Intelligent diagnosis method for galloping and disconnection of overhead line

Abstract

The invention relates to the technical field of overhead line fault detection, in particular to an intelligent diagnosis method for galloping and disconnection of an overhead line, and the method comprises the steps: collecting acceleration data and attitude angle data of the overhead line in different operation states through an attitude sensor, forming a pre-labeled training sample set, and calculating an acceleration change rate and an angle change amount; establishing a time sequence feature set constructed based on a time window, and calculating extension features in the time window; applying a feature weight matrix; adopting a radial basis kernel function to calculate a sample minimum distance to obtain an adaptive kernel width parameter; setting a class imbalance compensation factor according to the sample quantity proportion; a weighted feature vector, a self-adaptive kernel width parameter and the class imbalance compensation factor obtained in the fourth step serve as input, a radial basis kernel function is adopted to construct a multistage support vector machine based on a binary tree structure, and the state of the overhead line is diagnosed; and establishing a confidence backtracking mechanism, and automatically checking a low-confidence classification result.

Description

Technical Field

[0001] This application relates to the field of fault detection technology for overhead power distribution lines, specifically an intelligent diagnostic method for overhead power line galloping and breakage. Background Technology

[0002] Overhead lines, as a crucial component of power transmission, are vital to the entire power grid's safe and stable operation. Galloping and line breaks are two major issues affecting the safe operation of overhead lines. Galloping can lead to faults such as phase-to-phase flashover and hardware damage, while line breaks can cause serious consequences such as line tripping, power outages, and tower collapses. Timely detection of galloping and line breaks in overhead lines allows for rapid implementation of appropriate maintenance measures, reducing outage time and frequency, and improving the power grid's reliability and operational efficiency.

[0003] For example, the Chinese authorized patent application CN114485897B, entitled "A Method and Device for Determining the Galloping State of an Overhead Line," involves placing sensors on the overhead line and collecting wavelength data output by the sensors according to a preset acquisition cycle. Based on the wavelength data collected in each acquisition cycle, the system can determine the position data of the fiber optic gratings that sequentially contact the weighted ball when the overhead line gallops and causes the weighted ball to roll. Based on these position data, the galloping state of the overhead line can be determined, thereby enabling the monitoring of the galloping state of the overhead line and providing effective data reference for prevention and warning.

[0004] While the aforementioned existing technologies have made some progress, they rely on single-dimensional monitoring data acquired by sensors for judgment, which often results in insufficient feature information, high sensitivity to external interference, and weak state differentiation capabilities. This leads to misjudgment or missed judgment between the dancing state and the disconnected state, making it difficult to meet the requirements for diagnostic accuracy and reliability in real-time monitoring scenarios.

[0005] In view of the above, in order to overcome the aforementioned technical problems, this invention designs an intelligent diagnostic method for overhead line galloping and line breakage, thus solving the above-mentioned technical problems. Summary of the Invention

[0006] To address the aforementioned issues, this application proposes an intelligent diagnostic method for overhead line galloping and line breakage. By constructing a hierarchical multi-classification model, it achieves accurate differentiation between normal, mild galloping, severe galloping, and line breakage states of overhead lines, thereby improving the stability and accuracy of line breakage diagnosis.

[0007] An intelligent diagnostic method for overhead power line galloping and breakage includes the following steps:

[0008] Step 1: Collect acceleration and attitude angle data of the overhead line under different operating conditions by using attitude sensors installed on the line to form a pre-labeled training sample set including normal state, galloping state and disconnection state, and calculate the rate of change of acceleration and the amount of change of angle.

[0009] Step 2: By extracting the rate of change of acceleration and the amount of change of angle, a time-series feature set based on a time window is established, and the mean, variance, energy and peak factor within the time window are calculated as extended features, which simultaneously constitute the feature vectors of the training samples and the feature vectors of the samples to be diagnosed.

[0010] Step 3: Apply feature weight matrices to the feature vectors of the training samples and the samples to be diagnosed to obtain the weights of each feature in different dimensions; based on the weighted sum of the feature vectors of the training samples and the feature vectors of the samples to be diagnosed, use the radial basis kernel function to calculate the minimum distance between the samples and obtain the adaptive kernel width parameter.

[0011] Step 4: Based on the pre-labeled training sample set obtained in Step 1, count the number of samples in each category and set a class imbalance compensation factor according to the sample quantity ratio.

[0012] Step 5: Using the weighted feature vector and adaptive kernel width parameter obtained in Step 3, and the class imbalance compensation factor obtained in Step 4 as input, construct a multi-level support vector machine based on a binary tree structure using the radial basis kernel function. The first layer of the support vector machine is used to distinguish between normal and abnormal states; the second layer of the support vector machine is used to distinguish between dancing state and broken line state; and the third layer of the support vector machine is used to distinguish between mild dancing state and severe dancing state.

[0013] Step 6: Establish a confidence backtracking mechanism to automatically verify low-confidence classification results.

[0014] Preferably, the attitude angle data includes the combination of roll angle, pitch angle, and yaw angle, based on the roll angle. Pitch angle Heading angle Build by The coordinate transformation matrix formed by, through The acceleration in its own coordinate system is transformed to the geographic coordinate system for gravity compensation, whereby... This represents the acceleration vector of the carrier measured by the attitude sensor; G=[0,0,g]T.

[0015] Preferably, the rate of change of acceleration is calculated based on the acceleration at adjacent time points in a geographic coordinate system, using the following formula: ,in, This represents the acceleration in the geographic coordinate system at the previous sampling time. It represents the magnitude of the vector, that is, the magnitude of the change in acceleration between adjacent time points;

[0016] The angle change is calculated based on the time difference between adjacent attitude angles. , and .

[0017] Preferably, the time-series feature set ,in Indicates from Time's up The dynamic acceleration sequence in geographic coordinate system within a continuous time window of time; This represents the roll angle sequence within the same time window.

[0018] Preferably, the feature weight matrix ;in, Represents the original feature vector; This represents a diagonal weight matrix, where only the diagonal lines have values. arrive These are the weights corresponding to each feature, with off-diagonal lines being 0. This represents the weighted feature vector, where different features are multiplied by their corresponding weights to adjust their importance in classification.

[0019] And by minimizing Obtain the weight parameters, where, , , , These are the variances of the rate of change of acceleration, the change of roll angle, the change of pitch angle, and the rate of change of yaw angle, respectively. arrive , representing the weighting coefficients used to balance the optimization priorities of different features; the optimal weights are determined by minimizing the combination of variances of these features. arrive .

[0020] Preferably, the radial basis function maps the original features to a higher-dimensional space, allowing the originally non-linear features to be linearly segmented in the higher-dimensional space, thereby enabling the support vector machine to distinguish different states, as shown in the following formula: ;in, Indicates sample and The kernel function value between the two samples is used to measure their similarity. Indicates sample and eigenvector distance; This represents the kernel width parameter.

[0021] Preferably, the formula for calculating the adaptive kernel width is: ; ;in, express The adaptive kernel width at any given time; This is the adjustment coefficient; Indicates the current sample The minimum distance to all samples in the training set;

[0022] when Hours indicate violent dancing or a broken wire. Increasing the kernel size widens the effective range of the kernel function, improving the sensitivity of anomaly detection; when When the value is large, it indicates that the line is operating normally. Making the kernel smaller narrows the effective range of the kernel function, filtering out noise from small perturbations.

[0023] Preferably, the class imbalance compensation factor includes: first, counting the number of samples in the training set, and then denoting the majority class sample number, i.e., the total number of samples in the normal or dancing state. The minority class sample size, i.e., the total number of samples in the disconnected state, is denoted as... Set according to the ratio of the number of the two types of samples Formula reference: ;

[0024] An enhanced penalty factor is applied to minority class samples: ,in The basic penalty factor for support vector machines: A penalty factor is set for the majority class samples. ;Will , When applied to the training process of a support vector machine, the model imposes a heavier penalty on the classification error of broken line samples.

[0025] Preferably, the training samples for the first layer of the support vector machine are normal samples and all abnormal samples, as shown in the following formula: ;

[0026] The training samples for the second-layer support vector machine are dancing samples and broken line samples, as shown in the following formula: ;

[0027] The training samples for the third-layer support vector machine are samples of slight dancing and samples of violent dancing, as shown in the following formula: ;

[0028] in, , and This represents the output function of the corresponding three-layer support vector machine classifier; , , This represents the weight vector of each layer of the support vector machine; , , This represents the bias term of each layer of the support vector machine, used to adjust the position of the classification boundary; each layer , The training samples at the corresponding level are combined with the previously optimized weights and class imbalance compensation to obtain the training. This represents the sign function, which outputs +1 for positive input and -1 for negative input, and is used to output classification results. This indicates that multiple conditions are met simultaneously.

[0029] Preferably, the confidence backtracking mechanism outputs the line status based on the decision results of each layer, and performs confidence detection based on the classification output distance. The confidence detection formula is as follows: ;like If the confidence level is low, the system will automatically backtrack to the previous classifier and recalculate its confidence level. If the confidence level is high, the classification will be adjusted based on the results of the previous classifier. If the confidence level is low in the third layer, the system will return to the second layer to confirm that the condition is agitated and then re-determine whether it is mild or severe based on historical data. If the confidence level is low for L consecutive times, the system will initiate a reclassification process.

[0030] The beneficial effects of this invention are as follows:

[0031] 1. This invention constructs a multi-dimensional feature system including acceleration change rate, attitude angle change, and statistics, and uses a feature weight optimization mechanism to enhance the influence of key features in classification decisions. This allows the difference between galloping and line breakage to be reflected in a clear feature pattern. Through the hierarchical discriminative structure of a multi-layer support vector machine, different states are modeled separately in independent classifiers, avoiding discrimination interference caused by mixed training of multiple states. This makes it easier to distinguish between normal, galloping, mild galloping, severe galloping, and line breakage states of the line.

[0032] 2. This invention enhances the influence of minority class samples during training through a class imbalance compensation strategy, avoiding bias caused by the model favoring the majority class. This allows for the formation of clear discrimination boundaries even for rare states such as line breaks. Simultaneously, through an adaptive kernel width adjustment mechanism, the radial basis kernel function can dynamically change according to the distribution density of input features, thereby maintaining a relatively stable classification ability when facing line movements of different amplitudes, frequencies, or changing trends, and improving the adaptability of the diagnostic model in complex environments.

[0033] 3. This invention continuously evaluates the stability of the output results by setting up a confidence detection and backtracking correction mechanism based on classification distance. When uncertainty occurs in the classification result, the previous classifier is automatically triggered to re-discriminate, enabling the model to avoid misjudgments caused by critical states or sudden interference. This mechanism gives the diagnostic results self-correction capabilities, meeting the requirements for stable output in the real-time monitoring scenario of overhead lines. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a schematic diagram of the overhead line galloping and line breakage diagnosis framework of the present invention;

[0036] Figure 2 This is a flowchart of the SVM model training process of the present invention;

[0037] Figure 3 This is a flowchart of the confidence backtracking mechanism of the present invention;

[0038] Figure 4 This is a three-axis curve of acceleration before conversion during the dancing motion of the present invention;

[0039] Figure 5 This is a triaxial curve of acceleration after conversion during the dancing motion of the present invention;

[0040] Figure 6 This is a triaxial curve of acceleration before conversion during wire breakage according to the present invention;

[0041] Figure 7 This is a triaxial acceleration curve diagram after conversion during wire breakage according to the present invention;

[0042] Figure 8 This is an acceleration data diagram of an overhead line during galloping, as described in this invention.

[0043] Figure 9 This is an angle data diagram of the overhead line galloping during the present invention;

[0044] Figure 10 This is an acceleration data diagram of an overhead line breaking during the present invention.

[0045] Figure 11 This is an angular data diagram showing the occurrence of a broken overhead line according to the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0047] The application principle of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0048] like Figure 1-11 As shown, an intelligent diagnostic method for overhead power line galloping and breakage utilizes attitude sensors installed on the line to collect triaxial acceleration and attitude angle information of the line in its own coordinate system. Since the acceleration output by the sensor is based on a coordinate system according to its installation direction, and its coordinate axis direction is inconsistent with the actual geographic coordinate system of the line, the acceleration data in the sensor's own coordinate system is transformed to the local geographic coordinate system based on the combination relationship of roll, pitch, and yaw angles included in the attitude angles. This yields an acceleration signal that accurately reflects the direction and amplitude of the line's movement. During this transformation process, the gravitational acceleration component is effectively eliminated, ensuring that the obtained data represents the line's true dynamic acceleration.

[0049] Specifically, the geographic coordinate system represents the local coordinate system, adopting a northeast-oriented coordinate system model, with D as the geocenter, the Z-axis pointing to the sky, the X-axis pointing due east, and the Y-axis pointing due north. The coordinate system model of the attitude sensor itself is shown, with the alarm device placed horizontally as a reference, the Z-axis pointing upwards and perpendicular to the XOY plane. The subscript 'e' indicates the geographic coordinate system, and 'b' is the attitude sensor's own coordinate system. Let the geographic coordinate system of the device be OX. e Y e Z e The device's own coordinate system is OX. b Y b Z b .

[0050] Based on the rotation analysis of the three-dimensional coordinate system, the rotation of the three-dimensional coordinate system can be expressed as:

[0051] The standard orthogonal rotation matrix obtained by rotating about the X-axis by an angle r (roll angle) is:

[0052]

[0053] The standard orthogonal rotation matrix obtained by rotating around the Y-axis by an angle p (pitch angle) is:

[0054]

[0055] The orthogonal rotation matrix obtained by rotating about the Z-axis by an angle y (yaw heading angle) is:

[0056]

[0057] Since matrix multiplication is not commutative, let OX e Y e Z e To OX b Y b Z b The transformations are accomplished by rotating the body by angles r, p, and y around the three single axes Z, Y, and X, respectively. Therefore, the coordinate transformation matrix from Body to Earth is: ,in The coordinate transformation matrix from the body coordinate system to the earth coordinate system consists of three basic rotation matrices. The product is obtained by multiplying in order;

[0058] Since all objects on Earth experience gravity, and coordinate transformations primarily consider the object's real-time acceleration, gravitational acceleration needs to be removed during the transformation. Let the acceleration vector of the carrier measured by the attitude sensor be... Theoretically, the acceleration vector of the geographic coordinate system is The acceleration of the geographic coordinate system plus the effect of gravitational acceleration, multiplied by the rotation matrix from the geographic coordinate system to the vehicle coordinate system, yields the acceleration data of the vehicle after gravity compensation. The corresponding formula is as follows:

[0059]

[0060] Where G = [0,0,g]T, It has orthogonality. The orthogonal transformation matrix between the two coordinate systems, with parameters r, p, and y, can be obtained from the attitude sensor. In summary, the acceleration of the target object in the geographic coordinate system can be obtained as follows:

[0061]

[0062] Simultaneously, the acceleration sequences in the geographic coordinate system at different times were obtained:

[0063]

[0064] Obtained through gravity compensation combined with coordinate system transformation It is the acceleration of the pure dynamic motion of the route. When calculating the rate of change of acceleration and the change of attitude angle, it can avoid misjudging the sensor tilt as line galloping or line breakage.

[0065] After acquiring the acceleration signal in the geographic coordinate system, the rate of change of acceleration between consecutive sampling times is further calculated to reflect the sensitivity of the line to acceleration changes under wind load or disturbance. Simultaneously, based on the difference in attitude angle signals between adjacent sampling times, the change in the line's attitude angle is calculated to characterize the dynamic degree of spatial attitude deviation of the line. The aforementioned rate of change of acceleration and angle change constitute the multi-feature input data used for classification and recognition.

[0066] Specifically, the normal, galloping, or broken state of the railway line will be reflected in the intensity of its motion. If the line is normal, the changes in acceleration and attitude angle will be minimal; if the line gallops, especially violently, the acceleration will fluctuate rapidly and the attitude angle will deviate significantly; if the line is broken, the acceleration will change abruptly, such as with a sudden, large sway, and the attitude angle will also change drastically. By calculating the rate of change of acceleration and the change of attitude angle between adjacent time points, the intensity of the line's motion is quantified into specific numerical features. These numerical features are the basis for subsequent classification models to identify the state, including:

[0067] The formula for the rate of change of acceleration is as follows:

[0068]

[0069] in, This represents the acceleration in the geographic coordinate system at the previous sampling time. It represents the magnitude of the vector, that is, the magnitude of the change in acceleration between adjacent time points;

[0070] The formula for the change in angle is as follows:

[0071]

[0072]

[0073]

[0074] in, , , They represent The roll angle, pitch angle, and yaw angle at any given moment; , , This represents the attitude angle at the previous sampling time. , , This represents the rate of change of roll, pitch, and yaw angles between adjacent moments. The absolute value is used to show only the magnitude of change and not the direction.

[0075] Constructing extended features based on time windows: By extracting the rate of change of acceleration and the amount of change of angle, a connection is built between the original data and the state recognition, transforming the basic data into features that can reflect the state of the line.

[0076] To depict the continuity of the dance process, a time window is introduced. At every moment structure:

[0077]

[0078] express The feature set within the time window corresponding to the previous moment includes the previous moment. The roll angle at each moment, and the pitch and yaw angles at the current moment; This indicates the length of the time window, for example, taking 10 consecutive sampling moments as a window;

[0079] Calculate statistical characteristics within the window:

[0080] Mean: ,in Indicates the first time window Feature data at each moment, such as acceleration or angle;

[0081] variance: This indicates the degree of fluctuation in the data within the window;

[0082] energy: , indicating the cumulative intensity of the data within the window;

[0083] Peak factor: This represents the ratio of the largest data point within the window to the mean, reflecting the degree of abrupt change.

[0084] Line galloping is a continuous swaying process, while line breakage is an instantaneous abrupt change. Focusing only on the acceleration and angle changes at a single moment may miss the filtering of continuous fluctuations and is easily disturbed by occasional small disturbances. Introducing a time window can help mitigate this. , to make continuous Data from each time point is packaged, and the mean, variance, and kurtosis factor within this window are calculated. By combining the time window with statistical characteristics, the patterns of continuous processes are characterized. For example, during mild dancing, the variance, energy, and kurtosis factor are small; during severe dancing, the variance, energy, and kurtosis factor are large; and during discontinuity, the kurtosis factor shows a sudden surge. These statistical characteristics enable subsequent classification models to more accurately distinguish different states, avoiding misjudging a single disturbance as an anomaly or missing continuous anomalies.

[0085] Meanwhile, acceleration, attitude angle, time window, rate of change of acceleration, change of angle, and statistical features constitute the training feature vector and the feature vector to be diagnosed.

[0086] Constructing a feature weight optimization mechanism: Different features contribute differently to classification; this is addressed by introducing a feature weight matrix.

[0087]

[0088] in, This represents the original feature vector, such as the previously extracted rate of change of acceleration or change of angle. This represents a diagonal weight matrix, where only the diagonal lines have values. arrive These are the weights corresponding to each feature, with off-diagonal lines being 0. This represents the weighted feature vector, where different features are multiplied by their corresponding weights to adjust their importance in classification.

[0089] The weights are determined using the following optimization model:

[0090]

[0091] in, , , , These are the variances of the rate of change of acceleration, the change of roll angle, the change of pitch angle, and the rate of change of yaw angle, respectively. arrive , representing the weighting coefficients used to balance the optimization priorities of different features; the optimal weights are determined by minimizing the combination of variances of these features. arrive ;

[0092] By assigning different importance to different features, previously extracted features, such as the rate of change of acceleration and the amount of change of angle, have different effects on identifying broken lines and galloping movements. When a line is broken, the fluctuation of the rate of change of acceleration is very large; when galloping, the continuous fluctuation of the amount of change of angle is more obvious. Therefore, a feature weight matrix is ​​introduced. By optimizing the model by minimizing the combination of feature variances, the weight of each feature is automatically adjusted, giving greater weight to acceleration features that are more important for line breakage and greater weight to angle features that are more important for dancing, thus enabling the subsequent classification model to utilize features more accurately.

[0093] During the model training phase, the collected multi-dimensional feature samples are standardized to eliminate the influence of different feature dimensions. A Gaussian radial basis function kernel is used during support vector machine training, and a reasonable balance between discriminative and generalization abilities is ensured by setting kernel width parameters and penalty factors. The processed training samples are then input into the corresponding support vector machine model for learning, resulting in a classification model that reflects the boundaries of each state feature. After training, test samples are input into the model to verify its classification accuracy and apply it to real-time identification of line status during actual monitoring.

[0094] Specifically, a Gaussian radial basis function kernel is used to map the original features to a higher-dimensional space, allowing the originally non-linear features to be linearly segmented in the higher-dimensional space, thus enabling SVM to distinguish different states. The formula is as follows:

[0095]

[0096] in, Indicates sample and The kernel function value between the two samples is used to measure their similarity. Indicates sample and eigenvector distance; This parameter represents the kernel width and is used to control the scope of the kernel function. This kernel function is characterized by its simple form, few parameters, and high diagnostic accuracy.

[0097] Adaptive scaling of the kernel function: due to the Gaussian kernel's... It is a fixed value, but the feature density of samples in different states is different; samples that are violently dancing or broken have large feature fluctuations and are close to other samples; samples that are operating normally have stable features and are far from other samples. Therefore, the kernel width needs to be adaptively changed with the sample distance. The specific formula is:

[0098]

[0099]

[0100] in, express The adaptive kernel width at any given time; This is the adjustment coefficient (a hyperparameter set manually). Indicates the current sample The minimum distance to all samples in the training set. When Hours indicate violent dancing or a broken wire. Increasing the kernel size widens the effective range of the kernel function, improving its sensitivity in identifying anomalous samples; when When the value is large, it indicates that the line is operating normally. The smaller kernel size narrows the effective range of the kernel function, filtering out the noise effects of small perturbations;

[0101] This approach solves the problem of nonlinear separability of features, enabling SVM to handle complex features of line status. The adaptive kernel width allows the kernel function to adjust its sensitivity as needed, making it more sensitive to abnormal states and more resistant to noise in normal states. This improves the accuracy of anomaly identification and reduces the false alarm rate.

[0102] Constructing a class imbalance compensation factor: In reality, there are many normal and dancing line samples, but very few broken line samples. Directly training SVM will cause the model to be biased towards the majority class, that is, it is better at recognizing normal and dancing lines, but it is easy to misclassify broken lines. Therefore, by setting a larger penalty factor for the broken line class, the SVM will impose a heavier penalty on the classification error of fault class samples during training, forcing the model to pay more attention to the few broken line samples and improve the ability to recognize broken lines.

[0103] First, count the number of samples in the training set. The majority class sample size, i.e., the total number of samples in either the normal or dancing state, is denoted as [sample size]. The minority class sample size, i.e., the total number of samples in the disconnected state, is denoted as... The ratio of the two types of samples is set. Formula reference: ;

[0104] An enhanced penalty factor is applied to minority class samples: ,in The basic penalty factor for support vector machines: A penalty factor is set for the majority class samples. ;Will , By incorporating the training process of the support vector machine, the model imposes a heavier penalty on the classification error of the broken line sample, thereby optimizing and ensuring the recognition accuracy of the broken line sample.

[0105] Based on the above characteristics, a multi-level classification model using support vector machines is established. This model employs a binary tree-like hierarchical structure for state recognition: the first layer of support vector machines distinguishes between normal and abnormal states of the line; if determined to be in an abnormal state, it proceeds to the second layer, where the second support vector machine further distinguishes between a galloping state and a broken line state; if determined to be in a galloping state, it proceeds to the third layer, where the third support vector machine further classifies the galloping into mild galloping and severe galloping.

[0106] The training samples for the first layer of SVM are normal samples and all abnormal samples (dancing and broken lines), as shown in the following formula:

[0107]

[0108] The training samples for the second layer of SVM are dancing samples and broken line samples, as shown in the following formula:

[0109]

[0110] The training samples for the third layer of SVM are samples of slight dancing and samples of violent dancing, as shown in the following formula:

[0111]

[0112] Final classification results:

[0113]

[0114] in, , and This represents the output function of the corresponding three-layer SVM classifier; , , This represents the weight vector of each SVM layer, learned from the training samples, and represents the degree of influence of the features on the classification. , , This represents the bias term for each SVM layer, used to adjust the position of the classification boundary; each layer , It is obtained by training samples at the corresponding level, combined with previous feature optimization weights and class imbalance compensation. This represents the sign function, which outputs +1 for positive input and -1 for negative input, and is used to output classification results. This indicates that multiple conditions are met simultaneously;

[0115] Each layer of the support vector machine in the binary tree structure is trained independently based on the training samples corresponding to that layer. During training, a feature weight optimization mechanism and a class imbalance compensation strategy are introduced simultaneously. By assigning differentiated weights to different feature dimensions and setting an enhanced penalty factor for minority class samples, the hyperplane parameters of each layer of the support vector machine are optimized. With bias It can focus on key features while avoiding the model being dominated by majority class samples, thereby obtaining a more robust and discriminative classification boundary and achieving more accurate hierarchical identification of different line states.

[0116] By first quickly distinguishing between normal and abnormal, then further distinguishing between dancing and broken lines within the abnormal category, and finally refining the degree of dancing, the complexity of a single classification is reduced, the accuracy of identifying subdivided states is improved, and the confusion of distinguishing multiple categories at once is avoided.

[0117] A confidence backtracking mechanism is established to address the problem of misjudgment of boundary states. For example, when the line is at the boundary between mild and severe galloping, the classifier is prone to giving incorrect results. The formula is as follows:

[0118]

[0119] like If the confidence level is low, the system will automatically backtrack to the previous classifier and recalculate its confidence level. If the confidence level is high enough, the classification will be adjusted based on the results of the previous classifier. If the confidence level of the third layer is low, the system will return to the second layer to confirm that the condition is dancing, and then re-determine whether it is mild or severe based on historical data. If the confidence level is low for L consecutive times, the system will start reclassification.

[0120] in, Indicates the first The confidence score of the layer classifier, i.e., the distance function value; This represents the Lagrange multipliers and sample labels obtained through SVM training. ); Indicates sample (Training samples) and current samples The kernel function value; Indicates the first Bias terms of the layer classifier; The confidence threshold is set based on the average distance between the training samples and the classification boundary, such as taking half of the boundary distance; L represents the threshold for the number of consecutive low confidence levels.

[0121] By using confidence backtracking, low-confidence classification results can be automatically verified, avoiding misjudgments and making diagnostic results more stable and reliable.

Claims

1. An intelligent diagnostic method for overhead line galloping and breakage, characterized in that, Includes the following steps: Step 1: Collect acceleration and attitude angle data of the overhead line under different operating conditions by using attitude sensors installed on the line to form a pre-labeled training sample set including normal state, galloping state and disconnection state, and calculate the rate of change of acceleration and the amount of change of angle. Step 2: By extracting the rate of change of acceleration and the amount of change of angle, a time-series feature set based on a time window is established, and the mean, variance, energy and peak factor within the time window are calculated as extended features, which simultaneously constitute the feature vectors of the training samples and the feature vectors of the samples to be diagnosed. Step 3: Apply feature weight matrices to the feature vectors of both the training samples and the samples to be diagnosed, obtaining the weights for different dimensions of each feature; then, using a radial basis function, calculate the minimum distance between the samples based on the weighted sum of the feature vectors of the training samples and the feature vectors of the samples to be diagnosed. ,Will Substitute into the formula for calculating the adaptive kernel width: The adaptive kernel width parameter is obtained. in, express The adaptive kernel width at any given time; This is the adjustment coefficient; Indicates the current sample The minimum distance to all samples in the training set; Step 4: Based on the pre-labeled training sample set obtained in Step 1, count the number of samples in each class, set a class imbalance compensation factor according to the sample size ratio, and set an enhancement penalty factor for minority class samples. Set a penalty factor for the majority class samples. This penalizes the classification errors of faulty samples, making the model prioritize the few broken samples. Step 5: Using the weighted feature vector and adaptive kernel width parameter obtained in Step 3, and the class imbalance compensation factor obtained in Step 4 as input, construct a multi-level support vector machine based on a binary tree structure using the radial basis kernel function. The first layer of the support vector machine is used to distinguish between normal and abnormal states; the second layer of the support vector machine is used to distinguish between dancing state and broken line state; and the third layer of the support vector machine is used to distinguish between mild dancing state and severe dancing state. Step 6: Establish a confidence backtracking mechanism to automatically verify low-confidence classification results.

2. The intelligent diagnostic method for overhead line galloping and line breakage according to claim 1, characterized in that: The attitude angle data includes the combination of roll angle, pitch angle, and yaw angle, based on the roll angle... Pitch angle Heading angle Build by The coordinate transformation matrix formed by, through The acceleration in its own coordinate system is transformed to the geographic coordinate system for gravity compensation, whereby... This represents the acceleration vector of the carrier measured by the attitude sensor; G=[0,0,g]T.

3. The intelligent diagnostic method for overhead line galloping and line breakage according to claim 2, characterized in that: The rate of change of acceleration is calculated based on the acceleration at adjacent moments in the geographic coordinate system, using the following formula: ,in, This represents the acceleration in the geographic coordinate system at the previous sampling time. It represents the magnitude of the vector, that is, the magnitude of the change in acceleration between adjacent time points; The angle change is calculated based on the time difference between adjacent attitude angles. , and .

4. The intelligent diagnostic method for overhead line galloping and line breakage according to claim 1, characterized in that: The time-series feature set ,in Indicates from Time's up The dynamic acceleration sequence in geographic coordinate system within a continuous time window of time; This represents the roll angle sequence within the same time window.

5. The intelligent diagnostic method for overhead line galloping and line breakage according to claim 1, characterized in that: The feature weight matrix ;in, Represents the original feature vector; This represents a diagonal weight matrix, where only the diagonal lines have values. arrive These are the weights corresponding to each feature, with off-diagonal lines being 0. This represents the weighted feature vector, where different features are multiplied by their corresponding weights to adjust their importance in classification. And by minimizing Obtain the weight parameters, where, , , , These are the variances of the rate of change of acceleration, the change of roll angle, the change of pitch angle, and the rate of change of yaw angle, respectively. arrive , representing the weighting coefficients used to balance the optimization priorities of different features; the optimal weights are determined by minimizing the combination of variances of these features. arrive .

6. The intelligent diagnostic method for overhead line galloping and line breakage according to claim 1, characterized in that: The radial basis function maps the original features to a higher-dimensional space, allowing the originally non-linear features to be linearly segmented in the higher-dimensional space, thus enabling the support vector machine to distinguish different states. The formula is as follows: ;in, Indicates sample and The kernel function value between the two samples is used to measure their similarity. Indicates sample and eigenvector distance; This represents the kernel width parameter.

7. The intelligent diagnostic method for overhead line galloping and line breakage according to claim 6, characterized in that: In step three, when Hours indicate violent dancing or a broken wire. Increasing the kernel size widens the effective range of the kernel function, improving the sensitivity of anomaly detection; when When the value is large, it indicates that the line is operating normally. Making the kernel smaller narrows the effective range of the kernel function, filtering out noise from small perturbations.

8. The intelligent diagnostic method for overhead line galloping and line breakage according to claim 1, characterized in that: The class imbalance compensation factor includes: first, counting the number of samples in the training set, and then denoting the majority class sample number, i.e., the total number of samples in the normal or dancing state. The minority class sample size, i.e., the total number of samples in the disconnected state, is denoted as... Set according to the ratio of the number of the two types of samples Formula reference: ; Will , When applied to the training process of a support vector machine, the model imposes a heavier penalty on the classification error of broken line samples.

9. The intelligent diagnostic method for overhead line galloping and line breakage according to claim 1, characterized in that: The training samples for the first layer of the support vector machine are normal samples and all abnormal samples, as shown in the following formula: ; The training samples for the second-layer support vector machine are dancing samples and broken line samples, as shown in the following formula: ; The training samples for the third-layer support vector machine are samples of slight dancing and samples of violent dancing, as shown in the following formula: ; in, , and This represents the output function of the corresponding three-layer support vector machine classifier; , , This represents the weight vector of each layer of the support vector machine; , , This represents the bias term of each layer of the support vector machine, used to adjust the position of the classification boundary; each layer , The training samples at the corresponding level are combined with the previously optimized weights and class imbalance compensation to obtain the training. This represents the sign function, which outputs +1 for positive inputs and -1 for negative inputs, and is used to output classification results.

10. The intelligent diagnostic method for overhead line galloping and line breakage according to claim 1, characterized in that: The confidence backtracking mechanism outputs the line status based on the decision results of each layer, and performs confidence detection based on the classification output distance. The confidence detection formula is as follows: ;like If the confidence level is low, the system will automatically backtrack to the previous classifier and recalculate its confidence level. If the confidence level is high, the classification will be adjusted based on the results of the previous classifier. If the confidence level is low in the third layer, the system will return to the second layer to confirm that the condition is agitated and then re-determine whether it is mild or severe based on historical data. If the confidence level is low for L consecutive times, the system will initiate a reclassification process.

Images