Green certification evaluation method for renewable energy hydrogen-based energy

Abstract

The application provides a renewable energy hydrogen production-based energy green certification evaluation method, relates to the technical field of renewable energy hydrogen production, and comprises the following steps: collecting hydrogen production system operation data to form a time series data set, performing wavelet decomposition on the time series data set to extract features and generate a characteristic fingerprint spectrum, performing dimensionality reduction reconstruction by using a variational autoencoder to form a dynamic evaluation benchmark, fusing gradient strategy optimization and time series causal inference to predict evaluation parameters, finally calculating a green certification score, and generating an evaluation report and recording the evaluation report on a block chain. The application realizes intelligent evaluation of hydrogen production system green certification, and improves the accuracy and reliability of the evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of renewable energy hydrogen production technology, in particular to a green certification evaluation method for renewable energy hydrogen production based energy. BACKGROUND

[0002] With the promotion of energy transformation and carbon neutralization goals, renewable energy hydrogen production, as an important form of clean energy carrier, is increasingly concerned. Renewable energy hydrogen production system drives water electrolysis to produce hydrogen through renewable energy such as photovoltaic and wind energy, realizing clean conversion and storage of energy. In order to ensure the green property of hydrogen energy, a scientific and complete green certification evaluation system needs to be established to quantitatively evaluate the renewable energy utilization rate and system efficiency of the whole hydrogen production process.

[0003] However, the existing renewable energy hydrogen production evaluation method mainly has the following shortcomings: the evaluation index system is static and fixed, which is difficult to adapt to the dynamic characteristic changes of the system under different working conditions; the collection and processing of evaluation data lack systematicness, which cannot effectively identify and eliminate the interference caused by abnormal fluctuations; the construction method of evaluation benchmark is simple, and the adaptability problem under the working condition migration scene is not fully considered; the credibility and traceability of the evaluation result are poor, which is difficult to support the public credit requirement 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 working condition migration capability, integrating gradient optimization and time series causal inference evaluation parameter prediction method, and combining blockchain technology to ensure the credible evidence of the evaluation process, the scientific, dynamic and credible green certification evaluation of renewable energy hydrogen production system is realized. The present application can solve the problems in the prior art. SUMMARY

[0005] The present application provides a green certification evaluation method for renewable energy hydrogen production based energy, which can solve the problems in the prior art.

[0006] The first aspect of the embodiment of the present application is,

[0007] A green certification evaluation method for renewable energy hydrogen production based energy is provided, comprising:

[0008] Collecting operation data of the renewable energy hydrogen production system, including hydrogen production process parameters, renewable energy input parameters, and energy conversion efficiency parameters, to form a time series data set;

[0009] Wavelet decomposition is performed on the time series data set to extract multi-scale time-frequency features, the dynamic characteristics of the system are identified by combining entropy analysis, and abnormal fluctuations are removed by adaptive threshold to generate a feature fingerprint spectrum of the system;

[0010] The feature fingerprint spectrum is reconstructed by using a variational autoencoder for dimensionality reduction, a latent feature vector is extracted, the feature vector of a known working condition is migrated to a new working condition through adversarial training to form a dynamic evaluation benchmark;

[0011] Based on the dynamic evaluation benchmark, the gradient strategy optimization and the time series causal inference are fused to predict the optimal evaluation parameter, the prediction deviation is corrected by confidence interval calibration, and the authentication score standard is output;

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

[0013] In an optional embodiment,

[0014] The time series data set is wavelet decomposed to extract multi-scale time-frequency features, the dynamic characteristics of the system are identified by combining entropy analysis, and abnormal fluctuations are removed by adaptive threshold to generate the feature fingerprint spectrum of the system, including:

[0015] The time series data is discrete wavelet transformed by using a Daubechies wavelet basis function, wavelet coefficients and scale coefficients are generated based on the wavelet mother function, and the time series data is decomposed into a plurality of frequency sub-signals according to the wavelet coefficients and the scale coefficients;

[0016] The sum of squares of the wavelet coefficients is calculated to obtain energy eigenvalues, the mean, standard deviation, skewness and kurtosis of the frequency sub-signals are extracted to determine statistical eigenvalues, the derivative values of the phase function of the frequency sub-signals are calculated to obtain instantaneous frequency characteristics, and the frequency sub-signals are integrated in the time-frequency domain to obtain edge spectrum characteristics;

[0017] A multi-dimensional vector and a multi-resolution tree structure are constructed for the frequency sub-signals, sample entropy and permutation entropy are calculated respectively, the frequency sub-signals are coarsely granulated to obtain multi-scale sequences, and multi-scale entropy feature vectors are obtained by sparse adaptive dictionary optimization;

[0018] The abnormal scores of the frequency sub-signals are calculated by using a local outlier factor, a dynamic threshold is determined based on the mean and standard deviation of the abnormal scores, and abnormal fluctuations are removed according to the dynamic threshold;

[0019] The energy eigenvalues, the statistical eigenvalues, the instantaneous frequency characteristics, the edge spectrum characteristics and the multi-scale entropy feature vectors are combined to form a feature matrix, the feature matrix is normalized by maximum and minimum value to obtain a normalized feature matrix, and the normalized feature matrix is dimensionally reduced by using a t-SNE method to generate a feature fingerprint spectrum.

[0020] In an optional embodiment,

[0021] constructing a multi-dimensional vector and a multi-resolution tree structure for the frequency sub-signals, respectively calculating sample entropy and permutation entropy, performing coarse-grained processing on the frequency sub-signals to obtain a multi-scale sequence, and obtaining a multi-scale entropy feature vector through sparse adaptive dictionary optimization, comprising:

[0022] reconstructing the frequency sub-signals into a multi-dimensional vector, the multi-dimensional vector containing consecutive data points under a preset time delay, calculating a Chebyshev distance based on the multi-dimensional 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 counting;

[0023] dividing the frequency sub-signals into sub-sequences of different scales using a recursive bisection method, establishing a hierarchical tree structure, wherein each node contains a sequence fragment at a corresponding scale, forming a multi-resolution sub-sequence tree structure, performing local sorting at each node of the multi-resolution sub-sequence tree structure to obtain sequence permutation patterns, and calculating a permutation entropy value based on the conditional probability distribution of the sequence permutation patterns at different levels of nodes;

[0024] performing coarse-grained processing on the frequency sub-signals 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] initializing an atomic basis matrix, performing block processing on the multi-scale entropy feature vector to obtain a training sample set, iteratively optimizing the atomic basis matrix using the training sample set, updating sparse coefficients through least squares method, judging convergence based on the change amount of the sparse coefficients, obtaining a sparse adaptive dictionary, and performing sparse decomposition on the multi-scale entropy feature vector using the sparse adaptive dictionary to obtain an optimized multi-scale entropy feature vector.

[0026] In an alternative embodiment,

[0027] reducing and reconstructing the feature fingerprint spectrum using a variational autoencoder to extract a latent feature vector, and migrating the feature vector of a known working condition to a new working condition through adversarial training to form a dynamic evaluation benchmark, comprising:

[0028] inputting the feature fingerprint spectrum into the encoder of the variational autoencoder, decomposing the feature fingerprint spectrum into feature components of different frequency bands using wavelet transform, performing nonlinear function mapping on the feature components, obtaining an encoding mean vector and an encoding variance vector through weighted average aggregation, and sampling to obtain a latent feature vector, inputting the latent feature vector into the decoder for reconstruction, optimizing reconstruction error and KL divergence until the parameters of the variational autoencoder converge, and obtaining a reduced latent feature vector;

[0029] inputting the dimension-reduced latent feature vectors into a feature generator to generate new working condition feature vectors, performing working condition attribute discrimination by a domain discriminator, calculating an adversarial training loss, re-inputting the new working condition feature vectors into the feature generator to calculate a cycle consistency loss, optimizing parameters of the feature generator based on the adversarial training loss and the cycle consistency loss, and finally obtaining stable working condition feature vectors;

[0030] constructing a Mahalanobis distance metric function for the stable working condition feature vectors, calculating a feature sequence evolution trajectory based on a symbolic dynamics method, extracting working condition state transition features, obtaining a distance mean and a distance standard deviation of the working condition, taking a weighted sum of the distance mean and the distance standard deviation as an adaptive threshold, and determining a dynamic evaluation benchmark.

[0031] In an alternative embodiment,

[0032] constructing a Mahalanobis distance metric function for the stable working condition feature vectors, calculating a feature sequence evolution trajectory based on a symbolic dynamics method, extracting working condition state transition features, obtaining a distance mean and a distance standard deviation of the working condition, taking a weighted sum of the distance mean and the distance standard deviation as an adaptive threshold, and determining a dynamic evaluation benchmark.

[0033] calculating Mahalanobis distances between the stable working condition feature vectors based on the Mahalanobis distance metric function, and constructing a working condition feature sequence based on the Mahalanobis distances;

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

[0035] inputting the working condition slope sequence into a symbolic mapping function for symbolic processing to obtain a working condition state sequence, the symbolic mapping function adaptively determining a slope threshold based on a variance of the working condition slope sequence;

[0036] constructing a Markov state transition matrix based on the working condition state sequence, and mapping it into a working condition state transition network, a node of the working condition state transition network representing a working condition state combination, and an edge of the working condition state transition network representing a working condition state transition relationship;

[0037] calculating a duration distribution of working condition states in the working condition state transition network, extracting a working condition state average duration, a working condition state duration standard deviation, and a working condition state transition entropy from the duration distribution;

[0038] performing weighted combination on the working condition state average duration, the working condition state duration standard deviation, and the working condition state transition entropy to obtain a working condition feature distance metric, and extracting a distance mean and a distance standard deviation of the working condition from the working condition feature distance metric.

[0039] In an alternative embodiment,

[0040] Based on the dynamic evaluation benchmark, the gradient strategy optimization and the time sequence causal inference are fused to predict the optimal evaluation parameter, the prediction deviation is corrected by the confidence interval calibration, and the authentication score standard is output, including:

[0041] A time sequence feature vector is constructed based on the dynamic evaluation benchmark, and the time sequence feature vector includes 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] The time sequence feature vector is input into a gradient strategy optimization unit to obtain an optimized evaluation parameter by constructing a discount reward function and iteratively optimizing strategy parameters; and the time sequence feature vector is input into a time sequence causal inference unit to obtain a causal evaluation parameter by constructing a time sequence causal graph and calculating a causal effect strength;

[0043] The optimized evaluation parameter and the causal evaluation parameter are adaptively weighted and combined to obtain a predicted evaluation parameter;

[0044] A multi-dimensional parameter constraint set is established based on the predicted evaluation parameter, a benchmark sensitivity index is introduced to construct a benchmark influence degree function, the predicted evaluation parameter is optimized by gradient iteration of the benchmark influence degree, and the optimal evaluation parameter is obtained by taking the benchmark influence degree change rate as an iteration termination condition;

[0045] The confidence interval of the predicted evaluation parameter is calculated, and the predicted evaluation parameter is calibrated by a deviation calibration coefficient calculated based on historical data to obtain the optimal evaluation parameter;

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

[0047] In an optional embodiment,

[0048] A multi-dimensional parameter constraint set is established based on the predicted evaluation parameter, a benchmark sensitivity index is introduced to construct a benchmark influence degree function, the predicted evaluation parameter is optimized by gradient iteration of the benchmark influence degree, and the optimal evaluation parameter is obtained by taking the benchmark influence degree change rate as an iteration termination condition, including:

[0049] A parameter constraint set is constructed based on the predicted evaluation parameter, and the parameter constraint set includes a parameter value domain constraint, a parameter change rate constraint and a parameter correlation constraint;

[0050] A first calculation result is obtained by calculating the partial derivative of the predicted evaluation parameter with respect to the evaluation benchmark, a second calculation result is obtained by calculating the ratio of the predicted evaluation parameter to the evaluation benchmark, the benchmark sensitivity index is obtained by multiplying the first calculation result and the second calculation result, the benchmark influence degree function is constructed based on the benchmark sensitivity index, and the benchmark influence degree function is a weighted sum of the benchmark sensitivity indexes in each dimension.

[0051] The prediction evaluation parameter is taken as an initial parameter value, whether the initial parameter value satisfies the constraint condition corresponding to the parameter constraint set is verified, and a benchmark influence degree corresponding to the initial parameter value is calculated through a benchmark influence degree function;

[0052] Based on the gradient direction of the benchmark influence degree, the initial parameter value is iteratively updated to obtain an iterative parameter value, a first benchmark influence degree corresponding to the iterative parameter value at the current iteration number is calculated, a second benchmark influence degree corresponding to the iterative parameter value at the last iteration number is calculated, a difference value between the first benchmark influence degree and the second benchmark influence degree is calculated, and a benchmark influence degree change rate is obtained by dividing the second benchmark influence degree, and when the benchmark influence degree change rate is less than a preset change threshold, the current iterative parameter value is determined as the optimal evaluation parameter.

[0053] In the embodiment of the present application, by performing multi-scale time-frequency analysis and entropy value feature extraction on the renewable energy hydrogen production system operation data, combining with self-adaptive threshold to remove abnormal fluctuations, a system characteristic fingerprint spectrum is established, the accuracy and robustness of data feature expression are improved, and a reliable data basis is provided for green certification evaluation; the variational autoencoder is used for feature dimension reduction reconstruction, the working condition feature migration is realized through adversarial training, the dynamic evaluation benchmark is established, and the gradient strategy optimization and time sequence causal inference are combined to predict the optimal evaluation parameter, the adaptability of the evaluation method to different working conditions and the accuracy of the evaluation result are improved; the evaluation whole process data and result are recorded in the blockchain to form the certification chain, the traceability and non-tamperability of the evaluation process and result are realized, the credibility of the green certification is improved, and a scientific and reliable technical support is provided for the green certification evaluation of the renewable energy hydrogen production system. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 It is a flowchart of the renewable energy hydrogen production green energy source certification evaluation method of the embodiment of the present application;

[0055] Figure 2 It is a radar chart for feature extraction performance and calculation complexity analysis;

[0056] Figure 3 It is a working condition state duration distribution curve chart;

[0057] Figure 4 It is a working condition state duration distribution analysis table;

[0058] Figure 5 It is a benchmark influence degree change scatter plot. DETAILED DESCRIPTION

[0059] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0060] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and some embodiments may not be described again for the same or similar concepts or processes.

[0061] Figure 1 The flowchart of the renewable energy hydrogen production-based green certification evaluation method of the embodiments of the present application is shown in FIG. 1, which comprises the following steps. Figure 1

[0062] Collecting operation data of the renewable energy hydrogen production system, including hydrogen production process parameters, renewable energy input parameters, and energy conversion efficiency parameters, to form a time series data set;

[0063] Performing wavelet decomposition on the time series data set to extract multi-scale time-frequency features, combining entropy analysis to identify the dynamic characteristics of the system, and removing abnormal fluctuations through an adaptive threshold to generate a characteristic fingerprint spectrum of the system;

[0064] Using a variational autoencoder to reduce and reconstruct the characteristic fingerprint spectrum to extract a latent feature vector, and migrating the feature vector of a known working condition to a new working condition through adversarial training to form a dynamic evaluation benchmark;

[0065] Based on the dynamic evaluation benchmark, combining gradient strategy optimization and time series causal inference to predict optimal evaluation parameters, correcting the prediction deviation through a confidence interval, and outputting a certification score standard;

[0066] According to the certification score standard, calculating the green certification score of the system, generating an evaluation report containing the time series data set, the characteristic fingerprint spectrum, the dynamic evaluation benchmark, the certification score standard and the green certification score, and recording in the blockchain to form a certification chain.

[0067] In one specific embodiment, a data acquisition device is deployed to monitor the renewable energy hydrogen production system in real time. The collected hydrogen production process parameters include cell voltage, current, temperature, pressure, hydrogen production, and other operation data; the collected renewable energy input parameters include photovoltaic power generation power, wind turbine power generation power, grid power compensation power, and other energy input data; and the collected energy conversion efficiency parameters include electrolysis efficiency, system comprehensive efficiency, and energy utilization rate. All the collected parameters are sorted according to a unified time stamp to form a time series data set containing multi-dimensional parameters.​

[0068] The parameters are decomposed into multiple layers using discrete wavelet transform to obtain approximation coefficients and detail coefficients at different frequency scales, which reflect the characteristics of the parameters at different time scales. Then, the sample entropy of each scale coefficient is calculated to quantify the complexity and uncertainty of the parameter sequence and identify the dynamic characteristics of the system. An adaptive kernel density estimation method is used to determine the anomaly detection threshold to eliminate abnormal values outside the normal fluctuation range. Finally, the processed multi-scale features are combined to generate a feature fingerprint spectrum that can represent the system's running characteristics.

[0069] The feature fingerprint spectrum is reduced using a variational autoencoder. The encoder maps high-dimensional features to a low-dimensional latent space, and the decoder reconstructs the latent features to the original feature space, optimizing the network parameters through reconstruction error. The latent feature vector output by the encoder is extracted as a compact representation of the system. For the new working condition evaluation scenario, an adversarial training method is used to realize feature transfer. The discriminator judges the source of the features, and the generator maps the known working condition features to the new working condition distribution, realizing the construction of a cross-working condition dynamic evaluation benchmark.

[0070] Based on the constructed dynamic evaluation benchmark, a gradient strategy optimization algorithm is designed to project the evaluation parameters into the benchmark space, calculate the deviation of the parameters from the benchmark, and iteratively optimize the parameters along the gradient direction of the deviation. Meanwhile, a graph convolution network-based temporal causal inference module is introduced to model the dynamic dependency between parameters and constrain the optimization direction of the parameters. For the optimized evaluation parameters, the confidence interval is calculated, and a calibration coefficient is designed based on the historical evaluation deviation to correct it. Finally, a reliable authentication score standard is output.

[0071] According to the authentication score standard, the green authentication comprehensive score of the system is calculated by combining the weights of each parameter. A complete evaluation report is generated, which includes the original time series data set, the processed feature fingerprint spectrum, the constructed dynamic evaluation benchmark, the determined authentication score standard, and the final green authentication score. The evaluation report is hashed to generate a unique identifier, which is recorded in the blockchain network to form an unalterable authentication chain, ensuring the traceability and credibility of the evaluation results.

[0072] Each authentication record on the blockchain contains information such as timestamp, evaluation report hash, and evaluation agency signature. When verifying the authentication results, the complete authentication record can be queried through the blockchain, and the authenticity can be verified by comparing the evaluation report hash value, achieving the whole-process credible evidence of the authentication results.

[0073] In an alternative embodiment, the time series data set is decomposed using wavelet decomposition to extract multi-scale time-frequency features, combined with entropy analysis to identify the dynamic characteristics of the system, and abnormal fluctuations are removed through adaptive thresholding to generate a feature fingerprint spectrum of the system, including:

[0074] Discrete wavelet transform is performed on the time series data by using a Daubechies wavelet basis function, wavelet coefficients and scale coefficients are generated based on a wavelet mother function, and the time series data is decomposed into a plurality of frequency sub-signals according to the wavelet coefficients and the scale coefficients;

[0075] An energy eigenvalue is obtained by calculating the square sum of the wavelet coefficients, a mean value, a standard deviation, a skewness and a kurtosis of the frequency sub-signals are extracted to determine statistical eigenvalues, an instantaneous frequency feature is obtained by calculating a derivative value of a phase function of the frequency sub-signals, and an edge spectrum feature is obtained by integrating the frequency sub-signals in a time-frequency domain;

[0076] A multi-dimensional vector and a multi-resolution tree structure are constructed for the frequency sub-signals, sample entropy and permutation entropy are respectively calculated, a multi-scale sequence is obtained by performing coarse-grained processing on the frequency sub-signals, and a multi-scale entropy feature vector is obtained by optimization through a sparse adaptive dictionary;

[0077] An abnormal score of the frequency sub-signals is calculated by using a local outlier factor, a dynamic threshold is determined based on a mean value and a standard deviation of the abnormal score, and abnormal fluctuations are removed according to the dynamic threshold;

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

[0079] In a specific embodiment, data collection is performed on the voltage of an electrolytic cell of a renewable energy hydrogen production system, the collection interval is 10 minutes, continuous collection is performed for 24 hours, 144 time series sampling points are obtained, and the collected voltage values range from 1.8V to 2.2V. A db4 wavelet basis function in a Daubechies wavelet basis function family is selected as a decomposition tool, and three-layer discrete wavelet transform is performed on the collected voltage data. During the wavelet transform process, the signal is progressively decomposed by using a high-pass filter and a low-pass filter, and finally three detail coefficient sequences d1, d2 and d3 of different frequency bands and an approximation coefficient sequence a3 are obtained, which correspond to the characteristic performance of the original signal in different frequency ranges, respectively.

[0080] The energy features of each frequency band signal are calculated by extracting the features of the coefficient sequence obtained by decomposition, and the energy values of the high frequency band, the medium frequency band, the low frequency band and the base frequency band are obtained by square sum operation on the wavelet coefficient sequence, which are 0.15, 0.08, 0.05 and 0.02 respectively. The statistical features of each frequency band signal are calculated, the mean value is obtained by calculating the arithmetic mean value of the signal sequence, which is 2.0V, the standard deviation is obtained by calculating the mean value of the square sum of the difference between all data points and the mean value and taking the square root, which is 0.12V, the skewness is obtained by calculating the third central moment, which is 0.03, and the kurtosis is obtained by calculating the fourth central moment, which is 2.98. The signal is analyzed by Hilbert transform, the derivative of the phase function is calculated to obtain the instantaneous frequency feature, and it is found that the main frequency component of the signal is concentrated between 0.001Hz and 0.1Hz. The edge distribution features of the signal in time domain and frequency domain are obtained by integrating the signal along the time axis and the frequency axis respectively in the time-frequency plane, and the results show that the signal energy is mainly concentrated in the low frequency band of 0 to 0.01Hz.

[0081] The multi-dimensional phase space reconstruction vector and multi-resolution tree structure of the signal are constructed, the sample entropy value 1.25 is calculated based on the reconstruction vector, and the permutation entropy value 0.85 is calculated based on the symbol sequence. The signal is converted into sequences of different time scales by coarse graining processing, an overcomplete dictionary containing multiple atomic signals is established, and the orthogonal matching pursuit algorithm is used to optimize the dictionary, finally a ten-dimensional entropy feature vector is obtained which can represent the complexity of the signal at different time scales.

[0082] The abnormality degree of each data point is calculated by using the local outlier factor algorithm, which evaluates the abnormality degree by calculating the local density ratio of the data point and its neighbor samples. Statistical analysis is performed on the abnormal scores of all data points, and the mean value of the abnormal scores is 0.5 and the standard deviation is 0.2. Based on these statistical quantities, the dynamic threshold interval is set to 0.1 to 0.9. Four abnormal data points whose abnormal scores exceed this interval are removed, and the data points within the normal fluctuation range are retained.

[0083] The energy features, statistical features, frequency features and entropy features extracted in the foregoing steps are combined into a high-dimensional feature matrix. Each dimension of the feature matrix is normalized by maximum and minimum value, and the feature values of different dimensions are mapped to the uniform interval of zero to one. The t-SNE algorithm is used to reduce the dimensionality of the normalized high-dimensional feature matrix, which constructs the conditional probability distribution of high-dimensional space and low-dimensional space, minimizes the KL divergence between the two distributions, and finally reduces the high-dimensional features to a two-dimensional plane to form a feature fingerprint spectrum with visual interpretability. In the obtained feature fingerprint spectrum, data points with similar running states tend to cluster, showing a typical butterfly distribution structure, and different running states form relatively separate clusters on the two-dimensional plane.

[0084] In the embodiment, the time series data is decomposed into multiple frequency sub-signals using Daubechies wavelet basis function, and multi-dimensional features such as energy, statistics, instantaneous frequency, edge spectrum, etc. are extracted, so that the system dynamic characteristics are captured more comprehensively and accurately; by constructing a multi-dimensional vector, a multi-resolution tree structure and calculating sample entropy and permutation entropy, the signal is coarsely granulated and the sparse adaptive dictionary is optimized, which can deeply reveal the complex time-varying law inside the signal; the local outlier factor is used to calculate the abnormal score, and the abnormal fluctuations are removed through a dynamic threshold, thereby improving the 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 recognition.

[0085] In an optional embodiment, a multi-dimensional vector and a multi-resolution tree structure are constructed for the frequency sub-signals, sample entropy and permutation entropy are respectively calculated, the frequency sub-signals are coarsely granulated to obtain multi-scale sequences, and a multi-scale entropy feature vector is obtained through sparse adaptive dictionary optimization, including:

[0086] The frequency sub-signals are reconstructed into a multi-dimensional vector, the multi-dimensional vector contains consecutive data points under a preset time delay, the Chebyshev distance is calculated based on the multi-dimensional vector, the mode matching count is performed on the Chebyshev distance according to a preset similarity threshold, and the sample entropy value is calculated based on the mode matching count;

[0087] The frequency sub-signals are divided into sub-sequences of different scales using a recursive bisection method, a hierarchical tree structure is established, each node contains a sequence fragment at a corresponding scale, a multi-resolution sub-sequence tree structure is formed, a local order is performed at each node of the multi-resolution sub-sequence tree structure to obtain a sequence permutation pattern, the conditional probability distribution of the sequence permutation pattern at different hierarchical nodes is counted, and the permutation entropy value is calculated based on the conditional probability distribution;

[0088] The frequency sub-signals are coarsely granulated to obtain multi-scale sequences, the sample entropy value sequence and the permutation entropy value sequence of the multi-scale sequences are respectively calculated, and the sample entropy value sequence and the permutation entropy value sequence are combined into a multi-scale entropy feature vector;

[0089] An atomic basis matrix is initialized, the multi-scale entropy feature vector is block-processed to obtain a training sample set, the atomic basis matrix is iteratively optimized using the training sample set, sparse coefficients are updated by a least square method, convergence is judged based on the change amount of the sparse coefficients, a sparse adaptive dictionary is obtained, and the multi-scale entropy feature vector is sparsely decomposed using the sparse adaptive dictionary to obtain an optimized multi-scale entropy feature vector.

[0090] In a specific embodiment, the frequency sub-signals obtained by wavelet decomposition are reconstructed, the time delay parameter is set to 3 sampling points, and each time 5 consecutive data points are taken to construct a multi-dimensional vector. The Chebyshev distance is calculated for each pair of the constructed multi-dimensional vectors, and the distance is defined as the maximum value of the difference between the dimensions of the vectors. The similarity threshold is set to 0.2, and the number of vector pairs with a distance less than the threshold is counted as the pattern matching count. By calculating the ratio of the pattern matching counts in different dimensions, the sample entropy value is obtained, which is used to quantify the complexity of the signal.

[0091] The frequency sub-signals are divided into multiple levels using the recursive bisection method. Starting from the top level, each time the sequence is equally divided into two sub-sequences, and the division process is repeated until the preset minimum sequence length is reached. In this way, a multi-level tree structure is formed, where the root node contains the complete sequence, and each sub-node contains a partial sequence at the corresponding scale. At each node of the tree, the data points in the sequence segment are locally sorted to obtain ascending or descending arrangement patterns. The transition probabilities of these arrangement patterns between adjacent level nodes are counted, and the permutation entropy value is calculated based on these conditional probabilities to reflect the dynamic evolution characteristics of the signal.

[0092] The frequency sub-signals are subjected to coarse-grained processing, and the original sequence is resampled through a sliding window, with the window size gradually increasing from 2 to 20, to obtain coarse-grained sequences at different time scales. The sample entropy and permutation entropy are calculated for each sequence at each scale to obtain two groups of entropy value sequences. These two groups of entropy value sequences are combined in the order of time scale to form a feature vector representing the multi-scale complexity of the signal.

[0093] A basis matrix containing 100 atomic basis vectors is initialized, and the dimension of each atomic basis vector is the same as that of the multi-scale entropy feature vector. The multi-scale entropy feature vector is divided into several data blocks of equal length, which are used as training samples for dictionary learning. An iterative optimization method is used to update the atomic basis matrix. In each iteration, the atomic basis is first fixed, and the sparse coefficients are solved by least squares method. Then, the sparse coefficients are fixed, and the atomic basis is updated. When the relative change in sparse coefficients is less than 0.001, the optimization is considered to have converged, and the final sparse adaptive dictionary is obtained. The original multi-scale entropy feature vector is sparsely decomposed using the dictionary to obtain the optimized feature representation.

[0094] Exemplarily, the current signal of an electrolytic cell of a hydrogen production system is processed, the sampling frequency is 1 Hz, and 2048 data points are processed. 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 includes 128 sequence segments with a length of 16, and the permutation entropy value is calculated to be 2.15. Coarse-grained processing is performed to obtain 19 sequences of different scales, and 19-dimensional sample entropy sequences and 19-dimensional permutation entropy sequences are calculated, which are combined to form a 38-dimensional multi-scale entropy feature vector. The 100*38 atomic matrix is initialized, the feature vector is divided into 20 training samples, and the sparse adaptive dictionary is obtained after 50 iterations of optimization convergence. The original feature vector is decomposed using the dictionary to obtain a 30-dimensional optimized feature vector with better noise reduction effect and feature expression ability.

[0095] In the prior art, signal complexity analysis mainly relies on single-scale entropy measure methods, such as single-dimensional sample entropy or permutation entropy for feature extraction. This method has obvious limitations, and can only reflect the complexity 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 method usually directly performs coarse-grained processing on the original signal, and this processing method is easily disturbed by noise, and the calculation process is low in efficiency. In addition, the traditional method often uses sample entropy and permutation entropy as independent features, and does not fully utilize the complementary properties between the two entropy measures. In terms of feature optimization, the current method mainly uses fixed feature extraction algorithms, which lack necessary adaptability and are difficult to dynamically adjust and optimize the signal features under different working conditions.

[0096] The embodiment proposes a dual-path entropy measure framework based on the combination of multi-dimensional vector reconstruction and multi-resolution tree structure. The framework captures the local dynamic characteristics of the signal through the multi-dimensional vector reconstruction method, and describes the multi-scale evolution law of the signal through the multi-resolution tree structure, realizing the multi-angle comprehensive description of the signal complexity. In specific implementation, the application designs a sample entropy calculation method based on Chebyshev distance and a permutation entropy calculation method based on conditional probability. Among them, Chebyshev distance can better reflect the extreme difference between vectors, and conditional probability can effectively represent the state transition characteristics of the signal, and the combination of the two methods significantly improves the discrimination ability of the entropy measure.

[0097] The sparse adaptive dictionary learning mechanism is also innovatively introduced, and the adaptive extraction and optimization of features are realized by iteratively optimizing the atomic basis matrix. This mechanism can dynamically adjust the feature representation according to the signal characteristics, effectively improving the expression ability and noise resistance of the features. This improvement mainly aims at three aspects: first, to solve the problem that a single entropy measure cannot comprehensively represent the signal complexity, a multi-dimensional and multi-scale entropy measure combination is used to realize complete description of the signal features; second, to solve the problem of weak noise resistance of traditional methods, a sparse adaptive optimization mechanism is introduced to improve the robustness of feature extraction; and finally, a dynamically adjustable feature optimization framework is designed to make the feature representation better adapt to changes in different working conditions.

[0098] The embodiment has significant technical advantages in practical application. Taking a certain hydrogen production system as an example, compared with the traditional single entropy measure method, the features extracted by the present application have an accuracy rate of working condition recognition improved by 15% and an anti-noise performance improved by 20%. Through the combination of multi-dimensional vector reconstruction and multi-resolution tree structure, the distinguishing ability of the features is obviously improved, and the feature distribution under different working conditions is more clear and separable. After using the sparse adaptive dictionary optimization, the feature dimension is reduced from 38 to 30, while maintaining a high information expression ability, and the calculation efficiency is improved by 25%. Especially in the dynamic working condition change scene, the present application can adaptively adjust the feature extraction strategy, and the adaptability of the features is improved by 30% compared with the fixed feature extraction algorithm. These improvements enable the present application to more accurately and efficiently represent the running state of the hydrogen production system, providing reliable technical support for state evaluation and fault diagnosis.

[0099] As shown in the radar chart Figure 2 The radar chart shows the comprehensive performance of the four different methods in feature extraction. In the figure, different line types of polygons represent the performance of each method: solid line polygons represent the present technical solution, short dashed line polygons represent the traditional sample entropy method, dotted line polygons represent the single scale permutation entropy method, and long dashed line polygons represent the unoptimized multi-scale entropy method. Five key dimensions are evaluated: feature extraction efficiency, noise resistance, computational complexity, information retention, and classification accuracy (each with a maximum score of 10).

[0100] The technical solution (solid polygon) achieves 9.2 points in feature extraction performance, which is significantly better than the traditional sample entropy method (6.4 points, dashed polygon), the single scale permutation entropy method (7.1 points, dotted polygon) and the unoptimized multi-scale entropy method (8.0 points, long dashed polygon). In particular, in the noise resistance performance dimension, the technical solution obtains a high score of 9.5 points, reflecting its robustness in complex signal environment, while the other methods are only 5.8 points, 6.7 points and 7.9 points respectively. In terms of computational complexity, the technical solution is optimized by sparse adaptive dictionary, although it introduces additional calculation steps, the score is 7.1 points, lower than the traditional method of 8.6 points, but considering its efficient feature extraction capability, this calculation cost is acceptable. In terms of information retention, the technical solution achieves 8.9 points, retaining most of the effective information of the original signal, while the traditional method is only 6.5 points, the single scale permutation entropy method is 7.2 points, and the unoptimized multi-scale entropy method is 7.8 points. Most significantly, in terms of classification accuracy, the technical solution achieves a high performance of 9.3 points, with an accuracy of 96.7% in the actual hydrogen production system state classification task, while the other methods are 7.2 points (86.3%), 7.8 points (89.1%) and 8.3 points (92.4%) respectively. From the shape of the radar chart, it can be seen that the polygon area of the technical solution is the largest, and the shape is more balanced, indicating that it has superior performance in all aspects, especially in the two key indicators of noise resistance performance and classification accuracy.

[0101] In the embodiment, by reconstructing the frequency sub-signals into multi-dimensional vectors and using Chebyshev distance for pattern matching and sample entropy calculation, the dynamic change characteristics of the signals under the preset time delay can be effectively captured; the recursive dichotomy is used to divide the sub-sequences, a hierarchical tree structure is constructed, and the permutation entropy value is calculated based on the sequence permutation pattern, which is helpful to fully reveal the structure and complexity of the signals under different scales; the sample entropy and permutation entropy are combined to form a multi-scale entropy feature vector, which realizes multi-dimensional description of the uncertainty and complexity of the signals, enhances the sensitivity and discrimination ability of the system to small changes; by iteratively optimizing the atomic basis matrix and updating the sparse coefficients using the least square method, 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 alternative embodiment, the feature fingerprint is processed to achieve dimension reduction and feature extraction through a variational autoencoder. The feature fingerprint is input into an encoder network, which first uses 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, each layer of which includes a convolutional layer, a batch normalization layer, and an activation layer. The mapped features are weighted and averaged through an attention mechanism to generate an encoding mean vector and an encoding variance vector. Random sampling is performed based on these two vectors to obtain a low-dimensional latent feature vector. The latent feature vector is input into a decoder network for reconstruction, which gradually restores the low-dimensional features to the original dimension through inverse convolutional layers. By minimizing two objective functions, reconstruction error and KL divergence, the network parameters of the variational autoencoder are iteratively optimized until the training loss converges stably, and the latent feature vector after dimension reduction is finally obtained.

[0103] Then, condition transfer training is performed to build an adversarial network containing a feature generator and a domain discriminator. The feature generator adopts an encoder-decoder structure, inputs the latent feature vector after dimension reduction, and generates a feature vector in the target condition through multi-layer residual blocks. The domain discriminator adopts a multi-layer discriminant network to discriminate the generated feature vector according to the condition attribute, and outputs a discrimination probability. During training, the feature generator attempts to generate features that can deceive the discriminator, while the discriminator strives to accurately identify the condition attribute of the features, forming an adversarial game. Meanwhile, the generated feature vector in the new condition is re-input into the generator to calculate the cycle consistency loss with the original feature vector. By combining the adversarial loss and the cycle consistency loss, the generator parameters are optimized through backpropagation until the generated feature vector has stable target condition attributes.

[0104] A dynamic evaluation benchmark is constructed to calculate the Mahalanobis distance metric for the stable condition feature vector, which takes into account the covariance structure of the feature distribution. Then, the symbolic dynamics method is used to analyze the evolution law of the feature sequence, dividing the continuous feature space into a finite number of symbolic intervals, and calculating the transition probability of feature points between different symbolic intervals to extract the dynamic transition features of the condition state. Based on the Mahalanobis distance, the distance mean and distance standard deviation of the condition feature are calculated, and these two statistics are weighted and summed to obtain an adaptive threshold that dynamically adjusts with the distribution of the condition feature, and finally determines the dynamic evaluation benchmark.

[0105] Exemplarily, the operation data of a certain hydrogen production system under standard working conditions and high load working conditions is processed. A 1000-dimensional feature fingerprint spectrum is input into a variational autoencoder, and 5 feature components in different frequency bands are obtained through 4-layer wavelet decomposition. After nonlinear mapping by a 3-layer convolutional neural network, a 128-dimensional encoding mean vector and variance vector are obtained. A 64-dimensional latent feature vector is obtained by sampling, and the original feature is reconstructed by a decoder. After 1000 rounds of iterative training, the reconstruction error is reduced to below 0.05, and the KL divergence converges to about 0.1.

[0106] The reduced-dimensional feature vector is input into a feature generator, and the generator includes 6 residual blocks, each of which is composed of two convolutional layers and a skip connection. The domain discriminator adopts a 4-layer fully connected network, and outputs a working condition discrimination probability. After 500 rounds of adversarial training, the generator can stably migrate the standard working condition features to the high load working condition, and the discrimination accuracy of the discriminator is stably around 50%, indicating that the generated features have the true target working condition attribute.

[0107] The Mahalanobis distance of the generated working condition features is calculated, and the distance mean is 0.8 and the standard deviation is 0.2. Through symbolic dynamics analysis, the feature space is divided into 8 symbol intervals, and the state transition matrix is extracted. The distance mean and the standard deviation are combined with a weight of 6:4 to obtain an adaptive threshold of 0.56, which is used as a dynamic evaluation benchmark. The benchmark can adaptively adjust with the change of the working condition feature distribution, providing a reliable reference standard for subsequent evaluation.

[0108] In this embodiment, the variational autoencoder is used to reduce and reconstruct the feature fingerprint spectrum, and the wavelet transform decomposition and nonlinear mapping are used to extract and optimize the latent feature vector, thereby improving the compactness and effectiveness of data representation; the known working condition feature vectors are migrated to the new working condition through adversarial training, and the stability of the generated features is ensured through the cycle consistency loss, thereby constructing a reliable dynamic evaluation benchmark; the Mahalanobis distance metric function is constructed, and the working condition state transition features are extracted by combining symbolic dynamics, and the adaptive threshold is obtained by weighting the distance mean and the standard deviation, thereby realizing accurate evaluation of the dynamic change of the system.

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

[0110] Based on the Mahalanobis distance metric function, the Mahalanobis distance between the stable working condition feature vectors is calculated, and a working condition feature sequence is constructed based on the Mahalanobis distance.

[0111] An adaptive segmentation process is performed on the working condition feature sequence to obtain a plurality of working condition sequence segments, and a working condition slope sequence is obtained by linear fitting of the working condition sequence segments.

[0112] inputting the working condition slope sequence into a symbol mapping function to obtain a working condition state sequence, the symbol mapping function adaptively determining a slope threshold value based on a variance of the working condition slope sequence;

[0113] constructing a Markov state transition matrix based on the working condition state sequence, and mapping the Markov state transition matrix into a working condition state transition network, a node of the working condition state transition network representing a working condition state combination, and an edge of the working condition state transition network representing a working condition state transition relationship;

[0114] calculating a duration distribution of the working condition states in the working condition state transition network, extracting a working condition state average duration, a working condition state duration standard deviation, and a working condition state transition entropy from the duration distribution;

[0115] performing a weighted combination of the working condition state average duration, the working condition state duration standard deviation, and the working condition state transition entropy to obtain a working condition feature distance metric, and extracting a distance mean and a distance standard deviation of the working condition from the working condition feature distance metric.

[0116] In a specific embodiment, a distance calculation is performed on a stable working condition feature vector, and a Mahalanobis distance metric function is used to calculate the distance between the feature vectors. The Mahalanobis distance takes into account the covariance structure of the feature distribution, and can reflect the correlation of the features in different dimensions. The calculated Mahalanobis distance is arranged in chronological order to form a continuous feature sequence representing the evolution process of the working condition.

[0117] An adaptive segmentation process is performed on the working condition feature sequence, and a sliding window method is used to detect the mutation points of the sequence. The window size is dynamically adjusted according to the local variation characteristics of the sequence, and a smaller window is used in the area with sharp changes, and a larger window is used in the stable area. Based on the detected mutation points, the sequence is divided into multiple continuous sequence segments. A least squares method is performed on each sequence segment to obtain a slope sequence representing the trend of the working condition change.

[0118] A symbol mapping function is constructed to perform a symbolization process on the slope sequence. The variance of the slope sequence is calculated, and the classification threshold of the slope is adaptively determined based on the variance value. According to the relationship between the slope value and the threshold value, the slope is mapped to different symbol states, such as the rising, stable and falling states, so as to obtain a discrete working condition state sequence.

[0119] A Markov state transition matrix is constructed based on the working condition state sequence, and the transition probabilities between different states are counted. The state transition matrix is converted into a network representation, in which the nodes represent different working condition state combinations, and the connecting edges between the nodes represent the transition relationships between the states, and the weights of the edges correspond to the transition probabilities. This network representation can intuitively show the dynamic evolution characteristics of the working condition states.

[0120] The duration characteristics of each state in the state transition network are analyzed, and the duration distribution of each state is counted. Three key features are extracted from the duration distribution: the average duration reflects the stability of the state, the duration standard deviation represents the fluctuation degree 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 feature distance metric. Statistical features are extracted from the metric to obtain the distance mean and distance standard deviation, which reflect the overall characteristics of the operating condition.

[0121] For example, the data of a hydrogen production system running under rated operating conditions for 4 hours is analyzed. First, the Mahalanobis distance between the 64-dimensional feature vectors is calculated, obtaining an operating condition feature sequence containing 1440 data points. The adaptive segmentation algorithm is used to process the sequence, and the window size is dynamically adjusted between 10 and 30. Finally, the sequence is divided into 25 segments. Linear fitting is performed on these segments to obtain the slope sequence reflecting the trend of operating condition changes.

[0122] The variance of the slope sequence is 0.015, and the state classification threshold is set to ±0.012 accordingly. The slope sequence is mapped to three symbol states: greater than 0.012 is an upward state (represented by "1"), less than -0.012 is a downward state (represented by "-1"), and the rest is a stable state (represented by "0"), obtaining the operating condition state sequence.

[0123] A 3x3 state transition matrix is constructed, which is converted into a state transition network containing 9 nodes. The state duration characteristics in the network are analyzed: the average duration is 15 minutes, the duration standard deviation is 3.5 minutes, and the state transition entropy is 1.2. These three features are combined according to the weights of 4:3:3 to obtain the operating condition feature distance metric. From the metric, the distance mean 0.75 and the distance standard deviation 0.18 are extracted, which comprehensively reflect the stability and fluctuation characteristics of the operating condition.

[0124] In the prior art, operating condition state feature analysis mainly uses simple distance metric methods such as Euclidean distance or cosine similarity, which do not consider the correlation between features and are difficult to accurately depict the data distribution characteristics in high-dimensional feature space. At the same time, traditional time series data analysis methods often use fixed window size segmentation processing, which is difficult to adapt to the dynamic changes of local signal characteristics. In terms of state recognition, existing methods mostly use fixed threshold discrimination rules, lack of 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 cannot fully exploit the dynamic evolution information contained in the state sequence.

[0125] The embodiment proposes a working condition feature analysis method based on Mahalanobis distance and symbolic dynamics. The method first introduces Mahalanobis distance measurement into the distance calculation of working condition feature vectors, and more accurately measures the similarity of data in high-dimensional feature space by considering the covariance structure of feature distribution. An adaptive segmentation processing mechanism is innovatively designed, which dynamically adjusts the window size according to the local change characteristics of the sequence, and realizes the fine segmentation of the working condition feature sequence.

[0126] An adaptive symbol mapping method based on variance is also introduced, which automatically determines the state classification threshold through statistical analysis of the slope sequence, and improves the adaptability and robustness of state recognition. In the aspect of state transition analysis, a Markov state transition network is innovatively constructed, which shows the state transition relationship through the network topology structure, and extracts multi-dimensional features from the duration distribution, realizing the comprehensive characterization of the dynamic characteristics of working conditions. These improvements mainly aim at three aspects: first, to improve the accuracy of feature distance measurement, and to better reflect the correlation of features through Mahalanobis distance; second, to enhance the adaptability of the analysis method, and to adapt to different working condition features by dynamically adjusting the parameters; and third, to deepen the extraction of state transition features, and to provide more rich working condition information through network representation and multi-dimensional feature combination.

[0127] The embodiment shows significant performance advantages in practical application. Taking a hydrogen production system as an example, compared with the traditional Euclidean distance method, the use of Mahalanobis distance measurement improves the feature similarity recognition accuracy by 25%. The adaptive segmentation processing mechanism makes the sequence segmentation more reasonable, and the mutation point detection accuracy is improved by 30%. The adaptive symbol mapping method based on variance improves the anti-noise ability of state recognition by 20% and the adaptability by 35% compared with the fixed threshold method.

[0128] In the aspect of state transition analysis, the Markov state transition network can provide more rich dynamic feature information compared with the traditional statistical method, and the representation ability of working condition evolution features is improved by 40%. Through the weighted combination of multi-dimensional features, the working condition feature distance measurement obtained is more comprehensive than a single statistical index, and the evaluation accuracy is improved by 28%. Especially in the scene where the working condition fluctuates greatly, the method shows stronger stability, and the consistency of the evaluation result is improved by 32%. These improvements make the invention more accurate and reliable in analyzing and evaluating the running state of the hydrogen production system, and provide strong technical support for system monitoring and optimization control.

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

[0130] As Figure 3 and Figure 4As shown, the duration distribution characteristics of different working conditions (rising, stable, and falling) in the hydrogen production system are detailed, and the differences in state duration patterns under normal working conditions, slightly abnormal working conditions, and severely abnormal working conditions are compared. It can be clearly seen from the duration distribution curve that the state duration distribution under normal working conditions is more concentrated and smooth, and the average duration of the stable state (0) is 15.0 minutes, and the standard deviation is only 3.5 minutes, indicating that the system is stable. In contrast, the state duration under abnormal working conditions is significantly shortened, and the distribution is more discrete, especially the average duration of the stable state under severe abnormal working conditions is only 3.2 minutes, which is less than 1 / 4 of the normal value. From the state transition entropy index, the stable state transition entropy of the normal working condition is the lowest (1.20), and the falling state transition entropy of the severe abnormal working condition is the highest (2.79), with an increase of 132.5%, which indicates that the uncertainty of system state transition under abnormal working conditions increases significantly. By combining these characteristics with a weight of 4:3:3 (average duration: duration standard deviation: state transition entropy), the average of the comprehensive feature distance under normal, slightly abnormal, and severely abnormal working conditions is 0.75, 1.28, and 1.96, respectively, and the standard deviation is 0.18, 0.35, and 0.68, respectively, forming a clear hierarchical structure, which provides strong support for accurate identification of abnormal working conditions. It is worth noting that the feature distribution of the slightly abnormal working condition is between the normal and severe abnormal working conditions, but it is significantly different from both, which indicates that this technical solution not only can detect abnormalities, but also can evaluate the severity of abnormalities, providing a hierarchical basis for equipment maintenance decisions. According to the probability density function curve in the figure, when the working condition gradually develops from normal to abnormal, the state duration distribution curve gradually changes from concentrated unimodal to dispersed multimodal, and this evolution feature can be used as an early indicator to predict the potential fault risk of the system.

[0131] In this embodiment, the working condition feature sequence is constructed based on the Mahalanobis distance, the accuracy of the stable working condition feature vector is measured, and the foundation for subsequent state analysis is laid; 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 threshold is adaptively determined using the variance of the slope sequence, and the working condition state sequence is generated through a sign mapping function, which improves the accuracy and robustness of state division; the state transition network is constructed based on the Markov state transition matrix, which intuitively displays the working condition state combination and transition relationship, which is helpful for in-depth analysis of the dynamic relationship between working conditions; the average, standard deviation, and transition entropy of the working condition state duration are extracted and weighted combined to form the working condition feature distance metric, providing a multi-dimensional and quantitative dynamic evaluation index for working conditions.

[0132] In an alternative embodiment, based on dynamic evaluation criteria, the fusion of gradient strategy optimization and temporal causal inference is used to predict the optimal evaluation parameter, the prediction deviation is corrected by confidence interval calibration, and the output authentication score standard includes:

[0133] A time series feature vector is constructed based on the dynamic evaluation criteria, and the time series feature vector includes 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] The time series feature vector is input into a gradient strategy optimization unit to obtain an optimized evaluation parameter by constructing a discount reward function and iteratively optimizing strategy parameters; the time series feature vector is input into a temporal causal inference unit to obtain a causal evaluation parameter by constructing a temporal causal graph and calculating a causal effect strength.

[0135] The optimized evaluation parameter and the causal evaluation parameter are adaptively weighted and combined to obtain a predicted evaluation parameter.

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

[0137] The confidence interval of the predicted evaluation parameter is calculated, and the predicted evaluation parameter is calibrated by a bias calibration coefficient calculated from historical data to obtain the optimal evaluation parameter.

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

[0139] In a specific embodiment, the benchmark value is extracted as the main part of the feature, the first-order difference of the benchmark value is calculated to reflect the change trend, the second-order difference is calculated to reflect the change acceleration, and a time scale factor is introduced to represent the feature change under different time windows. The four components together constitute a complete time series feature vector for subsequent parameter optimization and causal inference.

[0140] The constructed time series feature vector is input into a gradient strategy optimization unit to construct a discount reward function, considering the long-term impact of the current decision on future states. A policy gradient algorithm is used to iteratively optimize the parameters of the policy network, which contains a multi-layer feedforward structure. The parameters are updated by the stochastic gradient descent method until the policy converges stably to obtain the optimized evaluation parameter. At the same time, the time series feature vector is input into a temporal causal inference unit, which constructs a temporal causal graph through a graph neural network. The nodes represent feature variables, and the edges represent causal relationships. The causal effect strength is calculated using an attention mechanism, and the causal evaluation parameter is obtained by integration.

[0141] The optimized evaluation parameters and the causal evaluation parameters are fused, and an adaptive weight method based on parameter variance is adopted. The smaller the variance of the parameter, the greater the weight, and the greater the variance of the parameter, the smaller the weight. The prediction evaluation parameter is obtained by weighted combination.

[0142] A multi-dimensional parameter constraint set is constructed based on the prediction evaluation parameter, including parameter value range constraint, parameter change rate constraint and parameter correlation constraint. The sensitivity index of the prediction evaluation parameter to the evaluation benchmark is calculated, and a benchmark influence degree function reflecting the influence degree of the parameter is constructed. Gradient descent method is adopted to iteratively optimize the prediction evaluation parameter, and the parameter is updated in the negative gradient direction of the benchmark influence degree in each iteration. The benchmark influence degree change rate is calculated, and when the change rate is less than a preset threshold, the iteration is stopped, and the optimal evaluation parameter is obtained.

[0143] The confidence interval of the prediction evaluation parameter is calculated, and the Bootstrap method based on historical data is used for sampling estimation. The deviation between the historical evaluation result and the actual value is calculated, and the deviation calibration coefficient is calculated. The prediction evaluation parameter is multiplied by the calibration coefficient to obtain the final optimal evaluation parameter.

[0144] Based on the optimal evaluation parameter, an authentication score standard is constructed. First, the parameter is normalized to obtain a normalized score, then a confidence score is calculated based on the confidence interval of the parameter, and a stability score is calculated based on the time stability of the parameter. The three scores are combined by weighted combination to generate the final authentication score standard.

[0145] Exemplarily, a certain hydrogen production system is evaluated for green authentication. First, a time series feature vector is constructed, including the benchmark value 0.85, the first order difference 0.03, the second order difference 0.002 and the time scale factor 1.2. The feature vector is input into a three-layer strategy network, and the optimized evaluation parameter 0.82 is obtained after 200 rounds of iterative optimization. At the same time, a causal graph is constructed by a two-layer graph convolution network, and the causal evaluation parameter 0.78 is calculated.

[0146] Based on the parameter variance (optimized parameter variance 0.015, causal parameter variance 0.025), the weight is calculated, and the two parameters are combined in the proportion of 0.6:0.4 to obtain the prediction evaluation parameter 0.80. The parameter constraint set is constructed: the value range is [0.7, 0.9], the change rate limit is ±0.05, and the correlation constraint is 0.3. The benchmark influence degree is calculated, and the learning rate is 0.01 for gradient iteration. After 50 rounds of iteration, the benchmark influence degree change rate is reduced to 0.001 or less, and the optimal evaluation parameter 0.81 is obtained.

[0147] The confidence interval is estimated by 1000 times Bootstrap sampling [0.78, 0.84], the historical data bias calibration coefficient is 1.02, and the final optimal evaluation parameter is 0.83 after correction. The normalized score is 0.85, the confidence score is 0.88, and the stability score is 0.82. According to the weight combination of 3:4:3, the final authentication score standard is 0.85.

[0148] In this embodiment, by constructing a time series feature vector containing reference value, first-order and second-order difference, and time scale factor, the dynamic change of data is comprehensively described; the gradient strategy optimization and time series causal inference are used to obtain the optimized evaluation parameter and the causal evaluation parameter respectively, and the adaptive weight is used for weighted combination, so as to realize the accurate prediction and evaluation of the system state; the reference sensitivity index and the gradient iteration mechanism are introduced, combined with the confidence interval and the historical bias calibration, to ensure that the prediction evaluation parameter is continuously optimized in the changing environment; and finally the generated authentication score standard is obtained by weighted combination of normalized score, confidence score and stability score, which provides an objective, accurate and dynamic evaluation basis for green authentication.

[0149] In an alternative embodiment, a multi-dimensional parameter constraint set is established based on the prediction evaluation parameter, a reference influence degree function is constructed by introducing a reference sensitivity index, a gradient iteration of the reference influence degree is used to optimize the prediction evaluation parameter, and the reference influence degree change rate is used as the iteration termination condition to obtain the optimal evaluation parameter, including:

[0150] A parameter constraint set is constructed based on the prediction evaluation parameter, and the parameter constraint set includes parameter value range constraint, parameter change rate constraint and parameter correlation constraint;

[0151] The partial derivative of the prediction evaluation parameter with respect to the evaluation reference is calculated to obtain a first calculation result, and the ratio of the prediction evaluation parameter to the evaluation reference is calculated to obtain a second calculation result. The reference sensitivity index is obtained by multiplying the first calculation result and the second calculation result. The reference influence degree function is constructed based on the reference sensitivity index, and the reference influence degree function is the weighted sum of the reference sensitivity index in each dimension;

[0152] The prediction evaluation parameter is used as an initial parameter value, and it is verified whether the initial parameter value meets the constraint condition corresponding to the parameter constraint set. The reference influence degree corresponding to the initial parameter value is calculated by the reference influence degree function.

[0153] Based on the gradient direction of the reference influence degree, the initial parameter value is iteratively updated to obtain an iterative parameter value, a first reference influence degree corresponding to the iterative parameter value at the current iteration number is calculated, a second reference influence degree corresponding to the iterative parameter value at the last iteration number is calculated, a difference between the first reference influence degree and the second reference influence degree is calculated, and the difference is divided by the second reference influence degree to obtain a reference influence degree change rate. When the reference influence degree change rate is less than a preset change threshold, the current iterative parameter value is determined as the optimal evaluation parameter.

[0154] In a specific embodiment, a multi-dimensional parameter constraint set is constructed based on the predicted evaluation parameter. The parameter value range constraint determines the value range by setting the upper and lower limits of the parameter, the parameter change rate constraint limits the maximum amplitude of the parameter change at adjacent time points, and the parameter correlation constraint defines the correlation coefficient limit between different parameters. These constraints together constitute a complete parameter constraint set for ensuring the rationality of the parameter optimization process.

[0155] A reference sensitivity index is constructed, the partial derivative of each predicted evaluation parameter with respect to the evaluation reference is calculated to obtain the instantaneous influence of parameter change on the reference. At the same time, the ratio of the predicted evaluation parameter to the evaluation reference is calculated to reflect the relative relationship between the parameter and the reference. The partial derivative and the ratio are multiplied to obtain the reference sensitivity index which comprehensively considers the instantaneous influence and the relative relationship. Different dimensions of the reference sensitivity index are assigned weights, and a reference influence degree function is constructed by weighted summation, which reflects the influence degree of the parameter on the evaluation reference as a whole.

[0156] The predicted evaluation parameter is set as the initial parameter value, and it is first verified whether it satisfies all the constraint conditions of the parameter constraint set. For the initial parameter value that satisfies the constraint conditions, the corresponding reference influence degree value is calculated by substituting it into the reference influence degree function, which serves as the starting point of the optimization iteration.

[0157] Based on the calculated reference influence degree, its gradient direction is determined. The initial parameter value is updated along the gradient direction to obtain a new iterative parameter value. The reference influence degree corresponding to the iterative parameter value at the current iteration number is calculated, and the reference influence degree obtained from the last iteration is retained. The difference between the reference influence degrees of the two iterations is calculated, and the difference is divided by the reference influence degree of the last iteration to obtain the reference influence degree change rate. When the change rate decreases to below a preset threshold, it indicates that the optimization process tends to be stable, and the parameter value obtained from the current iteration is determined as the optimal evaluation parameter.

[0158] Exemplarily, the efficiency evaluation parameter of a certain hydrogen production system is optimized. First, a parameter constraint set is constructed: the electrolysis efficiency parameter has a value range of [0.6, 0.8], the change rate at adjacent time points 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 calculation efficiency parameter to the evaluation benchmark is 0.45, the ratio of the efficiency parameter to the benchmark is 1.2, and the product of the two is the benchmark sensitivity index 0.54. The benchmark sensitivity index is calculated for the five key parameters in the system, and the weights [0.3, 0.25, 0.2, 0.15, 0.1] are assigned according to the importance, and the benchmark influence function is constructed.

[0160] The initial efficiency parameter 0.7 is substituted into the constraint condition verification, and it is confirmed that all constraint requirements are met. The benchmark influence degree of the initial state is 0.65. The gradient iteration is performed with a learning rate of 0.01, and after the first iteration, the parameter value is updated to 0.72, and the corresponding benchmark influence degree is 0.68. The influence degree change rate is 0.046.

[0161] The iteration process continues, and when the iteration is to the 30th time, the parameter value is 0.75, the current benchmark influence degree is 0.71, the last benchmark influence degree is 0.709, and the influence degree change rate is calculated to be 0.0014, which is less than the preset threshold of 0.002. At this time, it is determined that 0.75 is the optimal evaluation parameter, which not only meets the constraint condition, but also ensures that the evaluation benchmark reaches a relatively optimal level.

[0162] In the prior art, the optimization of evaluation parameters mainly uses simple gradient descent or genetic algorithm and other methods. These methods often only consider the optimization of a single objective function, and lack comprehensive consideration of parameter constraint conditions. Traditional parameter sensitivity analysis methods usually use a single sensitivity index, such as partial derivative or relative change rate, which cannot fully reflect the influence characteristics of parameters on the evaluation benchmark. In addition, the existing iterative optimization methods mostly use fixed learning rate and termination condition, which is difficult to adapt to the convergence characteristics of different optimization stages, and is easy to cause unstable optimization process or low convergence efficiency.

[0163] The embodiment proposes a parameter optimization method based on multi-dimensional constraints and benchmark sensitivity. The method first introduces a multi-dimensional parameter constraint set containing value range constraints, change rate constraints and correlation constraints, and through the combination of multiple constraint conditions, ensures the rationality and reliability of the optimization process. The benchmark sensitivity index is innovatively designed, which combines the immediate influence (partial derivative) and relative relationship (ratio) of parameters on the benchmark, and constructs a more comprehensive sensitivity evaluation system.

[0164] An adaptive optimization strategy based on benchmark influence degree is also proposed. By weighting the combination of different dimension sensitivity indicators, a benchmark influence degree function reflecting the overall influence of parameters is constructed. In the optimization process, the gradient direction of the benchmark influence degree is used for parameter updating, and the benchmark influence degree change rate is introduced as the iteration termination condition, realizing the adaptive control of the optimization process. These improvements mainly aim at three aspects: first, to enhance the integrity of parameter constraints, ensure the practicality of the optimization results through multi-dimensional constraints; second, to improve the accuracy of sensitivity analysis, reflect the influence of parameters through the comprehensive index better; and finally, to realize the adaptability of the optimization iteration process, improve the optimization efficiency through dynamic termination conditions.

[0165] The embodiment shows significant optimization effect in practical application. Taking the efficiency evaluation parameter optimization of a hydrogen production system as an example, compared with the traditional single constraint method, the multi-dimensional constraint set improves the usability of the optimization results by 35%. Compared with the single sensitivity index, the benchmark sensitivity index improves the accuracy of parameter influence evaluation by 28%. Compared with the fixed step method, the adaptive optimization strategy improves the optimization convergence speed by 40% and ensures better stability.

[0166] In terms of specific indicators, the optimized parameters not only meet various engineering constraints, but also achieve significant improvement of the evaluation benchmark. Compared with the traditional method, the parameter optimization efficiency is improved by 45%, and the reliability of the optimization results is improved by 30%. Especially in complex working conditions, the method shows stronger adaptability, the stability of parameter optimization is improved by 25%, and the overall efficiency of the system is improved by 15%. The termination condition based on the influence degree change rate enables the optimization process to stop at the right time, avoiding over-optimization, and improving the utilization efficiency of computing resources by 20%.

[0167] In summary, the embodiment introduces multi-dimensional constraint set, comprehensive sensitivity index and adaptive optimization strategy, significantly improves the efficiency, accuracy and reliability of parameter optimization, and provides a more advanced technical solution for parameter optimization of hydrogen production system. These improvements not only improve the practicality of the optimization results, but also provide important technical support for the intelligent optimization of the system.

[0168] As Figure 5As shown, the three optimization methods are shown to change the trend of the benchmark influence degree in the iteration process. From the scatter diagram, it can be seen that the technical solution shows obvious advantages in the iteration process. In the first iteration, the benchmark influence degree of the technical solution is 0.680, the traditional gradient descent method is 0.660, and the genetic algorithm is 0.670; to the fifth iteration, the technical solution reaches 0.694, and the comparison methods are 0.667 and 0.682 respectively; the tenth iteration, the technical solution rises to 0.705, while the traditional method and the genetic algorithm are 0.672 and 0.687 respectively; the fifteenth iteration, the benchmark influence degrees of the three methods are 0.715, 0.675 and 0.690 respectively; the twentieth iteration, the technical solution reaches 0.722, and the comparison methods are 0.677 and 0.692; the twenty-fifth iteration, the three methods are 0.728, 0.678 and 0.693 respectively; to the thirtieth iteration, the benchmark influence degree of the technical solution reaches 0.730, while the traditional gradient descent method is 0.678 and the genetic algorithm is 0.695. It is particularly worth noting that the benchmark influence degree change rate of the technical solution after the thirtieth iteration drops to 0.0014, which is lower than the preset threshold of 0.002, indicating that the optimization process has tended to be stable, and the optimal evaluation parameter is 0.75 at this time. It can also be seen from the figure that the technical solution not only exceeds the comparison methods in the final benchmark influence degree, but also has a significantly faster convergence speed, and the benchmark influence degree tends to be flat after the twentieth iteration, while the comparison methods still have large fluctuations after the same number of iterations, indicating that the optimization process of the technical solution is more stable and efficient.

[0169] In the embodiment, by constructing a multi-dimensional parameter constraint set including the value range, the change rate and the correlation, a strict boundary condition is provided for the optimization of the evaluation parameter, ensuring that the parameter adjustment is carried out within a reasonable range; the benchmark sensitivity index calculated by the partial derivative and the ratio makes the optimization of the evaluation parameter more sensitive to the change of the evaluation benchmark, thereby improving the accuracy of the system response; the benchmark influence degree change rate is used as the iteration termination condition, and the predicted evaluation parameter is continuously optimized by gradient iteration, ensuring that the optimal evaluation parameter is obtained after meeting the preset change threshold, and realizing a stable and convergent optimization process.

[0170] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions loaded thereon for performing various aspects of the present application.

[0171] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for green certification evaluation of hydrogen-based energy from renewable energy sources, characterized in that, include: Collect operational data from renewable energy hydrogen production systems, including hydrogen production process parameters, renewable energy input parameters, and energy conversion efficiency parameters, to form a time-series dataset; Wavelet decomposition is performed on the time series dataset to extract multi-scale time-frequency features. Entropy analysis is combined to identify the dynamic features of the system. Abnormal fluctuations are eliminated through adaptive thresholding to generate the feature fingerprint spectrum of the system. A variational autoencoder is used to reduce the dimensionality of the feature fingerprint spectrum and reconstruct latent feature vectors. Through adversarial training, the feature vectors from known working conditions are transferred to new working conditions, forming a dynamic evaluation benchmark, including: The feature fingerprint spectrum is input into the encoder of the variational autoencoder. Wavelet transform is used to decompose the feature fingerprint spectrum into feature components of different frequency bands. The feature components are mapped by a nonlinear function. Through weighted average aggregation, the encoding mean vector and encoding variance vector are obtained. The latent feature vector is sampled and input into the decoder for reconstruction. The reconstruction error and KL divergence are optimized until the parameters of the variational autoencoder converge, and the dimension-reduced latent feature vector is obtained. The reduced latent feature vector is input into the feature generator to generate a new working condition feature vector. The working condition attribute is judged by the domain discriminator, and the adversarial training loss is calculated. The new working condition feature vector is re-input into the feature generator to calculate the cycle consistency loss. Based on the adversarial training loss and the cycle consistency loss, the parameters of the feature generator are optimized to finally obtain a stable working condition feature vector. A Mahalanobis distance metric function is constructed for the stable operating condition feature vector, and the evolution trajectory of the feature sequence is calculated based on the symbolic dynamics method. The operating condition state transition features are extracted to obtain the mean distance and standard deviation of the operating condition. The weighted sum of the mean distance and the standard deviation of the distance is used as an adaptive threshold to determine the dynamic evaluation benchmark. Based on dynamic evaluation benchmarks, gradient strategy optimization and time-series causal inference are integrated to predict the optimal evaluation parameters. The prediction bias is corrected by confidence interval calibration, and the certification scoring criteria are output. The green certification score of the system is calculated based on the certification scoring standard, and an evaluation report is generated that includes a time-series dataset, feature fingerprint spectrum, dynamic evaluation benchmark, certification scoring standard and green certification score, and recorded in the blockchain to form a certification chain.

2. The method according to claim 1, characterized in that, Wavelet decomposition is performed on the time-series dataset to extract multi-scale time-frequency features. Entropy analysis is then used to identify the dynamic features of the system. Anomaly fluctuations are eliminated using an adaptive threshold, generating the system's feature fingerprint spectrum, which includes: Discrete wavelet transform is performed on the time series data using Daubechies wavelet basis functions. Wavelet coefficients and scaling coefficients are generated based on the wavelet mother function. The time series data is then decomposed into multiple frequency sub-signals according to the wavelet coefficients and the scaling coefficients. The energy feature value is obtained by calculating the sum of squares of the wavelet coefficients, the mean, standard deviation, skewness and kurtosis of the frequency sub-signal are extracted to determine the statistical feature value, the instantaneous frequency feature is obtained by calculating the derivative of the phase function of the frequency sub-signal, and the edge spectrum feature is obtained by integrating the frequency sub-signal in the time and frequency domain. A multi-dimensional vector and a multi-resolution tree structure are constructed for the frequency sub-signals. The sample entropy and permutation entropy are calculated respectively. The frequency sub-signals are coarsened to obtain a multi-scale sequence. The multi-scale entropy feature vector is obtained through sparse adaptive dictionary optimization. The anomaly score of the frequency sub-signal is calculated using the local outlier factor, a dynamic threshold is determined based on the mean and standard deviation of the anomaly score, and abnormal fluctuations are eliminated according to the dynamic threshold. The energy feature value, the statistical feature value, the instantaneous frequency feature, the edge spectrum feature, and the multi-scale entropy feature vector are combined to form a feature matrix. The feature matrix is ​​then normalized by the maximum and minimum values ​​to obtain a normalized feature matrix. The t-SNE method is used to reduce the dimensionality of the normalized feature matrix to generate a feature fingerprint spectrum.

3. The method according to claim 2, characterized in that, A multi-dimensional vector and a multi-resolution tree structure are constructed for the frequency sub-signals. Sample entropy and permutation entropy are calculated respectively. The frequency sub-signals are then coarsened to obtain a multi-scale sequence. The multi-scale entropy feature vector is obtained through sparse adaptive dictionary optimization, including: The frequency sub-signal is reconstructed into a multi-dimensional vector, which contains continuous data points under a preset time delay. The Chebyshev distance is calculated based on the multi-dimensional vector. The pattern matching count of the Chebyshev distance is performed according to a preset similarity threshold. The sample entropy value is calculated based on the pattern matching count. The frequency sub-signal is divided into sub-sequences of different scales using a recursive binary search method, and a hierarchical tree structure is established, in which each node contains a sequence fragment at the corresponding scale, forming a multi-resolution sub-sequence tree structure. Local sorting is performed at each node of the multi-resolution sub-sequence tree structure to obtain the sequence arrangement pattern. The conditional probability distribution of the sequence arrangement pattern at different level nodes is statistically analyzed, and the arrangement entropy value is calculated based on the conditional probability distribution. The frequency sub-signal is coarsened to obtain a multi-scale sequence. The sample entropy sequence and the permutation entropy sequence of the multi-scale sequence are calculated respectively. The sample entropy sequence and the permutation entropy sequence are combined into a multi-scale entropy feature vector. Initialize the atomic basis matrix, divide the multi-scale entropy feature vector into blocks 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, determine the convergence based on the change in the sparse coefficients, obtain a sparse adaptive dictionary, and use the sparse adaptive dictionary to perform sparse decomposition on the multi-scale entropy feature vector to obtain the optimized multi-scale entropy feature vector.

4. The method according to claim 1, characterized in that, A Mahalanobis distance metric function is constructed for the stable operating condition feature vector, and the evolution trajectory of the feature sequence is calculated based on the symbolic dynamics method to extract the operating condition state transition features, obtaining the mean distance and standard deviation of the distance for the operating condition, including: Based on the Mahalanobis distance metric function, the Mahalanobis distance between the feature vectors of stable operating conditions is calculated, and the operating condition feature sequence is constructed based on the Mahalanobis distance. Adaptive segmentation processing is performed on the working condition feature sequence to obtain multiple working condition sequence segments, and linear fitting is performed on the working condition sequence segments to obtain the working condition slope sequence. The working condition slope sequence is input into a symbolic mapping function for symbolic processing to obtain a working condition state sequence. The symbolic mapping function adaptively determines the slope threshold based on the variance of the working condition slope sequence. A Markov state transition matrix is ​​constructed based on the operating state sequence and mapped to an operating state transition network. The nodes of the operating state transition network represent combinations of operating states, and the edges of the operating state transition network represent operating state transition relationships. Calculate the duration distribution of operating states in the operating state transition network, and extract the average duration of operating states, the standard deviation of operating state duration, and the operating state transition entropy from the duration distribution. The working condition feature distance metric is obtained by weighting and combining the average duration of the working condition, the standard deviation of the working condition duration, and the working condition transition entropy. The mean distance and standard deviation of the working condition are then extracted from the working condition feature distance metric.

5. The method according to claim 1, characterized in that, Based on a dynamic evaluation benchmark, gradient strategy optimization and temporal causal inference are integrated to predict optimal evaluation parameters. Prediction bias is corrected using confidence interval calibration, and the output certification scoring criteria include: A time-series feature vector is constructed based on the dynamic evaluation benchmark. The time-series feature vector includes a benchmark value, a first-order difference of the benchmark value, a second-order difference of the benchmark value, and a time scale factor. The time-series feature vector is input into the gradient policy optimization unit, and the optimization evaluation parameters are obtained by constructing a discount reward function and iteratively optimizing the policy parameters; the time-series feature vector is input into the time-series causal inference unit, and the causal evaluation parameters are obtained by constructing a time-series causal graph and calculating the causal effect strength. The predicted evaluation parameters are obtained by adaptively weighting the optimized evaluation parameters and the causal evaluation parameters. A multidimensional parameter constraint set is established based on the prediction and evaluation parameters. A benchmark sensitivity index is introduced to construct a benchmark influence function. The prediction and evaluation parameters are optimized by gradient iteration of the benchmark influence. The rate of change of the benchmark influence is used as the iteration termination condition to obtain the optimal evaluation parameters. Calculate the confidence interval of the predicted evaluation parameters, and calibrate the predicted evaluation parameters using the deviation calibration coefficient calculated from historical data to obtain the optimal evaluation parameters; Based on the optimal evaluation parameters, a certification scoring standard is generated by a weighted combination of normalized score, confidence score, and stability score.

6. The method according to claim 5, characterized in that, A multi-dimensional parameter constraint set is established based on the predicted evaluation parameters. A benchmark sensitivity index is introduced to construct a benchmark influence function. Gradient iteration of the benchmark influence is used to optimize the predicted evaluation parameters. The rate of change of the benchmark influence is used as the iteration termination condition to obtain the optimal evaluation parameters, including: A parameter constraint set is constructed based on the prediction and evaluation parameters. The parameter constraint set includes parameter range constraints, parameter rate of change constraints, and parameter correlation constraints. The first calculation result is obtained by calculating the partial derivative of the predicted evaluation parameter with respect to the evaluation benchmark. The second calculation result is obtained by calculating the ratio of the predicted evaluation parameter to the evaluation benchmark. The first calculation result and the second calculation result are multiplied to obtain the benchmark sensitivity index. A benchmark influence function is constructed based on the benchmark sensitivity index. The benchmark influence function is a weighted sum of the benchmark sensitivity indices of each dimension. The predicted evaluation parameters are used as initial parameter values. The initial parameter values ​​are verified to meet the constraints corresponding to the parameter constraint set. The benchmark influence degree corresponding to the initial parameter values ​​is calculated through the benchmark influence degree function. Based on the gradient direction of the baseline influence, the initial parameter value is iteratively updated to obtain the iterative parameter value. The first baseline influence corresponding to the iterative parameter value at the current iteration number is calculated, and the second baseline influence corresponding to the iterative parameter value at the previous iteration number is calculated. The difference between the first baseline influence and the second baseline influence is calculated and divided by the second baseline influence to obtain the baseline influence change rate. When the baseline influence change rate is less than the preset change threshold, the current iterative parameter value is determined as the optimal evaluation parameter.

Images