A transformer insulating oil chromatographic fault diagnosis method based on quantum feature fusion

Abstract

The present application relates to the technical field of power fault diagnosis, and particularly relates to a transformer insulation oil chromatographic fault diagnosis method based on quantum feature fusion, comprising the following steps: S1, obtaining transformer dissolved gas data, and constructing original gas features and gas ratio features; S2, using a parameterized quantum circuit to extract high-dimensional nonlinear features, obtaining quantum features, and fusing the quantum features with the original gas features and the gas ratio features; S3, constructing a fault diagnosis model based on ensemble learning, using a gradient boosting decision tree, a random forest and a multilayer perception neural network as basic classifiers, determining the weights of the classifiers through stratified cross-validation, and combining weighted soft voting prediction to perform fault diagnosis; and S4, inputting multi-source features into the fault diagnosis model based on ensemble learning, and outputting fault diagnosis results. The present application shows significant improvement in accuracy, stability and robustness, and is suitable for online monitoring and fault diagnosis applications in power systems.

Description

Technical Field

[0001] This invention relates to the field of power fault diagnosis technology, specifically to a method for fault diagnosis of transformer insulating oil based on quantum feature fusion using chromatography. Background Technology

[0002] Transformers are critical equipment in power systems, and their operating status directly affects the safety and stability of the power grid. During long-term operation, transformers are affected by factors such as thermal stress, electrical stress, and insulation aging. This causes the insulating oil and solid insulating materials to decompose, producing characteristic gases such as hydrogen, methane, ethane, ethylene, and acetylene. These gases dissolve in the transformer oil, and changes in their types and concentrations can reflect the internal operating status and potential faults of the transformer. Therefore, fault diagnosis methods based on dissolved gas analysis (DGA) have become a common tool in transformer condition monitoring.

[0003] Existing technologies for fault diagnosis of dissolved gases in transformer insulating oil mainly include rule-based diagnostic methods such as the IEC three-ratio method and the characteristic gas method, as well as machine learning or deep learning-based diagnostic methods such as support vector machines, random forests, and neural networks. While these methods can achieve fault identification in certain scenarios, they generally suffer from the following problems: First, relying solely on original gas concentration features or a small number of fixed ratio features limits their feature representation capabilities, making it difficult to fully characterize the complex nonlinear coupling relationships between different gases. Second, when facing multiple types of faults such as low-to-medium temperature overheating, high-temperature overheating, partial discharge, low-energy discharge, and high-energy discharge, the boundaries between categories easily overlap, making it difficult to accurately distinguish similar fault types. Third, single classification models are easily affected by sample size, noise disturbances, and changes in feature distribution, resulting in insufficient model stability and generalization ability, making it difficult to meet the high accuracy and high reliability requirements of practical engineering applications. Summary of the Invention

[0004] This invention addresses the shortcomings of existing transformer fault diagnosis methods based on dissolved gas analysis, which often rely on raw gas concentrations or limited ratio features for modeling. These methods suffer from limited feature representation, failing to adequately depict the complex nonlinear coupling relationships between various gases, resulting in insufficient model differentiation of different fault types. Furthermore, traditional machine learning or single deep learning models often exhibit poor generalization and stability when dealing with multi-class, small-sample, and noisy data, thus affecting overall diagnostic accuracy. This invention provides a quantum feature fusion-based transformer insulating oil chromatography fault diagnosis method. By fusing gas concentration features, gas ratio features, and multi-scale quantum features, a unified multi-source feature representation is formed. An ensemble learning model composed of gradient boosting decision trees, random forests, and multilayer perceptrons is then used for prediction, effectively improving the accuracy, stability, and robustness of fault diagnosis.

[0005] The technical solution adopted in this invention is to provide a method for fault diagnosis of transformer insulating oil based on quantum feature fusion, including the following steps: S1. Obtain dissolved gas data from the transformer, and construct the original gas characteristics and gas ratio characteristics after preprocessing; S2. High-dimensional nonlinear features are extracted from transformer dissolved gas data using parameterized quantum circuits to obtain quantum features, which are then fused with the original gas features and gas ratio features to form a unified multi-source feature. S3. Construct a fault diagnosis model based on ensemble learning, using gradient boosting decision tree, random forest and multilayer perceptron neural network as basic classifiers, determine the weight of each classifier through hierarchical cross-validation, and combine weighted soft voting prediction for fault diagnosis. S4. The fault diagnosis model is based on ensemble learning and inputs multiple features, and outputs fault diagnosis results.

[0006] The parameterized quantum circuit described in step S2 uses a three-way parallel quantum circuit, and forms a multi-scale structure using 2, 3 and 5 qubits respectively.

[0007] The parameterized quantum circuit described in step S2 consists of an encoding layer, a variable layer, and a measurement layer. The encoding layer is used to map the transformer dissolved gas data to the quantum state space. The variable layer performs nonlinear transformations on the quantum state through quantum gates. The measurement layer is used to project the transformed quantum state into quantum features.

[0008] The encoding layer described in step S2 uses an angle encoding method to map the transformer dissolved gas data to the quantum state space, as shown in the following expression. For the i-th qubit Apply the pink RY gate and the green RZ gate in sequence. Formula 1 In Equation 1, (·)and (·) represent rotation gates around the Y-axis and Z-axis, respectively. Through double rotation, the same eigenvalue xi simultaneously affects two degrees of freedom on the Bloch sphere, realizing nonlinear feature embedding. The RY gate and RZ gate of each qubit qi use its corresponding feature xi. The variable layering described in step S2 consists of alternating rotation layers and entanglement layers. In the rotation layers, a universal rotation gate (Rot) is applied to each qubit, as expressed below. Formula 2 In Equation 2, α, β, and γ are trainable parameters, and αi,l, βi,l, and γi,l are three independent trainable parameters on the i-th qubit of the l-th layer. The parameter matrix has a shape of (L×n×3) and is initialized with a uniform distribution on [−π,π]. It is updated through gradient optimization during training. In the entanglement layer, a controlled NOT gate (CNOT) and a controlled Z-gate (CZ) are introduced simultaneously. Information transfer between adjacent qubits is achieved through a ring-structured CNOT gate; a CZ gate is applied between the qubits, as shown in the following expression. Formula 3 In Equation 3, the i-th qubit is the control qubit, the (i+1) mod n-th qubit is the target qubit, and the last qubit qn−1 is connected back to q0 through a modulo operation to form a ring topology. When the number of qubits n>2, an additional controlled Z-gate CZ is applied to the even-numbered adjacent bit pairs, as expressed below. Formula 4 For a 5-qubit circuit, CZ gates are applied to (q0,q1) and (q2,q3) respectively; no CZ gates are applied to a 2-qubit circuit; the global unitary transform of the l-th layer is expressed as: Formula 5 In Equation 5, and These represent the rotation operation and entanglement operation of the l-th layer, respectively. The entire variable layer is obtained after L=4 iterations to reach the final quantum state. ; The measurement layer described in step S2 is related to the final quantum state. Perform Pauli-Z measurements; the single-bit measurement output for the i-th qubit is: Formula 6 In Equation 6, For the Pauli-Z operator acting on the i-th qubit, a total of n single-bit features are obtained; Joint measurements between adjacent qubits are introduced to capture entanglement information between qubits. Formula 7 The maximum number of joint measurements is min(n−1,3). The output dimension of a single n-qubit circuit is... Formula 8.

[0009] The hierarchical cross-validation described in step S3 determines the weights of each classifier. The specific process includes... For each base classifier, calculate the accuracy and macro-average F1 score on the validation set, and take the weighted average of the two as the comprehensive score. The expression is as follows: Formula 9 In Equation 9, k represents the base classifier index. Let k be the average accuracy of classifier k in K=5-fold validation. The macro-average F1 score of classifier k in K=5 fold validation is normalized to obtain the ensemble weights of each classifier. The expression is as follows: Formula 10.

[0010] The weighted soft-voting prediction process described in step S3 includes the following steps: Each base classifier outputs a posterior probability vector of the class for each input feature. C represents the number of fault categories; weighted fusion yields the final probability distribution: Formula 11 The final fault type is determined by the highest probability category, as shown in the following expression: Equation 12.

[0011] The preprocessing described in step S1 includes outlier handling, standardization, and normalization.

[0012] The beneficial effects of this invention are that it provides a quantum feature fusion-based fault diagnosis method for transformer insulating oil chromatography. By constructing a multi-source feature fusion mechanism that combines original gas features, gas ratio features, and quantum features, the model can retain the absolute information of dissolved gas concentration while using ratio features to reflect the relative relationships in the fault mechanism. Simultaneously, it utilizes multi-scale parameterized quantum circuits to achieve high-dimensional mapping of complex nonlinear fault modes, thereby significantly enhancing feature representation capabilities. Furthermore, it employs an ensemble learning model composed of gradient boosting decision trees, random forests, and multilayer perceptrons, and uses cross-validation to adaptively allocate classifier weights, allowing the advantages of different models to complement each other, ultimately improving the accuracy, stability, and robustness of fault diagnosis.

[0013] Furthermore, the three base classifiers used in this invention—GBDT, Random Forest, and MLP—are not simply stacked in parallel. They have a structurally complementary relationship in terms of learning paradigm, error structure, and decision boundary geometry. They also have an isomorphic mapping relationship between the original gas features, gas ratio features, and quantum features and the multi-source features. This results in a combined technical effect that surpasses that of a single classifier and a single feature system.

[0014] In terms of learning paradigms, GBDT employs a Boosting additive model, achieving sequential bias correction by progressively fitting the residuals from the previous round; Random Forest uses the Bagging approach, constructing mutually decorrelated base trees through Bootstrap sample perturbation and feature subset randomization; and MLP uses end-to-end gradient optimization based on backpropagation to perform continuously differentiable nonlinear mapping on the input features. These three approaches correspond to three typical learning mechanisms: "progressive error correction," "independent voting," and "end-to-end approximation," covering the mainstream paths of discriminative modeling and achieving multi-angle modeling of the feature space at the algorithmic level. Regarding the geometric shape of the decision boundary, GBDT and Random Forest generate axis-aligned piecewise constant boundaries, naturally adaptable to discriminative rules with clear physical meanings, such as gas concentration thresholds and IEC ratio upper and lower limits; MLP generates continuously differentiable smooth nonlinear manifold boundaries, with stronger approximation capabilities for higher-order entangled coupling relationships formed after quantum features are projected into Hilbert space. The two types of decision boundaries are geometrically complementary, eliminating the "decision blind spot" caused by a single boundary shape in the class boundary region.

[0015] Building upon this foundation, the combination of multi-source features and heterogeneous integration creates further combined technical effects, the core of which lies in the isomorphic mapping effect between features and models. The multi-source features constructed in this invention consist of three heterogeneous subspaces: the 5-dimensional original gas concentration feature has a clear correspondence between physical dimensions and mechanisms; the 5-dimensional gas ratio feature is essentially a threshold-type regular feature, maintaining geometric consistency with the IEC three-ratio method criterion; and the 16-dimensional multi-scale quantum feature is a continuous floating-point projection of the parameterized quantum circuit output at the measurement layer, reflecting high-order nonlinear coupling between gas molecules. GBDT and random forests are naturally adept at handling the first two types of threshold-type and mechanism-type features, while MLP excels at end-to-end fitting of the entanglement relationships inherent in high-dimensional continuous quantum features. The three base classifiers are respectively connected to their most proficient sub-feature spaces, avoiding the capacity mismatch and suboptimal fitting problems caused by using a single model to handle all heterogeneous features, thus fully releasing the discriminative information contained in each type of feature.

[0016] Meanwhile, due to the completely different objective functions, sampling strategies, and inductive biases of the three base classifiers, their distribution correlation on misclassified samples is low. According to the error decomposition of ensemble learning, the variance of the fusion model is equal to the weighted sum of the average variance of the base classifiers and the covariance among the base classifiers. The covariance among cross-paradigm base classifiers is significantly reduced, allowing the ensemble model to significantly compress variance without increasing bias, and it is robust to training sample perturbations, initial hyperparameter values, and gas concentration measurement noise. Furthermore, GBDT and Random Forest, by aligning the original gas features and ratio features, form discrimination rules that are highly isomorphic to traditional mechanism-based diagnostic rules such as the IEC three-ratio method and the Rogers ratio method, possessing good interpretability and engineering credibility; MLP, on the other hand, captures high-order nonlinear couplings that are difficult to represent by classical statistical methods through quantum feature channels, forming an effective supplement to the blind spots of mechanism. Under the weighted soft voting framework, the mechanistically interpretable discriminant components and the data-driven implicit representations are mutually verified and weighted in the same probability space, so that the final diagnostic results do not deviate from engineering experience and can identify atypical failure modes that are difficult to cover by traditional ratio methods.

[0017] The diagnostic architecture of this invention, which combines multi-source features with heterogeneous integration, produces a combined technical effect from multiple levels, including learning paradigm, feature geometry, error structure, and mechanism interpretability, fundamentally improving the overall performance of the transformer insulating oil chromatography fault diagnosis method. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the parameterized quantum circuit structure of 5 qubits in this invention; Figure 3 This is a schematic diagram of the diagnostic process of the integrated learning fault diagnosis model in this invention; Figure 4 This is a bar chart showing the accuracy of fault classification for each category in the test examples of this invention. Figure 5 This is a schematic diagram of the test set confusion matrix in this invention. Detailed Implementation

[0019] like Figure 1 As shown, this invention provides a method for fault diagnosis of transformer insulating oil based on quantum feature fusion, comprising the following steps: S1. Data Acquisition and Preprocessing Data on dissolved gases in transformers are acquired, and after data preprocessing, original gas characteristics and gas ratio characteristics are constructed.

[0020] Data preprocessing includes outlier handling, standardization, and normalization.

[0021] S2. Feature extraction of multi-scale parameterized quantum circuits High-dimensional nonlinear features are extracted from transformer dissolved gas data using multi-scale parameterized quantum circuits to obtain quantum features. These original features, ratio features, and quantum features are then fused to form a unified multi-source feature set. Utilizing the superposition and entanglement properties of qubits, classical input data is mapped to a high-dimensional Hilbert space, thereby achieving effective modeling of complex gas coupling relationships.

[0022] The parameterized quantum circuit employs three parallel quantum circuits, using 2, 3, and 5 qubits respectively, to form a multi-scale structure, thereby enabling feature extraction at different scales.

[0023] The parameterized quantum circuit consists of an encoding layer, a variable layer, and a measurement layer. The encoding layer maps the input data to the quantum state space; the variable layer performs nonlinear transformations on the quantum state by training quantum gates; and the measurement layer projects the quantum state into classical output features.

[0024] 1. Coding layer The role of the encoding layer is to map classical input features to quantum state space. This invention uses angle encoding to map the normalized input features... Mapping to the quantum state space, specifically as follows: For the i-th qubit Apply the pink RY gate and the green RZ gate in sequence: Formula 1 In Equation 1, and These represent rotation gates around the Y and Z axes, respectively. Through double rotation, the same eigenvalue xi simultaneously affects two degrees of freedom on the Bloch sphere, achieving richer nonlinear feature embeddings. It's important to note that... Figure 2 As shown in the coding layer, the RY gate and RZ gate of each qubit qi use its corresponding feature xi, rather than sharing the same input.

[0025] 2. Variable layering After encoding, the quantum state enters the parameterized quantum circuit. The QCAN variable layer is composed of alternating rotational layers and entangled layers, repeated L=4 times to enhance the circuit's ability to characterize complex modes.

[0026] (1) Rotation layer In the rotation layer, a universal rotation gate (Rot) is applied to each qubit, which is essentially a combination of three fundamental rotations: Formula 2 In Equation 2, α, β, and γ are trainable parameters. αi,l, βi,l, and γi,l are three independent trainable parameters on the i-th qubit of the l-th layer. The parameter matrix has a shape of (L×n×3) and is initialized with a uniform distribution on [−π, π]. It is updated through gradient optimization during training.

[0027] (2) Entanglement layer In the entanglement layer, to establish correlation between different qubits, this invention simultaneously introduces a controlled NOT gate (CNOT) and a controlled Z-gate (CZ). First, information transfer between adjacent qubits is achieved through a ring-structured CNOT gate; second, a CZ gate is applied between some qubits to further enhance the entanglement relationship. This process can be represented as follows: Formula 3 In Equation 3, the i-th qubit is the control qubit, the (i+1) mod n-th qubit is the target qubit, and the last qubit qn−1 is connected back to q0 through a modulo operation, forming a ring topology. Figure 2 Vertical lines that cross each horizontal line in the variable layering.

[0028] Secondly, when the number of qubits n>2, an additional controlled Z-gate CZ is applied to the even-numbered adjacent bit pairs: Formula 4 For a 5-qubit circuit, CZ gates are applied to (q0, q1) and (q2, q3) respectively, corresponding to the two CZ boxes on the right side of the layered diagram in Figure X. For a 2-qubit circuit, since there is only a single pair of bits, no CZ gate is applied. The combination of CNOT and CZ enables the circuit to simultaneously capture the linear coupling and higher-order entanglement relationships between adjacent bits.

[0029] The global unitary transformation of the l-th layer is expressed as: Formula 5 In Equation 5, and These represent the rotation operation and entanglement operation of the l-th layer, respectively. The entire variable layer is obtained after L=4 iterations to reach the final quantum state. .

[0030] 3. Measurement layer like Figure 2 As shown in the measurement layer, for the final quantum state Pauli-Z measurements were performed to extract classical features.

[0031] The single-bit measurement output for the i-th qubit is: Formula 6 In Equation 6, The Pauli-Z operator, acting on the i-th qubit, yields a total of n single-bit features.

[0032] In addition, joint measurements between adjacent qubits are introduced to capture entanglement information between qubits: Formula 7 The joint measurement can take at most min(n−1, 3) pairs, and the output dimension of a single n-qubit circuit is: Formula 8.

[0033] S3. Construction of Fault Diagnosis Model Based on Ensemble Learning like Figure 3 As shown, this invention constructs a fault diagnosis model based on ensemble learning, which is used to classify and model fused features and output the final fault diagnosis result.

[0034] The fault diagnosis model consists of three parts: base classifier, adaptive weights, and weighted fusion decision.

[0035] 1. Base classifier construction unit To fully extract the discriminative information from the fused features, this invention selects three representative machine learning models as base classifiers: Gradient Boosting Decision Tree (GBDT), Random Forest (RF), and Multi-Layer Perceptron (MLP).

[0036] The three types of models have significant differences in structure and learning mechanism, and can model the feature space from different perspectives.

[0037] (1) GBDT: GBDT is an ensemble learning method based on the Boosting idea, which continuously optimizes the model performance by fitting the residuals of the previous model in each round. Its core idea is to combine multiple weak learners in an additive model so that the model gradually approximates the true objective function. Its iterative process is expressed as:

[0038] in, This represents the model after the m-th iteration. Let be the residual function fitted in the current round. This represents the learning rate or step size coefficient. By progressively correcting the prediction error of the previous round, GBDT can effectively capture local differences between samples and has a strong fitting ability for complex nonlinear decision boundaries.

[0039] (2) Random Forest: Random forest is an ensemble learning method based on the Bagging idea. It constructs multiple independent decision trees by bootstrap resampling of the training samples and randomly selecting a subset of features in each decision tree for partitioning. Finally, the classification result is obtained through a voting mechanism. Its prediction process is expressed as follows:

[0040] in, Let K represent the prediction result of the k-th decision tree, where K is the number of decision trees. Random forests effectively reduce model variance by introducing randomness into samples and features, exhibiting good resistance to overfitting and strong generalization performance.

[0041] (3) MLP: Multilayer Perceptron is a typical feedforward neural network structure that achieves complex mapping of input features through multiple fully connected networks and nonlinear activation functions. Its forward propagation process can be represented as:

[0042] Where x is the input feature vector, h is the hidden layer representation, and y is the output vector. , b1 and b2 are network parameters, respectively, and σ(·) is a nonlinear activation function (such as ReLU). MLP can learn high-order nonlinear combination relationships between features and has a strong expressive power for multi-source fusion features.

[0043] The three base classifiers are complementary in terms of bias-variance properties and feature modeling capabilities: GBDT focuses on local error correction, RF emphasizes model stability and anti-overfitting ability, and MLP excels at modeling high-dimensional nonlinear feature relationships. By integrating the advantages of the three through an ensemble learning strategy, the classification accuracy and robustness of the model can be effectively improved.

[0044] 2. Adaptive weight determination To avoid the subjectivity of manually setting weights, this invention adaptively determines the weights of each classifier using a K=5 fold hierarchical cross-validation method. For each base classifier, the accuracy and macro-average F1 score are calculated on the validation set, and their equally weighted average is taken as the comprehensive score. :

[0045] Where k represents the base classifier index, Let k be the average accuracy of classifier k in K=5-fold validation. The macro-average F1 score (Macro-F1) of classifier k in K=5-fold validation is given. The ensemble weights of each classifier are obtained after normalization. :

[0046] The weights are determined by the model’s actual performance on the validation set. The classifier with better performance automatically receives higher voting weights, achieving adaptive dynamic weight allocation.

[0047] 3. Weighted fusion decision-making Each base classifier outputs a posterior probability vector of the class for each input sample. Where C represents the number of fault categories; weighted fusion yields the final probability distribution:

[0048] The final failure type is determined by the most probable category:

[0049] This weighted soft voting method can combine the advantages of different models to improve diagnostic accuracy and robustness.

[0050] S4. The fault diagnosis model is based on ensemble learning and inputs multiple features, and outputs fault diagnosis results.

[0051] This invention uses a transformer fault example to illustrate the accuracy of the transformer fault diagnosis model based on multi-scale quantum feature fusion and ensemble learning.

[0052] This invention utilizes the DGA dataset for dissolved gas analysis in transformer oil from real-world industrial scenarios for experimental verification. The dataset contains 1140 samples, covering six typical fault types: low-to-medium temperature overheating, low-energy discharge, partial discharge, normal state, high-temperature overheating, and high-energy discharge. The samples are evenly distributed across categories, with 100 samples per category used for training and 90 samples per category for testing, resulting in a total of 600 training samples and 540 testing samples. The input features are the concentrations of five characteristic gases: H2, CH4, C2H6, C2H4, and C2H2, all derived from gas chromatography detection results.

[0053] In terms of feature construction, this invention adopts a multi-source feature fusion strategy, including standardized original gas features, gas ratio features constructed based on the IEC method, and quantum features extracted through multi-scale parameterized quantum circuits. Specifically, the quantum features are extracted by three parallel quantum circuits of 2 qubits, 3 qubits, and 5 qubits, ultimately forming a 16-dimensional quantum feature vector. The three types of features are concatenated to form a 26-dimensional fused feature, which serves as the input to the ensemble learning model.

[0054] For model training, the ensemble learning model consists of Gradient Boosting Decision Tree (GBDT), Random Forest (RF), and Multilayer Perceptron (MLP); the GBDT model is configured with 250 trees. The learning rate was 0.05, and the maximum depth was 4. The random forest model was set to 300 decision trees with a maximum depth of 18. The MLP adopted a three-layer fully connected structure. The weights of each base classifier were automatically determined through 5-fold cross-validation, using the combined score of classification accuracy and macro-average F1 score as the weighting basis, and a weighted soft voting method was used for the final decision.

[0055] Experimental results show that the method of this invention achieved an accuracy of 96.3% on the test set, with high recognition rates for various types of faults and stable overall classification performance. The classification performance for each category is as follows: Figure 4 As shown.

[0056] Test set confusion matrix as follows Figure 5 As shown in the confusion matrix results, most samples can be correctly classified, with only slight confusion between a few faults with similar physical characteristics, which is consistent with actual engineering practices.

[0057] Compared with Support Vector Machine (SVM), Random Forest (single model), and Multilayer Perceptron (single model), this invention has significant improvements in classification accuracy and stability, as shown in Table 1.

[0058]

[0059] The transformer fault diagnosis method proposed in this invention, based on multi-scale quantum feature fusion and ensemble learning, employs a multi-source feature fusion method involving original gas features, gas ratio features, and quantum features. By fusing multiple types of features, it enhances the characterization capability of transformer fault information. It utilizes multi-scale parameterized quantum circuits for feature extraction, employing parallel quantum circuits of different qubit scales to encode, evolve, and measure input data, extracting high-dimensional nonlinear features. An ensemble fault diagnosis model combining GBDT, random forest, and MLP is used, with parallel modeling through multiple classifiers to improve the accuracy and stability of fault identification. Adaptive weight allocation through cross-validation and weighted soft voting automatically determine weights based on the performance of each classifier and fuses them to output the final diagnostic result. This method demonstrates significant improvements in accuracy, stability, and robustness, making it suitable for online monitoring and fault diagnosis applications in practical power systems.

Claims

1. A method for fault diagnosis of transformer insulating oil based on quantum feature fusion, characterized in that: Includes the following steps, S1. Obtain dissolved gas data from the transformer, and construct the original gas characteristics and gas ratio characteristics after preprocessing; S2. High-dimensional nonlinear features are extracted from transformer dissolved gas data using parameterized quantum circuits to obtain quantum features, which are then fused with the original gas features and gas ratio features to form a unified multi-source feature. S3. Construct a fault diagnosis model based on ensemble learning, using gradient boosting decision tree, random forest and multilayer perceptron neural network as basic classifiers, determine the weight of each classifier through hierarchical cross-validation, and combine weighted soft voting prediction for fault diagnosis. S4. The fault diagnosis model is based on ensemble learning and inputs multiple features, and outputs fault diagnosis results.

2. The method for fault diagnosis of transformer insulating oil based on quantum feature fusion according to claim 1, characterized in that: The parameterized quantum circuit described in step S2 uses a three-way parallel quantum circuit, and forms a multi-scale structure using 2, 3 and 5 qubits respectively.

3. The method for fault diagnosis of transformer insulating oil based on quantum feature fusion according to claim 1, characterized in that: The parameterized quantum circuit described in step S2 consists of an encoding layer, a variable layer, and a measurement layer. The encoding layer is used to map the transformer dissolved gas data to the quantum state space. The variable layer performs nonlinear transformations on the quantum state through quantum gates. The measurement layer is used to project the transformed quantum state into quantum features.

4. The method for fault diagnosis of transformer insulating oil based on quantum feature fusion according to claim 3, characterized in that: The encoding layer described in step S2 uses an angle encoding method to map the transformer dissolved gas data to the quantum state space, as shown in the following expression. For the i-th qubit Apply the pink RY gate and the green RZ gate in sequence. Formula 1 In Equation 1, (·)and (·) represent rotation gates around the Y-axis and Z-axis, respectively. Through double rotation, the same eigenvalue xi simultaneously affects two degrees of freedom on the Bloch sphere, realizing nonlinear feature embedding. The RY gate and RZ gate of each qubit qi use its corresponding feature xi. The variable layering described in step S2 consists of alternating rotation layers and entanglement layers. In the rotation layers, a universal rotation gate (Rot) is applied to each qubit, as expressed below. Formula 2 In Equation 2, α, β, and γ are trainable parameters, and αi,l, βi,l, and γi,l are three independent trainable parameters on the i-th qubit of the l-th layer. The parameter matrix has a shape of (L×n×3) and is initialized with a uniform distribution on [−π,π]. It is updated through gradient optimization during training. In the entanglement layer, a controlled NOT gate (CNOT) and a controlled Z-gate (CZ) are introduced simultaneously. Information transfer between adjacent qubits is achieved through a ring-structured CNOT gate; a CZ gate is applied between the qubits, as shown in the following expression. Formula 3 In Equation 3, the i-th qubit is the control qubit, the (i+1) mod n-th qubit is the target qubit, and the last qubit qn−1 is connected back to q0 through a modulo operation to form a ring topology. When the number of qubits n>2, an additional controlled Z-gate CZ is applied to the even-numbered adjacent bit pairs, as expressed below. Formula 4 For a 5-qubit circuit, CZ gates are applied to (q0,q1) and (q2,q3) respectively; no CZ gates are applied to a 2-qubit circuit; the global unitary transform of the l-th layer is expressed as: Formula 5 In Equation 5, and These represent the rotation operation and entanglement operation of the l-th layer, respectively. The entire variable layer is obtained after L=4 iterations to reach the final quantum state. ; The measurement layer described in step S2 is related to the final quantum state. Perform Pauli-Z measurements; the single-bit measurement output for the i-th qubit is: Formula 6 In Equation 6, For the Pauli-Z operator acting on the i-th qubit, a total of n single-bit features are obtained; Joint measurements between adjacent qubits are introduced to capture entanglement information between qubits. Formula 7 The maximum number of joint measurements is min(n−1,3). The output dimension of a single n-qubit circuit is... Formula 8.

5. The method for fault diagnosis of transformer insulating oil based on quantum feature fusion according to claim 1, characterized in that: The hierarchical cross-validation described in step S3 determines the weights of each classifier. The specific process includes... For each base classifier, calculate the accuracy and macro-average F1 score on the validation set, and take the weighted average of the two as the comprehensive score. The expression is as follows: Formula 9 In Equation 9, k represents the base classifier index. Let k be the average accuracy of classifier k in K=5-fold validation. The macro-average F1 score of classifier k in K=5 fold validation is normalized to obtain the ensemble weights of each classifier. The expression is as follows: Formula 10.

6. The method for fault diagnosis of transformer insulating oil based on quantum feature fusion according to claim 1, characterized in that: The weighted soft-voting prediction process described in step S3 includes the following steps: Each base classifier outputs a posterior probability vector of the class for each input feature. C represents the number of fault categories; weighted fusion yields the final probability distribution: Formula 11 The final fault type is determined by the highest probability category, as shown in the following expression: Equation 12.

7. The method for fault diagnosis of transformer insulating oil based on quantum feature fusion according to claim 1, characterized in that: The preprocessing described in step S1 includes outlier handling, standardization, and normalization.

Images