Multi-parameter adaptive dynamic multilevel decision converter transformer fault diagnosis method and system

Abstract

The invention provides a multi-parameter adaptive dynamic multilevel decision converter transformer fault diagnosis method and system, and the method comprises the steps: constructing a multi-parameter fuzzy feature space based on temperature, voiceprint, current and oil soluble gas parameter data, carrying out the statistics of fuzzy features through sliding time sequence segmentation, generating a multi-parameter fuzzy time sequence feature vector, and carrying out the fault diagnosis of a converter transformer. Obtaining multi-parameter internal cross characteristics and multi-parameter external cross characteristics, and performing splicing fusion to form multi-parameter correlation characteristics; constructing a multi-stage aggregation classifier, dynamically distributing a weight for each classifier based on verification set performance, fusing probability output fault types of each layer, and adjusting the matching degree of a model prediction probability and a real fault occurrence probability; and dynamically adjusting the depth of the hierarchy according to the matching degree adjustment condition, if the improvement of the fault diagnosis precision of continuous multiple layers is smaller than a set threshold value, terminating the expansion and retaining the current hierarchy, and executing multi-level processing on a prediction result. According to the invention, various parameter information is effectively fused, and the prediction accuracy and reliability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of state monitoring and fault diagnosis of power transmission and transformation equipment in power systems, and in particular to a multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method and system, as well as a corresponding computer terminal and computer-readable storage medium. Background Art

[0002] As the core equipment of the HVDC transmission system, the operating status of the converter transformer directly affects the safety of the power grid. Traditional converter transformer fault diagnosis technology mainly relies on the threshold judgment of a single parameter or a fixed rule model, which has the following limitations:

[0003] Limitations of Single Parameters: Data from a single sensor cannot fully reflect complex fault mechanisms. For example, partial discharge can simultaneously cause an abnormal sound pattern and elevated hydrogen concentrations, but single-parameter analysis can easily lead to misjudgments. Fixed threshold settings rely on manual experience and cannot adapt to dynamic interference such as equipment aging and ambient temperature fluctuations, leading to increased false alarm rates.

[0004] Insufficient synergy between multi-source data: Existing methods have limited ability to fuse multimodal sensor data and lack a unified feature representation and correlation analysis mechanism. Key information, such as the temporal correlation between temperature and load current, and the nonlinear coupling between voiceprint spectrum and gas concentration, is not effectively utilized. Traditional feature engineering (such as mean and variance) struggles to capture dynamic characteristics under complex operating conditions, leading to feature redundancy or information loss.

[0005] Weak model adaptability: Fault diagnosis models based on static classifiers (such as support vector machines and random forests) have fixed weight distribution and layer depth, and cannot be dynamically optimized based on real-time data. This leads to performance degradation with equipment aging or sudden environmental changes. The lack of a probability calibration mechanism means that the predicted probability output by the model deviates from the actual probability of fault occurrence, affecting the reliability of confidence interval determinations. Summary of the Invention

[0006] In view of the above-mentioned deficiencies in the prior art, the present invention provides a multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method and system, and also provides a corresponding computer terminal and computer-readable storage medium.

[0007] According to one aspect of the present invention, a multi-parameter adaptive dynamic multi-stage decision-making method for converter transformer fault diagnosis is provided, comprising:

[0008] Based on the four types of parameter data obtained by sensors, namely temperature, soundprint, current and dissolved gas in oil, a multi-parameter fuzzy feature space is constructed. The fuzzy features are statistically analyzed through sliding time series segmentation to generate a multi-parameter fuzzy time series feature vector.

[0009] Based on the multi-parameter fuzzy time series feature vector, multi-parameter internal cross features and external cross features are respectively obtained, and multi-parameter correlation characteristics for characterizing the dynamic coupling relationship of multi-source data are formed by splicing and fusing the internal cross features and the external cross features;

[0010] Based on the multi-parameter correlation characteristics, a multi-level aggregate classifier model is constructed, and weights are dynamically assigned to each classifier based on the performance of the validation set. The predicted probabilities of each layer are fused to output the fault type, and calibration parameters are introduced to adjust the matching degree between the model predicted probability and the actual fault probability to obtain the calibrated fault probability;

[0011] According to the matching adjustment of the dynamic weighted multi-level aggregation classifier on the validation set, the hierarchical depth is dynamically adjusted. If the improvement of the fault diagnosis accuracy of multiple consecutive layers is less than the set threshold a, the expansion is terminated and the current hierarchical structure is locked. At the same time, multi-level processing is performed on the prediction results based on the calibrated fault probability.

[0012] According to another aspect of the present invention, a multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis system is provided, comprising:

[0013] The fuzzy feature acquisition module constructs a multi-parameter fuzzy feature space based on four types of parameter data acquired by sensors: temperature, voiceprint, current, and dissolved gas in oil. It then performs statistics on the fuzzy features through sliding time series segmentation to generate a multi-parameter fuzzy time series feature vector.

[0014] A correlation feature acquisition module, which obtains multi-parameter internal cross features and external cross features based on the multi-parameter fuzzy time series feature vector, and forms a multi-parameter correlation characteristic for characterizing the dynamic coupling relationship of multi-source data by splicing and fusing the internal cross features and the external cross features;

[0015] A fault probability prediction module, which builds a multi-level aggregate classifier model based on the multi-parameter correlation characteristics, dynamically assigns weights to each classifier based on the validation set performance, fuses the predicted probabilities of each layer to output the fault type, and introduces calibration parameters to adjust the matching degree between the model's predicted probability and the actual fault probability, thereby obtaining a calibrated fault probability;

[0016] A prediction result processing module dynamically adjusts the hierarchical depth according to the matching adjustment of the dynamically weighted multi-level aggregation classifier on the validation set. If the improvement in the fault diagnosis accuracy of multiple consecutive layers is less than the set threshold a, the expansion is terminated and the current hierarchical structure is locked. At the same time, multi-level processing is performed on the prediction results based on the calibrated fault probability.

[0017] According to a third aspect of the present invention, a computer terminal is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the terminal can be used to execute the method described above in the present invention, or to execute the system described above in the present invention.

[0018] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, it can be used to execute the method described above in the present invention, or to run the system described above in the present invention.

[0019] Due to the adoption of the above technical solution, the present invention has at least one of the following beneficial effects compared with the prior art:

[0020] The multi-parameter adaptive dynamic multi-level decision-making converter transformer fault diagnosis method and system provided by the present invention realizes efficient collaboration of multi-source heterogeneous data and adaptive optimization of model output through core technologies such as fuzzy feature space construction, internal and external cross-feature fusion, and verification set-driven dynamic calibration, significantly improving the accuracy and reliability of converter transformer fault diagnosis.

[0021] The multi-parameter adaptive dynamic multi-level decision-making converter transformer fault diagnosis method and system provided by this invention maps temperature, soundprint fundamental frequency amplitude, current, and oil-soluble gas parameters into a unified fuzzy feature sequence through piecewise linear membership functions, Gaussian membership functions, S-shaped membership functions, and trapezoidal membership functions. This solves the problem that traditional single-parameter threshold judgment is sensitive to dynamic interference. Combining sliding time series segmentation (with a window length of 10 seconds and a step length of 1 second) to generate time series correlation feature vectors effectively captures the time series correlation characteristics of temperature and load current, enhancing the collaborative expression capabilities of multi-source heterogeneous data.

[0022] The multi-parameter adaptive dynamic multi-stage decision-making method and system for commutation transformer fault diagnosis proposed in this paper generates inner cross-signals by calculating the product combinations of pairwise state memberships and the off-diagonal elements of the covariance matrix within the same sensor window. Kernel density estimation is used to quantify the strength of nonlinear correlations across sensor membership sequences, and outer cross-signals are generated by combining product combinations of key state pairs. This method effectively integrates coupled relationships across modal data, such as temperature-voiceprint and current-gas concentration, addressing the problem of traditional methods' inability to capture nonlinear correlations in multi-source data.

[0023] The multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method and system provided by this invention dynamically assigns classifier weights based on validation set data, integrates the output probabilities of multi-stage classifiers such as logistic regression, support vector machines, gradient boosting trees, and deep neural networks, and uses an exponential smoothing algorithm to calibrate prediction bias in real time. This technology overcomes the performance degradation of static classifiers due to equipment aging or sudden environmental changes, significantly improving the reliability and adaptability of the model output.

[0024] The multi-parameter adaptive dynamic multi-level decision-making converter transformer fault diagnosis method and system provided by the present invention dynamically terminates the layer expansion according to the accuracy improvement of the verification set (stops when the accuracy improvement of 4 consecutive layers is insufficient) to avoid model overfitting; at the same time, it divides the confidence intervals into high, medium and low according to the fusion probability, and performs automatic early warning, joint analysis (combining the oil-gas ratio and load-temperature coupling characteristics) and noise filtering operations respectively, thereby optimizing the efficiency of operation and maintenance decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0026] Figure 1 The figure is a working flow chart of a multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method in a preferred embodiment of the present invention.

[0027] Figure 2 Schematic diagram of the process of constructing inner cross features and outer cross features in a preferred embodiment of the present invention.

[0028] Figure 3 This is a structural diagram of a multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis system in a preferred embodiment of the present invention.

[0029] Figure 4 The figure is a working diagram of a multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method in a specific application example of the present invention.

[0030] Figure 5 This is a three-dimensional T-SNE diagram of the dynamic weighted multi-level aggregation classifier model in a specific application example of the present invention; in the figure, (a) is a loose core, (b) is a loose winding, (c) is a DC bias magnetization, and (d) is a loop fault.

[0031] Figure 6 Schematic diagram of the confusion matrix for core loosening, winding loosening, DC bias magnetization, and loop fault in a specific application example of the present invention; in the figure, (a) is core loosening, (b) is winding loosening, (c) is DC bias magnetization, and (d) is loop fault.

[0032] Figure 7This is a comparison chart of the diagnostic accuracy of the method provided by the present invention and other neural network models. DETAILED DESCRIPTION

[0033] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process. It should be noted that those skilled in the art may make various modifications and improvements without departing from the scope of the present invention, and these modifications and improvements fall within the scope of protection of the present invention.

[0034] Existing converter transformer fault diagnosis technologies often suffer from single-parameter limitations, insufficient multi-source data collaboration, and weak model adaptability. To address these issues, one embodiment of the present invention provides a multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method. This method leverages core technologies such as fuzzy feature space construction, internal and external cross-feature fusion, and validation set-driven dynamic calibration to achieve efficient collaboration of multi-source heterogeneous data and adaptive optimization of model output, significantly improving the accuracy and reliability of converter transformer fault diagnosis.

[0035] Specifically, if Figure 1 As shown, the multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method provided by this embodiment may include:

[0036] S1, based on the four types of parameter data obtained by the sensor, namely temperature, soundprint, current and dissolved gas in oil, constructs a multi-parameter fuzzy feature space, and performs statistics on the fuzzy features through sliding time series segmentation to generate a multi-parameter fuzzy time series feature vector.

[0037] S2, based on the multi-parameter fuzzy time series feature vector, obtains the multi-parameter internal cross features and external cross features respectively, and forms the multi-parameter correlation characteristics for characterizing the dynamic coupling relationship of multi-source data by splicing and fusing the internal cross features and the external cross features.

[0038] S3, based on the multi-parameter correlation characteristics, construct a multi-level aggregation classifier, dynamically assign weights to each classifier based on the performance of the validation set, fuse the predicted probabilities of each layer to output the fault type, and introduce calibration parameters to adjust the matching degree between the model prediction probability and the actual fault probability to obtain the calibrated fault probability.

[0039] S4, dynamically adjust the hierarchical depth according to the matching adjustment of the dynamically weighted multi-level aggregate classifier on the validation set (i.e., the accuracy improvement of the multi-level classifier on the validation set). If the improvement of the fault diagnosis accuracy of multiple consecutive layers is less than the set threshold a, terminate the expansion and lock the current hierarchical structure. At the same time, perform multi-level processing on the prediction results based on the calibrated fault probability.

[0040] In some preferred embodiments, the above S1, constructing a multi-parameter fuzzy feature space, may further include:

[0041] S111, define a piecewise linear membership function for each parameter data, and map the temperature value to three fuzzy sets: low temperature, normal temperature, and high temperature:

[0042]

[0043] Where μ 低温 (T) represents the membership function of the temperature belonging to the low temperature fuzzy set; μ 高温 (T) represents the membership function of the temperature belonging to the high temperature fuzzy set; T represents the real-time monitored temperature sensor value, which is the input variable of the membership function; T min represents the lowest temperature threshold of the low temperature fuzzy set; T normlow Indicates the lower limit of the normal temperature range; T normhigh Indicates the upper limit of the normal temperature range; T max represents the highest temperature threshold of the high temperature fuzzy set;

[0044] S112, using the Gaussian membership function, maps the voiceprint fundamental frequency amplitude A to normal and abnormal fuzzy sets:

[0045]

[0046] Where μ 异常 (A) represents the mapping of the fundamental frequency amplitude of the voiceprint A to the abnormal fuzzy set; A represents the mean fundamental frequency amplitude of the normal working condition in the training set; μ A Indicates the central value of the abnormal voiceprint fundamental frequency amplitude; σ A Indicates the standard deviation of the fundamental frequency amplitude of abnormal conditions in the training set.

[0047] S113, define the S-type membership function to map the current value to the normal and overcurrent fuzzy sets:

[0048]

[0049] Where μ 过流 (I) represents the mapping of current value to overcurrent fuzzy set; I represents the current value detected in real time; I rated represents the rated current of the device; k represents the slope factor, which is used to control the steepness of the transition interval. More preferably, k = 0.1; I th represents the overcurrent alarm threshold, and further preferably, I th =1.2*I rated .

[0050] S114, define a trapezoidal membership function to map gas concentrations to normal and warning fuzzy sets:

[0051]

[0052] Where μ 预警 (G) represents mapping gas concentration to warning fuzzy set; G represents collected gas concentration; G norm Indicates the normal concentration threshold; G alarm Indicates the warning threshold;

[0053] S115, construct the four types of parameter data after fuzzification into a unified fuzzy feature sequence F fused .

[0054] In some preferred embodiments, the above S1, which performs statistics on fuzzy features by sliding time series segmentation to generate a multi-parameter fuzzy time series feature vector for characterizing the dynamic coupling relationship of multi-source data, may further include:

[0055] S121 sets the window length to m seconds and the step length to n seconds, and segments the fuzzy feature sequence. The continuously collected fuzzy feature sequence is segmented into multiple time series segments, generating multiple window feature sets, thereby obtaining a multi-parameter fuzzy time series feature vector. Further preferably, m is 10 seconds and n is 1 second.

[0056] S122: Based on the dimensions of the generated window features for temperature, voiceprint, current, and gas concentration, a total window feature dimension is obtained, which is the sum of the dimensions of the window features for temperature, voiceprint, current, and gas concentration. Further preferably, the dimensions of the generated window features for temperature, voiceprint, current, and gas concentration are 3, 2, 2, and 2, respectively; and the total window feature dimension is the sum of the dimensions of the temperature, voiceprint, current, and gas window features, which is 9 dimensions.

[0057] In some preferred embodiments, the above S2, obtaining the inner intersection feature, may further include:

[0058] S211, calculate the membership product of each state in the same window of the same sensor to obtain a product combination. The state refers to the window characteristics of temperature, voiceprint, current, and gas concentration;

[0059] S212, calculating the covariance matrix of the membership of each state in the window, extracting the off-diagonal elements, and obtaining the covariance feature.

[0060] S213, concatenate the product combination and the covariance feature to obtain the final inner cross feature.

[0061] Further preferably, the temperature membership product combination feature number is 3, and the membership covariance matrix feature is 3; the voiceprint membership product combination feature number is 1, and the membership covariance matrix feature is 1; the current membership product combination feature number is 1, and the membership covariance matrix feature is 1; the gas concentration membership product combination feature number is 1, and the membership covariance matrix feature is 1; the product combination feature and the covariance feature are spliced ​​to obtain the final internal cross feature, and the internal cross feature is a total of 12.

[0062] In some preferred embodiments, the above S2, obtaining the outer intersection feature, may further include:

[0063] S221, use the kernel density estimation method to estimate the joint distribution of the membership of the i-th state (window characteristics of temperature, voiceprint, current, and gas concentration) of sensor A and the j-th state membership of sensor B in the sliding window, and obtain the marginal probability density as follows:

[0064]

[0065] Where, represents the marginal probability density; W represents the weight vector; h represents the bandwidth parameter, which controls the smoothness of the kernel density estimation; K represents the kernel function, which is used to estimate the joint distribution of the membership of sensors A and B, and t represents the current time.

[0066] S222, based on the marginal probability density, the nonlinear correlation strength of the membership sequence is obtained as:

[0067]

[0068] Where M(A,B) represents the nonlinear correlation strength; K A represents the kernel function of sensor A; K B represents the kernel function of sensor B; represents the marginal probability density of the i-th state of sensor A; represents the marginal probability density of the j-th state of sensor B;

[0069] S223, select M from sensor A A Key state membership, select M from sensor B B The key state membership is calculated, and the product of the real-time membership is formed to form M A ×M B Furthermore, the key state membership refers to the fuzzified membership state of sensor A / B, which is obtained by screening state combinations with high nonlinear correlation strength and frequent co-occurrence.

[0070] In step S224, the product of the nonlinear correlation strength and all key state pairs is combined to form the final outer cross feature. Furthermore, each state pair directly corresponds to the product of the nonlinear correlation strength and the real-time membership. These two are combined to form the outer cross feature and input into the classifier, achieving targeted enhancement of fault-sensitive state pairs.

[0071] Further preferably, the number of temperature external cross features is 3, the number of voiceprint external cross features is 1, the number of current external cross features is 1, and the number of gas concentration external cross features is 1, resulting in a total of 6 external cross features and internal cross features.

[0072] The construction process of the above inner cross features and outer cross features is as follows: Figure 2 shown.

[0073] In the above S2, the "multi-parameter correlation characteristics" form a dynamic coupling relationship representation by fusing the internal and external cross-features of multi-sensor data, which serves as the input feature of the dynamic classifier to optimize the model weight distribution and fault prediction; it supports the probability output after calibration, triggering graded treatments such as early warnings, joint analysis or noise filtering with different confidence levels; at the same time, through the accuracy feedback of the verification set, the classifier hierarchical structure is dynamically adjusted to achieve adaptive closed-loop optimization from feature extraction to diagnostic decision-making, thereby improving the coordination and reliability of fault diagnosis.

[0074] In some preferred embodiments, the above S3 constructs a multi-level aggregate classifier, dynamically assigns weights to each classifier based on the performance of the validation set, and fuses the predicted probabilities of each layer to output the fault type, including:

[0075] S311, construct a validation set, which includes time samples of multiple consecutive days in the historical monitoring data and covers all fault types.

[0076] S312: Based on the internal and external cross features of temperature, voiceprint, current and gas concentration, multiple classifiers are used to classify the faults, and the classification accuracy of each classifier is output.

[0077] S313, using the validation set, dynamically assign weights based on the classification accuracy of each classifier, then:

[0078]

[0079] Where, represents the initial weight of the classifier; represents the initial validation set accuracy of the classifier; represents the initial accuracy of classifier j on the validation set, represents the sum of the initial accuracies of the four classifiers; α represents an exponential amplification factor, which is used to enhance the weight ratio of the high-precision classifier. More preferably, α=1.5.

[0080] S314, the fault diagnosis probability predicted by each classifier is fused through weights to obtain the model prediction probability, and the final fault type is output.

[0081] In some preferred embodiments, the above S3, introducing calibration parameters to adjust the matching degree between the model prediction probability and the actual fault occurrence probability, may further include:

[0082] S321, based on the statistical deviation between the model prediction probability and the actual fault occurrence probability, a dynamic mapping function is established. The output probability distribution of the classifier prediction is corrected in real time using an exponential smoothing algorithm to obtain a correction model, thereby adjusting the matching degree between the model prediction probability and the actual fault occurrence probability. The correction model is expressed as:

[0083]

[0084] Where, Represents the probability of the classifier after calibration at the current moment; represents the probability after calibration at the previous moment; represents the original output probability of the classifier at the current moment; β represents the smoothing coefficient, which is adaptively adjusted according to the historical prediction error. Further preferably, the value of β is between [0.2, 0.5].

[0085] The above correction model uses the exponential smoothing algorithm to correct the classifier output probability in real time to ensure that the prediction result is consistent with the actual fault probability.

[0086] In some preferred embodiments, the above S4, dynamically adjusting the level depth according to the matching adjustment of the dynamic weighted multi-level aggregation classifier on the validation set, may further include:

[0087] S411, in the multi-level structure of the stochastic decision ensemble of the multi-level aggregate classifier model, if the improvement in validation set accuracy over successive layers of iteration is less than a set threshold a, then the iteration is terminated. Otherwise, the iteration is continued until the requirement is met, and the final classifier weights and fusion fault diagnosis probability are output. More preferably, the number of iteration layers is 4, and the threshold a is set to a value between 0 and 1%, more preferably 0.5%.

[0088] In some preferred embodiments, the above S4, performing multi-level processing on the prediction result based on the calibrated failure probability, may further include:

[0089] Multi-level handling includes: automatic early warning with high confidence probability, joint analysis with medium confidence probability, and automatic ignoring of low confidence probability as noise.

[0090] Further:

[0091] S421: When the calibrated failure probability is within a high confidence interval, it is automatically determined to be a confirmed failure and an early warning is initiated. Furthermore, the high confidence interval is a confidence interval of ≥90%.

[0092] S422: When the calibrated fault probability is within a low confidence interval, it is marked as environmental noise interference and ignored. Further, the low confidence interval is a confidence interval of ≤15%.

[0093] S423: When the calibrated fault probability is within the medium confidence interval, it is pushed to the joint diagnosis interface, where a joint analysis is performed based on the oil-gas ratio and the load-temperature coupling characteristics. Furthermore, the medium confidence interval is a confidence interval greater than 15% and less than 90%. Among them:

[0094] The gas-in-oil ratio is: If this value is greater than a set threshold value b, a fault occurs. Further preferably, the threshold value b is set to 0.1.

[0095] The load-temperature coupling characteristics are: If the feature is greater than the set threshold c, it is determined to be a fault condition; where:

[0096] I load Indicates the effective value of the commutation load current, which is obtained from the existing current sensor data;

[0097] T hotspot Indicates the hot spot temperature of the winding, taken from the temperature sensor;

[0098] Cov() represents the covariance between load current and hotspot temperature, which measures the linear correlation between the two;

[0099] σ I Indicates the standard deviation of the load current and characterizes the load fluctuation amplitude;

[0100] σ T It represents the standard deviation of hot spot temperature and characterizes the amplitude of temperature fluctuation.

[0101] The threshold value b is set to a value between 0 and 1, and more preferably 0.1.

[0102] The threshold value c is set to a value of 0 to 5, and more preferably 1.

[0103] Based on the same inventive concept, an embodiment of the present invention further provides a multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis system.

[0104] Specifically, if Figure 3 As shown, the multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis system provided by this embodiment may include

[0105] The fuzzy feature acquisition module constructs a multi-parameter fuzzy feature space based on four types of parameter data acquired by sensors: temperature, voiceprint, current, and dissolved gas in oil. It then performs statistics on the fuzzy features through sliding time series segmentation to generate a multi-parameter fuzzy time series feature vector.

[0106] A correlation feature acquisition module, which obtains multi-parameter internal cross features and external cross features based on the multi-parameter fuzzy time series feature vector, and forms a multi-parameter correlation characteristic for characterizing the dynamic coupling relationship of multi-source data by splicing and fusing the internal cross features and the external cross features;

[0107] A fault probability prediction module, which builds a multi-level aggregate classifier model based on the multi-parameter correlation characteristics, dynamically assigns weights to each classifier based on the validation set performance, fuses the predicted probabilities of each layer to output the fault type, and introduces calibration parameters to adjust the matching degree between the model's predicted probability and the actual fault probability, thereby obtaining a calibrated fault probability;

[0108] A prediction result processing module dynamically adjusts the hierarchical depth according to the matching adjustment of the dynamically weighted multi-level aggregation classifier on the validation set. If the improvement in the fault diagnosis accuracy of multiple consecutive layers is less than the set threshold a, the expansion is terminated and the current hierarchical structure is locked. At the same time, multi-level processing is performed on the prediction results based on the calibrated fault probability.

[0109] The specific contents of the functional modules constituting the system provided by the above embodiment of the present invention are further described below.

[0110] The multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis system provided by the above embodiment of the present invention mainly includes the following functional modules

[0111] The fuzzy feature acquisition module constructs a multi-parameter fuzzy feature space based on the temperature, voiceprint, current, and dissolved gas in oil parameter data collected by temperature sensors, ultrasonic sensors, current sensors, and MEMS photoacoustic sensors, and generates a multi-parameter fuzzy time series feature vector by sliding time series segmentation statistical features.

[0112] The correlation feature acquisition module forms multi-parameter correlation features by splicing and fusing multi-parameter internal cross features and external cross features.

[0113] The fault probability prediction module constructs a dynamically weighted multi-level aggregate classifier. Through a multi-level structure of random decision ensemble, it dynamically assigns weights to each classifier based on the performance of the validation set, fuses the probabilities of each layer to output the fault type, and introduces calibration parameters to adjust the matching degree between the model prediction probability and the actual fault probability.

[0114] The prediction result processing module dynamically adjusts the layer depth according to the improvement of the verification set accuracy (matching adjustment). If the fault diagnosis accuracy of four consecutive layers is improved by less than 0.5%, the expansion is terminated. At the same time, multi-level processing is performed on the prediction results based on the calibrated probability.

[0115] Furthermore, the fuzzy feature acquisition module further includes:

[0116] Fuzzy feature construction unit, which constructs a multi-parameter fuzzy feature space, including:

[0117] First, define the piecewise linear membership function for multiple parameters:

[0118] Define a piecewise linear membership function to map temperature values ​​to three fuzzy sets: "low temperature", "normal", and "high temperature":

[0119]

[0120]

[0121] Use the Gaussian membership function to map the voiceprint fundamental frequency amplitude A to the "normal" and "abnormal" fuzzy sets:

[0122]

[0123] Where A is the mean amplitude of the fundamental frequency under normal working conditions in the training set; σ A : Standard deviation of the fundamental frequency amplitude under normal working conditions in the training set

[0124] Define the S-type membership function to map the current value to the "normal" and "overcurrent" fuzzy sets:

[0125]

[0126] I rated is the rated current of the equipment; k = 0.1 is the slope factor, which controls the steepness of the transition interval; I th =1.2*I rated Overcurrent alarm threshold

[0127] Define a trapezoidal membership function to map gas concentrations to "normal" and "warning" fuzzy sets:

[0128]

[0129] G norm is the normal concentration threshold, G alarm is the warning threshold;

[0130] The four types of fuzzified sensor data are constructed into a unified feature sequence: F fused .

[0131] The time series feature construction unit generates a multi-parameter fuzzy time series feature vector by sliding the time series segmentation statistical features, including:

[0132] The sliding time series segmentation process is as follows: set the window length to 10 seconds and the step length to 1 second, then segment the fuzzy feature sequence to generate multiple window sets:

[0133] Temperature, voiceprint, current, and gas concentration window features generate 3D, 2D, 2D, and 2D features respectively; the total window feature dimensions are temperature, voiceprint, current, and gas window feature dimensions and 9 dimensions;

[0134] Furthermore, the associated feature acquisition unit further includes:

[0135] The inner cross feature construction unit is used to construct the inner cross feature, including:

[0136] First, the membership product of each two states in the same window of the same sensor is calculated;

[0137] Secondly, the covariance matrix of the membership of each state in the window is calculated and the off-diagonal elements are extracted;

[0138] Finally, the product combination is concatenated with the covariance feature to obtain the final inner cross feature:

[0139] Among them, the number of temperature membership product combination features is 3, and the covariance matrix feature of membership is 3; the number of voiceprint membership product combination features is 1, and the covariance matrix feature of membership is 1; the number of current membership product combination features is 1, and the covariance matrix feature of membership is 1; the number of gas concentration membership product combination features is 1, and the covariance matrix feature of membership is 1; the product combination features and the covariance features are spliced ​​together to obtain the final internal cross features, and the total internal cross features are 12.

[0140] The outer cross feature construction unit is used to construct the outer cross feature, including:

[0141] First, the nonlinear correlation strength of the membership sequences of the two sensors is calculated;

[0142] The kernel density estimation method is used to estimate the joint distribution of the i-th state membership of sensor A and the j-th state membership of sensor B in the sliding window, and the marginal probability density is obtained:

[0143]

[0144] in,

[0145] Thus, the nonlinear correlation strength of the membership sequence is obtained:

[0146]

[0147] Next, we calculate the cross-sensor product combination:

[0148] (1) Select M from sensor A A Select M from sensor B B A key state. Form M A ×M B Key state pairs;

[0149] (2) For each key state pair, calculate the product of the real-time membership;

[0150] Finally, the product of the nonlinear correlation strength and all key pairs is combined to form the final outer cross feature. The number of outer cross features for temperature is 3, the number of outer cross features for voiceprint is 1, the number of outer cross features for current is 1, and the number of outer cross features for gas concentration is 1, resulting in a total of 6 outer cross features and inner cross features.

[0151] Furthermore, the failure probability prediction module also includes:

[0152] The classifier unit is used to build a dynamic weighted multi-level aggregate classifier, including:

[0153] First, a validation set is constructed. The validation set is independent of the training set and contains 7 consecutive days of time samples in the historical monitoring data, covering all fault types.

[0154] Secondly, we use logistic regression, support vector machine, gradient boosting tree, and deep neural network as classifiers for fault classification based on the internal and external cross-features of temperature, voiceprint, current, and gas concentration, and output the classification accuracy of each classifier.

[0155] Then, dynamic weight allocation is performed based on the accuracy of each classifier:

[0156]

[0157] is the initial weight of the classifier, is the initial validation set accuracy of the classifier; Represents the sum of the initial accuracies of the four classifiers; α: exponential amplification factor, used to enhance the weight ratio of the high-precision classifier; and the final fusion fault diagnosis probability is calculated through the weight.

[0158] The calibration unit is used to introduce calibration parameters to correct the output probability distribution, including:

[0159] Based on the statistical deviation between the model prediction probability and the actual fault probability, a dynamic mapping function is established to correct the output probability distribution of the classifier in real time through the exponential smoothing algorithm. The correction formula is:

[0160]

[0161] The original output probability of the classifier at the current moment;

[0162] The value of α is between [0.2, 0.5], which is a smoothing coefficient and is adaptively adjusted according to the historical prediction error.

[0163] Furthermore, the prediction result processing module also includes:

[0164] Dynamic adjustment unit, which dynamically adjusts the layer depth according to the validation set, including

[0165] In multi-level decision making, if the accuracy improvement of the validation set for four consecutive iterations is less than 0.5%, the iteration is stopped. Otherwise, the iteration is continued until the requirements are met, and the final classifier weight and fusion fault diagnosis probability are output.

[0166] The processing unit is used to perform multi-level processing on the prediction results, including:

[0167] When the probability after calibration is in a high confidence interval (≥90%): it is automatically determined to be a fault and an early warning is initiated;

[0168] When the probability after calibration is in the low confidence interval (≤15%): mark it as environmental noise interference and ignore it;

[0169] When the probability after calibration is in the middle confidence interval (15% to 90%): push it to the joint diagnosis interface, and conduct joint analysis and judgment based on auxiliary information such as the oil-gas ratio and load-temperature coupling characteristics.

[0170] The gas-in-oil ratio is: If this value is greater than 0.1, a fault occurs;

[0171] Load-temperature coupling characteristics: If this value is greater than 1, a fault occurs; where:

[0172] I load : The effective value of the load current is taken from the existing current sensor data;

[0173] T hotspot : Winding hot spot temperature, taken from the temperature sensor;

[0174] Cov(): covariance between load current and hotspot temperature, which measures the linear correlation between the two;

[0175] σ I : Standard deviation of load current, representing the load fluctuation amplitude;

[0176] σ T : Standard deviation of hotspot temperature, representing the temperature fluctuation amplitude

[0177] It should be noted that the steps in the method provided by the present invention can be implemented by using the corresponding components in the system. Those skilled in the art can refer to the technical solution of the system to implement the step flow of the method, and can also refer to the technical solution of the method to implement the composition of the system. That is, the embodiments in the system and the embodiments in the method can be understood as preferred examples of each other, and will not be elaborated here.

[0178] The technical solution provided by the above embodiment of the present invention is further described in detail below with reference to a specific application example.

[0179] This application example uses a converter transformer at a ±800kV converter station as an example. Real-time monitoring parameters include: winding hotspot temperature, acoustic fundamental frequency amplitude, load current, and oil dissolved gas concentration. Fuzzy feature mapping uses a piecewise linear membership function for temperature, a Gaussian membership function for acoustic fundamental frequency, an S-shaped membership function for current, and a trapezoidal membership function for gas concentration. Typical converter transformer faults, including loose core, loose winding, DC bias, and circuit faults, are diagnosed.

[0180] like Figure 4 As shown, the multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method involved in this specific application example includes the following steps:

[0181] Step S1: Construct a multi-parameter fuzzy feature space for the temperature, soundprint, current, and dissolved gas in oil parameter data collected by the temperature sensor, ultrasonic sensor, current sensor, and MEMS photoacoustic sensor, and generate a multi-parameter fuzzy time series feature vector by sliding time series segmentation statistical features.

[0182] Step S2: Multi-parameter correlation features are formed by splicing and fusing the multi-parameter internal cross features and the external cross features.

[0183] Step S3: Construct a dynamic weighted multi-level aggregate classifier. Through the multi-level structure of random decision ensemble, dynamically assign weights to each classifier based on the performance of the validation set, fuse the probabilities of each layer to output the fault type, and introduce calibration parameters to adjust the matching degree between the model prediction probability and the actual fault probability. Figure 5 shown.

[0184] Step S4: Dynamically adjust the layer depth according to the accuracy improvement of the validation set. If the fault diagnosis accuracy of four consecutive layers is improved by less than 0.5%, the expansion is terminated and multi-level processing is performed on the calibrated probability prediction results.

[0185] In step S1, a multi-parameter fuzzy feature space is constructed.

[0186] First, define the piecewise linear membership function for multiple parameters:

[0187] Define a piecewise linear membership function to map temperature values ​​to three fuzzy sets: "low temperature", "normal", and "high temperature":

[0188]

[0189] Use the Gaussian membership function to map the voiceprint fundamental frequency amplitude A to the "normal" and "abnormal" fuzzy sets:

[0190]

[0191] Where A is the mean amplitude of the fundamental frequency under normal working conditions in the training set; σ A : Standard deviation of the fundamental frequency amplitude under normal working conditions in the training set

[0192] Define the S-type membership function to map the current value to the "normal" and "overcurrent" fuzzy sets:

[0193]

[0194] I rated is the rated current of the equipment; k = 0.1 is the slope factor, which controls the steepness of the transition interval; I th =1.2*I rated Overcurrent alarm threshold

[0195] Define a trapezoidal membership function to map gas concentrations to "normal" and "warning" fuzzy sets:

[0196]

[0197] G norm is the normal concentration threshold, G alarm is the warning threshold;

[0198] The four types of fuzzified sensor data are constructed into a unified feature sequence: F fused

[0199] In step S1, a multi-parameter fuzzy time series feature vector is generated by sliding the time series segmentation statistical features.

[0200] The sliding time series segmentation process is as follows: set the window length to 10 seconds and the step length to 1 second, then segment the fuzzy feature sequence to generate multiple window sets:

[0201] Temperature, voiceprint, current, and gas concentration window features generate 3D, 2D, 2D, and 2D features respectively; the total window feature dimensions are temperature, voiceprint, current, and gas window feature dimensions and 9 dimensions;

[0202] In step S2, inner cross features are proposed.

[0203] Inner crossover features:

[0204] First, the membership product of each two states in the same window of the same sensor is calculated;

[0205] Secondly, the covariance matrix of the membership of each state in the window is calculated and the off-diagonal elements are extracted;

[0206] Finally, the product combination is concatenated with the covariance feature to obtain the final inner cross feature:

[0207] Among them, the number of temperature membership product combination features is 3, and the covariance matrix feature of membership is 3; the number of voiceprint membership product combination features is 1, and the covariance matrix feature of membership is 1; the number of current membership product combination features is 1, and the covariance matrix feature of membership is 1; the number of gas concentration membership product combination features is 1, and the covariance matrix feature of membership is 1; the product combination features and the covariance features are spliced ​​together to obtain the final internal cross features, and the total internal cross features are 12.

[0208] In step S2, outer cross features are extracted.

[0209] Outer Cross Features:

[0210] First, the nonlinear correlation strength of the membership sequences of the two sensors is calculated;

[0211] The kernel density estimation method is used to estimate the joint distribution of the i-th state membership of sensor A and the j-th state membership of sensor B in the sliding window, and the marginal probability density is obtained:

[0212]

[0213] in,

[0214] Thus, the nonlinear correlation strength of the membership sequence is obtained:

[0215]

[0216] Next, we calculate the cross-sensor product combination:

[0217] (1) Select M from sensor A A Select M from sensor B B A key state. Form M A ×M B Key state pairs;

[0218] (2) For each key state pair, calculate the product of the real-time membership;

[0219] Finally, the product of the nonlinear correlation strength and all key pairs is combined to form the final outer cross feature. The number of outer cross features for temperature is 3, the number of outer cross features for voiceprint is 1, the number of outer cross features for current is 1, and the number of outer cross features for gas concentration is 1, resulting in a total of 6 outer cross features and inner cross features.

[0220] Dynamically weighted multi-level aggregate classifier in step S3.

[0221] First, a validation set is constructed. The validation set is independent of the training set and contains 7 consecutive days of time samples in the historical monitoring data, covering all fault types.

[0222] Secondly, we use logistic regression, support vector machine, gradient boosting tree, and deep neural network as classifiers for fault classification based on the internal and external cross-features of temperature, voiceprint, current, and gas concentration, and output the classification accuracy of each classifier.

[0223] Then, dynamic weight allocation is performed based on the accuracy of each classifier:

[0224]

[0225] is the initial weight of the classifier, is the initial validation set accuracy of the classifier; α: exponential amplification factor, used to enhance the weight ratio of the high-precision classifier; and the final fusion fault diagnosis probability is calculated through the weight.

[0226] Calibration parameters in step S3.

[0227] Based on the statistical deviation between the model prediction probability and the actual fault probability, a dynamic mapping function is established to correct the output probability distribution of the classifier in real time through the exponential smoothing algorithm. The correction formula is:

[0228]

[0229] The original output probability of the classifier at the current moment;

[0230] The value of α is between [0.2, 0.5], which is a smoothing coefficient and is adaptively adjusted according to the historical prediction error.

[0231] In step S4, the layer depth is dynamically adjusted according to the validation set.

[0232] In multi-level decision making, if the accuracy improvement of the validation set for four consecutive iterations is less than 0.5%, the iteration is stopped. Otherwise, the iteration is continued until the requirements are met, and the final classifier weight and fusion fault diagnosis probability are output.

[0233] In step S4, multi-level processing is performed on the prediction results.

[0234] When the corrected probability is in a high confidence interval (≥90%): it is automatically determined to be a confirmed fault and an early warning is initiated;

[0235] When the corrected probability is in the low confidence interval (≤15%): mark it as environmental noise interference and ignore it;

[0236] When the corrected probability is in the middle confidence interval (15% to 90%): the data is pushed to the joint diagnosis interface and jointly analyzed and judged in combination with auxiliary information such as the oil-gas ratio and load-temperature coupling characteristics.

[0237] The gas-in-oil ratio is: If this value is greater than 0.1, a fault occurs;

[0238] Load-temperature coupling characteristics: If this value is greater than 1, a fault occurs; where:

[0239] I load : The effective value of the load current is taken from the existing current sensor data;

[0240] T hotspot : Winding hot spot temperature, taken from the temperature sensor;

[0241] Cov(): covariance between load current and hotspot temperature, which measures the linear correlation between the two;

[0242] σ I : Standard deviation of load current, representing the load fluctuation amplitude;

[0243] σ T : Standard deviation of hotspot temperature, representing the temperature fluctuation amplitude

[0244] The verification results of the above specific application examples show that the diagnostic accuracy is improved: the traditional method has an accuracy rate of 72% and a false alarm rate of 28%; the method of the present invention has an accuracy rate of 96.48% and a false alarm rate of 3.52%. Dynamic adaptability verification: Simulating the equipment aging scenario (after 3 years of operation), the calibration parameters automatically adjust the weights, and the diagnostic accuracy only decreases by 2%, while the traditional method decreases by 15%. Disposal efficiency optimization: The automatic warning response time for high confidence intervals (≥90%) is <1 second; the joint analysis and judgment of medium confidence intervals takes an average of 0.5 minutes. Specifically, the accuracy comparison results of the fault diagnosis technology provided by the above embodiment of the present invention and other existing models are as follows: Figure 7 As shown; the fault diagnosis technology provided by the above embodiment of the present invention is compared with the confusion matrix of other existing models of loose core, loose winding, DC bias, and loop fault, as shown Figure 6 shown.

[0245] An embodiment of the present invention further provides a computer terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the processor can be used to execute any one of the methods described in the foregoing embodiments of the present invention, or to execute any one of the systems described in the foregoing embodiments of the present invention.

[0246] Optionally, the memory is used to store programs; the memory may include volatile memory (English: volatile memory), such as random-access memory (English: random-access memory, abbreviated: RAM), such as static random-access memory (English: static random-access memory, abbreviated: SRAM), double data rate synchronous dynamic random access memory (English: Double Data Rate Synchronous Dynamic Random Access Memory, abbreviated: DDR SDRAM), etc.; the memory may also include non-volatile memory (English: non-volatile memory), such as flash memory (English: flash memory). The memory is used to store computer programs (such as applications, functional modules, etc. that implement the above-mentioned methods), computer instructions, etc., and the above-mentioned computer programs, computer instructions, etc. can be partitioned and stored in one or more memories. In addition, the above-mentioned computer programs, computer instructions, data, etc. can be called by the processor.

[0247] The processor is configured to execute the computer program stored in the memory to implement the various steps of the method or various modules of the system involved in the above embodiments. For details, please refer to the relevant descriptions in the above method and system embodiments.

[0248] The processor and memory can be independent structures or integrated structures. When the processor and memory are independent structures, the memory and processor can be coupled via a bus.

[0249] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it can be used to execute any method of the above embodiments of the present invention, or to run any system of the above embodiments of the present invention.

[0250] Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transfer of computer programs from one location to another. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer. An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and storage medium can be located in an ASIC. Alternatively, the ASIC can be located in a user device. Of course, the processor and storage medium can also exist as discrete components in a communication device.

[0251] The multi-parameter adaptive dynamic multi-level decision-making converter transformer fault diagnosis method and system provided by the above-mentioned embodiments of the present invention constructs a multi-parameter fuzzy feature space based on temperature, voiceprint, current, and oil-soluble gas parameter data collected by temperature sensors, ultrasonic sensors, current sensors, and MEMS photoacoustic sensors. This space is then used to generate a multi-parameter fuzzy time series feature vector by segmenting and statistically analyzing the sliding time series. Multi-parameter internal and external cross-features are then combined to form multi-parameter correlation features. A dynamically weighted multi-level aggregate classifier is constructed. Using a multi-level structure of randomized decision ensembles, weights are dynamically assigned based on validation set performance. The probabilities of each layer are integrated to output the fault type, and calibration parameters are introduced to adjust the match between the model's predicted probability and the actual fault probability. The layer depth is dynamically adjusted based on the validation set. If the improvement in fault diagnosis accuracy for four consecutive layers is less than 0.5%, the expansion is terminated, and multi-level processing is performed on the prediction results. This method effectively integrates information from various parameters, significantly improving the accuracy and reliability of predictions.

[0252] Matters not mentioned in the above embodiments of the present invention are well known in the art.

[0253] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art may make various variations or modifications within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A multi-parameter adaptive dynamic multi-stage decision-making method for fault diagnosis of converter transformer, characterized in that: include: Based on the four types of parameter data obtained by sensors, namely temperature, soundprint, current and dissolved gas in oil, a multi-parameter fuzzy feature space is constructed. The fuzzy features are statistically analyzed through sliding time series segmentation to generate a multi-parameter fuzzy time series feature vector. Based on the multi-parameter fuzzy time series feature vector, multi-parameter internal cross features and external cross features are respectively obtained, and multi-parameter correlation characteristics for characterizing the dynamic coupling relationship of multi-source data are formed by splicing and fusing the internal cross features and the external cross features; Based on the multi-parameter correlation characteristics, a multi-level aggregate classifier model is constructed, and weights are dynamically assigned to each classifier based on the performance of the validation set. The predicted probabilities of each layer are fused to output the fault type, and calibration parameters are introduced to adjust the matching degree between the model predicted probability and the actual fault probability to obtain the calibrated fault probability; According to the matching adjustment of the dynamic weighted multi-level aggregation classifier on the validation set, the hierarchical depth is dynamically adjusted. If the improvement of the fault diagnosis accuracy of multiple consecutive layers is less than the set threshold a, the expansion is terminated and the current hierarchical structure is locked. At the same time, multi-level processing is performed on the prediction results based on the calibrated fault probability.

2. The multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method according to claim 1 is characterized in that: The construction of a multi-parameter fuzzy feature space includes: A piecewise linear membership function is defined for each parameter data, mapping the temperature value to three fuzzy sets: low temperature, normal temperature, and high temperature: Where μ 低温 (T) represents the membership function of the temperature belonging to the low temperature fuzzy set; μ 高温 (T) represents the membership function of the temperature belonging to the high temperature fuzzy set; T represents the real-time monitored temperature sensor value, which is the input variable of the membership function; T min represents the lowest temperature threshold of the low temperature fuzzy set; T normlow Indicates the lower limit of the normal temperature range; T normhigh Indicates the upper limit of the normal temperature range; T max represents the highest temperature threshold of the high temperature fuzzy set; Use the Gaussian membership function to map the voiceprint fundamental frequency amplitude A to normal and abnormal fuzzy sets: m 正常 (A)=1-μ 异常 (A) Where μ 异常 (A) represents the mapping of the fundamental frequency amplitude of the voiceprint A to the abnormal fuzzy set; A represents the mean fundamental frequency amplitude of the normal working condition in the training set; μ A Indicates the central value of the abnormal voiceprint fundamental frequency amplitude; σ A Indicates the standard deviation of the fundamental frequency amplitude of abnormal conditions in the training set; Define the S-type membership function to map the current value to the normal and overcurrent fuzzy sets: m 正常 (I)=1-μ 过流 (I) Where μ 过流 (I) represents the mapping of current value to overcurrent fuzzy set; I represents the current value detected in real time; I rated Indicates the rated current of the equipment; k represents the slope factor, which is used to control the steepness of the transition interval; I th Indicates the overcurrent alarm threshold; Define a trapezoidal membership function to map gas concentrations to normal and warning fuzzy sets: Where μ 预警 (G) represents mapping gas concentration to warning fuzzy set; G represents collected gas concentration; G norm Indicates the normal concentration threshold; G alarm Indicates the warning threshold; The four types of parameter data after fuzzification are constructed into a unified fuzzy feature sequence F fused .

3. The multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method according to claim 1 is characterized in that: The method of performing statistics on fuzzy features through sliding time series segmentation to generate a multi-parameter fuzzy time series feature vector for characterizing the dynamic coupling relationship of multi-source data includes: Set the window length to m seconds and the step length to n seconds, segment the fuzzy feature sequence, and generate multiple window feature sets, that is, obtain the multi-parameter fuzzy time series feature vector; where: According to the dimensions of the generated window features of temperature, voiceprint, current, and gas concentration, the total window feature dimension is obtained as the sum of the dimensions of the window features of temperature, voiceprint, current, and gas concentration.

4. The multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method according to claim 1 is characterized in that: Obtaining the inner intersection feature includes: Calculate the membership product of the two states in the same window of the same sensor to obtain the product combination; Calculate the covariance matrix of each state membership in the window, extract the off-diagonal elements, and obtain the covariance features; The product combination is concatenated with the covariance feature to obtain the final inner cross feature.

5. The multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method according to claim 1 is characterized in that: Obtaining the outer intersection feature includes: The kernel density estimation method is used to estimate the state membership of the i-th sensor A in the sliding window. Membership degree with the jth state of sensor B The joint distribution of is estimated, and the marginal probability density is obtained as: Where, represents the marginal probability density; W represents the weight vector; h represents the bandwidth parameter, which is used to control the smoothness of the kernel density estimation; K represents the kernel function, which is used to estimate the joint distribution of the membership of sensors A and B, and t represents the current time; Based on the marginal probability density, the nonlinear correlation strength of the membership sequence is obtained as: Where M(A,B) represents the nonlinear correlation strength; K A represents the kernel function of sensor A; K B represents the kernel function of sensor B; represents the marginal probability density of the i-th state of sensor A; represents the marginal probability density of the j-th state of sensor B; Select M from sensor A A Key state membership, select M from sensor B B The key state membership is calculated, and the product of the real-time membership is formed to form M A ×M B Key state pairs; The product of the nonlinear correlation strength and all key state pairs is combined and concatenated into the final outer cross feature.

6. The multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method according to claim 1, characterized in that: The multi-level aggregate classifier model is constructed, and weights are dynamically assigned to each classifier based on the performance of the validation set. The predicted probabilities of each layer are integrated to output the fault type, including: Constructing a validation set, which includes time samples of multiple consecutive days in historical monitoring data and covers all fault types; Based on the internal and external cross-features of temperature, voiceprint, current, and gas concentration, multiple classifiers are used to classify faults and the classification accuracy of each classifier is output. Using the validation set, dynamic weight allocation is performed according to the classification accuracy of each classifier, then: Where, represents the initial weight of classifier i; represents the initial validation set accuracy of classifier i; represents the initial accuracy of classifier j on the validation set, represents the sum of the initial accuracies of the four classifiers; α represents the exponential amplification factor; The fault diagnosis probability predicted by each classifier is fused by weight to obtain the model prediction probability and output the final fault type.

7. The multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method according to claim 1, characterized in that: The introduction of calibration parameters to adjust the matching degree between the model prediction probability and the actual fault occurrence probability includes: Based on the statistical deviation between the model prediction probability and the actual fault probability, a dynamic mapping function is established. The output probability distribution of the classifier prediction is corrected in real time through the exponential smoothing algorithm to obtain a correction model, which adjusts the matching degree between the model prediction probability and the actual fault probability. The correction model is expressed as: Where, Represents the probability of the classifier after calibration at the current moment; Represents the probability after classifier calibration at the previous moment; represents the original output probability of the classifier at the current moment; β represents the smoothing coefficient as a calibration parameter, which is adaptively adjusted according to the historical prediction error.

8. The multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method according to claim 1 is characterized in that: The dynamically adjusting the level depth according to the matching adjustment of the dynamic weighted multi-level aggregation classifier on the validation set includes: In the multi-level structure of the random decision ensemble of the multi-level aggregate classifier model, if the accuracy improvement of the validation set in consecutive multi-layer iterations is less than the set threshold a, the iteration is stopped; otherwise, the iteration is continued until the requirements are met, and the final classifier weight and fusion fault diagnosis probability are output.

9. The multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis method according to claim 1, characterized in that: The performing multi-level processing on the prediction result based on the calibrated failure probability includes: When the calibrated failure probability is within a high confidence interval, it is automatically determined to be a confirmed failure and an early warning is initiated; When the calibrated failure probability is in the low confidence interval, it is marked as environmental noise interference and ignored; When the calibrated fault probability is within the medium confidence interval, it is pushed to the joint diagnosis interface, where it is combined with the oil-gas ratio and load-temperature coupling characteristics for joint analysis and judgment; The gas-in-oil ratio is: If this value is greater than the set threshold b, a fault occurs; The load-temperature coupling characteristics are: If the feature is greater than the set threshold c, it is determined to be a fault condition; where: I load Indicates the effective value of the commutation load current, which is obtained from the existing current sensor data; T hotspot Indicates the hot spot temperature of the winding, taken from the temperature sensor; Cov() represents the covariance between load current and hotspot temperature, which measures the linear correlation between the two; σ I Indicates the standard deviation of the load current and characterizes the load fluctuation amplitude; σ T It represents the standard deviation of hot spot temperature and characterizes the amplitude of temperature fluctuation.

10. A multi-parameter adaptive dynamic multi-stage decision-making converter transformer fault diagnosis system, characterized in that: include: The fuzzy feature acquisition module constructs a multi-parameter fuzzy feature space based on four types of parameter data acquired by sensors: temperature, voiceprint, current, and dissolved gas in oil. It then performs statistics on the fuzzy features through sliding time series segmentation to generate a multi-parameter fuzzy time series feature vector. A correlation feature acquisition module, which obtains multi-parameter internal cross features and external cross features based on the multi-parameter fuzzy time series feature vector, and forms a multi-parameter correlation characteristic for characterizing the dynamic coupling relationship of multi-source data by splicing and fusing the internal cross features and the external cross features; A fault probability prediction module, which builds a multi-level aggregate classifier model based on the multi-parameter correlation characteristics, dynamically assigns weights to each classifier based on the validation set performance, fuses the predicted probabilities of each layer to output the fault type, and introduces calibration parameters to adjust the matching degree between the model's predicted probability and the actual fault probability, thereby obtaining a calibrated fault probability; A prediction result processing module dynamically adjusts the hierarchical depth according to the matching adjustment of the dynamically weighted multi-level aggregation classifier on the validation set. If the improvement in the fault diagnosis accuracy of multiple consecutive layers is less than the set threshold a, the expansion is terminated and the current hierarchical structure is locked. At the same time, multi-level processing is performed on the prediction results based on the calibrated fault probability.

11. A computer terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When executing the computer program, the processor can be used to perform the method according to any one of claims 1 to 9, or run the system according to claim 10.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it can be used to perform the method according to any one of claims 1 to 9, or to run the system according to claim 10.

Images