Transformer fault diagnosis method, device and system, and storage medium

Abstract

The invention discloses a transformer fault diagnosis method, device and system, and a storage medium. The method comprises the following steps: synchronously collecting heterogeneous sensor data; according to the collected parameters, complexity analysis and feature extraction are carried out through an improved LZC-LMSE algorithm; an improved space-time attention convolutional network is introduced, and space-time features of multi-source heterogeneous data such as vibration and temperature fields are fused through a multi-head attention mechanism; according to the obtained fused features, fault association rule mining is carried out through an incremental # imgabs0 # algorithm; and a fault diagnosis result generated by using the optimal parameter is compared with a set threshold value, so that fault diagnosis is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment status monitoring, and in particular relates to a transformer fault diagnosis method and device, a system, and a storage medium. Background Art

[0002] The access of new users during business expansion, especially the high-proportion and large-scale access of new energy, will lead to complex working conditions such as sudden dynamic load changes and harmonic pollution in the transformer. The existing transformer fault diagnosis technology relies too much on a single type of sensor, and the collected data is too single, making it difficult to effectively capture the coupling effects of multi-source heterogeneous parameters such as mechanical vibration, temperature, and chemical parameters. It cannot effectively characterize the correlation between multi-dimensional parameters, and the multi-dimensional parameters of the transformer are highly nonlinear, making effective feature extraction difficult, and insufficient consideration of timing. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a transformer fault diagnosis method and device, system and storage medium.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A transformer fault diagnosis method, comprising:

[0006] Step S1, synchronously collecting heterogeneous sensor data;

[0007] Step S2: Improve the LZC-LMSE algorithm to extract fault features from heterogeneous sensor data;

[0008] Step S3: Based on the extracted fault features, multi-source data fusion is performed by improving the spatiotemporal attention convolutional network;

[0009] Step S4: Based on the fused features, use incremental The base algorithm generates fault association rules;

[0010] Step S5: Dynamically adjust the diagnosis model parameters based on the fault association rules to generate dynamic fault diagnosis results.

[0011] Preferably, the heterogeneous sensor data includes mechanical parameters, electrical parameters, thermodynamic parameters and chemical data.

[0012] Preferably, in step S5, based on the fault association rules, the improved Bayesian optimization algorithm is used to dynamically adjust the diagnosis model parameters and output the fault diagnosis result.

[0013] The present invention also provides a transformer fault diagnosis device, comprising:

[0014] A first processing module is used to synchronously collect heterogeneous sensor data;

[0015] The second processing module improves the LZC-LMSE algorithm to extract fault features from heterogeneous sensor data;

[0016] The third processing module performs multi-source data fusion based on the extracted fault features by improving the spatiotemporal attention convolutional network;

[0017] The fourth processing module uses incremental The base algorithm generates fault association rules;

[0018] The fifth processing module dynamically adjusts the diagnosis model parameters based on the fault association rules to generate dynamic fault diagnosis results.

[0019] Preferably, the heterogeneous sensor data includes mechanical parameters, electrical parameters, thermodynamic parameters and chemical data.

[0020] Preferably, the fifth processing module dynamically adjusts the diagnosis model parameters based on the fault association rules using an improved Bayesian optimization algorithm and outputs the fault diagnosis results.

[0021] The present invention also provides a transformer fault diagnosis system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a transformer fault diagnosis method when executed by the processor.

[0022] The present invention also provides a storage medium, on which a computer program is stored. When the computer program is run, the method for diagnosing transformer faults is executed.

[0023] The present invention solves the problem of single data collection by synchronously collecting heterogeneous sensor data, including mechanical parameters, electrical parameters, thermodynamic parameters and chemical parameters; uses the improved LZC-LMSE algorithm for complexity analysis and feature extraction to achieve nonlinear feature extraction; introduces an improved spatiotemporal attention convolutional network to achieve multi-source data fusion; and through incremental The algorithm mines fault association rules, taking into account temporal sequence and avoiding missed detection of latent faults. Finally, it generates optimal parameters based on improved Bayesian optimization and uses the fault diagnosis results generated by the optimal parameters to achieve dynamic and real-time fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0025] Figure 1 This is a flow chart of a transformer fault diagnosis method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0027] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] Example 1:

[0029] like Figure 1 As shown, an embodiment of the present invention provides a transformer fault diagnosis method including:

[0030] Step 1: Synchronously collect heterogeneous sensor data

[0031] In the transformer fault diagnosis system, the synchronous acquisition process requires multiple types of heterogeneous sensors as follows: Real-time monitoring of multiple physical field parameters is achieved by deploying a multi-type sensor network. These parameters include mechanical parameters (vibration accelerometers, acoustic emission sensors, strain gauges), electrical parameters (current transformers, voltage transformers, partial discharge sensors), thermodynamic parameters (fiber optic temperature sensors, infrared thermal imagers, humidity sensors), chemical parameters (oil chromatographs, dielectric testers), and new sensors (high-precision MEMS sensors, Doppler ultrasonic sensors). All sensors achieve microsecond-level time synchronization (error <1μs) using the IEEE 1588 precision clock protocol to form a unified time series dataset. This ensures the consistency of multi-source heterogeneous data (such as high-frequency vibration signals and low-frequency temperature fields) in the temporal and spatial dimensions, providing high-quality input for subsequent feature extraction and fault correlation analysis.

[0032] Step 2: Complexity analysis and feature extraction by improving the LZC-LMSE algorithm

[0033] Based on the multi-source heterogeneous sensor data collected in step 1, the improved Lempel-Ziv complexity (LZC) is used to quantify the nonlinear self-similarity and local structural complexity of the data, targeting the complex time-varying characteristics of the transformer multi-source heterogeneous data. This simplifies the time-varying characteristics of the multi-source heterogeneous data and achieves temporal unification. The coding length statistical process is optimized by introducing a dynamic threshold strategy. Compared with the traditional LZC quantization process based on a fixed threshold binary sequence, the improved method introduces a dynamic threshold strategy:

[0034]

[0035] Among them, the dynamic threshold δ t According to the statistical characteristics of the data in the sliding window (such as the mean μ t and standard deviation σ t ) Adaptive generation:

[0036] δ t =μ t +α*σ t (α is the adjustment factor)

[0037] Through the dynamic threshold δ t Adaptive binary coding, for binary sequence S = {S1, S2, ..., S N} Perform segmented compression coding, complexity C LZC for:

[0038]

[0039] Where K is the number of unique sub-patterns and N is the sequence length. The dynamic threshold strategy is implemented by δ t The time-varying property improves the adaptability to non-stationary data. LZC , reflecting the overall random change trend of the signal.

[0040] Complexity analysis at a single scale is difficult to capture the differences in fault characteristics at multiple time resolutions. To this end, the improved multi-scale entropy (LMSE) analysis is introduced to map the signal to different scale spaces through wavelet packet decomposition and reconstruction, and independently calculate the sample entropy (SSE) and cross entropy (CE) in multiple frequency bands. Among them, wavelet packet decomposition realizes the multi-resolution expression of the signal, ensuring the separation of high-frequency details and low-frequency trends. The signal x(t) is decomposed into multi-scale self-signals through wavelet packets. Scale j corresponds to frequency band k, and the reconstruction formula is:

[0041]

[0042] Sample entropy (SSE) quantifies the signal irregularity within a single scale and identifies the fault-sensitive mode in a specific frequency band; reconstructs the signal x at scale jj (t), calculate the sample entropy:

[0043]

[0044] Among them, m is the embedding dimension, r is the similarity domain value, A m (r) is the m-dimensional vector matching probability.

[0045] Cross entropy (CE) measures the difference in entropy distribution across scales, revealing the multi-scale coupling effect caused by faults. Comparing the differences in sample entropy distribution at different scales j1 and j2:

[0046]

[0047] Among them, P j (i) is the probability of entropy distribution at scale j, which is used to capture cross-scale fluctuation characteristics.

[0048] Although the features extracted by LZC and LMSE have physical interpretations, their sensitivity to fault modes varies. Direct concatenation can lead to redundancy and noise amplification. To this end, a nonlinear weighted feature extraction method is used. To construct fault-sensitive feature vectors, a nonlinear weighted fusion of LZC global complexity and LMSE improved multi-scale entropy is performed:

[0049] F (1) =[ω1*C LZC ,ω1*SSE(j),ω3*CE(j1,j2)]

[0050] F (1) The weight coefficients ω1, ω2, and ω3 are optimized by the entropy weight method to reflect the contribution of the feature to the fault mode:

[0051]

[0052] H i is the information entropy of feature i, p i is the normalized probability.

[0053] Step 3: Implement multi-dimensional data fusion by improving the spatiotemporal attention convolution algorithm

[0054] We further introduce an improved spatiotemporal attention convolutional network, and realize the spatiotemporal correlation feature fusion of transformer multi-source heterogeneous data through the collaborative design of multi-head attention mechanism and layered convolution structure. First, according to the feature vector F extracted in step 2 (1) Use parallel temporal convolution branches and spatial convolution branches to generate time dimension feature maps F respectively t and spatial dimension feature vector graph F s , highlighting the key spatiotemporal information related to the fault:

[0055] A t =softmax(W t *F t +b t )

[0056] A s =softmax(W s *F s +b s )

[0057] Among them, A t and A s They represent the time dimension and space dimension attention weights, respectively. t and W t They represent the attention weights calculated in the time domain and the attention weights calculated in the spatial domain, respectively, and b t and b t They represent the attention bias calculated in the time domain and the attention bias calculated in the spatial domain, respectively.

[0058] The temporal and spatial attentions are combined through the outer product to construct the joint attention matrix A:

[0059]

[0060] in, Represents the outer product operation, that is, the two weight vectors are combined into a matrix. Then the preliminary features are weighted using the joint attention matrix to obtain the fused features:

[0061] F f =A⊙F

[0062] Among them, ⊙ represents element-by-element multiplication. Based on the weighted fusion features, higher-level features are extracted through subsequent convolutional layers:

[0063] F (2) =σ(W*F f +b)

[0064] Among them, W and b are the convolution kernel and bias; σ(·) is the activation function. (2) It is the final representation after the fusion of multi-source data, which can be used for subsequent fault rule mining and dynamic diagnosis.

[0065] Step 4: Use incremental Generate fault association rules based on the basic algorithm

[0066] After feature extraction in step 2 and data fusion in step 3, the fused feature F is obtained. (2) , then incrementally Incremental The fault association rules are dynamically mined using the Inference Based Basis (IGB) algorithm to discover the deep fault relationships between different sensor data, adapt to the evolution of equipment status over time, and establish a model that can be used for fault diagnosis.

[0067] The multi-source feature vector F obtained in step 2 and step 3 is (2) Map it into a polynomial and construct a polynomial equation system based on the fused features:

[0068] g i (F (2) )=g i ([f1, f2, ..., f N ])=0, j=1,2,...,M

[0069] Among them, f i Represents the eigenvalues of the i sensors after fusion. g i (·) is the optimal expression found by data-driven approach.

[0070] In the initial training phase, samples D = {D1, D2, ..., D K}, and construct the characteristic polynomial equations constructed above The basic equation system G is calculated using the Buchberger algorithm. base:

[0071] G={g1,g2,……g M}

[0072] G0=GB(G)

[0073] During the training process, as time goes by, when the new data of the extracted samples reaches D K+1 When , construct the new equation:

[0074] g M+1 (F (2) )=0

[0075] And check whether the new equation is independent of the current Base, if g M+1 If it can be generated by G0, no update is required; if g M+1 If it cannot be generated by G0, it needs to be updated base; updated The basis is expressed as:

[0076] G k+1 =GB(G0∪{g M+1})

[0077] pass Base G kThe generated polynomial rules are used to extract association rules, and the mapping relationship between fault variables can be obtained:

[0078]

[0079] Among them, R j represents the jth fault association rule, f i (·) is passed The characteristic mapping function obtained after the basis simplification; τ i is the failure threshold of the rule; and the rule set is expressed as:

[0080] R={R1,R2,……,R M}

[0081] Step 5: Generate dynamic fault diagnosis results

[0082] After mining the fault association rules in step 4, the improved Bayesian optimization algorithm is used to dynamically adjust the diagnosis model parameters to output more accurate diagnosis results.

[0083] According to step 4, based on the fault association rules between different sensors, define the objective function, that is, the loss function of the diagnosis model, set the fault label as y, and construct the diagnosis model f(F (2) ,Θ), the prediction result of the model is In order to find the optimal parameter Θ * , minimize the loss function, define the loss function, and its mean square error:

[0084]

[0085] The improved Bayesian optimization algorithm is used to model the distribution of the objective function ρ(Θ) using Gaussian process (GP):

[0086] p(ρ(Θ))~gp(θ(Θ),k(Θ,Θ'))

[0087] Among them, θ(Θ) is the mean function, and k(Θ, Θ') is the kernel function, which is used to measure the correlation between different parameters.

[0088] After building the surrogate model, it is necessary to select an appropriate acquisition function to determine the parameter points for evaluation. Taking Expected Improvement (EI) as an example, the calculation formula is as follows:

[0089]

[0090] Among them, ρ *is the currently known optimal loss value. Based on the known loss value, the improved Bayesian optimization is performed, and the iterative optimization process is as follows: ① Initialization, randomly select several parameters Θ1, Θ2, ..., Θ k And calculate the corresponding loss value ρ(Θ i ); ② Train the Gaussian process model based on the existing data; ③ Select the next parameter point according to the acquisition function: ④Calculate the loss value of the new point and update the proxy model; ⑤When the predetermined number of iterations is reached or convergence is achieved, determine the final optimal parameter Θ * ⑥Finally, the optimized parameters Θ are used * To troubleshoot:

[0091]

[0092] By comparing with the set threshold, it is determined whether a fault has occurred, thereby achieving dynamic and real-time fault diagnosis.

[0093] The present invention is based on the microsecond-level synchronous acquisition and wavelet noise reduction preprocessing of the IEEE 1588 protocol, which solves the difficulty of spatiotemporal alignment of multi-source heterogeneous data such as mechanical, electrical, thermodynamic, and chemical data. Based on the proposed improved LZC-LMSE algorithm combined with a dynamic threshold strategy and multi-scale entropy, nonlinear fault features are effectively extracted. By improving the spatiotemporal attention convolutional network through a multi-head attention mechanism, the spatiotemporal correlation feature fusion of multi-source heterogeneous data of transformers is realized. The introduced incremental Grobner basis (IGB) algorithm can dynamically update the fault rule base, and at the same time, combined with the improved Bayesian optimization algorithm, the new fault diagnosis model has adaptive learning capabilities.

[0094] Example 2:

[0095] The embodiment of the present invention also provides a transformer fault diagnosis device, comprising:

[0096] A first processing module is used to synchronously collect heterogeneous sensor data;

[0097] The second processing module improves the LZC-LMSE algorithm to extract fault features from heterogeneous sensor data;

[0098] The third processing module performs multi-source data fusion based on the extracted fault features by improving the spatiotemporal attention convolutional network;

[0099] The fourth processing module uses incremental The base algorithm generates fault association rules;

[0100] The fifth processing module dynamically adjusts the diagnosis model parameters based on the fault association rules to generate dynamic fault diagnosis results.

[0101] As an implementation manner of an embodiment of the present invention, the heterogeneous sensor data includes: mechanical parameters, electrical parameters, thermodynamic parameters and chemical data.

[0102] As an implementation method of the embodiment of the present invention, the fifth processing module dynamically adjusts the diagnosis model parameters based on the fault association rules using an improved Bayesian optimization algorithm and outputs the fault diagnosis result.

[0103] Example 3:

[0104] The present invention also provides a transformer fault diagnosis system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a transformer fault diagnosis method when executed by the processor.

[0105] Example 4:

[0106] The present invention also provides a storage medium, on which a computer program is stored. When the computer program is run, the method for diagnosing transformer faults is executed.

[0107] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A transformer fault diagnosis method, characterized in that: include: Step S1, synchronously collecting heterogeneous sensor data; Step S2: Improve the LZC-LMSE algorithm to extract fault features from heterogeneous sensor data; Step S3: Based on the extracted fault features, multi-source data fusion is performed by improving the spatiotemporal attention convolutional network; Step S4: Based on the fused features, use incremental The base algorithm generates fault association rules; Step S5: Dynamically adjust the diagnosis model parameters based on the fault association rules to generate dynamic fault diagnosis results.

2. The transformer fault diagnosis method according to claim 1, wherein: Heterogeneous sensor data includes: mechanical parameters, electrical parameters, thermodynamic parameters and chemical data.

3. The transformer fault diagnosis method according to claim 2, characterized in that: In step S5, based on the fault association rules, the improved Bayesian optimization algorithm is used to dynamically adjust the diagnosis model parameters and output the fault diagnosis results.

4. A transformer fault diagnosis device, characterized in that: include: A first processing module is used to synchronously collect heterogeneous sensor data; The second processing module improves the LZC-LMSE algorithm to extract fault features from heterogeneous sensor data; The third processing module performs multi-source data fusion based on the extracted fault features by improving the spatiotemporal attention convolutional network; The fourth processing module uses incremental The base algorithm generates fault association rules; The fifth processing module dynamically adjusts the diagnosis model parameters based on the fault association rules to generate dynamic fault diagnosis results.

5. The transformer fault diagnosis device according to claim 4, characterized in that: Heterogeneous sensor data includes: mechanical parameters, electrical parameters, thermodynamic parameters and chemical data.

6. The transformer fault diagnosis device according to claim 5, characterized in that: The fifth processing module dynamically adjusts the diagnosis model parameters based on the fault association rules using the improved Bayesian optimization algorithm and outputs the fault diagnosis results.

7. A transformer fault diagnosis system, characterized in that: include: A memory and a processor, wherein the memory stores a computer program executed by the processor, and when the computer program is executed by the processor, the transformer fault diagnosis method according to any one of claims 1 to 3 is executed.

8. A storage medium, characterized in that: The storage medium stores a computer program, which, when running, executes the transformer fault diagnosis method according to any one of claims 1 to 3.

Images