Renewable energy hydrogen production-based energy green authentication evaluation method

Abstract

The invention provides a renewable energy hydrogen production-based energy green authentication evaluation method, and relates to the technical field of renewable energy hydrogen production, and the method comprises the steps: collecting the operation data of a hydrogen production system to form a time series data set, carrying out the wavelet decomposition of the time series data set, extracting features, and generating a feature fingerprint spectrum; a variational auto-encoder is used to carry out dimension reduction reconstruction and form a dynamic evaluation benchmark, gradient strategy optimization and time sequence causal inference are fused to predict evaluation parameters, and finally a green authentication score is calculated and an evaluation report is generated and recorded in a block chain. According to the invention, intelligent evaluation of green authentication of the hydrogen production system is realized, and the accuracy and credibility of evaluation are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrogen production from renewable energy, and in particular to a green certification evaluation method for hydrogen production from renewable energy. Background Art

[0002] With the advancement of energy transition and carbon neutrality goals, renewable energy hydrogen production is gaining increasing attention as a key form of clean energy carrier. Renewable energy hydrogen production systems use renewable energy sources such as photovoltaics and wind power to produce hydrogen through water electrolysis, achieving clean energy conversion and storage. To ensure the green attributes of hydrogen energy, a comprehensive and scientific green certification and evaluation system is needed to quantitatively assess indicators such as renewable energy utilization and system efficiency throughout the hydrogen production process.

[0003] However, the existing renewable energy hydrogen production evaluation methods have the following main shortcomings: the evaluation index system is static and rigid, and it is difficult to adapt to the dynamic characteristics of the system under different operating conditions; the collection and processing of evaluation data lacks systematicity, and cannot effectively identify and eliminate the interference caused by abnormal fluctuations; the construction method of the evaluation benchmark is simple and does not fully consider the adaptability issues under the operating condition migration scenario; the credibility and traceability of the evaluation results are poor, which is difficult to support the credibility requirements of green certification.

[0004] In summary, by constructing a feature extraction method based on multi-scale time-frequency analysis, designing a dynamic evaluation benchmark with operating condition migration capabilities, integrating an evaluation parameter prediction method with gradient optimization and temporal causal inference, and integrating blockchain technology to ensure credible evidence of the evaluation process, this method achieves a scientific, dynamic, and reliable green certification evaluation of renewable energy hydrogen production systems. This invention can solve the problems of the existing technology. Summary of the Invention

[0005] The embodiment of the present invention provides a green certification evaluation method for hydrogen-based energy produced from renewable energy, which can solve the problems in the existing technology.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provided is a green certification evaluation method for renewable energy hydrogen-based energy, including:

[0008] Collect operational data of renewable energy hydrogen production systems, including hydrogen production process parameters, renewable energy input parameters, and energy conversion efficiency parameters, to form a time series data set;

[0009] Perform wavelet decomposition on the time series data set to extract multi-scale time-frequency features, combine entropy analysis to identify the dynamic characteristics of the system, eliminate abnormal fluctuations through adaptive thresholds, and generate the system's characteristic fingerprint spectrum;

[0010] A variational autoencoder is used to reduce the dimension of the feature fingerprint spectrum and reconstruct it, extracting the potential feature vector. Through adversarial training, the feature vector of the known working condition is transferred to the new working condition to form a dynamic evaluation benchmark.

[0011] Based on dynamic evaluation benchmarks, it integrates gradient strategy optimization and temporal causal inference to predict optimal evaluation parameters, corrects prediction deviations through confidence interval calibration, and outputs certification scoring standards.

[0012] The green certification score of the system is calculated according to the certification scoring standards, and an evaluation report is generated including a time series data set, a feature fingerprint spectrum, a dynamic evaluation benchmark, a certification scoring standard and a green certification score, and recorded in the blockchain to form a certification chain.

[0013] In an optional embodiment,

[0014] Perform wavelet decomposition on the time series data set to extract multi-scale time-frequency features. Combined with entropy analysis, the dynamic characteristics of the system are identified. Abnormal fluctuations are eliminated through adaptive thresholds to generate the system's characteristic fingerprint spectrum, including:

[0015] Performing discrete wavelet transform on the time series data using Daubechies wavelet basis functions, generating wavelet coefficients and scaling coefficients based on a wavelet mother function, and decomposing the time series data into multiple frequency sub-signals according to the wavelet coefficients and scaling coefficients;

[0016] Calculating the sum of squares of the wavelet coefficients to obtain an energy eigenvalue, extracting the mean, standard deviation, skewness, and kurtosis of the frequency sub-signal to determine a statistical eigenvalue, calculating the derivative of the phase function of the frequency sub-signal to obtain an instantaneous frequency feature, and integrating the frequency sub-signal in the time-frequency domain to obtain an edge spectrum feature;

[0017] Constructing a multidimensional vector and a multi-resolution tree structure for the frequency sub-signal, calculating sample entropy and permutation entropy respectively, performing coarse-graining processing on the frequency sub-signal to obtain a multi-scale sequence, and obtaining a multi-scale entropy feature vector through sparse adaptive dictionary optimization;

[0018] Calculating an anomaly score of the frequency sub-signal using a local outlier factor, determining a dynamic threshold based on a mean and a standard deviation of the anomaly score, and eliminating abnormal fluctuations according to the dynamic threshold;

[0019] The energy eigenvalue, the statistical eigenvalue, the instantaneous frequency feature, the edge spectrum feature and the multi-scale entropy eigenvector are combined to form a feature matrix, the feature matrix is normalized by maximum and minimum values to obtain a normalized feature matrix, and the t-SNE method is used to perform dimensionality reduction processing on the normalized feature matrix to generate a feature fingerprint spectrum.

[0020] In an optional embodiment,

[0021] Constructing a multidimensional vector and a multi-resolution tree structure for the frequency sub-signal, calculating the sample entropy and permutation entropy respectively, performing coarse-graining processing on the frequency sub-signal to obtain a multi-scale sequence, and obtaining a multi-scale entropy feature vector through sparse adaptive dictionary optimization includes:

[0022] Reconstructing the frequency sub-signal into a multidimensional vector, the multidimensional vector comprising continuous data points at a preset time delay, calculating a Chebyshev distance based on the multidimensional vector, performing pattern matching counting on the Chebyshev distance according to a preset similarity threshold, and calculating a sample entropy value based on the pattern matching count;

[0023] A recursive bisection method is used to divide the frequency sub-signal into sub-sequences of different scales, and a hierarchical tree structure is established, in which each node contains a sequence fragment at a corresponding scale, forming a multi-resolution sub-sequence tree structure. Local sorting is performed on each node of the multi-resolution sub-sequence tree structure to obtain a sequence arrangement pattern, and the conditional probability distribution of the sequence arrangement pattern at nodes at different levels is statistically analyzed, and the arrangement entropy value is calculated based on the conditional probability distribution;

[0024] Performing coarse-graining processing on the frequency sub-signal to obtain a multi-scale sequence, respectively calculating a sample entropy value sequence and a permutation entropy value sequence of the multi-scale sequence, and combining the sample entropy value sequence and the permutation entropy value sequence into a multi-scale entropy feature vector;

[0025] Initialize the atomic basis matrix, block-process the multi-scale entropy feature vector to obtain a training sample set, iteratively optimize the atomic basis matrix using the training sample set, update the sparse coefficients using the least squares method, judge the convergence based on the change of the sparse coefficients to obtain a sparse adaptive dictionary, and use the sparse adaptive dictionary to perform sparse decomposition on the multi-scale entropy feature vector to obtain an optimized multi-scale entropy feature vector.

[0026] In an optional embodiment,

[0027] The variational autoencoder is used to reduce the dimension of the feature fingerprint spectrum and reconstruct the potential feature vector. The feature vector of the known working condition is transferred to the new working condition through adversarial training to form a dynamic evaluation benchmark including:

[0028] Input the characteristic fingerprint spectrum into the encoder of the variational autoencoder, use wavelet transform to decompose the characteristic fingerprint spectrum into characteristic components of different frequency bands, perform nonlinear function mapping on the characteristic components, obtain the encoding mean vector and the encoding variance vector through weighted average aggregation, and sample to obtain the potential feature vector, input the potential feature vector into the decoder for reconstruction, optimize the reconstruction error and KL divergence until the parameters of the variational autoencoder converge, and obtain the reduced-dimensional potential feature vector;

[0029] Inputting the reduced-dimensional potential feature vector into a feature generator to generate a new operating condition feature vector, performing operating condition attribute discrimination through a domain discriminator, calculating an adversarial training loss, re-inputting the new operating condition feature vector into a feature generator to calculate a cycle consistency loss, optimizing the parameters of the feature generator based on the adversarial training loss and the cycle consistency loss, and finally obtaining a stable operating condition feature vector;

[0030] A Mahalanobis distance metric function is constructed for the stable working condition feature vector, and the feature sequence evolution trajectory is calculated based on the symbolic dynamics method. The working condition state transition characteristics are extracted to obtain the distance mean and distance standard deviation of the working condition. The weighted sum of the distance mean and the distance standard deviation is used as the adaptive threshold to determine the dynamic evaluation benchmark.

[0031] In an optional embodiment,

[0032] The Mahalanobis distance metric function is constructed for the stable working condition feature vector, and the feature sequence evolution trajectory is calculated based on the symbolic dynamics method to extract the working condition state transition characteristics. The distance mean and distance standard deviation of the working condition are obtained, including:

[0033] Calculating the Mahalanobis distance between stable operating condition feature vectors based on the Mahalanobis distance metric function, and constructing an operating condition feature sequence based on the Mahalanobis distance;

[0034] performing adaptive segmentation processing on the operating condition feature sequence to obtain a plurality of operating condition sequence segments, and performing linear fitting on the operating condition sequence segments to obtain an operating condition slope sequence;

[0035] Inputting the operating condition slope sequence into a symbol mapping function for symbolization processing to obtain an operating condition state sequence, wherein the symbol mapping function adaptively determines a slope threshold based on the variance of the operating condition slope sequence;

[0036] Constructing a Markov state transfer matrix based on the operating state sequence and mapping it into an operating state transfer network, wherein the nodes of the operating state transfer network represent operating state combinations, and the edges of the operating state transfer network represent operating state transfer relationships;

[0037] Calculating the duration distribution of the operating state in the operating state transition network, and extracting the average duration of the operating state, the standard deviation of the operating state duration, and the operating state transition entropy from the duration distribution;

[0038] The average duration of the operating condition, the standard deviation of the operating condition duration and the operating condition state transition entropy are weightedly combined to obtain an operating condition characteristic distance metric, and the distance mean and distance standard deviation of the operating condition are extracted from the operating condition characteristic distance metric.

[0039] In an optional embodiment,

[0040] Based on dynamic evaluation benchmarks, we integrate gradient strategy optimization and temporal causal inference to predict optimal evaluation parameters. We use confidence interval calibration to correct prediction deviations and output certification scoring criteria including:

[0041] Constructing a time series feature vector based on the dynamic evaluation benchmark, the time series feature vector including a benchmark value, a first-order difference of the benchmark value, a second-order difference of the benchmark value, and a time scale factor;

[0042] Input the time series feature vector into the gradient strategy optimization unit, and obtain the optimization evaluation parameter by constructing the discounted reward function and iteratively optimizing the strategy parameters; input the time series feature vector into the time series causal inference unit, and obtain the causal evaluation parameter by constructing the time series causal graph and calculating the causal effect strength;

[0043] Performing adaptive weighted combination on the optimization evaluation parameter and the causal evaluation parameter to obtain a prediction evaluation parameter;

[0044] A multi-dimensional parameter constraint set is established based on the prediction evaluation parameters. The benchmark sensitivity index is introduced to construct the benchmark influence function. The gradient iteration of the benchmark influence is used to optimize the prediction evaluation parameters. The benchmark influence change rate is used as the iteration termination condition to obtain the optimal evaluation parameters.

[0045] Calculating the confidence interval of the predicted evaluation parameter and calibrating the predicted evaluation parameter using the deviation calibration coefficient calculated from historical data to obtain the optimal evaluation parameter;

[0046] Based on the optimal evaluation parameters, a weighted combination of normalized score, confidence score and stability score is used to generate the certification score standard.

[0047] In an optional embodiment,

[0048] A multi-dimensional parameter constraint set is established based on the prediction evaluation parameters. The benchmark sensitivity index is introduced to construct the benchmark influence function. The gradient iteration of the benchmark influence is used to optimize the prediction evaluation parameters. The benchmark influence change rate is used as the iteration termination condition. The optimal evaluation parameters include:

[0049] Constructing a parameter constraint set based on the prediction evaluation parameters, wherein the parameter constraint set includes parameter range constraints, parameter change rate constraints, and parameter correlation constraints;

[0050] Calculating the partial derivative of the predicted evaluation parameter with respect to the evaluation benchmark to obtain a first calculation result, calculating the ratio of the predicted evaluation parameter to the evaluation benchmark to obtain a second calculation result, multiplying the first calculation result by the second calculation result to obtain a benchmark sensitivity index, and constructing a benchmark influence function based on the benchmark sensitivity index, where the benchmark influence function is a weighted sum of the benchmark sensitivity indexes of each dimension;

[0051] Taking the predicted evaluation parameter as an initial parameter value, verifying whether the initial parameter value satisfies the constraint condition corresponding to the parameter constraint set, and calculating the benchmark influence corresponding to the initial parameter value using a benchmark influence function;

[0052] Based on the gradient direction of the benchmark influence, the initial parameter value is iteratively updated to obtain the iterative parameter value, the first benchmark influence corresponding to the iterative parameter value under the current number of iterations is calculated, the second benchmark influence corresponding to the iterative parameter value under the previous number of iterations is calculated, the difference between the first benchmark influence and the second benchmark influence is calculated, and the result is divided by the second benchmark influence to obtain the benchmark influence change rate. When the benchmark influence change rate is less than the preset change threshold, the current iterative parameter value is determined as the optimal evaluation parameter.

[0053] In an embodiment of the present invention, by performing multi-scale time-frequency analysis and entropy feature extraction on the operating data of the renewable energy hydrogen production system, and combining adaptive thresholds to eliminate abnormal fluctuations, a system feature fingerprint spectrum is established, which improves the accuracy and robustness of the data feature expression and provides a reliable data basis for green certification evaluation; a variational autoencoder is used for feature dimensionality reduction and reconstruction, and adversarial training is used to achieve working condition feature migration, a dynamic evaluation benchmark is established, and gradient strategy optimization and time series causal inference are integrated to predict the optimal evaluation parameters, thereby improving the adaptability of the evaluation method to different working conditions and the accuracy of the evaluation results; the data and results of the entire evaluation process are recorded in the blockchain to form a certification chain, which realizes the traceability and non-tamperability of the evaluation process and results, improves the credibility of green certification, and provides scientific and reliable technical support for the green certification evaluation of renewable energy hydrogen production systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flow chart of a green certification and evaluation method for renewable energy hydrogen-based energy according to an embodiment of the present invention;

[0055] Figure 2 Analyze radar charts for feature extraction performance and computational complexity;

[0056] Figure 3 It is the working condition duration distribution curve;

[0057] Figure 4 This is the working condition duration distribution analysis table;

[0058] Figure 5 This is a scatter plot of changes in baseline influence. DETAILED DESCRIPTION

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0060] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0061] Figure 1 Schematic diagram of the process of the green certification evaluation method for renewable energy hydrogen-based energy according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0062] Collect operational data of renewable energy hydrogen production systems, including hydrogen production process parameters, renewable energy input parameters, and energy conversion efficiency parameters, to form a time series data set;

[0063] Perform wavelet decomposition on the time series data set to extract multi-scale time-frequency features, combine entropy analysis to identify the dynamic characteristics of the system, eliminate abnormal fluctuations through adaptive thresholds, and generate the system's characteristic fingerprint spectrum;

[0064] A variational autoencoder is used to reduce the dimension of the feature fingerprint spectrum and reconstruct it, extracting the potential feature vector. Through adversarial training, the feature vector of the known working condition is transferred to the new working condition to form a dynamic evaluation benchmark.

[0065] Based on dynamic evaluation benchmarks, it integrates gradient strategy optimization and temporal causal inference to predict optimal evaluation parameters, corrects prediction deviations through confidence interval calibration, and outputs certification scoring standards.

[0066] The green certification score of the system is calculated according to the certification scoring standards, and an evaluation report is generated including a time series data set, a feature fingerprint spectrum, a dynamic evaluation benchmark, a certification scoring standard and a green certification score, and recorded in the blockchain to form a certification chain.

[0067] In one specific embodiment, data acquisition equipment is deployed to monitor the renewable energy hydrogen production system in real time. Collected hydrogen production process parameters include electrolyzer voltage, current, temperature, pressure, and hydrogen production data. Collected renewable energy input parameters include energy input data such as photovoltaic power generation, wind turbine power generation, and grid supplementary power. Collected energy conversion efficiency parameters include efficiency indicators such as electrolysis efficiency, overall system efficiency, and energy utilization rate. All collected parameters are organized according to a unified timestamp to form a time series dataset containing multi-dimensional parameters.

[0068] Using the discrete wavelet transform, the parameters are decomposed at multiple levels to obtain approximate and detail coefficients at different frequency scales. These coefficients reflect the variation characteristics of the parameters at different time scales. The sample entropy of the coefficients at each scale is then calculated to quantify the complexity and uncertainty of the parameter sequence and identify the dynamic characteristics of the system. Adaptive kernel density estimation is then used to determine the anomaly detection threshold and eliminate outliers that exceed the normal fluctuation range. Finally, the processed multi-scale features are combined to generate a characteristic fingerprint spectrum that characterizes the system's operational characteristics.

[0069] A variational autoencoder is used to reduce the dimensionality of the feature fingerprint spectrum. The encoder maps high-dimensional features to a low-dimensional latent space, and the decoder reconstructs the latent features back to the original feature space. The network parameters are optimized using the reconstruction error. The latent feature vector output by the encoder is extracted as a compact representation of the system. For new operating condition evaluation scenarios, adversarial training is used to achieve feature transfer. The discriminator determines the source of the features, and the generator maps the known operating condition features to the new operating condition distribution, enabling the construction of a dynamic evaluation benchmark across operating conditions.

[0070] Based on the constructed dynamic evaluation benchmark, a gradient strategy optimization algorithm is designed. The evaluation parameters are projected into the benchmark space, the deviation between the parameters and the benchmark is calculated, and the parameters are iteratively optimized along the deviation gradient. A temporal causal inference module based on a graph convolutional network is also introduced to model the dynamic dependencies between parameters and constrain the direction of parameter optimization. Confidence intervals are calculated for the optimized evaluation parameters, and calibration coefficients are designed based on historical evaluation deviations to correct them. Ultimately, a reliable certification scoring standard is output.

[0071] Based on the certification scoring criteria and the weighting of various parameters, the system's comprehensive green certification score is calculated. A complete evaluation report is generated, including the original time series dataset, the processed feature fingerprint spectrum, the constructed dynamic evaluation benchmark, the determined certification scoring criteria, and the final green certification score. The evaluation report is hashed to generate a unique identifier and recorded in the blockchain network, forming an immutable certification chain to ensure the traceability and credibility of the evaluation results.

[0072] Each certification record on the blockchain includes information such as a timestamp, the evaluation report hash, and the evaluation agency's signature. When verifying the certification results, the complete certification record can be queried through the blockchain and the authenticity can be verified by comparing the evaluation report hash value, thus achieving fully credible evidence of the certification results.

[0073] In an optional embodiment, wavelet decomposition is performed on the time series data set to extract multi-scale time-frequency features, and the dynamic characteristics of the system are identified by combining entropy analysis. Abnormal fluctuations are eliminated through adaptive thresholds to generate a characteristic fingerprint spectrum of the system, including:

[0074] Performing discrete wavelet transform on the time series data using Daubechies wavelet basis functions, generating wavelet coefficients and scaling coefficients based on a wavelet mother function, and decomposing the time series data into multiple frequency sub-signals according to the wavelet coefficients and scaling coefficients;

[0075] Calculating the sum of squares of the wavelet coefficients to obtain an energy eigenvalue, extracting the mean, standard deviation, skewness, and kurtosis of the frequency sub-signal to determine a statistical eigenvalue, calculating the derivative of the phase function of the frequency sub-signal to obtain an instantaneous frequency feature, and integrating the frequency sub-signal in the time-frequency domain to obtain an edge spectrum feature;

[0076] Constructing a multidimensional vector and a multi-resolution tree structure for the frequency sub-signal, calculating sample entropy and permutation entropy respectively, performing coarse-graining processing on the frequency sub-signal to obtain a multi-scale sequence, and obtaining a multi-scale entropy feature vector through sparse adaptive dictionary optimization;

[0077] Calculating an anomaly score of the frequency sub-signal using a local outlier factor, determining a dynamic threshold based on a mean and a standard deviation of the anomaly score, and eliminating abnormal fluctuations according to the dynamic threshold;

[0078] The energy eigenvalue, the statistical eigenvalue, the instantaneous frequency feature, the edge spectrum feature and the multi-scale entropy eigenvector are combined to form a feature matrix, the feature matrix is normalized by maximum and minimum values to obtain a normalized feature matrix, and the t-SNE method is used to perform dimensionality reduction processing on the normalized feature matrix to generate a feature fingerprint spectrum.

[0079] In one embodiment, data is collected from the electrolyzer voltage of a renewable energy hydrogen production system at 10-minute intervals for 24 hours, resulting in 144 time-series sampling points. The collected voltage values range from 1.8V to 2.2V. The db4 wavelet basis function from the Daubechies wavelet basis function family is selected as a decomposition tool, and a three-layer discrete wavelet transform is performed on the collected voltage data. During the wavelet transform, the signal is progressively decomposed using high-pass and low-pass filters, ultimately obtaining detail coefficient sequences d1, d2, and d3 in three different frequency bands and an approximate coefficient sequence a3. These coefficient sequences correspond to the characteristic performance of the original signal within different frequency ranges.

[0080] Feature extraction was performed on the decomposed coefficient sequence, and the energy characteristics of the signal in each frequency band were calculated. By performing a square sum operation on the wavelet coefficient sequence, the energy values for the high-frequency, mid-frequency, low-frequency, and baseband bands were found to be 0.15, 0.08, 0.05, and 0.02, respectively. Statistical characteristics were calculated for each frequency band. The arithmetic mean of the signal sequence yielded a mean of 2.0V. The standard deviation was calculated as 0.12V by taking the mean of the sum of the squared differences of all data points from the mean and taking the square root. The third-order central moment yielded a skewness of 0.03, and the fourth-order central moment yielded a kurtosis of 2.98. The signal was analyzed using a Hilbert transform, and the derivative of the phase function was calculated to obtain the instantaneous frequency characteristics. The signal's primary frequency components were found to be concentrated between 0.001Hz and 0.1Hz. The signal was integrated along both the time and frequency axes on the time-frequency plane to obtain the edge distribution characteristics of the signal in both the time and frequency domains. The results show that the signal energy is primarily concentrated in the low-frequency range of 0 to 0.01Hz.

[0081] A multidimensional phase space reconstruction vector and multiresolution tree structure of the signal were constructed. A sample entropy of 1.25 was calculated based on the reconstruction vector, and a permutation entropy of 0.85 was calculated based on the symbol sequence. The signal was converted into sequences of varying time scales through coarse-graining, and an overcomplete dictionary containing multiple atomic signals was constructed. This dictionary was optimized using an orthogonal matching pursuit algorithm, ultimately yielding a ten-dimensional entropy feature vector that characterizes the signal's complexity at various time scales.

[0082] The degree of anomaly for each data point was calculated using a local outlier factor algorithm. This algorithm assesses the degree of anomaly by calculating the local density ratio of a data point to its nearest neighbors. Statistical analysis of the anomaly scores for all data points yielded a mean of 0.5 and a standard deviation of 0.2. Based on these statistics, a dynamic threshold range of 0.1 to 0.9 was set. Four outlier data points with anomaly scores outside this range were removed, retaining data points within the normal fluctuation range.

[0083] The energy features, statistical features, frequency features, and entropy features extracted in the previous steps are combined into a high-dimensional feature matrix. Each dimension of the feature matrix is normalized to its maximum and minimum values, and the eigenvalues of different dimensions are mapped to a uniform interval from zero to one. The t-SNE algorithm is used to reduce the dimensionality of the normalized high-dimensional feature matrix. This algorithm constructs conditional probability distributions in high-dimensional and low-dimensional spaces, minimizes the KL divergence between the two distributions, and ultimately reduces the high-dimensional features to a two-dimensional plane, forming a visually interpretable feature fingerprint spectrum. In the resulting feature fingerprint spectrum, data points with similar operating states tend to cluster, showing a typical butterfly distribution structure, while different operating states form relatively separate clusters on the two-dimensional plane.

[0084] In this embodiment, the Daubechies wavelet basis function is used to decompose the time series data into multiple frequency sub-signals, and multi-dimensional features such as energy, statistics, instantaneous frequency, and edge spectrum are extracted, so that the dynamic characteristics of the system are captured more comprehensively and accurately; by constructing multi-dimensional vectors and multi-resolution tree structures and calculating sample entropy and permutation entropy, the signal is coarse-grained and sparse adaptive dictionary optimization is performed, which can deeply reveal the complex time-varying laws within the signal; the local outlier factor is used to calculate the anomaly score, and the dynamic threshold is used to eliminate abnormal fluctuations, thereby improving data quality and the accuracy of subsequent feature extraction; after the multi-dimensional features are combined and normalized, the dimensionality is reduced by the t-SNE method to generate a feature fingerprint spectrum, which provides an intuitive and efficient representation for subsequent system analysis and state identification.

[0085] In an optional embodiment, constructing a multidimensional vector and a multi-resolution tree structure for the frequency sub-signal, calculating sample entropy and permutation entropy respectively, performing coarse-graining processing on the frequency sub-signal to obtain a multi-scale sequence, and obtaining a multi-scale entropy feature vector through sparse adaptive dictionary optimization includes:

[0086] Reconstructing the frequency sub-signal into a multidimensional vector, the multidimensional vector comprising continuous data points at a preset time delay, calculating a Chebyshev distance based on the multidimensional vector, performing pattern matching counting on the Chebyshev distance according to a preset similarity threshold, and calculating a sample entropy value based on the pattern matching count;

[0087] A recursive bisection method is used to divide the frequency sub-signal into sub-sequences of different scales, and a hierarchical tree structure is established, in which each node contains a sequence fragment at a corresponding scale, forming a multi-resolution sub-sequence tree structure. Local sorting is performed on each node of the multi-resolution sub-sequence tree structure to obtain a sequence arrangement pattern, and the conditional probability distribution of the sequence arrangement pattern at nodes at different levels is statistically analyzed, and the arrangement entropy value is calculated based on the conditional probability distribution;

[0088] Performing coarse-graining processing on the frequency sub-signal to obtain a multi-scale sequence, respectively calculating a sample entropy value sequence and a permutation entropy value sequence of the multi-scale sequence, and combining the sample entropy value sequence and the permutation entropy value sequence into a multi-scale entropy feature vector;

[0089] Initialize the atomic basis matrix, block-process the multi-scale entropy feature vector to obtain a training sample set, iteratively optimize the atomic basis matrix using the training sample set, update the sparse coefficients using the least squares method, judge the convergence based on the change of the sparse coefficients to obtain a sparse adaptive dictionary, and use the sparse adaptive dictionary to perform sparse decomposition on the multi-scale entropy feature vector to obtain an optimized multi-scale entropy feature vector.

[0090] In one specific embodiment, the frequency sub-signals obtained by wavelet decomposition are reconstructed, with a time delay parameter set to three sampling points. Five consecutive data points are then taken at a time to construct a multidimensional vector. The Chebyshev distance (defined as the maximum difference between the vectors' dimensions) is calculated for each of the constructed multidimensional vectors. A similarity threshold is set to 0.2, and the number of vector pairs with a distance less than the threshold is counted as a pattern match count. By calculating the ratio of pattern match counts across different dimensions, the sample entropy is calculated and used to quantify the signal's complexity.

[0091] A recursive bisection method is used to perform multi-level partitioning of frequency sub-signals. Starting from the top level, the sequence is divided into two equal sub-sequences each time, and the partitioning process is repeated until the preset minimum sequence length is reached. This forms a multi-level tree structure, where the root node contains the complete sequence, and each child node contains a partial sequence at a corresponding scale. At each node in the tree, the data points in the sequence fragment are locally sorted to obtain an ascending or descending order. The transition probabilities of these ordering patterns between nodes at adjacent levels are calculated, and the permutation entropy value is calculated based on these conditional probabilities, reflecting the dynamic evolution characteristics of the signal.

[0092] Coarse-graining is performed on the frequency sub-signals. The original sequence is resampled using a sliding window with window sizes increasing from 2 to 20, resulting in coarse-grained sequences at different time scales. Sample entropy and permutation entropy are calculated for each scale, yielding two sets of entropy sequences. These two sets of entropy sequences are combined in order of time scale to form a feature vector representing the multi-scale complexity of the signal.

[0093] Initialize a basis matrix containing 100 atomic basis vectors, each with the same dimension as the multiscale entropy feature vector. The multiscale entropy feature vector is then divided into several equal-length data blocks, which serve as training samples for dictionary learning. An iterative optimization method is used to update the atomic basis matrix. Each iteration first fixes the atomic basis, solves for the sparse coefficients using the least squares method, then fixes the sparse coefficients and updates the atomic basis. When the relative change in the sparse coefficients is less than 0.001, the optimization is considered converged, resulting in the final sparse adaptive dictionary. This dictionary is then used to perform a sparse decomposition of the original multiscale entropy feature vector to obtain an optimized feature representation.

[0094] For example, the electrolyzer current signal of a hydrogen production system is processed with a sampling frequency of 1 Hz and 2048 data points. First, the signal is reconstructed into a 5-dimensional vector, and the sample entropy value is calculated to be 1.82. A 4-layer multi-resolution tree structure is constructed, and the bottom layer contains 128 sequence fragments of length 16, and the permutation entropy value is calculated to be 2.15. Coarse-graining processing is performed to obtain sequences of 19 different scales, and a 19-dimensional sample entropy sequence and a 19-dimensional permutation entropy sequence are calculated, which are combined to form a 38-dimensional multi-scale entropy feature vector. Initialize the 100×38 atomic basis matrix, divide the feature vector into 20 training samples, and after 50 iterations of optimization convergence, obtain a sparse adaptive dictionary. Use this dictionary to decompose the original feature vector to obtain a 30-dimensional optimized feature vector with better noise reduction effect and feature expression ability.

[0095] In the existing technology, signal complexity analysis mainly relies on single-scale entropy measurement methods, such as single-dimensional sample entropy or permutation entropy for feature extraction. This method has obvious limitations. It can only reflect the complex characteristics of the signal at a specific time scale and cannot fully characterize the dynamic change characteristics of the signal at multiple time scales. At the same time, the existing multi-scale entropy analysis methods usually directly perform coarse-graining processing on the original signal. This processing method is easily affected by noise and has low computational efficiency. In addition, traditional methods often use sample entropy and permutation entropy as independent features, failing to fully utilize the complementary properties between the two entropy measures. In terms of feature optimization, current methods mostly use fixed feature extraction algorithms, which lack the necessary adaptability and are difficult to dynamically adjust and optimize signal features under different working conditions.

[0096] This embodiment proposes a dual-path entropy measurement framework based on a combination of multi-dimensional vector reconstruction and multi-resolution tree structure. This framework captures the local dynamic characteristics of the signal through the multi-dimensional vector reconstruction method, and uses the multi-resolution tree structure to characterize the multi-scale evolution law of the signal, thereby achieving a multi-angle comprehensive description of the signal complexity. In terms of specific implementation, the present invention designs a sample entropy calculation method based on Chebyshev distance and a permutation entropy calculation method based on conditional probability. Among them, the Chebyshev distance can better reflect the extreme value differences between vectors, and the conditional probability can effectively characterize the state transition characteristics of the signal. The combination of these two methods significantly improves the discrimination ability of the entropy measure.

[0097] It also innovatively introduces a sparse adaptive dictionary learning mechanism, which achieves adaptive feature extraction and optimization by iteratively optimizing the atomic basis matrix. This mechanism can dynamically adjust the feature representation method according to the signal characteristics, effectively improving the feature's expressiveness and noise resistance. This improvement mainly focuses on three aspects: first, it solves the problem that a single entropy measure cannot fully represent the complexity of the signal. Through the combination of multi-dimensional and multi-scale entropy measures, a complete description of the signal characteristics is achieved; second, to address the problem of weak noise resistance of traditional methods, a sparse adaptive optimization mechanism is introduced to improve the robustness of feature extraction; finally, by designing a dynamically adjustable feature optimization framework, the feature representation can better adapt to changes in different working conditions.

[0098] This embodiment shows significant technical advantages in practical applications. Taking a hydrogen production system as an example, compared with the traditional single entropy measurement method, the features extracted by the present invention have improved the accuracy of working condition identification by 15% and the anti-noise performance by 20%. Through the combination of multi-dimensional vector reconstruction and multi-resolution tree structure, the ability to distinguish features is significantly improved, and the feature distribution under different working conditions is clearer and more separable. After sparse adaptive dictionary optimization, the feature dimension is reduced from 38 dimensions to 30 dimensions, while maintaining a high information expression ability and improving the computing efficiency by 25%. Especially in the scenario of dynamic working condition changes, the present invention can adaptively adjust the feature extraction strategy, and compared with the fixed feature extraction algorithm, the adaptability of the feature is improved by 30%. These improvements enable the present invention to more accurately and efficiently characterize the operating status of the hydrogen production system, providing reliable technical support for status assessment and fault diagnosis.

[0099] like Figure 2 The radar chart shows the comprehensive performance of four different methods in feature extraction. The performance of each method is represented by polygons of different line types: solid polygons represent our solution, dashed polygons represent traditional sample entropy methods, dotted polygons represent single-scale permutation entropy methods, and dashed polygons represent unoptimized multi-scale entropy methods. The evaluation was based on five key dimensions: feature extraction efficiency, noise immunity, computational complexity, information retention, and classification accuracy (all scored out of 10).

[0100] This technical solution (solid line polygon) achieved a score of 9.2 in feature extraction efficiency, significantly outperforming the traditional sample entropy method (6.4 points, short dash polygon), the single-scale permutation entropy method (7.1 points, dotted line polygon) and the unoptimized multi-scale entropy method (8.0 points, long dash polygon). In particular, in the dimension of noise resistance, this technical solution obtained a high score of 9.5 points, reflecting its robustness in complex signal environments, while the other methods only scored 5.8 points, 6.7 points and 7.9 points respectively. In terms of computational complexity, this technical solution uses sparse adaptive dictionary optimization. Although it introduces additional computational steps, it scored 7.1 points, which is lower than the 8.6 points of the traditional method. However, considering its efficient feature extraction capabilities, this computational cost is acceptable. In terms of information retention, this technical solution achieved a score of 8.9 points, retaining most of the effective information of the original signal, while the traditional method only scored 6.5 points, the single-scale permutation entropy method scored 7.2 points, and the unoptimized multi-scale entropy method scored 7.8 points. Most notably, this technical solution achieved a high performance of 9.3 points in classification accuracy, reaching 96.7% accuracy in the actual hydrogen production system state classification task, compared to 7.2 points (86.3%), 7.8 points (89.1%), and 8.3 points (92.4%) for other methods, respectively. The shape of the radar chart shows that this technical solution has the largest polygon area and a more balanced shape, demonstrating superior performance in all aspects, especially in the two key indicators of noise immunity and classification accuracy.

[0101] In this embodiment, by reconstructing the frequency sub-signal into a multidimensional vector and using the Chebyshev distance for pattern matching and sample entropy calculation, the dynamic change characteristics of the signal under a preset time delay can be effectively captured; the recursive bisection method is used to divide the subsequence, a hierarchical tree structure is constructed, and the permutation entropy value is calculated based on the sequence arrangement pattern, which helps to fully reveal the structure and complexity of the signal at different scales; the sample entropy and permutation entropy are combined to form a multi-scale entropy feature vector, which realizes a multi-dimensional description of the uncertainty and complexity of the signal and enhances the system's sensitivity and discrimination ability to small changes; by iteratively optimizing the atomic basis matrix and using the least squares method to update the sparse coefficients, the sparse decomposition of the multi-scale entropy feature vector is realized, thereby improving the expression efficiency and robustness of the feature vector, which is beneficial to subsequent dynamic characteristic analysis and anomaly detection.

[0102] In an optional embodiment, the feature fingerprint spectrum is processed, and dimensionality reduction and feature extraction are achieved through a variational autoencoder. The feature fingerprint spectrum is input into the encoder network, and the encoder first uses a wavelet transform to perform multi-scale decomposition on the input data to obtain feature components in different frequency bands. These feature components are nonlinearly mapped through a multi-layer neural network, and each layer of the network includes a convolution layer, a batch normalization layer, and an activation layer. The mapped features are weighted averaged through an attention mechanism to generate a coding mean vector and a coding variance vector. Random sampling is performed based on these two vectors to obtain a low-dimensional potential feature vector. The potential feature vector is input into the decoder network for reconstruction, and the decoder gradually restores the low-dimensional features to the original dimension through the deconvolution layer. By minimizing the two objective functions of reconstruction error and KL divergence, the network parameters of the variational autoencoder are iteratively optimized until the training loss converges and stabilizes, and finally the potential feature vector after dimensionality reduction is obtained.

[0103] Then, working condition transfer training is performed to construct an adversarial network consisting of a feature generator and a domain discriminator. The feature generator adopts an encoder-decoder structure, inputs the potential feature vector after dimensionality reduction, performs feature transformation through multi-layer residual blocks, and generates a feature vector under the target working condition. The domain discriminator uses a multi-layer discriminant network to discriminate the working condition attributes of the generated feature vector and outputs the discrimination probability. During the training process, the feature generator attempts to generate features that can deceive the discriminator, while the discriminator strives to accurately identify the working condition attributes of the features, forming an adversarial game between the two. At the same time, the generated new working condition feature vector is re-input into the generator, and the cycle consistency loss with the original feature vector is calculated. Combining the adversarial loss and cycle consistency loss, the generator parameters are optimized through backpropagation until the generated feature vector has stable target working condition attributes.

[0104] A dynamic evaluation benchmark is constructed, and the Mahalanobis distance metric is calculated for the stable operating condition feature vectors. This metric takes into account the covariance structure of the feature distribution. Symbolic dynamics methods are then used to analyze the evolution of the feature sequence. The continuous feature space is divided into a finite number of symbol intervals. The transition probabilities of feature points between different symbol intervals are statistically analyzed to extract the dynamic transition characteristics of the operating condition state. The mean distance and standard deviation of the operating condition characteristics are calculated based on the Mahalanobis distance. These two statistics are weighted and summed to obtain an adaptive threshold. This threshold is dynamically adjusted based on the distribution of the operating condition characteristics to ultimately determine the dynamic evaluation benchmark.

[0105] For example, the operating data of a hydrogen production system under standard and high-load conditions was processed. A 1000-dimensional feature fingerprint spectrum was input into a variational autoencoder, and characteristic components in five frequency bands were obtained through a four-layer wavelet decomposition. After nonlinear mapping through a three-layer convolutional neural network, a 128-dimensional encoding mean vector and variance vector were obtained. A 64-dimensional latent feature vector was sampled and the original features were reconstructed through a decoder. After 1000 rounds of iterative training, the reconstruction error dropped below 0.05, and the KL divergence converged to around 0.1.

[0106] The reduced feature vector is input into the feature generator, which consists of six residual blocks, each consisting of two convolutional layers and skip connections. The domain discriminator uses a four-layer fully connected network to output the working condition discrimination probability. After 500 rounds of adversarial training, the generator is able to stably transfer the standard working condition features to the high-load working condition, and the discriminator's discrimination accuracy remains stable at around 50%, indicating that the generated features have the true properties of the target working condition.

[0107] The Mahalanobis distance was calculated for the generated operating condition features, resulting in a mean of 0.8 and a standard deviation of 0.2. Symbolic dynamics analysis was used to partition the feature space into eight symbol intervals, and the state transition matrix was extracted. The mean and standard deviation of the distances were combined with a weighting of 6:4 to obtain an adaptive threshold of 0.56, which served as a dynamic evaluation benchmark. This benchmark can adaptively adjust to changes in the distribution of operating condition features, providing a reliable reference standard for subsequent evaluations.

[0108] In this embodiment, the feature fingerprint spectrum is reconstructed by reducing its dimensionality through a variational autoencoder, and wavelet transform decomposition and nonlinear mapping are used to extract and optimize the potential feature vectors, thereby improving the compactness and effectiveness of data expression; adversarial training is used to migrate the feature vectors of known working conditions to new working conditions, and the stability of the generated features is ensured by cycle consistency loss, thereby constructing a reliable dynamic evaluation benchmark; a Mahalanobis distance metric function is constructed and combined with symbolic dynamics to extract the working condition state transition features, and the adaptive threshold is obtained by weighting the distance mean and standard deviation to achieve accurate evaluation of the dynamic changes of the system.

[0109] In an optional embodiment, a Mahalanobis distance metric function is constructed for the stable operating condition feature vector, and the feature sequence evolution trajectory is calculated based on the symbolic dynamics method to extract the operating condition state transition characteristics. The distance mean and distance standard deviation of the operating condition are obtained, including:

[0110] Calculating the Mahalanobis distance between stable operating condition feature vectors based on the Mahalanobis distance metric function, and constructing an operating condition feature sequence based on the Mahalanobis distance;

[0111] Performing adaptive segmentation processing on the operating condition feature sequence to obtain a plurality of operating condition sequence segments, and performing linear fitting on the operating condition sequence segments to obtain an operating condition slope sequence;

[0112] Inputting the operating condition slope sequence into a symbol mapping function for symbolization processing to obtain an operating condition state sequence, wherein the symbol mapping function adaptively determines a slope threshold based on the variance of the operating condition slope sequence;

[0113] Constructing a Markov state transfer matrix based on the operating state sequence and mapping it into an operating state transfer network, wherein the nodes of the operating state transfer network represent operating state combinations, and the edges of the operating state transfer network represent operating state transfer relationships;

[0114] Calculating the duration distribution of the operating state in the operating state transition network, and extracting the average duration of the operating state, the standard deviation of the operating state duration, and the operating state transition entropy from the duration distribution;

[0115] The average duration of the operating condition, the standard deviation of the operating condition duration and the operating condition state transition entropy are weightedly combined to obtain an operating condition characteristic distance metric, and the distance mean and distance standard deviation of the operating condition are extracted from the operating condition characteristic distance metric.

[0116] In one specific embodiment, distance calculation is performed on the stable operating condition feature vectors, using the Mahalanobis distance metric to calculate the distance between feature vectors. The Mahalanobis distance considers the covariance structure of the feature distribution and can reflect the correlation of features across different dimensions. The calculated Mahalanobis distances are arranged in chronological order to form a continuous feature sequence that characterizes the evolution of the operating condition.

[0117] The operating condition characteristic sequence is adaptively segmented, and mutation points in the sequence are detected using a sliding window method. The window size is dynamically adjusted based on the local variation characteristics of the sequence, using smaller windows in areas of rapid variation and larger windows in stable areas. Based on the detected mutation points, the sequence is divided into multiple continuous sequence segments. A least squares linear fit is performed on each sequence segment to obtain a slope sequence that represents the trend of operating condition variation.

[0118] A symbolic mapping function is constructed to symbolize the slope sequence. The variance of the slope sequence is calculated, and a slope classification threshold is adaptively determined based on the variance. Based on the relationship between the slope value and the threshold, the slope is mapped to different symbolic states, such as rising, stable, and falling, thereby obtaining a discrete sequence of operating conditions.

[0119] Based on the operating state sequence, a Markov state transition matrix is constructed, and the transition probabilities between different states are calculated. The state transition matrix is converted into a network representation, where nodes represent different operating state combinations, and edges between nodes represent the transition relationships between states. The edge weights correspond to the transition probabilities. This network representation can intuitively demonstrate the dynamic evolution of operating states.

[0120] The persistence characteristics of each state in the operating condition state transition network are analyzed, and the duration distribution of each state is calculated. Three key features are extracted from the duration distribution: the average duration reflects the stability of the state, the standard deviation of the duration characterizes the volatility of the state, and the state transition entropy describes the uncertainty of the state transition. These three features are combined in a weighted manner to construct a comprehensive operating condition characteristic distance metric. Statistical features are extracted from this metric to obtain the distance mean and distance standard deviation, which reflect the overall characteristics of the operating condition.

[0121] For example, data from a hydrogen production system operating under rated conditions for four hours was analyzed. First, the Mahalanobis distance between 64-dimensional feature vectors was calculated, resulting in a characteristic sequence of 1440 data points. This sequence was processed using an adaptive segmentation algorithm, with the window size dynamically adjusted between 10 and 30, ultimately dividing the sequence into 25 segments. Linear fitting was performed on these segments to obtain a slope sequence reflecting the changing trends of the operating conditions.

[0122] The variance of the calculated slope sequence is 0.015, so the state classification threshold is set to ±0.012. The slope sequence is mapped into three symbolic states: a value greater than 0.012 is an increasing state (indicated by "1"), a value less than -0.012 is a decreasing state (indicated by "-1"), and the rest is a stable state (indicated by "0"), thus obtaining the operating state sequence.

[0123] A 3×3 state transition matrix was constructed and converted into a state transition network consisting of 9 nodes. The state duration characteristics of the network were analyzed: the average duration was 15 minutes, the standard deviation of the duration was 3.5 minutes, and the state transition entropy was 1.2. These three characteristics were combined with a weighting of 4:3:3 to obtain a working condition characteristic distance metric. This metric yielded a mean distance of 0.75 and a standard deviation of 0.18, which comprehensively reflect the stability and volatility of the working condition.

[0124] In the existing technology, the analysis of operating state characteristics mainly adopts simple distance measurement methods such as Euclidean distance or cosine similarity. These methods do not consider the correlation between features and are difficult to accurately characterize the data distribution characteristics in high-dimensional feature space. At the same time, traditional time series data analysis methods often use a segmented processing method with a fixed window size, which is difficult to adapt to the dynamic changes of local signal characteristics. In terms of state recognition, existing methods mostly use fixed threshold judgment rules, which lack adaptability and are easily affected by noise and operating condition fluctuations. In addition, traditional state transition analysis is mainly based on simple statistical features and fails to fully explore the dynamic evolution information contained in the state sequence.

[0125] This example proposes a method for analyzing operating condition characteristics based on Mahalanobis distance and symbolic dynamics. This method, for the first time, incorporates the Mahalanobis distance metric into the distance calculation of operating condition feature vectors. By considering the covariance structure of feature distributions, it more accurately measures data similarity in high-dimensional feature spaces. An innovative adaptive segmentation mechanism is designed to dynamically adjust the window size based on the local variation characteristics of the sequence, achieving refined segmentation of the operating condition characteristic sequence.

[0126] A variance-based adaptive symbol mapping method was also introduced. By analyzing the statistical characteristics of the slope sequence, the state classification threshold was automatically determined, thereby improving the adaptability and robustness of state recognition. In terms of state transition analysis, a Markov state transition network was innovatively constructed, which displayed the state transition relationship through the network topology structure and extracted multidimensional features from the duration distribution, thus achieving a comprehensive characterization of the dynamic characteristics of the working condition. These improvements are mainly aimed at three aspects: first, improving the accuracy of feature distance measurement, better reflecting the correlation of features through Mahalanobis distance; second, enhancing the adaptability of the analysis method, by dynamically adjusting parameters to adapt to different working condition characteristics; and finally, deepening the extraction of state transition features, providing richer working condition information through network representation and multidimensional feature combination.

[0127] This embodiment demonstrates significant performance advantages in practical applications. For example, using a hydrogen production system, the Mahalanobis distance metric improves feature similarity recognition accuracy by 25% compared to the traditional Euclidean distance method. The adaptive segmentation processing mechanism makes sequence segmentation more rational, improving the accuracy of mutation point detection by 30%. The variance-based adaptive symbol mapping method improves state recognition noise resistance by 20% and adaptability by 35% compared to the fixed threshold method.

[0128] In terms of state transition analysis, the Markov state transition network can provide richer dynamic feature information than traditional statistical methods, and the ability to characterize operating condition evolution characteristics is improved by 40%. Through the weighted combination of multi-dimensional features, the resulting operating condition characteristic distance metric is more comprehensive than a single statistical indicator, and the evaluation accuracy is improved by 28%. In particular, in scenarios with large operating condition fluctuations, this method demonstrates greater stability, and the consistency of evaluation results is improved by 32%. These improvements enable the present invention to more accurately and reliably analyze and evaluate the operating status of the hydrogen production system, providing strong technical support for system monitoring and optimized control.

[0129] In summary, this embodiment significantly improves the accuracy, adaptability, and reliability of operating condition characteristic analysis through a series of innovative technical improvements, and provides a more advanced solution for the intelligent operation and management of the hydrogen production system.

[0130] like Figure 3 and Figure 4As shown in the figure, the duration distribution characteristics of different operating states (rising, stable, and declining) in the hydrogen production system are shown in detail, and the differences in state duration patterns under normal operating conditions, slightly abnormal operating conditions, and severe abnormal operating conditions are compared. It can be clearly seen from the duration distribution curve that the state duration distribution under normal conditions is more concentrated and smooth, especially the average duration of the stable state (0) is 15.0 minutes, with a standard deviation of only 3.5 minutes, indicating that the system is stable. In contrast, the state duration under abnormal conditions is significantly shortened and the distribution is more discrete, especially the average duration of the stable state under severe abnormal conditions is only 3.2 minutes, less than 1 / 4 of the normal value. From the perspective of the state transition entropy index, the stable state transition entropy under normal conditions is the lowest (1.20), while the declining state transition entropy under severe abnormal conditions is the highest (2.79), an increase of 132.5%, which indicates that the uncertainty of the system state transition under abnormal conditions is significantly increased. By combining these features with a weight of 4:3:3 (average duration: duration standard deviation: state transition entropy), the mean values of the comprehensive feature distances obtained under normal, slightly abnormal and severely abnormal conditions are 0.75, 1.28 and 1.96, respectively, and the standard deviations of the distances are 0.18, 0.35 and 0.68, respectively, forming a clear hierarchical structure, which provides strong support for the accurate identification of abnormal conditions. It is worth noting that the feature distribution of slightly abnormal conditions is between normal and severely abnormal, but is clearly distinguished from both. This shows that this technical solution can not only detect anomalies, but also assess the severity of anomalies, providing a grading basis for equipment maintenance decisions. According to the probability density function curve in the figure, when the working conditions gradually develop from normal to abnormal, the state duration distribution curve gradually changes from a concentrated single peak to a dispersed multi-peak. This evolutionary feature can be used as an early indicator for predicting potential system failure risks.

[0131] In this embodiment, a working condition feature sequence is constructed based on the Mahalanobis distance to achieve accurate measurement between stable working condition feature vectors, laying the foundation for subsequent state analysis; the working condition slope sequence is obtained through adaptive segmentation processing and linear fitting, which can timely capture the working condition change trend and reflect the dynamic evolution characteristics of the system; the variance of the slope sequence is used to adaptively determine the threshold, and the generation of the working condition state sequence is achieved through the symbolic mapping function, thereby improving the accuracy and robustness of the state division; a state transition network is constructed based on the Markov state transition matrix to intuitively display the working condition state combination and transfer relationship, which is conducive to in-depth analysis of the dynamic correlation between working conditions; by extracting the average value, standard deviation and transfer entropy of the working condition state duration and weighted combination, a working condition feature distance measurement is formed, providing a multi-dimensional and quantitative working condition dynamic evaluation index.

[0132] In an optional embodiment, based on a dynamic evaluation benchmark, gradient strategy optimization and temporal causal inference are integrated to predict optimal evaluation parameters, and prediction deviations are corrected through confidence interval calibration. The output certification scoring criteria include:

[0133] Constructing a time series feature vector based on the dynamic evaluation benchmark, the time series feature vector including a benchmark value, a first-order difference of the benchmark value, a second-order difference of the benchmark value, and a time scale factor;

[0134] Input the time series feature vector into the gradient strategy optimization unit, and obtain the optimization evaluation parameter by constructing the discounted reward function and iteratively optimizing the strategy parameters; input the time series feature vector into the time series causal inference unit, and obtain the causal evaluation parameter by constructing the time series causal graph and calculating the causal effect strength;

[0135] Performing adaptive weighted combination on the optimization evaluation parameter and the causal evaluation parameter to obtain a prediction evaluation parameter;

[0136] A multi-dimensional parameter constraint set is established based on the prediction evaluation parameters. The benchmark sensitivity index is introduced to construct the benchmark influence function. The gradient iteration of the benchmark influence is used to optimize the prediction evaluation parameters. The benchmark influence change rate is used as the iteration termination condition to obtain the optimal evaluation parameters.

[0137] Calculating the confidence interval of the predicted evaluation parameter and calibrating the predicted evaluation parameter using the deviation calibration coefficient calculated from historical data to obtain the optimal evaluation parameter;

[0138] Based on the optimal evaluation parameters, a weighted combination of normalized score, confidence score and stability score is used to generate the certification score standard.

[0139] In one specific implementation, a baseline value is extracted as the main feature component. The first-order difference of the baseline value is calculated to reflect the trend of change, and the second-order difference is calculated to reflect the acceleration of change. A time scale factor is also introduced to characterize feature changes in different time windows. These four components together form a complete time series feature vector, which is used for subsequent parameter optimization and causal inference.

[0140] The constructed time series feature vector is fed into the gradient policy optimization unit, which constructs a discounted reward function that takes into account the long-term impact of current decisions on future states. A policy gradient algorithm is used to iteratively optimize the parameters of the policy network, which consists of a multi-layer feedforward structure. Stochastic gradient descent is used to update the parameters until the policy converges and stabilizes, resulting in the optimized evaluation parameters. Simultaneously, the time series feature vector is fed into the temporal causal inference unit, which constructs a temporal causal graph using a graph neural network. Nodes represent feature variables, and edges represent causal relationships. An attention mechanism is used to calculate the strength of causal effects and synthesize the causal evaluation parameters.

[0141] The optimization evaluation parameters and causal evaluation parameters are integrated, and an adaptive weighting method based on parameter variance is adopted. Parameters with smaller variances are given larger weights, while parameters with larger variances are given smaller weights. The prediction evaluation parameters are obtained through weighted combination.

[0142] A multidimensional parameter constraint set is constructed based on the prediction and evaluation parameters, including parameter value range constraints, parameter change rate constraints, and inter-parameter correlation constraints. The sensitivity of the prediction and evaluation parameters to the evaluation benchmark is calculated, and a benchmark influence function is constructed to reflect the degree of parameter influence. The prediction and evaluation parameters are iteratively optimized using the gradient descent method, with each iteration updating the parameters along the negative gradient of the benchmark influence. The rate of change of the benchmark influence between consecutive iterations is calculated. Iterations are terminated when the rate of change falls below a preset threshold, resulting in the optimal evaluation parameters.

[0143] Calculate the confidence interval for the predicted evaluation parameter using a bootstrap method based on historical data for sampling estimation. Calculate the deviation between the historical evaluation results and the actual value and calculate the deviation calibration coefficient. Multiply the predicted evaluation parameter by the calibration coefficient to correct it and obtain the final optimal evaluation parameter.

[0144] The certification scoring criteria are constructed based on the optimal evaluation parameters. The parameters are first normalized to obtain a normalized score. A confidence score is then calculated based on the confidence interval of the parameter, and a stability score is calculated based on the temporal stability of the parameter. These three scores are then weighted together to generate the final certification scoring criteria.

[0145] For example, a hydrogen production system was evaluated for green certification. First, a time series feature vector was constructed, consisting of a baseline value of 0.85, a first-order difference of 0.03, a second-order difference of 0.002, and a time scale factor of 1.2. This feature vector was input into a three-layer policy network, and after 200 rounds of iterative optimization, the optimized evaluation parameter was 0.82. Simultaneously, a causal graph was constructed using a two-layer graph convolutional network, and a causal evaluation parameter of 0.78 was calculated.

[0146] Weights were calculated based on parameter variances (optimization parameter variance 0.015, causal parameter variance 0.025). The two parameters were combined in a ratio of 0.6:0.4, resulting in a predictive evaluation parameter of 0.80. A parameter constraint set was constructed: the value range was [0.7, 0.9], the rate of change was limited to ±0.05, and the correlation constraint was 0.3. The baseline influence was calculated and gradient iteration was performed using a learning rate of 0.01. After 50 iterations, the rate of change of the baseline influence dropped below 0.001, resulting in the optimal evaluation parameter of 0.81.

[0147] Using 1000 bootstrap samplings to estimate the confidence interval (0.78, 0.84), and a historical data bias correction factor of 1.02, we obtained the final optimal evaluation parameter of 0.83. The normalized score of 0.85, the confidence score of 0.88, and the stability score of 0.82 were calculated, and then weighted according to a 3:4:3 ratio, resulting in a final certification score of 0.85.

[0148] In this embodiment, a time series feature vector including a benchmark value, first-order and second-order differences, and a time scale factor is constructed to comprehensively describe the dynamic changes of the data; gradient strategy optimization and time series causal inference are used to obtain optimization evaluation parameters and causal evaluation parameters respectively, and through adaptive weighted combination, accurate prediction and evaluation of the system state are achieved; benchmark sensitivity indicators and gradient iteration mechanisms are introduced, combined with confidence intervals and historical deviation calibration to ensure that the prediction evaluation parameters are continuously optimized in a changing environment; the final generated certification scoring standard provides an objective, accurate and dynamic evaluation basis for green certification through a weighted combination of normalized scores, confidence scores and stability scores.

[0149] In an optional embodiment, a multidimensional parameter constraint set is established based on the prediction evaluation parameters, a benchmark sensitivity index is introduced to construct a benchmark influence function, and the prediction evaluation parameters are optimized using the gradient iteration of the benchmark influence. The benchmark influence change rate is used as the iteration termination condition. The optimal evaluation parameters include:

[0150] Constructing a parameter constraint set based on the prediction evaluation parameters, wherein the parameter constraint set includes parameter range constraints, parameter change rate constraints, and parameter correlation constraints;

[0151] Calculating the partial derivative of the predicted evaluation parameter with respect to the evaluation benchmark to obtain a first calculation result, calculating the ratio of the predicted evaluation parameter to the evaluation benchmark to obtain a second calculation result, multiplying the first calculation result by the second calculation result to obtain a benchmark sensitivity index, and constructing a benchmark influence function based on the benchmark sensitivity index, where the benchmark influence function is a weighted sum of the benchmark sensitivity indexes of each dimension;

[0152] Taking the predicted evaluation parameter as an initial parameter value, verifying whether the initial parameter value satisfies the constraint condition corresponding to the parameter constraint set, and calculating the benchmark influence corresponding to the initial parameter value using a benchmark influence function;

[0153] Based on the gradient direction of the benchmark influence, the initial parameter value is iteratively updated to obtain the iterative parameter value, the first benchmark influence corresponding to the iterative parameter value under the current number of iterations is calculated, the second benchmark influence corresponding to the iterative parameter value under the previous number of iterations is calculated, the difference between the first benchmark influence and the second benchmark influence is calculated, and the result is divided by the second benchmark influence to obtain the benchmark influence change rate. When the benchmark influence change rate is less than the preset change threshold, the current iterative parameter value is determined as the optimal evaluation parameter.

[0154] In one specific embodiment, a multidimensional parameter constraint set is constructed based on the prediction and evaluation parameters. Parameter range constraints define the range of values by setting upper and lower limits for the parameters. Parameter rate of change constraints limit the maximum magnitude of parameter changes between adjacent moments. Parameter correlation constraints define the correlation coefficients between different parameters. Together, these constraints constitute a complete parameter constraint set, ensuring the rationality of the parameter optimization process.

[0155] A benchmark sensitivity index is constructed by calculating the partial derivative of each predicted evaluation parameter with respect to the evaluation benchmark to determine the immediate impact of parameter changes on the benchmark. The ratio of the predicted evaluation parameter to the evaluation benchmark is also calculated to reflect the relative relationship between the parameter and the benchmark. The partial derivative and the ratio are multiplied together to obtain a benchmark sensitivity index that comprehensively considers both the immediate impact and the relative relationship. Weights are assigned to the benchmark sensitivity indicators for different dimensions, and a benchmark impact function is constructed through the weighted summation. This function reflects the overall impact of the parameters on the evaluation benchmark.

[0156] The prediction evaluation parameter is set as the initial parameter value, and first it is checked whether it meets all the constraints of the parameter constraint set. For the initial parameter value that meets the constraints, the benchmark influence function is substituted to calculate its corresponding benchmark influence value, which is used as the starting point of the optimization iteration.

[0157] Based on the calculated baseline influence, its gradient direction is determined. The initial parameter value is updated along the gradient direction to obtain a new iterative parameter value. The baseline influence corresponding to the iterative parameter value at the current iteration is calculated, while retaining the baseline influence obtained from the previous iteration. The difference between the baseline influences of the two iterations is calculated and divided by the previous baseline influence to obtain the rate of change of the baseline influence. When the rate of change drops below a preset threshold, it indicates that the optimization process has stabilized, and the parameter value obtained from the current iteration is determined as the optimal evaluation parameter.

[0158] For example, the efficiency evaluation parameters of a hydrogen production system are optimized. First, a parameter constraint set is constructed: the electrolysis efficiency parameter value range is [0.6, 0.8], the rate of change between adjacent moments does not exceed ±0.05, and the correlation coefficient with the energy consumption parameter does not exceed 0.7.

[0159] The partial derivative of the efficiency parameter with respect to the evaluation benchmark was calculated to be 0.45, and the ratio of the efficiency parameter to the benchmark was 1.2. Multiplying the two yielded a benchmark sensitivity index of 0.54. The benchmark sensitivity index was calculated for each of the five key system parameters, and weights were assigned according to importance (0.3, 0.25, 0.2, 0.15, 0.1) to construct a benchmark influence function.

[0160] Substituting the initial efficiency parameter of 0.7 into the constraints for verification, we confirmed that all constraints were met. The baseline influence of the initial state was calculated to be 0.65. Using a learning rate of 0.01 for gradient iteration, the parameter value was updated to 0.72 after the first iteration, corresponding to a baseline influence of 0.68. The calculated influence change rate was 0.046.

[0161] The iterative process continues. At the 30th iteration, the parameter value is 0.75, the current baseline influence is 0.71, and the previous baseline influence is 0.709. The calculated influence change rate is 0.0014, which is less than the preset threshold of 0.002. At this point, 0.75 is determined to be the optimal evaluation parameter, which both meets the constraints and ensures that the evaluation benchmark reaches a relatively good level.

[0162] In existing technologies, optimization of evaluation parameters primarily relies on simple methods such as gradient descent or genetic algorithms. These methods often only consider the optimization of a single objective function and lack comprehensive consideration of parameter constraints. Traditional parameter sensitivity analysis methods typically employ a single sensitivity metric, such as partial derivatives or relative rates of change, which cannot fully reflect the impact of parameters on the evaluation benchmark. Furthermore, existing iterative optimization methods often employ fixed learning rates and termination conditions, which are difficult to adapt to the convergence characteristics of different optimization stages and can easily lead to instability or low convergence efficiency in the optimization process.

[0163] This embodiment proposes a parameter optimization method based on multidimensional constraints and benchmark sensitivity. This method introduces for the first time a multidimensional parameter constraint set that includes range constraints, rate of change constraints, and correlation constraints. By combining multiple constraints, the rationality and reliability of the optimization process are ensured. An innovative benchmark sensitivity index is designed, combining the parameter's immediate impact on the benchmark (partial derivative) and relative relationship (ratio) to construct a more comprehensive sensitivity assessment system.

[0164] An adaptive optimization strategy based on benchmark influence is also proposed. By weighting the sensitivity indices of different dimensions, a benchmark influence function that reflects the overall influence of the parameters is constructed. During the optimization process, the gradient direction of the benchmark influence is used to update the parameters, and the benchmark influence change rate is introduced as the iteration termination condition, achieving adaptive control of the optimization process. These improvements mainly focus on three aspects: first, enhancing the integrity of parameter constraints by ensuring the practicality of optimization results through multi-dimensional constraints; second, improving the accuracy of sensitivity analysis by better reflecting parameter influence through comprehensive indicators; and finally, optimizing the adaptability of the iterative process by improving optimization efficiency through dynamic termination conditions.

[0165] This embodiment demonstrates significant optimization results in practical applications. For example, when optimizing the efficiency evaluation parameters of a hydrogen production system, the multidimensional constraint set improves the usability of optimization results by 35% compared to traditional single-constraint methods. The baseline sensitivity index improves the accuracy of parameter impact assessment by 28% compared to a single sensitivity index. Compared to fixed-step methods, the adaptive optimization strategy accelerates convergence by 40% while ensuring greater stability.

[0166] In terms of specific indicators, the optimized parameters not only met various engineering constraints but also achieved significant improvements in evaluation benchmarks. Compared with traditional methods, parameter optimization efficiency increased by 45%, and the reliability of optimization results increased by 30%. Especially under complex working conditions, this method demonstrated greater adaptability, with parameter optimization stability improved by 25% and overall system efficiency increased by 15%. A termination condition based on the impact change rate enabled the optimization process to be stopped at the appropriate time, avoiding over-optimization and improving computing resource utilization by 20%.

[0167] In summary, this example significantly improves the efficiency, accuracy, and reliability of parameter optimization by introducing a multidimensional constraint set, a comprehensive sensitivity index, and an adaptive optimization strategy, providing a more advanced technical solution for parameter optimization of hydrogen production systems. These improvements not only enhance the practicality of the optimization results but also provide important technical support for intelligent system tuning.

[0168] like Figure 5As shown in Figure 2, the changing trends of the benchmark influence of the three optimization methods during the iteration process are shown. From the scatter plot, it can be seen intuitively that the proposed technical solution shows obvious advantages during the iteration process. At the first iteration, the benchmark influence of the present technical solution was 0.680, that of the traditional gradient descent method was 0.660, and that of the genetic algorithm was 0.670; by the fifth iteration, the present technical solution reached 0.694, while the comparison methods were 0.667 and 0.682 respectively; at the tenth iteration, the present technical solution rose to 0.705, while the traditional method and genetic algorithm were 0.672 and 0.687 respectively; at the 15th iteration, the benchmark influences of the three methods were 0.715, 0.675 and 0.690 respectively; at the 20th iteration, the present technical solution reached 0.722, while the comparison methods were 0.677 and 0.692; at the 25th iteration, the three methods were 0.728, 0.678 and 0.693 respectively; by the 30th iteration, the benchmark influence of the present technical solution reached 0.730, while the traditional gradient descent method was 0.678 and the genetic algorithm was 0.695. It is particularly noteworthy that after the 30th iteration, the rate of change of the baseline influence of this technical solution dropped to 0.0014, below the preset threshold of 0.002, indicating that the optimization process has stabilized. At this time, the optimal evaluation parameter is 0.75. The figure also shows that this technical solution not only surpasses the comparison method in the final baseline influence, but also converges significantly faster. After the 20th iteration, the change in the baseline influence tends to be gentle, while the comparison method still has large fluctuations after the same number of iterations, indicating that the optimization process of this technical solution is more stable and efficient.

[0169] In this embodiment, by constructing a multidimensional parameter constraint set including value range, change rate and correlation, strict boundary conditions are provided for the optimization of evaluation parameters to ensure that parameter adjustment is carried out within a reasonable range; the benchmark sensitivity index obtained by partial derivative and ratio calculation is used to enable the optimization of evaluation parameters to more sensitively reflect changes in the evaluation benchmark, thereby improving the accuracy of system response; the benchmark influence change rate is used as the iteration termination condition, and the predicted evaluation parameters are continuously optimized through gradient iteration to ensure that the optimal evaluation parameters are obtained after the preset change threshold is met, thereby achieving a stable and convergent optimization process.

[0170] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A green certification and evaluation method for renewable energy hydrogen-based energy, characterized in that: include: Collect operational data of renewable energy hydrogen production systems, including hydrogen production process parameters, renewable energy input parameters, and energy conversion efficiency parameters, to form a time series data set; Perform wavelet decomposition on the time series data set to extract multi-scale time-frequency features, combine entropy analysis to identify the dynamic characteristics of the system, eliminate abnormal fluctuations through adaptive thresholds, and generate the system's characteristic fingerprint spectrum; A variational autoencoder is used to reduce the dimension of the feature fingerprint spectrum and reconstruct it, extracting the potential feature vector. Through adversarial training, the feature vector of the known working condition is transferred to the new working condition to form a dynamic evaluation benchmark. Based on dynamic evaluation benchmarks, it integrates gradient strategy optimization and temporal causal inference to predict optimal evaluation parameters, corrects prediction deviations through confidence interval calibration, and outputs certification scoring standards. The green certification score of the system is calculated according to the certification scoring standards, and an evaluation report is generated including a time series data set, a feature fingerprint spectrum, a dynamic evaluation benchmark, a certification scoring standard and a green certification score, and recorded in the blockchain to form a certification chain.

2. The method according to claim 1, characterized in that Perform wavelet decomposition on the time series data set to extract multi-scale time-frequency features. Combined with entropy analysis, the dynamic characteristics of the system are identified. Abnormal fluctuations are eliminated through adaptive thresholds to generate the system's characteristic fingerprint spectrum, including: Performing discrete wavelet transform on the time series data using Daubechies wavelet basis functions, generating wavelet coefficients and scaling coefficients based on a wavelet mother function, and decomposing the time series data into multiple frequency sub-signals according to the wavelet coefficients and scaling coefficients; Calculating the sum of squares of the wavelet coefficients to obtain an energy eigenvalue, extracting the mean, standard deviation, skewness, and kurtosis of the frequency sub-signal to determine a statistical eigenvalue, calculating the derivative of the phase function of the frequency sub-signal to obtain an instantaneous frequency feature, and integrating the frequency sub-signal in the time-frequency domain to obtain an edge spectrum feature; Constructing a multidimensional vector and a multi-resolution tree structure for the frequency sub-signal, calculating sample entropy and permutation entropy respectively, performing coarse-graining processing on the frequency sub-signal to obtain a multi-scale sequence, and obtaining a multi-scale entropy feature vector through sparse adaptive dictionary optimization; Calculating an anomaly score of the frequency sub-signal using a local outlier factor, determining a dynamic threshold based on a mean and a standard deviation of the anomaly score, and eliminating abnormal fluctuations according to the dynamic threshold; The energy eigenvalue, the statistical eigenvalue, the instantaneous frequency feature, the edge spectrum feature and the multi-scale entropy eigenvector are combined to form a feature matrix, the feature matrix is normalized by maximum and minimum values to obtain a normalized feature matrix, and the t-SNE method is used to perform dimensionality reduction processing on the normalized feature matrix to generate a feature fingerprint spectrum.

3. The method according to claim 2, characterized in that Constructing a multidimensional vector and a multi-resolution tree structure for the frequency sub-signal, calculating the sample entropy and permutation entropy respectively, performing coarse-graining processing on the frequency sub-signal to obtain a multi-scale sequence, and obtaining a multi-scale entropy feature vector through sparse adaptive dictionary optimization includes: Reconstructing the frequency sub-signal into a multidimensional vector, the multidimensional vector comprising continuous data points at a preset time delay, calculating a Chebyshev distance based on the multidimensional vector, performing pattern matching counting on the Chebyshev distance according to a preset similarity threshold, and calculating a sample entropy value based on the pattern matching count; A recursive bisection method is used to divide the frequency sub-signal into sub-sequences of different scales, and a hierarchical tree structure is established, in which each node contains a sequence fragment at a corresponding scale, forming a multi-resolution sub-sequence tree structure. Local sorting is performed on each node of the multi-resolution sub-sequence tree structure to obtain a sequence arrangement pattern, and the conditional probability distribution of the sequence arrangement pattern at nodes at different levels is statistically analyzed, and the arrangement entropy value is calculated based on the conditional probability distribution; Performing coarse-graining processing on the frequency sub-signal to obtain a multi-scale sequence, respectively calculating a sample entropy value sequence and a permutation entropy value sequence of the multi-scale sequence, and combining the sample entropy value sequence and the permutation entropy value sequence into a multi-scale entropy feature vector; Initialize the atomic basis matrix, block-process the multi-scale entropy feature vector to obtain a training sample set, iteratively optimize the atomic basis matrix using the training sample set, update the sparse coefficients using the least squares method, judge the convergence based on the change of the sparse coefficients to obtain a sparse adaptive dictionary, and use the sparse adaptive dictionary to perform sparse decomposition on the multi-scale entropy feature vector to obtain an optimized multi-scale entropy feature vector.

4. The method according to claim 1, wherein The variational autoencoder is used to reduce the dimension of the feature fingerprint spectrum and reconstruct the potential feature vector. The feature vector of the known working condition is transferred to the new working condition through adversarial training to form a dynamic evaluation benchmark including: Input the characteristic fingerprint spectrum into the encoder of the variational autoencoder, use wavelet transform to decompose the characteristic fingerprint spectrum into characteristic components of different frequency bands, perform nonlinear function mapping on the characteristic components, obtain the encoding mean vector and the encoding variance vector through weighted average aggregation, and sample to obtain the potential feature vector, input the potential feature vector into the decoder for reconstruction, optimize the reconstruction error and KL divergence until the parameters of the variational autoencoder converge, and obtain the reduced-dimensional potential feature vector; Inputting the reduced-dimensional potential feature vector into a feature generator to generate a new operating condition feature vector, performing operating condition attribute discrimination through a domain discriminator, calculating an adversarial training loss, re-inputting the new operating condition feature vector into a feature generator to calculate a cycle consistency loss, optimizing the parameters of the feature generator based on the adversarial training loss and the cycle consistency loss, and finally obtaining a stable operating condition feature vector; A Mahalanobis distance metric function is constructed for the stable working condition feature vector, and the feature sequence evolution trajectory is calculated based on the symbolic dynamics method. The working condition state transition characteristics are extracted to obtain the distance mean and distance standard deviation of the working condition. The weighted sum of the distance mean and the distance standard deviation is used as the adaptive threshold to determine the dynamic evaluation benchmark.

5. The method according to claim 4, characterized in that The Mahalanobis distance metric function is constructed for the stable working condition feature vector, and the feature sequence evolution trajectory is calculated based on the symbolic dynamics method to extract the working condition state transition characteristics. The distance mean and distance standard deviation of the working condition are obtained, including: Calculating the Mahalanobis distance between stable operating condition feature vectors based on the Mahalanobis distance metric function, and constructing an operating condition feature sequence based on the Mahalanobis distance; Performing adaptive segmentation processing on the operating condition feature sequence to obtain a plurality of operating condition sequence segments, and performing linear fitting on the operating condition sequence segments to obtain an operating condition slope sequence; Inputting the operating condition slope sequence into a symbol mapping function for symbolization processing to obtain an operating condition state sequence, wherein the symbol mapping function adaptively determines a slope threshold based on the variance of the operating condition slope sequence; Constructing a Markov state transfer matrix based on the operating state sequence and mapping it into an operating state transfer network, wherein the nodes of the operating state transfer network represent operating state combinations, and the edges of the operating state transfer network represent operating state transfer relationships; Calculating the duration distribution of the operating state in the operating state transition network, and extracting the average duration of the operating state, the standard deviation of the operating state duration, and the operating state transition entropy from the duration distribution; The average duration of the operating condition, the standard deviation of the operating condition duration and the operating condition state transition entropy are weightedly combined to obtain an operating condition characteristic distance metric, and the distance mean and distance standard deviation of the operating condition are extracted from the operating condition characteristic distance metric.

6. The method according to claim 1, characterized in that Based on dynamic evaluation benchmarks, we integrate gradient strategy optimization and temporal causal inference to predict optimal evaluation parameters. We use confidence interval calibration to correct prediction deviations and output certification scoring criteria including: Constructing a time series feature vector based on the dynamic evaluation benchmark, the time series feature vector including a benchmark value, a first-order difference of the benchmark value, a second-order difference of the benchmark value, and a time scale factor; Input the time series feature vector into the gradient strategy optimization unit, and obtain the optimization evaluation parameter by constructing the discounted reward function and iteratively optimizing the strategy parameters; input the time series feature vector into the time series causal inference unit, and obtain the causal evaluation parameter by constructing the time series causal graph and calculating the causal effect strength; Performing adaptive weighted combination on the optimization evaluation parameter and the causal evaluation parameter to obtain a prediction evaluation parameter; A multi-dimensional parameter constraint set is established based on the prediction evaluation parameters. The benchmark sensitivity index is introduced to construct the benchmark influence function. The gradient iteration of the benchmark influence is used to optimize the prediction evaluation parameters. The benchmark influence change rate is used as the iteration termination condition to obtain the optimal evaluation parameters. Calculating the confidence interval of the predicted evaluation parameter and calibrating the predicted evaluation parameter using the deviation calibration coefficient calculated from historical data to obtain the optimal evaluation parameter; Based on the optimal evaluation parameters, a weighted combination of normalized score, confidence score and stability score is used to generate the certification score standard.

7. The method according to claim 6, characterized in that A multi-dimensional parameter constraint set is established based on the prediction evaluation parameters. The benchmark sensitivity index is introduced to construct the benchmark influence function. The gradient iteration of the benchmark influence is used to optimize the prediction evaluation parameters. The benchmark influence change rate is used as the iteration termination condition. The optimal evaluation parameters include: Constructing a parameter constraint set based on the prediction evaluation parameters, wherein the parameter constraint set includes parameter range constraints, parameter change rate constraints, and parameter correlation constraints; Calculating the partial derivative of the predicted evaluation parameter with respect to the evaluation benchmark to obtain a first calculation result, calculating the ratio of the predicted evaluation parameter to the evaluation benchmark to obtain a second calculation result, multiplying the first calculation result by the second calculation result to obtain a benchmark sensitivity index, and constructing a benchmark influence function based on the benchmark sensitivity index, where the benchmark influence function is a weighted sum of the benchmark sensitivity indexes of each dimension; Taking the predicted evaluation parameter as an initial parameter value, verifying whether the initial parameter value satisfies the constraint condition corresponding to the parameter constraint set, and calculating the benchmark influence corresponding to the initial parameter value using a benchmark influence function; Based on the gradient direction of the benchmark influence, the initial parameter value is iteratively updated to obtain the iterative parameter value, the first benchmark influence corresponding to the iterative parameter value under the current number of iterations is calculated, the second benchmark influence corresponding to the iterative parameter value under the previous number of iterations is calculated, the difference between the first benchmark influence and the second benchmark influence is calculated, and the result is divided by the second benchmark influence to obtain the benchmark influence change rate. When the benchmark influence change rate is less than the preset change threshold, the current iterative parameter value is determined as the optimal evaluation parameter.

Images