Block chain-based power equipment supply chain data sharing method

Abstract

The invention relates to the technical field of power equipment state monitoring and data security, and discloses a power equipment supply chain data sharing method based on a block chain, and the method comprises the following steps: S1, data collection and standardization: collecting sensor data of power equipment in real time, and carrying out the standardization processing; s2, dynamic entropy threshold calculation: dynamically calculating a data entropy threshold based on an information geometry theory; s3, verification mode decision making, wherein quantum chaos random walk data sampling or full-amount zero-knowledge proof verification is triggered according to entropy threshold decision making; and S4, anti-tamper certificate construction: constructing an anti-quantum hash chain for the sampled data and carrying out distributed anti-tamper coding. According to the method, through dynamic entropy threshold adaptation and quantum chaos sampling, accurate improvement of anomaly detection sensitivity is realized, based on Reed-Solomon fragmentation and algebraic verification, it is ensured that data can still be completely recovered when being partially tampered, and security and throughput efficiency are balanced by using block chain evidence storage and improvement of PBFT consensus.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment status monitoring and data security, and specifically to a method for sharing power equipment supply chain data based on blockchain. Background Art

[0002] In recent years, power equipment status monitoring technology has developed rapidly, mainly relying on solutions such as fixed threshold anomaly detection, software random sampling verification, and full-volume data centralized storage. Existing technologies trigger alarms by setting parameter thresholds such as vibration and temperature, verify data integrity using structures such as Merkle trees, and rely on regular manual parameter adjustment to maintain system performance. Such methods have certain effects in the initial operation stage of the equipment, but as the equipment ages, the environment changes, and the attack means are upgraded, they gradually expose the problem of insufficient adaptability.

[0003] Specifically, fixed threshold detection cannot capture the gradual change of equipment status, with a high false alarm rate for new equipment and a high risk of missed alarms for aging equipment; the software random sampling path is predictable, and attackers can target and tamper with the undetected data; full-volume verification has a huge computational overhead and is difficult to meet the real-time requirement; the centralized storage architecture has a single point of failure risk, and the traditional erasure code has limited fault tolerance; static parameter configuration leads to continuous degradation of system performance over the operation time. These defects form a chain reaction. For example, threshold deviation triggers ineffective full-volume verification, further exacerbating resource consumption and security risks. Summary of the Invention

[0004] In view of the deficiencies of the prior art, the present invention provides a method for sharing power equipment supply chain data based on blockchain, which solves the problems of high false alarm and missed alarm rates, weak anti-attack ability, and long-term operation performance decline in the existing equipment monitoring system due to static thresholds, predictable sampling, and centralized storage.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for sharing power equipment supply chain data based on blockchain, including the following steps:

[0006] S1. Data collection and standardization: Real-time collect sensor data of power equipment and perform standardization processing;

[0007] S2. Dynamic entropy threshold calculation: Dynamically calculate the data entropy value threshold based on information geometry theory;

[0008] S3. Verification mode decision: Decide to trigger quantum chaotic random walk data sampling or full-volume zero-knowledge proof verification according to the entropy value threshold;

[0009] S4. Anti-tampering voucher construction: Construct an anti-quantum hash chain for the sampled data and perform distributed anti-tampering encoding;

[0010] S5, Blockchain Evidence Storage: Upload the hash value and verification certificate to the blockchain network;

[0011] S6, Distributed Verification: Execute distributed fault-tolerant verification through a smart contract and output the result;

[0012] S7, Dynamic Parameter Optimization: Monitor the false negative rate and false positive rate of historical verification results; When the false negative rate exceeds the threshold, retrain the statistical manifold parameters; Adjust the chaotic evolution intensity according to the cumulative running time of the device.

[0013] Preferably, the S2 dynamic entropy threshold calculation includes:

[0014] Calculate the local Shannon entropy using a sliding window;

[0015] Establish a statistical manifold model of the sensor data distribution;

[0016] Dynamically adjust the mean and standard deviation weight parameters based on the Riemannian gradient flow equation of the statistical manifold to generate an adaptive threshold.

[0017] Preferably, the S3 quantum chaotic random walk data sampling includes:

[0018] Construct a chaotic Hamiltonian according to the physical characteristics of the device;

[0019] Embed a chaotic coin operator during the quantum walk evolution;

[0020] Generate an unpredictable sampling path based on the quantum state probability distribution.

[0021] Preferably, the construction of the chaotic Hamiltonian includes:

[0022] Adopt a harmonic oscillator model as the integrable ground state Hamiltonian;

[0023] Introduce a Gaussian random matrix to simulate the device's nonlinear perturbation;

[0024] Dynamically adjust the chaotic perturbation intensity according to the device aging degree.

[0025] Preferably, the construction of the S4 anti-quantum hash chain includes:

[0026] Collect the device physical noise and generate an initial entropy source through analog-to-digital conversion;

[0027] Adopt an anti-quantum hash function to iteratively generate a chain certificate;

[0028] During verification, only need to compare the consistency of the hash values at both ends of the chain.

[0029] Preferably, the S4 distributed anti-tampering coding includes:

[0030] Use Reed-Solomon codes to perform redundant coding on the sampled data;

[0031] Store the encoded shards at the edge nodes of the supply chain associated enterprises;

[0032] During verification, the original data is restored by decoding the polynomial equation, and it tolerates tampered shards not exceeding half of the total number of shards.

[0033] Preferably, the S5 blockchain network is of a consortium chain architecture:

[0034] The core nodes adopt a practical Byzantine fault tolerance consensus algorithm introducing pipeline batch processing;

[0035] The edge nodes participate in local consensus according to business shards;

[0036] The regulatory agency accesses the network as an observer node for compliance auditing.

[0037] Preferably, the S6 distributed fault tolerance verification includes:

[0038] Randomly select multiple edge nodes to obtain encoded shards;

[0039] Tolerate tampered shards not exceeding half of the total number of shards through algebraic decoding;

[0040] Determine that the data is valid when the decoding result matches the on-chain hash.

[0041] Preferably, the hardware implementation of the method includes:

[0042] Deploy edge computing nodes supporting GPU acceleration on the power equipment side;

[0043] Configure a hardware security module at the blockchain node to perform quantum-resistant signatures;

[0044] Integrate a streaming data processing engine at the application layer server to clean the data uploaded to the chain in real time.

[0045] The present invention provides a method for sharing power equipment supply chain data based on blockchain. It has the following beneficial effects:

[0046] 1. The present invention adopts a dynamic entropy threshold calculation technology optimized by statistical manifolds and gradient flows, which can adapt to the changes in the device operation state in real time. Compared with the existing fixed threshold scheme that is prone to misjudgment due to device aging, this scheme can automatically capture the data distribution shift through information geometric mapping, significantly improving the sensitivity and accuracy of anomaly detection.

[0047] 2. The present invention generates random paths through the evolution of the quantum chaotic Hamiltonian, and combines physical noise to enhance unpredictability. Traditional methods rely on software random numbers that are easily reverse-cracked. This scheme integrates the resonance characteristics of the device into the chaotic model to achieve hardware-level random source-driven sampling, fundamentally blocking the possibility for attackers to predict the path.

[0048] 3. Through the piecewise storage and distributed verification mechanism of algebraic coding, the present invention allows the recovery of complete information even when part of the data is tampered with or lost. The prior art has strict requirements for data integrity, and the lack of a piece causes verification failure. However, the present invention ensures the continuous reliability of the system under extreme interference through coding redundancy and dynamic decoding strategies.

[0049] 4. The present invention continuously iterates the core parameters through a closed-loop parameter optimization system based on historical data and verification results. The traditional scheme relies on manual experience adjustment, resulting in a response lag. The adaptive mechanism of the present invention incorporates multiple factors such as device status and network load into the optimization target to achieve stable performance throughout the life cycle of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flowchart of the method steps of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0052] Please refer to the attached Figure 1 , the embodiment of the present invention provides a method for sharing power equipment supply chain data based on blockchain, including the following steps:

[0053] S1. Data acquisition and standardization: Real-time acquisition of sensor data of power equipment and standardization processing;

[0054] In this embodiment, an integrated triaxial MEMS acceleration sensor (such as AD IADXL1002) is used, with a range of ±50g and a bandwidth range of 0.5Hz - 12kHz, to adapt to the broadband vibration characteristics of power equipment. The sensor is installed at the middle axial position of the transformer oil tank wall, bonded with epoxy resin glue, and the surface roughness Ra of the contact surface ≤ 3.2μm to avoid mechanical impedance mismatch. For example, the main vibration frequency of a 500kV transformer oil tank surface measurement point is about 80Hz, and the sensor bandwidth needs to cover more than three times this frequency band to ensure harmonic acquisition.

[0055] As an option, for the partial discharge monitoring scenario, a high-frequency current transformer (such as a Rogowski coil, bandwidth 1MHz) can be synchronously deployed. After its output signal is converted into a voltage signal through an integrating circuit, it is synchronously acquired with the vibration signal.

[0056] Signal conditioning and analog-to-digital conversion

[0057] Specifically, the output signal of the sensor is amplified by an instrumentation amplifier (such as INA826) and then input into a 24-bit Δ-Σ ADC (such as ADS1278). The conversion formula is as follows:

[0058]

[0059] where V ref = 4.096V is the reference voltage, Code is the 24-bit two's complement value output by the ADC (range -8,388,608 to 8,388,607), and G amp = 800 is the amplification gain. For example, when the sensor output is 1.2V, the corresponding digital quantity is calculated as follows:

[0060]

[0061] Baseline Calibration and Noise Suppression Implementation

[0062] In a possible implementation, after the device is powered on for the first time, baseline calibration is performed: the sensor output in the no-load state is collected continuously for 600 seconds, and the noise power of each frequency band is calculated. Taking the vibration signal of a certain circuit breaker operating mechanism as an example, there is periodic interference (peak -40dB / Hz) at 2kHz in its background noise. Based on this, an IIR notch filter is designed:

[0063]

[0064] where f c = 2000Hz is the center frequency, T s = 1μs is the sampling period, and r = 0.95 controls the stopband width. This filter can attenuate the interference amplitude to below -80dB.

[0065] Details of Dynamic Normalization Processing

[0066] In some embodiments, overlapping sliding windows are used for dynamic normalization. The window length T = 10ms and the overlap rate is 50%. The signal normalization formula within the window is as follows:

[0067]

[0068] where μ x is the mean value of the signal within the window, and σ x is the standard deviation. For example, at a certain moment, the mean value of the window data is 0.2V and the standard deviation is 0.15V. Then, the input signal of 0.35V corresponds to the normalized value (0.35 - 0.2) / 0.15 = 1.0. This processing makes the data distribution have zero mean and unit variance, improving the stability of subsequent entropy value calculation.

[0069] Wavelet Denoising Parameter Settings

[0070] Specifically, the db6 wavelet basis is selected for six-layer decomposition. The thresholds for each layer are calculated according to the following formula:

[0071]

[0072] where d j is the high-frequency coefficient of the j-th layer, and N j is the number of coefficients. After hard threshold processing, the signal is reconstructed through inverse transformation. Taking the data of a certain transformer core looseness fault as an example, the peak signal-to-noise ratio (PSNR) of the denoised signal is increased from 28.7 dB to 45.3 dB.

[0073] After being processed by step S1, the dynamic normalization window length T = 10 ms of the signal needs to be strictly matched with the entropy calculation window of subsequent S2. For example, if S2 uses 100 entropy updates per second, the sliding step of the normalization window in S1 is set to 10 ms. In addition, the noise power spectrum extracted by baseline calibration will be used as the initial covariance parameter of the statistical manifold model (Fisher information matrix) in S2 to ensure the rapid convergence of the model. The group delay of the filter needs to be compensated to avoid time sequence misalignment between S1 and S2.

[0074] In one embodiment, this step is applied to the partial discharge signal of a certain GIS device. After the original signal is conditioned, denoised, and normalized, the measurement error of the rising edge time of its pulse waveform is reduced from ±15 ns to ±3 ns, and the spectral leakage rate is reduced by 72%, meeting the accuracy requirements of S2 for the input signal.

[0075] S2. Dynamic entropy threshold calculation: Dynamically calculate the entropy value threshold of the data based on information geometry theory;

[0076] In the embodiment, the entropy calculation window size is determined according to the normalization window length T = 20 ms in step S1. For example, when the sampling rate f s = 1 kHz, the window contains w = 20 ms × 1000 Hz = 20 sampling points. The overlapping rate of adjacent windows is set to 50%, that is, 100 entropy value sequences are generated per second.

[0077] As an option, for the partial discharge monitoring scenario, when the main frequency f pulse of the pulse signal is 1 MHz, the window length is adjusted to w = 10 sampling points (10 μs) to capture transient characteristics.

[0078] Local Shannon entropy calculation

[0079] Specifically, the probability distribution is calculated for the data in each window. Let the window data set be X w = {x1, x2,..., x 20}, which is divided into N b = 10 equal-width intervals, and the frequency n i, the probability estimate is:

[0080]

[0081] Among them, α = 0.01 is the Laplace smoothing factor to prevent the zero - probability problem. The calculation formula of Shannon entropy is:

[0082]

[0083] For example, if the data distribution in a certain window is uniform (p(x i ) = 0.1), then H t = 3.32 bits, while the entropy value can be as low as 0.5 bits when the distribution is concentrated.

[0084] Statistical Manifold Modeling and Parameter Optimization

[0085] In a possible implementation, assume that the window data follows a Gaussian distribution, and construct a statistical manifold with coordinate parameters θ = (μ, σ 2 ). The Fisher information matrix on the manifold is defined as:

[0086]

[0087] Update the parameters based on the Riemannian gradient flow equation:

[0088]

[0089] Among them, J(θ) is the objective function, reflecting the trade - off relationship between the mean μ H of the historical entropy value and the standard deviation σ H . For example, set the iteration step size η = 0.01, and update the parameters every 100 windows.

[0090] Adaptive Threshold Generation Rule

[0091] In some embodiments, the dynamic threshold expression is:

[0092] τ t = β(t)·μ H + γ(t)·σ H ;

[0093] Among them, μ H is the exponentially weighted average of the entropy values in the past 300 seconds, and the decay factor λ = 0.995; σ H is the standard deviation in the same period. The parameters β(t) and γ(t) are optimized through the gradient flow, and the initial values are set to β = 1.2 and γ = 0.8. For example, at the initial stage of a certain transformer operation, τ t = 2.1 bits, and it is adjusted to 2.8 bits after three years of aging.

[0094] The dynamic threshold τ output by step S2 t directly controls the verification mode of step S3: when H t > τ t it triggers a full - scale zero - knowledge proof, otherwise it executes quantum chaotic sampling (S3). For example, at the moment when a certain circuit breaker trips, H t suddenly increases to 3.5 bits (exceeding the threshold of 2.6 bits), then a full - scale verification is started to ensure data integrity.

[0095] The update periods of parameters β(t) and γ(t) need to be synchronized with the dynamic optimization in step S7 (retraining monthly). The historical entropy value database is stored in the edge - node SSD and interacts with the blockchain evidence - storing (S5) module through shared memory, with a latency ≤ 1 ms.

[0096] In one embodiment, this step is implemented for the oil chromatogram data of a certain 500 kV transformer. After dynamic threshold adjustment, the verification trigger frequency under normal conditions drops from 12.3 times per hour to 1.8 times, and the computing resource consumption is reduced by 85%. When a latent fault occurs, the threshold τ t rises from 2.4 bits to 3.1 bits within 10 seconds, and the false - negative rate drops from 9.7% to 0.3%.

[0097] S3. Verification mode decision: Decide whether to trigger quantum chaotic random - walk data sampling or full - scale zero - knowledge proof verification according to the entropy value threshold;

[0098] In the embodiment, when the real - time entropy value H t ≤ τ t it executes quantum chaotic random - walk data sampling; when H t > τ t it switches to full - scale zero - knowledge proof verification. For example, at a certain moment, H t = 2.5 bits, r t = 3.0 bits, then the quantum sampling mode is triggered, and only about 0.01% of the data volume needs to be verified.

[0099] As an option, for high - value equipment (such as UHV converter transformers), a forced full - scale verification period (such as once a day) is set, which is not restricted by τ t

[0100] Method for constructing the chaotic Hamiltonian

[0101] Specifically, the ground - state Hamiltonian H0 is determined by the physical parameters of the device:

[0102]

[0103] ​Where: m is the equivalent mass (such as the mass of the transformer core, unit: kg). For example, for a certain 500 kV transformer, m = 1500 kg; ω is the angular frequency (unit: rad / s), which is converted from the rated vibration frequency f of the device (ω = 2πf). For example, when f = 60 Hz, ω = 377 rad / s; x is the displacement coordinate (unit: m), corresponding to the sensor measurement value; p is the momentum operator (unit: kg·m / s).

[0104] The chaotic perturbation term V is generated by a Gaussian random matrix:

[0105]

[0106] Where: ∈ is the perturbation intensity coefficient, and the calculation formula is For example, when the device has been running for 3 years (26280 hours), ∈ = 2.628, is a Gaussian distribution with a mean of 0 and a variance of ∈ 2 .

[0107] The total Hamiltonian is H = H0 + V.

[0108] Quantum chaotic walk evolution process

[0109] In a possible implementation, the quantum state evolution is achieved through a time-stepping operator:

[0110]

[0111] Where: Δt is the evolution time step (unit: s), and Δt = 1 ms is taken; is the reduced Planck constant; |ψ(t)> is the quantum state vector at time t, and the initial state is a uniform superposition state Where d is the dimension of the Hilbert space, which is determined by the total amount of data L

[0112] The number of evolution steps N is set according to the data scale:

[0113]

[0114] For example, when the total amount of data L = 10 6 , d = 20, N = 20 steps.

[0115] Chaotic coin operator design

[0116] In some embodiments, the coin operator C adopts a parameterized unitary matrix:

[0117]

[0118] where: θ is the rotation angle, generated by a chaotic map (such as the Logistic map), and θ n+1 = 4θ n (1 - θ n ), with the initial value θ0 = 0.3 as the phase angle, taken from the LSB (least significant bit) of the sensor thermal noise. For example, when the last bit of the ADC sampling value at a certain moment is 1, φ = 0.7π.

[0119] After each evolution, θ and φ are updated to ensure that the path is unpredictable.

[0120] Sampling path generation rule

[0121] Specifically, after the quantum state collapses, a sampling index sequence is generated. The probability distribution is:

[0122] P(i) = |<i|ψ N >| 2 ;

[0123] where: |ψ N > is the quantum state after N-step evolution; |i> is the position basis state, corresponding to the data storage index.

[0124] The number of samples k is determined by the following formula:

[0125]

[0126] For example, when L = 10 6 and N = 20, that is, 1 data point is sampled per batch.

[0127] The output sampling index of step S3 is directly passed to the anti-tampering credential construction module in step S4. For example, the data points corresponding to the sampling index [153, 8942, 45107] will be extracted and used to generate a hash chain and RS-encoded shards. The update period of the chaotic parameter ∈ (increasing by 0.05 per quarter) is synchronized with the dynamic optimization in step S7.

[0128] Quantum evolution operator The matrix calculation of is executed in parallel on the edge node GPU (such as NVIDIA A100), and the single-step calculation time ≤ 0.1 ms to ensure real-time performance.

[0129] In a vibration monitoring scenario of a converter transformer, the Hamming distance of the quantum chaotic sampling path is 98.7% (i.e., the difference degree between adjacent sampling positions), which is significantly higher than 45.2% generated by traditional pseudo-random numbers. The probability that an attacker successfully predicts the next sampling position < 0.03%, which is two orders of magnitude lower than the traditional scheme.

[0130] S4. Anti-tampering credential construction: Construct an anti-quantum hash chain for the sampled data and perform distributed anti-tampering encoding;

[0131] In this embodiment, the fault feature recognition steps based on the hybrid intelligent model aim to perform intelligent analysis on the signal data obtained by quantum chaotic random walk sampling, so as to extract key fault features and improve the accuracy and real-time performance of equipment anomaly detection. Generally, traditional feature extraction and classification methods are prone to noise interference and difficult to adapt to various working conditions when dealing with complex non-linear vibration signals. Therefore, this embodiment introduces a hybrid intelligent model, combines deep learning with statistical analysis methods, and constructs a high-precision fault recognition system based on the sampling data obtained in step S3.

[0132] As an option, this embodiment combines a deep neural network (DNN) with a support vector machine (SVM) to make full use of the feature extraction ability of deep learning and introduce the generalization ability of the SVM to enhance the model. Specifically, in a possible implementation, first, a convolutional neural network (CNN) is used to extract the time-frequency features of the signal, and an LSTM network is used to model the time series features, and then the SVM is used to classify the extracted features to improve the discriminant ability of fault types. The following details the specific implementation of this step.

[0133] Feature Extraction and Data Preprocessing

[0134] In this embodiment, to ensure the quality of the input data, the signal data sampled by quantum chaotic random walk is first preprocessed, including denoising, normalization, and data augmentation. Generally, due to the strong environmental noise in the vibration signals of power equipment, this embodiment uses the wavelet transform denoising method to smooth the original signal.

[0135] As an option, assuming the original signal is x[n], multi-scale decomposition is performed using the wavelet transform W(x) to extract the low-frequency component x low [n] and suppress the high-frequency noise component x lnigh [n]:

[0136] x demoised [n] = W -1 (x low [n]);

[0137] where W -1 (·) is the inverse wavelet transform to ensure the fidelity of the signal after denoising.

[0138] In some embodiments, to enhance the generalization ability of the model, this embodiment adopts data augmentation techniques, including methods such as random time shift and signal mirror flipping, to generate multiple different versions of training samples, so as to improve the adaptability of the model under different working conditions. For example, in the vibration analysis of high-voltage circuit breakers, random time offset is applied to the sampled signal to simulate the signal changes under different contact wear states, and the robustness of the model to the actual working conditions is improved.

[0139] Deep learning model construction

[0140] After the feature extraction is completed, this embodiment uses a hybrid intelligent model for fault feature recognition, specifically including the combination of CNN, LSTM, and SVM. Generally, CNN is suitable for extracting time-frequency domain features, while LSTM can model temporal relationships. Therefore, this embodiment adopts the following structure:

[0141] F CNN-1STM (x) = g LSTM (g CNN (x));

[0142] Among them, g CNN (·) represents the CNN feature extraction module, and g LSTM (·) represents the LSTM temporal modeling module.

[0143] Specifically, the CNN part uses a one-dimensional convolutional layer Conv1D for temporal feature extraction:

[0144]

[0145] Among them, w j is the convolutional kernel weight, K is the convolutional kernel size, and x[i] is the input signal.

[0146] In some embodiments, to improve the adaptability of the model to time-varying signals, this embodiment adopts a bidirectional LSTM (Bi-LSTM) network to enhance the ability to capture historical information. The state update formula of Bi-LSTM is as follows:

[0147]

[0148] Among them, and respectively represent the hidden states of the forward and backward LSTM units, W f , W b , U f , U b are network parameters, and σ(·) is the activation function.

[0149] In some embodiments, to enhance the generalization ability of the model, in this embodiment, an SVM classifier is connected after the output layer of the LSTM to optimize the decision boundary of fault classification. The classification decision function of the SVM is as follows:

[0150]

[0151] where α i is the weight of the support vector, K(x i , x) is the kernel function, and b is the bias term.

[0152] Fault Classification and Model Optimization

[0153] After the model training is completed, in this embodiment, the fault signals are classified, and the model parameters are optimized based on the cross-entropy loss function. Generally, the cross-entropy loss function is used to measure the matching degree between the predicted probability distribution and the true distribution, and is defined as follows:

[0154]

[0155] where C is the number of fault categories, y c is the true category label, is the model prediction probability.

[0156] In some embodiments, to further improve the classification accuracy, this embodiment adopts a transfer learning strategy and uses a pre-trained deep learning model to fine-tune the data. For example, in the analysis of the breaker vibration signal of a 500 kV GIS device, in this embodiment, the CNN-LSTM model is first trained using the historical data of the high-voltage substation, and then fine-tuned on the data of the new device to improve the applicability of the model in the new scenario.

[0157] The fault feature recognition in this embodiment depends on the quantum chaotic random walk sampling in step S3. If the path adjustment mechanism in S3 is inaccurate, it may lead to insufficient sampling density of key fault signals, thereby affecting the classification accuracy of this step. Therefore, the sampling strategy of S3 needs to match the feature extraction requirements of this step to ensure the quality of the data input.

[0158] In some embodiments, to improve the data consistency, the CNN-LSTM model in this step needs to be optimized according to the signal features sampled in S3. For example, if S3 adopts a high-density sampling strategy, the size of the CNN filter in this step should be appropriately reduced to enhance the local feature extraction ability; otherwise, the receptive field of the filter can be increased to capture long-time sequence dependencies.

[0159] In addition, the fault classification result of this step can be further used in step S5 (intelligent fault diagnosis and health assessment) to guide the operation and maintenance strategy of the equipment. For example, in the vibration analysis of the circuit breaker of a 500 kV GIS device, when this step detects an abnormal pattern, the health assessment mechanism in S5 can be triggered to perform trend prediction in combination with the equipment operation history to assist in operation and maintenance decision-making.

[0160] S5. Blockchain evidence storage: Upload the hash value and verification certificate to the blockchain network;

[0161] The intelligent fault diagnosis and health assessment step in this embodiment aims to construct an intelligent analysis model based on the fault characteristics obtained in step S4 to achieve comprehensive assessment of the equipment operation status and fault trend prediction. Generally, traditional health assessment methods rely on expert experience or rule-based analysis methods and are difficult to adapt to multi-source data fusion and long-term trend prediction under complex working conditions. Therefore, this embodiment combines deep learning and probability statistics methods to propose an intelligent diagnosis method based on time series prediction and health factor calculation to ensure diagnosis accuracy and assessment reliability.

[0162] As an option, this embodiment introduces a health factor (HealthIndex, HI) model to fuse multi-source information such as the vibration characteristics, temperature data, and current waveform of the equipment and construct a data-driven health status assessment system. Specifically, in a possible implementation, a long short-term memory network (LSTM) is used to predict the future evolution trend of the health factor, and the Bayesian inference method is combined to calculate the probability of fault occurrence to improve the credibility of equipment status assessment. The specific implementation method of this step is described in detail below.

[0163] Health factor calculation and fusion analysis

[0164] In this embodiment, in order to quantify the equipment health status, the health factor HI(t) is first defined as an index of the equipment operation status, and HI(t) is calculated by weighted fusion of multiple key parameters. Generally, the health factor is comprehensively calculated from multiple sub-features. Let x i (t) represent the i-th key health feature, then the calculation formula of the health factor is as follows:

[0165]

[0166] where w i is the weight of each feature, and generally the analytic hierarchy process (AHP) or principal component analysis (PCA) is used for optimization to ensure reasonable contribution degrees of each parameter.

[0167] As an option, the health factor in this embodiment may include the following key indicators:

[0168] Vibration energy feature E v : Calculate the energy distribution of the signal based on the vibration features extracted in step S4.

[0169] Temperature drift feature T d : Model the internal temperature change of the device to reflect the heat loss during long-term operation.

[0170] Current fluctuation feature I f : Evaluate the possibility of internal electrical faults of the device by analyzing the harmonic components of the current signal.

[0171] Specifically, in the health assessment of high-voltage GIS circuit breakers, the energy change of the vibration signal can be used to detect mechanical wear, while the temperature drift feature reflects the aging of the contact points. Through the weighted fusion of health factors, the current operating state of the device can be comprehensively evaluated.

[0172] LSTM prediction and fault trend analysis

[0173] After calculating the health factors, this embodiment uses a long short-term memory (LSTM) network to predict the future health state. Generally, the LSTM network can effectively capture long-term time dependencies, so it is suitable for device health trend modeling. Let HI(t) be the historical health factor data sequence, then the state update equation of the LSTM is as follows:

[0174] f t = σ(W f x t + U f h t-1 + b f );

[0175] i t = σ(W t x t + U t u t-1 + b t );

[0176] o t = σ(W s x t + U s h t-1 + b s );

[0177] c t = f t · c t-1 + i t · tanh(W s x t + U s h t-1 + bs );

[0178] h t = a t ·tanh(a t );

[0179] Among them, f t , i t , a t are the forget gate, input gate, and output gate respectively, W f , W t , W o , W c are network parameters, and c t is the cell state.

[0180] In some embodiments, in order to improve the prediction accuracy, this embodiment uses a double-layer LSTM network and combines it with an attention mechanism (Attention Mechanism) to enhance the attention to key health factors. Specifically, the attention weight is calculated as follows:

[0181] e t = v T tanh(W a b t + b a );

[0182] Among them, α t is the attention weight, and W a and v are attention layer parameters.

[0183] In the health assessment of 500 kV GIS equipment, the change trend of health factors in the next 6 months can be predicted through LSTM, and equipment anomalies can be detected in advance, and preventive maintenance measures can be taken to reduce the suddenness of equipment failures.

[0184] Bayesian Inference and Fault Probability Calculation

[0185] This embodiment further combines the Bayesian inference method to calculate the equipment health probability distribution under different fault modes. Generally, Bayesian inference can use historical data to update the equipment health state probability, thereby improving the credibility of the diagnosis results.

[0186] Let θ be the health state parameter and D be the observed data. Then the posterior probability is calculated as follows:

[0187]

[0188] Among them, P(D|θ) is the likelihood function, P(θ) is the prior distribution, and P(D) is the evidence.

[0189] In some embodiments, to improve the computational efficiency, the Markov Chain Monte Carlo (MCMC) method is adopted in this embodiment for Bayesian inference to generate the probability distribution of the health state. Specifically, in the analysis of high-voltage GIS circuit breakers, the probabilities of different fault modes such as insulation faults, mechanical wear, and electrical aging can be calculated through the MCMC method, and based on the long-term change trend of the health factors, the occurrence time of equipment faults can be predicted.

[0190] The intelligent fault diagnosis and health assessment of this embodiment rely on the fault feature recognition results of step S4. If the classification accuracy in S4 is low, it may lead to large feature deviations in the calculation of health factors, thus affecting the accuracy of subsequent fault prediction. Therefore, the classification model in S4 needs to match the health assessment algorithm in this step to ensure the reliability of the input data.

[0191] In some embodiments, to improve the continuity of the assessment, the calculation period of the health factors in this step should be consistent with the feature extraction period in S4. For example, in the analysis of high-voltage GIS equipment, if S4 uses a 10-ms time window for feature extraction, the calculation of the health factors in this step can adopt a 1-s sliding window to balance the calculation cost and assessment accuracy.

[0192] In addition, the assessment results of this step can be further used in step S6 (intelligent operation and maintenance decision-making) to optimize the operation and maintenance plan by combining the equipment historical data and expert knowledge. For example, in the health assessment of a 500-kV GIS circuit breaker, when the health factor drops to the critical threshold, the intelligent decision-making system in S6 can be triggered to recommend preventive maintenance or replacement of key components to avoid sudden failures.

[0193] S6. Distributed verification: Execute distributed fault-tolerant verification through smart contracts and output the results;

[0194] The intelligent operation and maintenance decision-making and optimization steps in this embodiment aim to construct an adaptive operation and maintenance strategy based on the health assessment results obtained in step S5 to optimize the equipment maintenance plan and improve the operation safety. Generally, traditional operation and maintenance decisions rely on fixed-period maintenance or manual experience judgment, which are difficult to adapt to the fault evolution under complex working conditions. Therefore, this embodiment combines reinforcement learning, fuzzy inference, and game analysis methods to propose a data-driven intelligent operation and maintenance optimization strategy to achieve precise maintenance and optimal resource allocation.

[0195] As an option, this embodiment introduces an operation and maintenance optimization framework based on reinforcement learning. Through the interaction and learning between the agent and the environment, the maintenance strategy is automatically adjusted. Specifically, in a possible implementation, a deep Q-network (DQN) is used to optimize the maintenance decision, and the Markov decision process (MDP) is combined to model the operation and maintenance state transition to improve the adaptability of the decision. The following details the specific implementation of this step.

[0196] Operation and Maintenance Strategy Modeling and Optimization Objectives

[0197] In this embodiment, to optimize the intelligent operation and maintenance decision, first, an association model between the device health status and the maintenance strategy is established, and the operation and maintenance optimization objective is defined. Generally, the intelligent operation and maintenance strategy aims to minimize the operation and maintenance cost C m while ensuring that the health status HI(t) of the device is maintained above the safety threshold HI th . The optimization objective can be expressed as:

[0198] s.t. HI(t) ≥ HI th ,

[0199] where T is the length of the operation and maintenance cycle, and C m (t) represents the operation and maintenance cost at the t-th moment, including labor costs, equipment replacement costs, etc.

[0200] As an option, this embodiment uses the Markov decision process (MDP) to model the evolution of the device health status. Let the system state S t be jointly described by the health factor HI(t) and the device operating environment parameter E t , and the control strategy A t represent different operation and maintenance decision options. Then the state transition function of the MDP can be expressed as:

[0201] S t+1 = f(S t , A t , E t ) + ∈;

[0202] where f(·) is the state transition model, and ∈ is a random perturbation term representing the uncertainty of the external environment.

[0203] In some embodiments, to improve the stability of the decision, this embodiment combines the fuzzy inference method to adjust the operation and maintenance strategy. For example, during the maintenance of high-voltage GIS equipment, based on the decline rate of the health factor and the historical failure mode, fuzzy rules are used to calculate the maintenance priority and dynamically adjust the maintenance cycle.

[0204] Intelligent Maintenance Strategy Driven by Reinforcement Learning

[0205] In this embodiment, the reinforcement learning method is further adopted to optimize the maintenance decision, and the decision-making accuracy is improved through the Deep Q-Network (DQN). Generally, DQN can learn the optimal maintenance strategy through the interaction between the agent and the environment to minimize the long-term operation and maintenance cost. The Q-value update formula is as follows:

[0206]

[0207] where α is the learning rate, γ is the discount factor, R t is the immediate reward, and A′ is all possible subsequent actions.

[0208] In some embodiments, to improve the adaptability of the model to complex operation and maintenance environments, this embodiment adopts the Double DQN (Double Deep Q-Network) architecture to reduce the overestimation problem. Specifically, the calculation method of the target value of Double DQN is as follows:

[0209] Y t =R t +γQ(S t+1 ,argmax A Q erwa (S t+1 ),A′));

[0210] where Q evel and Q tanget are the evaluation network and the target network respectively to ensure the stability of learning.

[0211] In the operation and maintenance optimization of 500 kV GIS circuit breakers, the DQN model can dynamically adjust the maintenance plan based on historical fault modes and real-time health assessment data, reduce unnecessary shutdown maintenance, and ensure the long-term safe operation of the equipment.

[0212] Multi-Agent Game and Cooperative Operation and Maintenance Optimization

[0213] In a complex power grid system, there may be resource competition in the maintenance plans of multiple devices. Therefore, this embodiment further introduces the multi-agent game method to optimize the global operation and maintenance resource allocation. Generally, a multi-agent system (MAS) can coordinate the operation and maintenance strategies between different devices through distributed learning.

[0214] Suppose N devices are controlled by agents A i respectively, and the objective function of each agent is as follows:

[0215] minC i (A i ,A -i ),s.t.HI i ≥HIth ;

[0216] Among them, C i (A i , A -i ) represents the operation and maintenance cost of agent A i and is related to the strategy A of other agents -i .

[0217] In some embodiments, this embodiment uses the Nash Equilibrium to solve the optimal collaborative operation and maintenance strategy. Specifically, by combining reinforcement learning with game optimization, each agent continuously adjusts its strategy during the training process and finally converges to an equilibrium solution, minimizing the overall operation and maintenance cost.

[0218] In the operation and maintenance optimization of high-voltage substations, through the multi-agent game method, the maintenance plans can be coordinated among different devices, resource conflicts can be avoided, and at the same time, the maintenance cost of the whole station can be optimized.

[0219] The intelligent operation and maintenance decision of this embodiment depends on the health assessment result of step S5. If the calculation deviation of the health factor in S5 is large, it may lead to the failure of the maintenance strategy in this step. Therefore, the prediction model in S5 needs to match the reinforcement learning algorithm in this step to ensure the reliability of the input data.

[0220] In some embodiments, to improve the real-time performance of the decision-making, the DQN model in this step needs to be dynamically adjusted according to the change rate of the health factor in S5. For example, in the analysis of GIS circuit breakers, if S5 predicts that the health factor will drop to the critical value within 3 months, the DQN in this step can trigger preventive maintenance in advance to avoid failures.

[0221] In addition, the operation and maintenance optimization result of this step can be further used in step S7 (intelligent alarm and feedback adjustment). By combining the device operation data with the expert experience, the alarm mechanism can be optimized to improve the adaptive ability of the system. For example, in the intelligent operation and maintenance system of a substation, when the optimization decision in this step triggers a maintenance plan, S7 can record the execution result and adjust the training parameters of the DQN through the feedback mechanism to optimize the long-term decision-making effect.

[0222] S7. Dynamic parameter optimization: Monitor the false negative rate and false positive rate of the historical verification results; retrain the statistical manifold parameters when the false negative rate exceeds the threshold; adjust the chaotic evolution intensity according to the cumulative operation time of the device;

[0223] In the intelligent alarm and feedback adjustment steps of this embodiment, the aim is to establish an intelligent alarm mechanism based on the operation and maintenance optimization results obtained in step S6, and adjust the feedback strategy according to the operating status and maintenance conditions of the device. Generally, traditional alarm systems often rely on fixed thresholds to issue alarms. However, with the increasing complexity of the device operation and maintenance environment, fixed thresholds often cannot meet the requirements of dynamic adjustment. Therefore, this embodiment combines machine learning and adaptive feedback adjustment methods to propose a data-driven intelligent alarm and feedback adjustment mechanism to improve the accuracy of alarms and optimize subsequent operation and maintenance strategies.

[0224] As an option, this embodiment introduces an intelligent alarm system based on a multi-layer perceptron (MLP), which automatically identifies device failure risks through training the model and issues real-time alarms. Specifically, in a possible implementation, historical operation data and health factors are used as inputs, combined with the device's historical alarm records, to generate an alarm model through training, and the alarm threshold is continuously adjusted based on real-time data. The following details the specific implementation of this step.

[0225] Design and Implementation of Intelligent Alarm Mechanism

[0226] In this embodiment, the intelligent alarm mechanism mainly predicts the change trend of the device state by real-time monitoring and analysis of device health factors, combined with historical data, and then dynamically triggers alarms. Generally, the judgment basis for alarm triggering not only depends on a single health factor but also needs to consider the changes of multiple relevant parameters. Let be the multi-dimensional health feature vector of the device, where x i (t) represents the value of the i-th feature at time t.

[0227] The judgment formula for intelligent alarm can be expressed as:

[0228]

[0229] Among them, f(·) is the alarm trigger function, and θ is the alarm threshold, indicating that when the health state of the device reaches a certain critical value, an alarm is triggered.

[0230] As an option, in this embodiment, the alarm trigger function f(·) is composed of a multi-layer perceptron (MLP), and its output is the alarm probability. Specifically, in the analysis of high-voltage GIS devices, features such as health factors, temperature, and vibration will be used as inputs, and learned and trained through a neural network model to automatically adjust the alarm threshold. The training objective of this model is to maximize the alarm accuracy and minimize the false alarm rate.

[0231] Feedback Adjustment and Adaptive Learning

[0232] In this embodiment, the feedback mechanism is further used to adjust the optimization processes of the alarm strategy and the operation and maintenance decision-making to achieve adaptive adjustment and continuous optimization. Generally, the health status of the device and its changing trend are dynamic. Therefore, it is necessary to continuously adjust the alarm model and strategy according to the actual alarm effect and the operation feedback of the device.

[0233] Let be the loss function, representing the deviation between the alarm trigger strategy and the actual maintenance result. Then the feedback adjustment process can be realized by minimizing the loss function:

[0234]

[0235] where is the alarm value predicted by the model, is the actual alarm value, and N is the number of samples.

[0236] In some embodiments, in order to improve the self-adaptability of the alarm, this embodiment combines an online learning mechanism to continuously optimize the performance of the model by regularly updating the training data. Specifically, when an alarm is triggered, the system will adjust the model according to the actual fault occurrence situation and optimize the feedback strategy, thereby reducing the probability of false alarms and missed alarms.

[0237] For example, in the operation and maintenance of 500kV GIS equipment, when the system detects that the health factor of a certain device exceeds the preset threshold, the alarm system will trigger an alarm and issue operation and maintenance suggestions. At this time, the system will input the actual fault information as a new training sample for online learning, adjust the alarm threshold and strategy to ensure timely response to future similar events.

[0238] Dynamic adjustment of the alarm threshold based on historical data

[0239] In this embodiment, in order to further improve the accuracy and reliability of the alarm mechanism, the alarm threshold will be dynamically adjusted according to the historical data of the device. Generally, traditional alarm systems use fixed thresholds as alarm conditions. However, for some devices, the health factor may change within the normal fluctuation range, which may lead to false alarms.

[0240] Therefore, this embodiment combines the statistical analysis of historical data to dynamically adjust the alarm threshold by modeling the changing trend of historical health factors. For example, in the maintenance of high-voltage GIS circuit breakers, the fluctuation situation of historical health factors and the fault mode of the device can be used as inputs, and the changing trend of the health factor can be calculated by the sliding window method to determine a dynamic threshold based on timeliness and trend. The formula for this dynamic threshold is as follows:

[0241] θ(t) = μ HI (t) + k·σ HI (t);

[0242] where, μ HI (t) is the mean value of the health factor at time t, and σ HI (t) is the standard deviation, and k is a constant factor used to adjust the threshold sensitivity.

[0243] In some embodiments, historical data analysis can also be combined with an autoregressive model (AR) or an autoregressive moving average model (ARMA) for trend prediction, so as to further optimize the threshold setting.

[0244] The intelligent alarm and feedback adjustment in this embodiment rely on the optimized operation and maintenance decision-making in step S6 and the health assessment results in step S5. If the operation and maintenance strategy in S6 fails to be updated in time or there are deviations in the calculation of the health factor in S5, it may lead to the failure of the alarm strategy. Therefore, the tight connection between S5, S6 and S7 is crucial, and it is necessary to ensure that the input data for each step is accurate and effective to ensure the real-time and accuracy of the alarm.

[0245] In some embodiments, in order to further improve the response speed of the alarm, this embodiment adopts a multi-level alarm mechanism. Specifically, during the operation of the device, different levels of alarms will be triggered according to different degrees of change of the health factor. For example, slight fluctuations in the device health may trigger a warning, while severe abnormalities in the health factor will trigger an emergency fault alarm.

[0246] In addition, the feedback mechanism in this step can also be combined with the reinforcement learning model in step S6. When the alarm strategy is triggered, the feedback information will be input into the reinforcement learning system to help optimize the operation and maintenance decision-making and adjust the maintenance priority.

[0247] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for sharing power equipment supply chain data based on blockchain, characterized in that, It includes the following steps: S1. Data acquisition and standardization: Real-time acquisition of sensor data of power equipment and performing standardization processing; S2. Dynamic entropy threshold calculation: Dynamically calculating the data entropy value threshold based on information geometry theory; S3. Verification mode decision: Making a decision to trigger quantum chaotic random walk data sampling or full-scale zero-knowledge proof verification according to the entropy value threshold; S4. Anti-tampering credential construction: Constructing an anti-quantum hash chain for the sampled data and performing distributed anti-tampering encoding; S5. Blockchain evidence storage: Uploading the hash value and verification credential to the blockchain network; S6. Distributed verification: Executing distributed fault-tolerant verification through a smart contract and outputting the result; S7. Dynamic parameter optimization: Monitoring the false negative rate and false positive rate of historical verification results; When the false negative rate exceeds the threshold, retraining the statistical manifold parameters; Adjusting the chaotic evolution intensity according to the cumulative running time of the device.

2. The method for sharing power equipment supply chain data based on blockchain according to claim 1, wherein The S2 dynamic entropy threshold calculation includes: Calculating the local Shannon entropy with a sliding window; Establishing a statistical manifold model of the sensor data distribution; Dynamically adjusting the mean and standard deviation weight parameters based on the Riemannian gradient flow equation of the statistical manifold to generate an adaptive threshold.

3. The method for sharing power equipment supply chain data based on blockchain according to claim 1, wherein The S3 quantum chaotic random walk data sampling includes: Constructing a chaotic Hamiltonian according to the physical characteristics of the device; Embedding a chaotic coin operator during the quantum walk evolution; Generating an unpredictable sampling path based on the quantum state probability distribution.

4. The method for sharing power equipment supply chain data based on blockchain according to claim 3, characterized in that The construction of the chaotic Hamiltonian includes: Using a harmonic oscillator model as the integrable ground state Hamiltonian; Introducing a Gaussian random matrix to simulate the nonlinear perturbation of the device; Dynamically adjusting the chaotic perturbation intensity according to the aging degree of the device.

5. The method for sharing data of the power equipment supply chain based on blockchain according to claim 1, wherein The construction of the S4 anti-quantum hash chain includes: Collecting the physical noise of the device and generating an initial entropy source through analog-to-digital conversion; Iteratively generating a chained credential using an anti-quantum hash function; Only the consistency of the hash values at the head and tail of the chain needs to be compared during verification.

6. The method for sharing power equipment supply chain data based on blockchain according to claim 1, characterized in that The S4 distributed anti-tampering encoding includes: Using Reed-Solomon codes to perform redundant encoding on the sampled data; Storing the encoded shards in the edge nodes of supply chain associated enterprises; During verification, the original data is restored by decoding the polynomial equation, and tampered shards not exceeding half of the total number of shards are tolerated.

7. The method for sharing power equipment supply chain data based on blockchain according to claim 1, wherein The S5 blockchain network is of a consortium chain architecture: The core nodes adopt a practical Byzantine fault-tolerant consensus algorithm introducing pipeline batch processing; The edge nodes participate in local consensus according to business shards; The regulatory agency is connected to the network as an observing node for compliance auditing.

8. The method for sharing data of the power equipment supply chain based on blockchain according to claim 1, characterized in that, The S6 distributed fault-tolerant verification includes: Randomly selecting multiple edge nodes to obtain encoded shards; Tolerating tampered shards not exceeding half of the total number of shards through algebraic decoding; Determining that the data is valid when the decoding result matches the hash on the chain.

9. The method for sharing power equipment supply chain data based on blockchain according to claim 1, wherein The hardware implementation of the method includes: Deploying edge computing nodes supporting GPU acceleration on the power equipment side; Configuring a hardware security module at the blockchain node to perform anti-quantum signatures; Integrating a streaming data processing engine at the application layer server to clean the data uploaded to the chain in real time.

Images