impact is generated by constructing a conditional input vector by aligning the environmental parameter correction factor with the standardized data sequence by timestamp, avoiding generation deviation caused by misplacement of input information; the channel dimension of the latent space modulation signal and the random noise vector in the generator is spliced, retaining the basic law of the standardized data and enhancing the diversity of extreme fault scenarios through random noise, effectively covering complex fault modes under multi-source interference; the time convolution decoding operation adapts to the characteristics of time series data, and the generator parameters are optimized in combination with the adversarial loss feedback of the discriminator, which can continuously calibrate the distribution of preliminary fault time series data and is closer to the statistical characteristics of real fault data; the extreme fault time series data meeting the requirements is determined by the comprehensive confidence threshold, ensuring the quality and reliability of the generated data, and solving the core problems of extreme fault sample scarcity in the industrial field and the disconnection between artificially generated data and actual physical laws.

[0098] In the embodiment of the present application, when specifically applied, the above steps can be realized by the following steps, for example:

[0099] The above step 2.1 is an accurate matching in the time dimension based on the environmental parameter correction factor and the standardized data sequence; specifically, the data corresponding to the same time in the environmental parameter correction factor and the standardized data sequence are associated and aligned based on the timestamp, so that each group of data contains not only the core state information of the device at that time, but also the corresponding environmental correction information; through the alignment processing, a conditional input vector that can comprehensively reflect the device state and environmental impact is constructed.

[0100] The above step 2.2 is full connection layer feature mapping; the full connection layer of the generator processes the conditional input vector, first multiplies the input vector with the weight matrix, then adds the bias value to obtain the linear transformation result; then through the activation function such as ReLU, i.e. retaining the positive value and setting the negative value to 0, the latent space modulation signal is obtained; a random noise vector is formed by combining the input data, the numerical value of which conforms to the standard normal distribution, i.e. the mean value is 0 and the standard deviation is 1, and the dimension is the same as that of the latent space modulation signal; the noise vector and the latent space modulation signal are spliced in parallel according to the feature dimension, for example, the first element of the modulation signal is adjacent to the first element of the noise, the second element is adjacent to the second element of the noise, and so on, to form the combined input data of the generator.

[0101] The step 2.3, time convolution decoding, the combined input data are processed through a plurality of time convolution layers, a convolution kernel slides along the time direction, a weighted sum in each window is calculated, after processing of each layer, batch normalization and an activation function such as Tanh are performed, and the output is limited between-1 and 1, and finally preliminary fault timing data with the same length as the input sequence are generated, containing information such as time, size and duration of the fault; the adversarial loss is used to measure the difference between the generated data and the real fault data, and the calculation method is as follows: first, the negative value of the expected value of the logarithm of the authenticity score of the discriminator for the real fault data is taken; then, the negative value of the expected value of the logarithm of the authenticity score of the discriminator for the generated data is taken; and the sum of the two is the adversarial loss, and the generator adjusts its parameters to minimize the loss, so that the generated data is closer to the distribution of the real fault data.

[0102] The step 2.4, the preliminary fault timing data optimized by the generator are input into the discriminator, and the validity of the preliminary fault timing data is verified by the discriminator according to the preset physical constraint rules; specifically, the verification process includes two main aspects, one is to verify the fluid dynamics coupling relationship reflected by the standardized data sequence contained in the preliminary fault timing data, to determine whether the fluid dynamics coupling relationship is consistent with the known physical law, so as to ensure that the generated data has rationality at the physical mechanism level; the other is to check whether there is a positive correlation between the fault amplitude and the environmental parameter correction factor value in the preliminary fault timing data, that is, whether the fault amplitude will increase correspondingly with the increase of the environmental parameter correction factor value, so as to further ensure that the correlation between the generated data and the environmental influence conforms to the actual situation.

[0103] The step 2.5, comprehensive confidence calculation, the comprehensive confidence is obtained by weighting two parts, one part is the score of the discriminator for the authenticity of the generated data distribution, ranging from 0 to 1, the closer to 1, the more real, the other part is the compliance degree of the physical constraint, such as the smaller the coupling verification deviation and the higher the correlation coefficient, the higher the compliance degree; the weights of the two parts are set according to the specific scene, and the weighted sum is the comprehensive confidence; when the comprehensive confidence is greater than a preset threshold, the generated data is determined to be valid, and is marked as extreme fault timing data under multi-source interference.

[0104] In a preferred embodiment of the present application, the step 3 can include:

[0105] Step 3.1, receiving the extreme fault timing data under multi-source interference, and inputting the extreme fault timing data into an initial multi-layer perception network;

[0106] Step 3.2, performing structured feature selection on the fully connected layer of the initial multi-layer perception network, and removing redundant neuron connections to generate a primary optimized network with compressed feature dimension;

[0107] Step 3.3, weight parameter dynamic tuning is performed on the primary optimization network, and the reconstruction error of compressed features and original fault data is taken as the fitness function, the inertia weight nonlinear decay mechanism and the global final disturbance strategy of the improved particle swarm algorithm are used to update the network parameters, and a parameter optimization network is generated;

[0108] Step 3.4, lightweight processing is performed on the parameter optimization network, the second-order derivative absolute value of the activation function is calculated as the neuron importance index, the neurons and connection edges with importance lower than the preset threshold are removed, and a pruned lightweight network is generated, and the output data of the lightweight network is marked as lightweight features.

[0109] In the embodiment of the application, by inputting the extreme fault time series data into the multilayer perceptron, the redundant neuron connections are removed by the final surgical algorithm to realize feature dimension compression and reduce invalid information interference; then the network parameters are optimized by the improved particle swarm algorithm dynamic parameter tuning to improve the accuracy of feature compression; finally, the low importance neurons are accurately removed based on the pruning strategy of the second-order derivative of the activation function, and the network is lightweighted under the premise of ensuring the integrity of the key features.

[0110] In the embodiment of the application, when specifically applied, the above steps can be realized by the following steps, for example:

[0111] In the above step 3.1, the extreme fault time series data under multi-source interference is received, which contains time series information such as fault occurrence time, amplitude, duration characteristics, etc., and is directly input into the initially constructed initial multilayer perceptron network, which is composed of an input layer, a plurality of fully connected hidden layers and an output layer, and signals are transmitted through neuron connections between layers.

[0112] In the above step 3.2, the importance of each neuron connection of the initial multilayer perceptron fully connected layer is calculated, and the importance value is , wherein, is the weight value of the neuron connection, is the second-order partial derivative of the loss function L with respect to the weight , that is, the second-order derivative; the neuron connections with importance values lower than the preset threshold are removed in order from small to large; this process is repeated, and the importance of the remaining connections is recalculated after each removal, until the feature dimension is compressed to the target range, a primary optimization network is generated, core features are retained, and redundant information is removed.

[0113] In the above step 3.3, the reconstruction error of the compressed features output by the primary optimization network and the original fault data is taken as the fitness, the reconstruction error usually adopts the mean square error, that is, the average value of the square of the difference between all samples of compressed features and corresponding values of original data; the inertia weight is nonlinearly reduced from the initial value to the minimum value with the number of iterations, for example, exponential decay, and the weight value is multiplied by a fixed decay coefficient after each iteration.

[0114] Velocity and position update, the velocity of each particle is updated as inertia weight x current velocity + cognitive factor x random number x (individual final parameter - current parameter) + social factor x random number x (global final parameter - current parameter); the position update is current position + new velocity, wherein the cognitive factor and the social factor are core parameters of the particle swarm optimization algorithm, and the essence is to quantify the importance of individual experience and group experience of the particle, the cognitive factor is used to regulate the learning strength of the particle to the historical final position of itself, and reflects the influence of individual cognition; the greater the value is, the more the particle tends to follow the final experience of itself in the past; the social factor is used to regulate the learning strength of the particle to the global final position of the group, and reflects the influence of social cooperation; the greater the value is, the more the particle tends to follow the common final experience of the group; every iteration for a certain number of times, a small amplitude random disturbance is added near the global final parameter, such as adding or subtracting a random number to the parameter value, the disturbance range decreases with the iteration number, avoiding falling into a local final, and the iteration optimization is performed until the fitness function converges, and the parameter optimization network is generated.

[0115] The above step 3.4, neuron importance calculation: the second derivative of the activation function output of each neuron in the parameter optimization network, such as ReLU, Sigmoid, etc., is calculated, and the absolute value thereof is taken as the importance index of the neuron; the greater the value is, the higher the sensitivity of the neuron to feature change is, and the more important the neuron is; pruning operation, set the importance threshold, remove all neurons whose importance index is lower than the threshold, and delete the input and output connections connected with these neurons; re-connect the remaining neurons and fine-tune the parameters to generate a lightweight network, and the output data of the lightweight network is a lightweight feature that retains core information.

[0116] In a preferred embodiment of the present application, the above step 4 can include:

[0117] Step 4.1, in the digital twin, the lightweight feature is fused with the fan topological structure data and the fan running state data acquired in real time according to the spatial position mapping relationship, and a space-time correlation matrix of the fault scene is constructed;

[0118] Step 4.2, inputting the space-time correlation matrix into a multi-physical field coupling simulator, taking the minimum power generation loss and the minimum repair cost as double optimization objectives, and through fluid-structure-power field coupling calculation, outputting the interaction influence value of the fault propagation path and the resource scheduling scheme;

[0119] Step 4.3, iteratively optimizing the scheduling parameters according to the interaction influence value; when the interaction influence value shows that the fault propagation risk is higher than a threshold, increasing the maintenance resource weight parameter; when the interaction influence value shows that the power generation loss exceeds a threshold, adjusting the power redistribution parameter; after dynamically updating the parameters, re-executing the coupling simulation until the double objectives achieve the Pareto optimal, and generating an adaptive scheduling strategy;

[0120] Step 4.4, load the adaptive scheduling strategy in the digital twin virtual mapping environment, inject historical extreme scene data and random disturbance variables for multiple rounds of deduction, and output the scheduling strategy that passes the verification when the reliability index of the strategy meets the failure suppression success rate greater than or equal to the preset threshold and the cost control deviation rate less than or equal to the preset threshold.

[0121] In the embodiment of the application, in the digital twin, by mapping the fusion lightweight features, fan topology structure and real-time running state data in the digital twin according to the spatial position, a space-time correlation matrix is constructed, ensuring the space-time consistency and correlation of the multi-source data of the fault scene; by means of multi-physical field coupling simulation, the interactive influence of fault propagation and resource scheduling is calculated, and the double objectives of minimizing power generation loss and repair cost are combined to realize comprehensive consideration of multi-factor interaction in complex scenarios; by dynamically iterating and optimizing the scheduling parameters, the strategy is converged to the Pareto optimal, and the adaptive ability of scheduling is improved; finally, the historical extreme scene and random disturbance are injected in the digital twin environment for multiple rounds of deduction verification, and the stability and actual applicability of the output scheduling strategy are guaranteed through strict reliability index screening, providing accurate and reliable decision support for the physical execution end.

[0122] In the embodiment of the application, when specifically applied, the above steps can be realized by the following steps, for example:

[0123] In the above step 4.1, in the virtual mapping environment of the digital twin, in order to comprehensively and accurately depict various types of associated information in the fault scene, the received lightweight features, the real-time obtained fan topology structure data and the fan running state data need to be fused, specifically, the fusion of the above data is carried out based on the spatial position mapping relationship, wherein the lightweight features are the core features of the compressed fault data, which can reflect the key characteristics of the fault; the fan topology structure data covers the spatial layout information of the spatial position coordinates, the mutual connection relationship and the cable direction of each component of the fan; the fan running state data includes real-time power, speed, temperature and other parameters that can reflect the current running condition of the equipment. Through the spatial position mapping relationship, each fan node is associated with a corresponding group of time series data, which contains the lightweight features and running state data of the node, and the topology structure data is taken as the weight parameter of the connection relationship between nodes, on this basis, the space-time correlation matrix of the fault scene is constructed.

[0124] In the process of processing the space-time correlation matrix by the multi-physical field coupling simulator, the core is to realize the dynamic simulation of the running state and the fault influence of the fan by constructing the correlation model among the fluid field, the structure field and the power field, and the specific calculation logic of each physical field is derived based on the correlation.

[0125] Specifically, the calculation of the fluid field starts from the interaction between the fan blade and the airflow, wherein the blade wind pressure of the fan at time τ is the key parameter for the fluid field to act on the structure field, and the calculation logic is as follows: first, according to the principle of aerodynamics, considering the air density and the wind speed at this time, the wind speed data is derived from the real-time running state data in the space-time correlation matrix, multiplying the square of the wind speed by 0.5 times the air density to obtain the basic wind pressure factor; then, combined with the current structure state of the blade, such as whether the blade has deformation, wear and the like, the relevant information is implied in the lightweight feature and the topological structure data, the blade drag coefficient is determined, and the structural parameter of the blade windward area is introduced, and finally the basic wind pressure factor, the drag coefficient and the windward area are multiplied to obtain the blade wind pressure at time τ.

[0126] Based on the blade wind pressure calculated by the fluid field, the calculation logic of the blade stress in the structure field is further derived. The blade stress of the fan i at time τ is essentially the mechanical response of the blade under the action of wind pressure, and its size is directly related to the blade wind pressure at this time: multiplying the blade wind pressure by the blade length to obtain the torque generated by the wind pressure on the blade; then, combined with the blade section bending modulus reflecting the ability of the blade structure to resist bending deformation, which is determined by the blade parameters in the fan topological structure data, the blade stress at time τ can be obtained by the ratio of the torque to the bending modulus.

[0127] The power generation calculation of the power field is closely related to the motion state of the blade in the structure field. The power generation of the fan i at time τ depends on the blade speed and the voltage, and the voltage data is derived from the real-time running state data. Specifically, the product of the blade speed and the voltage is converted into the actual power generation by the conversion coefficient. The blade speed is affected by both the wind speed and the stress: the higher the wind speed, the greater the driving force obtained by the blade, and the speed tends to rise; but the greater the blade stress, the closer the load borne by the structure to the limit, in order to avoid damage, the speed will be reduced accordingly, so the blade speed is proportional to the product of the wind speed and (1 minus the ratio of the current stress to the maximum allowable stress), which directly links the mechanical state of the structure field to the operating parameters of the power field. The vibration of the blade in the structure field will change the surrounding airflow state, resulting in a local wind speed lower than the original wind speed, and the greater the vibration displacement, the lower the wind speed, thereby inversely affecting the calculation of the fluid field; the change of the power generation in the power field will affect the motor counter-torque, the greater the power generation, the greater the counter-torque, which will further produce a braking effect on the rotation state of the blade in the structure field, forming a closed loop coupling among the multi-physical fields.

[0128] Dual optimization target quantification:

[0129] Power generation loss, the difference between actual power generation in a certain period of time under fault state and theoretical power generation in the same period of time under normal operation, that is, power generation loss = normal theoretical power generation - actual power generation under fault;

[0130] Repair cost, the total cost generated in the fault repair process, including labor cost, spare parts cost, transportation cost, etc., that is, repair cost = labor cost + spare parts cost + transportation cost + other miscellaneous expenses.

[0131] Interaction value calculation, the interaction value is a comprehensive quantitative index of the dual target, reflecting the interaction between the fault propagation path and the resource scheduling scheme. The calculation formula is interaction value = (power generation loss x a) + (repair cost x b), where a and b are preset weights, a + b = 1, and the initial value is set according to the importance of the scene, such as a > b for power priority.

[0132] The above step 4.3, when the interaction value shows that the fault propagation risk is higher than the preset threshold, it means that the current maintenance resource input is not enough to effectively suppress the fault diffusion, at this time, the maintenance resource weight parameter needs to be increased to preferentially allocate more maintenance resources; when the interaction value shows that the power generation loss exceeds the preset threshold, it means that the current power allocation cannot make up for the power gap caused by the fault, at this time, the power allocation parameter needs to be adjusted to increase the upper limit of the power allocation of the non-fault wind turbine. After dynamically updating the parameters, the multi-physical field coupling simulation is re-executed, and this is repeated until the dual target reaches the Pareto optimal, that is, it is impossible to reduce the power generation loss while not increasing the repair cost, or it is impossible to reduce the repair cost while not increasing the power generation loss, at this time, the obtained scheduling parameter combination is the adaptive scheduling strategy.

[0133] The above step 4.4 loads the adaptive scheduling strategy generated above in the digital twin virtual mapping environment. In order to verify the reliability of the adaptive scheduling strategy under different scenarios, historical extreme scenario data and random disturbance variables are injected for multiple rounds of deduction. In the deduction process, the reliability indicators of the strategy are calculated, wherein the fault suppression success rate is the ratio of the number of successfully suppressed faults to the total number of deduction faults, and successful suppression means that the fault does not spread and is repaired within the preset time; the cost control deviation rate is the absolute value of the difference between the actual repair cost and the planned repair cost divided by the planned repair cost. When the reliability indicators of the strategy meet the conditions that the fault suppression success rate is greater than or equal to the preset threshold and the cost control deviation rate is less than or equal to the preset threshold, it is determined that the adaptive scheduling strategy is effective, and the scheduling strategy that passes the verification is output. If any indicator does not meet the preset threshold, the strategy is adjusted again in the optimization link.

[0134] In a preferred embodiment of the present application, the above step 5 can include:

[0135] Step 5.1, receiving an adaptive scheduling strategy, encapsulating into an executable instruction set according to a physical execution end protocol;

[0136] Step 5.2, issuing the instruction set to the target fan physical execution end through a preset encrypted channel, and synchronously starting an end-to-end time delay monitoring clock;

[0137] Step 5.3, the physical execution end parses the instruction set and performs device start-stop, power adjustment and maintenance resource scheduling operations, and returns a status code in real time;

[0138] Step 5.4, judging the execution integrity according to the status code, if all operation success feedback codes are received and the time delay does not exceed the limit, marking the strategy execution success; if a failure feedback code is detected or the time delay exceeds the timeliness threshold, triggering the strategy rollback mechanism and generating a fault alarm;

[0139] Step 5.5, collecting actual power generation data and maintenance cost data of the physical execution end, comparing and verifying with the decision accuracy index, and generating an execution efficiency evaluation report.

[0140] In the embodiment of the present application, by encapsulating the instruction set according to the physical execution end protocol, the instruction compatibility and resolvability are ensured, by means of the preset encrypted channel, the safety of strategy transmission is ensured, and the synchronous start of the time delay monitoring realizes the accurate control of the execution timeliness; the physical execution end returns the status code in real time, which can dynamically track the operation progress, based on the double judgment mechanism of the status code and the time delay, the execution success is quickly confirmed, in the case of failure or timeout, the rollback and alarm are triggered, reducing the risk of fault expansion; finally, through the comparison and verification of the actual data and the decision index, the evaluation report is generated, forming a closed loop feedback from strategy issuing to execution efficiency evaluation.

[0141] In the embodiment of the present application, when specifically applied, the above steps can be implemented by the following steps, for example:

[0142] The above step 5.1, receiving an adaptive scheduling strategy that has passed the virtual environment verification, in order to ensure that the strategy can be accurately recognized and executed by the physical execution end, the adaptive scheduling strategy needs to be encapsulated according to the communication protocol and data format requirements followed by the physical execution end, forming a complete executable instruction set, the encapsulation process includes format conversion of each operation instruction in the strategy, adding necessary check information.

[0143] After the encapsulation of the instruction set is completed, the instruction set is issued to the physical execution end of the target fan through a pre-established encrypted channel in step 5.2 above, and the purpose of using the encrypted channel is to prevent the instruction from being illegally intercepted or tampered with during transmission, so as to ensure the security of the strategy transmission. At the same time, a start-to-end time delay monitoring clock is started synchronously, so as to record the time interval between the time when the instruction is issued and the time when the execution result is fed back from the physical execution end in real time.

[0144] After receiving the issued instruction set, the physical execution end parses the instruction set, extracts the specific operation information such as device start-stop, power adjustment and maintenance resource scheduling contained in the instruction set, and executes the corresponding operation according to the parsed operation requirement in step 5.3 above. During the execution process, the physical execution end will return the execution state of each operation to the control center in the form of a state code in real time.

[0145] The control center judges the execution integrity of the instruction set according to the state code returned by the physical execution end in step 5.4 above. Specifically, if the received state code is the successful feedback code of all operations, and the time delay recorded by the start-to-end time delay monitoring clock does not exceed the preset timeliness threshold, it is marked that the adaptive scheduling strategy is executed successfully. Otherwise, if the failure feedback code of any operation is detected, or the time delay exceeds the preset timeliness threshold, the strategy rollback mechanism is triggered immediately to cancel or correct the executed part of the operation, so as to avoid the expansion of the fault, and at the same time, the fault alarm information is generated to prompt the relevant personnel to handle it in time.

[0146] In step 5.5 above, the actual power generation data and maintenance cost data generated during the execution process of the physical execution end are collected, the actual data are compared and verified with the decision accuracy index relied on when the adaptive scheduling strategy is formulated, the deviation between the actual result and the expected target is analyzed, and finally, the execution efficiency evaluation report is generated according to the comparison and verification result.

[0147] As shown in Figure 2 The embodiment of the application also provides a scene information reverse transmission system based on digital twinning, which comprises:

[0148] The acquisition module is configured to perform space-time alignment and outlier processing on the collected multi-source data to form a standardized data sequence, select a reference data segment in the standardized data sequence, generate two feature sub-sequences based on the reference data segment, calculate the dynamic correlation of the two feature sub-sequences to form a feature interval, select one key data unit in the feature interval and two key data units outside the feature interval, construct a nonlinear data trajectory according to the time continuity of the three key data units, and generate a correction factor according to the nonlinear data trajectory.

[0149] The input module is configured to input the correction factor and the standardized data sequence into the generative adversarial network to generate extreme fault time series data under multi-source interference.

[0150] The calculation module is configured to perform feature selection and dimension compression according to the fault time series data, dynamically adjust parameters by combining the improved particle swarm algorithm, and realize calculation lightening by using the second-order derivative pruning to obtain light-weight features. In the digital twin, the light-weight features are fused with the fan topology structure and real-time running state data to establish a space-time correlation matrix of the fault scene. The interactive influence of the fault propagation path and the resource scheduling scheme is calculated by multi-physical field coupling simulation with the dual objectives of minimizing power generation loss and repair cost. The scheduling parameters are iteratively optimized based on the interactive influence to generate an adaptive scheduling strategy. The multi-scenario deduction verification is performed in a virtual mapping environment, and the scheduling strategy that passes the verification is output when the reliability index of the strategy reaches a preset threshold.

[0151] The processing module is configured to issue the scheduling strategy that passes the verification to a physical execution end for execution.

[0152] The above is the preferred embodiment of the present application. It should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should also be considered within the scope of protection of the present application.

Claims

1. A method for reverse transmission of scenario information based on digital twinning, characterized in that, The method comprises: Step 1, the collected multi-source data is subjected to space-time alignment and outlier processing to form a standardized data sequence, a reference data segment is selected in the standardized data sequence, two characteristic subsequences are generated based on the reference data segment, the dynamic correlation of the two characteristic subsequences is calculated to form a characteristic interval; a key data unit is selected within the characteristic interval, and two key data units are selected outside the characteristic interval; a nonlinear data trajectory is constructed according to the time continuity of the three key data units; and a correction factor is generated according to the nonlinear data trajectory; Wherein, based on the reference data segment, two characteristic subsequences are generated, including: the first characteristic subsequence extracts time domain statistical features through a sliding window, and the second characteristic subsequence extracts frequency energy distribution features through fast Fourier transform; Wherein, within the characteristic interval, a key data unit is selected, and two key data units are selected outside the characteristic interval; a nonlinear data trajectory is constructed according to the time continuity of the three key data units; and a correction factor is generated according to the nonlinear data trajectory, including: Within the characteristic interval, the data unit corresponding to the mutual information peak time is selected as the first key data unit; within a preset time period before the start time of the characteristic interval, the data unit with the lowest statistical feature fluctuation rate is selected as the second key data unit; within a preset time period after the end time of the characteristic interval, the extreme point of the frequency energy change gradient is selected as the third key data unit; Based on the time stamp continuity of the three key data units, a nonlinear data trajectory spanning the boundary of the characteristic interval is generated by fitting through a third order Bezier curve with the second key data unit as the starting point, the first key data unit as the relay point, and the third key data unit as the end point; For the nonlinear data trajectory, the Euclidean distance of adjacent points on the path is calculated and continuously accumulated to obtain the total length of the trajectory; the absolute value of the curvature of each point on the trajectory is also calculated; the product of the total length of the trajectory and the absolute value of the curvature is normalized to generate an environmental parameter correction factor; Step 2, the correction factor and the standardized data sequence are jointly input into a generative adversarial network to generate extreme fault time series data under multi-source interference; Step 3, according to the fault time series data, feature selection and dimension compression are performed, the improved particle swarm algorithm is used for dynamic parameter adjustment, and the second derivative pruning is used to realize calculation lightweight to obtain lightweight features; Step 4, in the digital twin, the lightweight features are fused with the fan topology structure and real-time running state data to establish a space-time correlation matrix of the fault scene; the interactive influence of the fault propagation path and the resource scheduling scheme is calculated through multi-physical field coupling simulation with the dual objectives of minimizing power generation loss and repair cost; the scheduling parameters are iteratively optimized based on the interactive influence to generate an adaptive scheduling strategy; the multi-scenario deduction verification is performed in the virtual mapping environment, and when the reliability index of the strategy reaches a preset threshold, the scheduling strategy that passes the verification is output; Step 5, the scheduling strategy that passes the verification is sent to the physical execution end for execution. 2.The digital-twin-based scene information reverse transmission method according to claim 1, characterized in that, The collected multi-source data is spatio-temporally aligned and outlier processed to form a standardized data sequence, a reference data segment is selected in the standardized data sequence, two feature subsequences are generated based on the reference data segment, and the dynamic correlation of the two feature subsequences is calculated to form a feature interval, including: Step 1.1, spatio-temporal alignment of received heterogeneous sensor data, alignment of data streams based on unified timestamps, and filling of missing position data through spatial interpolation; Step 1.2, processing outliers using quantile outlier detection algorithm, Gaussian smoothing correction for data points exceeding three times the standard deviation, and generating a standardized data sequence; Step 1.3, in the standardized data sequence, select a preset length of continuous and stable running data as a reference data segment; based on the reference data segment, perform double feature extraction, the first feature subsequence extracts time domain statistical features through sliding window, and the second feature subsequence extracts frequency energy distribution features through fast Fourier transform; Step 1.4, calculate the dynamic time-varying correlation of the two feature subsequences, use windowed mutual information algorithm to calculate local mutual information value with preset step size, and mark the time period with mutual information value continuously higher than the threshold as the feature interval. 3.The digital-twin-based scene information reverse transmission method according to claim 2, characterized in that, The modified factor and the standardized data sequence are jointly input into the generative adversarial network to generate extreme fault time series data under multi-source interference, including: Step 2.1, align the environmental parameter correction factor with the standardized data sequence according to the timestamp to construct a conditional input vector; Step 2.2, input the conditional input vector into the generator of the generative adversarial network, map it to the latent space modulation signal through the full connection layer, and concatenate the latent space modulation signal with the random noise vector in the channel dimension to form the combined input data of the generator; Step 2.3, perform time convolution decoding operation on the combined input data to generate preliminary fault time series data, and optimize the generator parameters based on the adversarial loss feedback of the discriminator to adjust the distribution of the preliminary fault time series data; Step 2.4, input the preliminary fault time series data into the discriminator, verify the data validity through the preset physical constraint rule, first verify whether the fluid dynamics coupling relationship of the standardized data sequence conforms to the physical law, and then check the positive correlation between the fault amplitude and the correction factor value; Step 2.5, when the comprehensive confidence of the discriminator output for the verified preliminary fault time series data reaches the preset threshold, it is determined that the data output by the current generator meets the requirements, and it is marked as extreme fault time series data under multi-source interference.

4. The digital-twin-based scene information reverse transmission method according to claim 3, characterized in that, According to the fault time series data, select features and compress dimensions, combine the improved particle swarm algorithm for dynamic parameter adjustment, and use the second derivative pruning to realize lightweight calculation to obtain lightweight features, including: Step 3.1, receive extreme fault time series data under multi-source interference, and input it into the initial multilayer perceptron network; Step 3.2, perform structured feature selection on the full connection layer of the initial multilayer perceptron network, and remove redundant neuron connections to generate a primary optimized network with compressed feature dimensions; Step 3.

3. Perform weight parameter dynamic optimization on the primary optimized network to compress the reconstruction error of the features and the original fault data as the fitness function, update the network parameters by improving the inertia weight nonlinear decay mechanism and global final disturbance strategy of the particle swarm algorithm, and generate a parameter optimized network; Step 3.

4. Perform lightweight processing on the parameter optimized network, calculate the second derivative absolute value of the activation function as the neuron importance indicator, remove the neurons and connection edges with importance lower than the preset threshold, and generate a pruned lightweight network. The output data of the lightweight network is marked as lightweight features.

5. The digital-twin-based scene information reverse transmission method according to claim 4, characterized in that, The step 4 comprises: Step 4.

1. In the digital twin, fuse the lightweight features with the fan topology structure data and the fan operation state data acquired in real time according to the spatial position mapping relationship to construct a space-time correlation matrix of the fault scene; Step 4.

2. Input the space-time correlation matrix into a multi-physical field coupling simulator to minimize the power generation loss and the repair cost as double optimization objectives, and output the interactive influence value of the fault propagation path and the resource scheduling scheme through fluid-structure-power field coupling calculation; Step 4.

3. Iteratively optimize the scheduling parameters according to the interactive influence value. When the interactive influence value shows that the fault propagation risk is higher than the threshold, increase the maintenance resource weight parameter. When the interactive influence value shows that the power generation loss exceeds the threshold, adjust the power redistribution parameter. Dynamically update the parameters and re-execute the coupling simulation to generate an adaptive scheduling strategy; Step 4.

4. Load the adaptive scheduling strategy in the digital twin virtual mapping environment, inject historical extreme scene data and random disturbance variables for multiple rounds of deduction. When the reliability indicators of the strategy meet the fault suppression success rate greater than or equal to the preset threshold and the cost control deviation rate less than or equal to the preset threshold, output the scheduling strategy that passes the verification.

6. The digital-twin-based scene information reverse transmission method according to claim 5, characterized in that, The scheduling strategy that passes the verification is issued to the physical execution end for execution, including: Step 5.

1. Receive the adaptive scheduling strategy and encapsulate it into an executable instruction set according to the physical execution end protocol; Step 5.

2. Issue the instruction set to the physical execution end of the target fan through a preset encryption channel, and start the end-to-end time delay monitoring clock synchronously; Step 5.

3. The physical execution end parses the instruction set and performs device start-stop, power adjustment and maintenance resource scheduling operations, and returns the status code in real time; Step 5.

4. Judge the execution integrity according to the status code. If the success feedback code of all operations is received and the time delay is not over the limit, mark the strategy execution as successful. If a failure feedback code is detected or the time delay exceeds the timeliness threshold, trigger the strategy rollback mechanism and generate a fault alarm; Step 5.

5. Collect the actual power generation data and maintenance cost data of the physical execution end, and compare them with the decision accuracy indicators to generate an execution efficiency evaluation report.

7. A digital-twin-based system for reverse transmission of scenario information, the system implementing the method of any one of claims 1 to 6, characterized in that, The method comprises: The acquisition module is configured to perform space-time alignment and outlier processing on the collected multi-source data to form a standardized data sequence, select a reference data segment in the standardized data sequence, generate two feature subsequences based on the reference data segment, and calculate the dynamic correlation of the two feature subsequences to form a feature interval. ​ select one key data unit in the feature interval, and select two key data units outside the feature interval; construct a nonlinear data track according to the time continuity of the three key data units; According to the nonlinear data track, a correction factor is generated; wherein, two feature subsequences are generated based on the reference data segment, and the two feature subsequences include: the first feature subsequence extracts time domain statistical features through a sliding window and the second feature subsequence extracts frequency energy distribution features through a fast Fourier transform; wherein, one key data unit is selected within the feature interval, and two key data units are selected outside the feature interval; a nonlinear data track is constructed according to the time continuity of the three key data units; a correction factor is generated according to the nonlinear data track, including: in the feature interval, the data unit corresponding to the time of the mutual information peak is selected as the first key data unit; in a preset time period before the start time of the feature interval, the data unit with the lowest statistical feature fluctuation rate is selected as the second key data unit; in a preset time period after the end time of the feature interval, the extreme point of the frequency energy change gradient is selected as the third key data unit; based on the time stamp continuity of the three key data units, taking the second key data unit as the starting point, the first key data unit as the relay point, and the third key data unit as the end point, a nonlinear data track spanning the boundary of the feature interval is generated through a third-order Bezier curve fitting; for the nonlinear data track, the Euclidean distance of adjacent points on the path is calculated and continuously accumulated to obtain the total length of the track; the absolute value of the curvature of each point on the track is also calculated; the product of the total length of the track and the absolute value of the curvature is normalized to generate an environmental parameter correction factor; The input module is used to input the correction factor and the standardized data sequence into the generative adversarial network to generate extreme fault time series data under multi-source interference; The calculation module is used to select features and compress dimensions according to the fault time series data, dynamically adjust parameters combined with the improved particle swarm algorithm, and realize lightweight calculation by using second-order derivative pruning to obtain lightweight features; in the digital twin, the lightweight features are fused with the fan topology structure and real-time running state data to establish a spatio-temporal correlation matrix of the fault scene; the interactive influence of the fault propagation path and the resource scheduling scheme is calculated through multi-physical field coupling simulation with the dual objectives of minimizing power generation loss and repair cost; the scheduling parameters are iteratively optimized based on the interactive influence to generate an adaptive scheduling strategy; in the virtual mapping environment, multiple scenarios are deduced and verified, and when the reliability index of the strategy reaches a preset threshold, the scheduling strategy that passes the verification is output. The processing module is used to issue the scheduling strategy that passes the verification to the physical execution end for execution.

8. A computing device, comprising: It includes: One or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, so that the one or more processors implement the method of any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a program which is executed by a processor to implement the method of any one of claims 1-6.

Images