A power transmission jumper wire state deviation measurement method based on multi-modal sensing

Abstract

The application provides a power transmission jumper state deviation measurement method based on multi-modal sensing, and relates to the technical field of power system monitoring. The method comprises the following steps: acquiring a multi-modal sensing data set of a power transmission jumper; integrating and generating a state feature matrix through a feature fusion processing unit; performing time sequence alignment based on a preset reference model, and outputting an alignment result; calculating a future operation state sequence according to the alignment result; calculating a trajectory deviation parameter, and constructing a service feature deviation vector; retrieving a historical operation and maintenance database to obtain a matched historical deviation modulus set; when the set variance exceeds a preset threshold, calculating a historical data storage time period ratio; generating a state optimization scheme and executing a dynamic adjustment instruction according to the state optimization scheme. The method realizes comprehensive and accurate measurement of the power transmission jumper state deviation by fusing multi-modal sensing data, generates an optimization scheme in combination with historical data, and is helpful to improving the effectiveness of power transmission jumper state monitoring and the timeliness of operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of power system monitoring, and in particular to a method for measuring the state deviation of transmission jumpers based on multimodal sensing. Background Technology

[0002] As a critical connecting component in power transmission lines, the operating status of transmission jumpers directly affects the stable operation of the power system. During long-term service, affected by various factors such as natural environment, mechanical stress, and material aging, transmission jumpers are prone to deviations in condition such as positional shift, deformation, and abnormal vibration. If these deviations are not detected and addressed in a timely manner, they may cause serious faults such as line short circuits and tripping, resulting in huge economic losses and social impact.

[0003] Traditional methods for measuring the condition of power transmission jumpers mostly rely on manual inspections or single-sensor technology. Manual inspections are not only labor-intensive but also limited by the experience of the inspectors and the external environment, making it difficult to achieve comprehensive, accurate, and real-time monitoring of the jumper's condition. While single-sensor technology can obtain some status information about the jumper to a certain extent, it has significant limitations: visual sensing is easily affected by factors such as lighting and obstruction, leading to data failure; acceleration sensing can only reflect vibration characteristics and cannot comprehensively characterize the jumper's spatial position and deformation state.

[0004] As power systems develop towards intelligence and precision, the requirements for monitoring the status of transmission line jumpers are increasing, and traditional methods can no longer meet the actual needs. Summary of the Invention

[0005] Purpose of the invention: To propose a method for measuring the state deviation of power transmission jumpers based on multimodal sensing, which can integrate multiple sensor data to achieve accurate measurement of the state deviation of power transmission jumpers, thereby improving the operational reliability and maintenance efficiency of the power system.

[0006] To achieve the above objectives, the present invention provides a method for measuring the state deviation of power transmission jumpers based on multimodal sensing, comprising the following steps:

[0007] Acquire multimodal sensor data sets from power transmission jumpers;

[0008] Integrate the multimodal sensing data set to generate a power transmission jumper state feature matrix;

[0009] Based on a preset standard feature template, the power transmission jumper state feature matrix is ​​subjected to time-series alignment processing, and the state feature alignment result is output.

[0010] Based on the alignment results of the aforementioned state features, the future operating state sequence of the power transmission jumper can be deduced;

[0011] Calculate the trajectory deviation parameter between the state feature matrix of the power transmission jumper and the future operating state sequence of the power transmission jumper;

[0012] Construct a service feature deviation vector that includes the current state features and the predicted state deviation;

[0013] Retrieve the historical operation and maintenance database to obtain a set of historical deviation magnitudes that match the service characteristic deviation vector;

[0014] When the variance of the historical deviation modulus set exceeds a preset variance threshold, the proportion of historical data storage periods is calculated.

[0015] A power transmission jumper status optimization scheme is generated based on the proportion of the historical data storage period; a dynamic adjustment instruction based on the power transmission jumper status optimization scheme is executed.

[0016] Preferably, the multimodal sensing data set includes image sensing data, deformation sensing data, and environmental sensing data; the generation of the power transmission jumper state feature matrix specifically includes:

[0017] Decompose the spatial feature vector of the image sensing data;

[0018] Extract the time-varying fluctuation characteristics of the deformation sensing data;

[0019] Environmental interference coefficient that integrates the environmental sensor data;

[0020] By integrating the aforementioned spatial feature vectors, time-varying fluctuation features, and environmental disturbance coefficients, a multidimensional feature tensor is constructed.

[0021] A dimensionality reduction operation is performed on the multidimensional feature tensor to generate the power transmission jumper state feature matrix.

[0022] Preferably, the output state feature alignment result specifically includes:

[0023] Load standard feature templates;

[0024] Match the timestamp sequence of the power transmission jumper status feature matrix;

[0025] The dimensional differences between the power transmission jumper state feature matrix and the standard feature template are determined;

[0026] Adjust the sampling frequency and feature dimension of the power transmission jumper state feature matrix;

[0027] Generate state feature alignment results that include time alignment parameters and dimension correction parameters.

[0028] Preferably, the calculation of the future operating state sequence of the transmission jumper specifically includes:

[0029] Parse the time alignment parameters in the state feature alignment result;

[0030] Construct a state transition probability model, and input the current observation window of the power transmission jumper state feature matrix;

[0031] Iteratively calculate the state transition path for future time windows and output a sequence of future operating states of the power transmission jumper containing multiple predicted time points.

[0032] Preferably, calculating the trajectory deviation parameter between the power transmission jumper state feature matrix and the future operating state sequence of the power transmission jumper specifically includes:

[0033] Extract the real-time feature values ​​of the power transmission jumper state feature matrix, obtain the predicted feature values ​​of the future operating state sequence of the power transmission jumper, and calculate the Euclidean distance parameter between the real-time feature values ​​and the predicted feature values;

[0034] The offset direction of the real-time feature value relative to the predicted feature value is determined, and a trajectory deviation parameter containing the Euclidean distance parameter and the offset direction is generated.

[0035] Preferably, the construction of the service feature deviation vector, which includes the deviation between the current state features and the predicted state, specifically includes:

[0036] Collect the current service time of the power transmission jumper, obtain the maximum value of the Euclidean distance parameter in the trajectory deviation parameters, and integrate the current service time and the maximum value of the distance parameter;

[0037] By adding the gradient of the environmental interference coefficient, a four-dimensional service characteristic deviation vector is constructed.

[0038] Preferably, the historical deviation magnitude set matching the service characteristic deviation vector is obtained by retrieving it from the historical operation and maintenance database, specifically including:

[0039] Decompose the dimensional components of the service characteristic deviation vector and set the matching tolerance threshold for each dimensional component;

[0040] Scan the index tags of the historical operation and maintenance database to extract historical records that meet the tolerance range of each dimension component;

[0041] The deviation modulus values ​​in the aggregated historical records are used to form a set of historical deviation modulus values.

[0042] Preferably, the calculation of the historical data storage period ratio specifically includes:

[0043] Determine the starting timestamp for storing historical data;

[0044] Get the timestamp of the current analysis operation;

[0045] Calculate the total time period from the storage start timestamp to the current analysis operation timestamp;

[0046] The effective coverage period of the calibration historical deviation modulus set;

[0047] Calculate the percentage of the effective coverage period to the total coverage period.

[0048] Preferably, the optimized scheme for generating power transmission jumper status specifically includes:

[0049] Analyze the percentage of the effective coverage period relative to the total time period;

[0050] By associating the offset direction with the trajectory deviation parameters, a state optimization decision tree is constructed.

[0051] Match the branch weight parameters in the state optimization decision tree, and output the power transmission jumper state optimization scheme that includes adjustment priority and adjustment magnitude.

[0052] Preferably, executing dynamic adjustment instructions based on the power transmission jumper status optimization scheme specifically includes:

[0053] The adjustment priority of the power transmission jumper status optimization scheme is decomposed, and the adjustment range is converted into a sequence of equipment control commands;

[0054] Verify the compatibility between the device control command sequence and the current environmental sensor data. If the verification is successful, send the verified device control commands to the execution terminal step by step and record the trajectory of state characteristic changes after the command is executed.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] (1) By acquiring a multimodal sensor data set, the limitations of single sensor technology are overcome, and the operating status information of power transmission jumpers can be captured comprehensively from different dimensions. The integration of multimodal data makes the originally scattered and isolated state parameters form an organic whole, providing a richer and more comprehensive foundation for subsequent state analysis.

[0057] (2) The feature fusion processing unit integrates multimodal sensor data to generate a power transmission jumper status feature matrix, realizing the effective association and fusion of different types of sensor data. The fusion is not a simple data superposition, but rather a deep processing of the data to extract key features that can characterize the overall status of the jumper, reducing data redundancy, improving the pertinence and effectiveness of feature information, and helping to more accurately reflect the actual operating status of the jumper.

[0058] (3) Timing alignment is performed based on a preset jumper state benchmark model to ensure the comparability of state characteristics at different times. Timing alignment can eliminate the deviation of state characteristics caused by time differences, making the analysis of jumper state changes more accurate and providing a reliable timing benchmark for subsequent state prediction and deviation calculation.

[0059] (4) By extrapolating the future operating state sequence of the transmission jumper, the monitoring of the jumper state is extended from the current state to the future trend, realizing a forward-looking grasp of state changes. By predicting the future state, possible deviation trends can be perceived in advance, avoiding the lag that may be caused by judging based solely on the current state, and providing a more timely reference for operation and maintenance decisions.

[0060] (5) Calculate the trajectory deviation parameters between the state feature matrix and the future operating state sequence, construct the service feature deviation vector, and intuitively quantify the difference between the current state and the future predicted state of the jumper; the quantified deviation information enables maintenance personnel to clearly understand the degree of deviation and specific characteristics of the jumper state, providing a clear basis for judging whether the jumper is abnormal and the degree of abnormality.

[0061] (6) The historical deviation modulus set matching the historical operation and maintenance database is retrieved, and an optimization scheme is generated by combining the proportion of historical data storage periods, thus realizing intelligent decision-making based on historical experience. The introduction of historical data makes full use of the practical experience accumulated in the past operation and maintenance process, making the generated optimization scheme more in line with the actual operation and maintenance needs, and enhancing the feasibility and effectiveness of the scheme.

[0062] (7) When the variance of the historical deviation modulus set exceeds the preset threshold, an optimization scheme is generated by calculating the proportion of historical data storage periods and executing dynamic adjustment instructions, thereby realizing dynamic response and adaptive adjustment of jumper status. The dynamic adjustment mechanism can flexibly adjust the operation and maintenance strategy according to the actual deviation situation, avoid the rigidity of fixed operation and maintenance mode, improve the pertinence and flexibility of operation and maintenance measures, and help to correct jumper status deviation in a timely manner and maintain the stable operation of jumpers. Attached Figure Description

[0063] Figure 1 This is a schematic diagram illustrating the working principle of the power transmission jumper state deviation measurement method based on multimodal sensing in an embodiment of the present invention.

[0064] Figure 2 A flowchart for generating the state feature matrix of power transmission jumpers.

[0065] Figure 3 A flowchart for predicting the future operating state sequence of power transmission jumpers.

[0066] Figure 4 A flowchart for constructing the service characteristic deviation vector.

[0067] Figure 5 A flowchart for calculating the proportion of historical data storage periods. Detailed Implementation

[0068] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0069] Please see Figure 1 This invention provides a method for measuring the state deviation of power transmission jumpers based on multimodal sensing, comprising the following steps:

[0070] A multimodal sensor dataset generated during the operation of the power transmission jumper is acquired. This dataset is collected in real time by various sensors deployed on the jumper. A feature fusion processing unit integrates the multimodal sensor dataset to generate a unified-dimensional power transmission jumper state feature matrix. Based on a pre-built jumper state baseline model, time-series alignment processing is performed on the power transmission jumper state feature matrix to obtain the state feature alignment result. The state feature alignment result is then analyzed, and a state prediction algorithm is used to predict the operating state sequence of the power transmission jumper in a future specified time window. The trajectory deviation parameter between the power transmission jumper state feature matrix (reflecting the current state) and the future operating state sequence is calculated. Based on the current state features and the predicted state deviation, a four-dimensional service feature deviation vector is constructed. A set of historical deviation moduli matching the service feature deviation vector is retrieved from the historical maintenance database. If the statistical variance of this set exceeds a preset variance threshold, the proportion of historical data storage periods is further calculated. Based on this proportion, a power transmission jumper state optimization scheme is generated, and corresponding dynamic adjustment instructions are executed.

[0071] Example 1:

[0072] See Figure 2The multimodal sensing dataset originates from real-time acquisition by three types of sensors: image sensing data is acquired through visible light and infrared dual-spectrum cameras deployed at specific locations on transmission towers; deformation sensing data is acquired by fiber optic strain sensor arrays attached to the surface of jumper conductors; and environmental sensing data comes from temperature and humidity sensors and ultrasonic anemometers installed on adjacent towers. Image sensing data processing includes: frame segmentation of the original video stream, Gaussian filtering for noise reduction of each frame, and extraction of gradient histogram distribution features of key point regions using a scale-invariant feature transform algorithm. This spatial feature vector contains a 128-dimensional descriptor. Deformation sensing data processing includes: receiving the micro-strain time series output from strain sensors, eliminating high-frequency noise using a Kalman filter, performing a five-level decomposition using the Daubechies wavelet basis function, and extracting the wavelet coefficient variance of the third-level detail components as time-varying fluctuation feature values. The processing of environmental sensor data requires combining the instantaneous values ​​of temperature, humidity, and wind speed with a preset historical environmental database (which stores the average environmental data of the same period in the past three years) to calculate the standardized difference between the current value and the historical average. This difference is then normalized by the Sigmoid function to generate the environmental interference coefficient.

[0073] Spatial feature vectors, time-varying fluctuation features, and environmental interference coefficients enter the feature integration stage. The process of constructing a multidimensional feature tensor is as follows: a feature container with dimensions of 128 (image features) + 1 (deformation features) + 1 (environmental features) is created, and the time axis is expanded at a sampling frequency of 1 frame per second. Feature values ​​are filled into tensor units according to timestamp alignment rules, where the 128 dimensions of image features are filled with missing time point data using bilinear interpolation. The tensor dimensionality reduction operation is implemented based on the principal component analysis algorithm: the covariance matrix of the feature tensor is calculated, eigenvalue decomposition is performed, the top 20 principal component directions with cumulative contribution rates exceeding a set threshold are selected, the original features are projected onto the dimensionality-reduced subspace, and finally a 20×N dimensional transmission jumper state feature matrix is ​​generated, where N represents the number of sampling time points.

[0074] The standard feature template stored in the jumper state baseline model originates from laboratory calibration data. This template includes the baseline values ​​of image features, deformation tolerance range, and environmental parameter thresholds of jumpers under standard operating conditions. The template time axis is divided in seconds. The time alignment operation consists of three steps: First, identify the acquisition timestamp sequence of the state feature matrix, compare it with the standard timeline of the baseline template, and align the time nodes using a dynamic time warping algorithm, outputting the time offset correction amount as the time alignment parameter. Second, compare the dimensional composition of the feature matrix and the standard template, identify the necessary dimensions in the feature matrix corresponding to the baseline template, delete duplicate redundant dimensions, and fill in the missing dimensions (such as conductor temperature) in the template using analogy with sensor data from nearby towers. The dimension correction record table records the addition and deletion operation identifiers. Third, adjust the sampling frequency to 10 frames per second, consistent with the baseline template. Low-frequency data is reconstructed using cubic spline interpolation, and high-frequency data is down-converted using moving average filtering. The frequency conversion coefficient is recorded in the dimension correction parameter. The output state feature alignment result includes: a time alignment parameter table containing time point mapping relationships and a dimension correction log file recording feature dimension change identifiers.

[0075] The storage structure of the multidimensional feature tensor employs a time-feature dual-level index. The time axis uses Unix timestamps as index keys, and the feature layer uses a hash table to store feature values ​​of different categories: image feature values ​​are stored as address pointers in the form of a 128-dimensional floating-point array, deformation feature values ​​are stored directly as single-precision floating-point numbers, and environmental interference coefficients are stored as double-precision floating-point numbers. Memory allocation for tensor units uses dynamic memory pool management; when new time node data arrives, the time axis length is automatically expanded and contiguous memory blocks are allocated.

[0076] The principal component analysis (PCA) process for feature dimensionality reduction specifically includes: initializing the feature tensor data as a two-dimensional matrix of 130 rows (original number of features) × N columns (number of time points). Calculating the covariance of the matrix column vectors, and obtaining eigenvalues ​​and eigenvectors through singular value decomposition. After sorting the eigenvalues, setting a cumulative contribution rate threshold of 95%, and selecting the top K eigenvectors satisfying this condition to form a projection matrix. Multiplying the original matrix by the projection matrix on the left to achieve dimensionality reduction, and storing the resulting 20×N matrix in the state feature storage area.

[0077] The dimensionality discrepancy calibration process employs an automated matching workflow: A feature dimension dictionary is created for the baseline template, with keys including "image gradient features," "deformation fluctuation," and "temperature and humidity interference factors." The dimension names of the state feature matrix are traversed, and similarity matching is performed with the dictionary keys. Dimensions with similarity scores below a threshold are marked as redundant dimensions to be deleted; dimensions present in the template but not matched in the matrix are marked as dimensions to be supplemented. Dimension supplementation utilizes cross-sensor data fusion. For example, when the conductor temperature dimension is missing, conductor surface temperature data collected by an infrared image sensor is used, converted to true temperature values ​​using an emissivity correction algorithm, and then inserted into the feature matrix.

[0078] Frequency-adjusted interpolation operations handle low-frequency data: When the input data sampling rate is less than 10 frames per second, virtual nodes are inserted between the original time points. A Lagrange interpolation polynomial is constructed using five consecutive real data points to calculate the feature estimates of the virtual nodes. Oversampled data processing employs a time window function for segmentation: a sliding time window with a width of 100 milliseconds is created, and the average value of the data within the window is taken as the central representative value of the time window, achieving a uniform output of 10 frames per second. Sampling rate conversion records are written to the dimension correction parameter log, including the original sampling rate, the target sampling rate, and the interpolation function coefficients.

[0079] Timestamp alignment employs an improved dynamic time warping algorithm: a time difference matrix is ​​constructed between the state feature matrix and the baseline template. A constrained zigzag path search (with local slope limited to 1-2) is used to find the path with the minimum cumulative distance. The time offset of this path is converted into a corrected timestamp record for each original time point in the state feature matrix. The time alignment parameter table stores the original acquisition timestamp, the corrected timestamp, and their corresponding relationships, with time correction values ​​accurate to the millisecond level.

[0080] The dimension correction log uses a structured storage format: each record includes the operation type (add / delete / modify), feature dimension name, original value (if it exists), correction value (if it exists), data source identifier, and operation timestamp. For the newly added temperature dimension, the record includes the infrared sensor device number, emissivity correction formula coefficient, and the final completed temperature value sequence.

[0081] Example 2:

[0082] See Figure 3The time alignment parameter in the state feature alignment result includes a timestamp correction mapping table, which records the correspondence between the original acquisition time and the standard time reference axis. The state prediction algorithm first parses this mapping table to determine the prediction time domain starting point as the time point at which the current system timestamp is offset backward by the correction amount. The state transition probability model adopts a hidden Markov model architecture, which includes three hidden state categories (normal, slight offset, and critical offset) and an observation state space (i.e., the dimension of the transmission jumper state feature matrix). The model input is the data of the transmission jumper state feature matrix in the current observation window: the observation window is set to the most recent 30 consecutive time points, and each time point corresponds to an observation sequence composed of 20-dimensional feature vectors.

[0083] The state transition probability matrix is ​​obtained through training on historical operational data and stores the transition probabilities between each hidden state. The observation probability matrix records the Gaussian distribution parameters (mean vector and covariance matrix) of the feature vectors in each hidden state. The prediction process uses a forward-backward recursive algorithm: the hidden state probability distribution at the initial time is the posterior probability at the end of the observation window. In the forward recursive phase, the state probability at the i-th future time point is calculated:

[0084] Here, represents the state transition probability, and represents the observation probability. This process iterates times, expanding the prediction time step by one time step each time. The probability distribution is corrected in the backward recursive phase, ultimately outputting the complete state probability distribution for future time windows (each window containing 5 seconds). The prediction result is transformed into a specific state sequence: for each prediction time point, the hidden state category with the highest probability is selected, and the mean vector of the features corresponding to that category is used as the predicted feature value, forming a 3D feature vector sequence.

[0085] The trajectory deviation parameter calculation module synchronously receives two input sources: a real-time data stream of the power transmission jumper state feature matrix and a future operating state sequence output by the prediction module. The calculation process strictly adheres to the time synchronization principle: when the system reaches a given time point, the actual feature vector of the state feature matrix at that moment is acquired in real time, and the predicted feature vectors at the same time point in the prediction sequence are extracted simultaneously. Deviation calculation is performed in two parallel threads:

[0086] The distance calculation thread iterates through the 20 dimensional components of the feature vector, performing a squared difference operation on each dimension. The square root of the cumulative sum of the squared dimensional differences is then calculated to generate a Euclidean distance scalar value. This value is recorded in a distance parameter set, which stores the distance values ​​for the most recent 100 time points in chronological order.

[0087] The orientation calibration thread creates a 20-dimensional orientation identifier array. For each dimension component, the algebraic difference between the real-time feature value and the predicted feature value is calculated. A zero-value tolerance range is set. When the difference exceeds the range: if the difference is greater than the upper limit of zero, the dimension is marked as a positive offset; if the difference is less than the lower limit of zero, it is marked as a negative offset; if it falls within the tolerance range, it is marked as no offset. The offset direction set is stored as a binary encoded sequence, with each dimension using two binary digits to represent the offset state.

[0088] The trajectory deviation parameters are stored using a two-layer time-stamp indexed record: timestamps are accurate to milliseconds, the upper layer stores Euclidean distance values, and the lower layer stores a 20-dimensional direction identifier array. Data updates employ a circular buffer mechanism, automatically overwriting the oldest record when the number of time points exceeds 100. The deviation parameter output port pushes data packets containing the distance parameter set and the direction identifier set in real time.

[0089] The training process of the state transition probability model runs independently: the training dataset comes from a historical operation and maintenance database, selecting state feature matrix fragments under normal operating conditions for three consecutive months. Data preprocessing includes: aligning the time series according to the standard time reference axis and filtering out anomalous data points. Hidden state partitioning uses the K-means clustering algorithm: after standardizing the 20-dimensional data of the feature matrix, the number of cluster centers is set to 3, and three hidden state categories are formed after iterative convergence. Transition probability statistics use the frequency counting method: the number of transitions between each category is counted, and the transition probability matrix is ​​generated after Laplace smoothing. Observation probability parameters are fitted using a Gaussian distribution: for the feature vector samples under each hidden state category, the mean and standard deviation of each dimension are calculated, and the covariance matrix is ​​set as a diagonal matrix.

[0090] The iterative computation process of the prediction algorithm employs block-parallel optimization: the 20-dimensional feature vector is divided into four 5-dimensional sub-vector blocks, each allocated an independent computational unit to perform forward recursion. The calculation of the state probability distribution uses logarithmic space operations to prevent floating-point underflow: all probability values ​​are converted to natural logarithmic form, and multiplication operations are converted to addition operations. The recursive results are restored to probability values ​​through exponential transformation, and after normalization, the final probability distribution is output.

[0091] The tolerance range for offset direction calibration varies across dimensions: a narrower tolerance range is set for image-related dimensions (the first 12 dimensions), a medium range for deformation-related dimensions (the middle 5 dimensions), and a wider range for environment-related dimensions (the last 3 dimensions). The upper and lower limits for zero values ​​are set based on historical data statistics: the upper limit is 10% of the historical maximum positive deviation for that dimension, and the lower limit is 10% of the historical maximum negative deviation. The direction identifier array uses bit-field compression storage: the two bits of each dimension's identifier (00 for no offset, 01 for negative, and 10 for positive) are combined into a 40-bit string, compressing the storage space to 1 / 8 of the original array.

[0092] The synchronization mechanism between real-time and predicted data employs a hardware timestamp counter: the sensor acquisition end and the prediction module share a high-precision clock source, and data packets include nanosecond-level timestamps. When the system time reaches a certain point, the real-time data buffer executes a snapshot of the latest data, and the prediction data buffer retrieves the prediction vector corresponding to the timestamp. A mutex lock is used in both buffers to prevent read / write conflicts, and the time tolerance window is set to ±5 milliseconds. If no matching data is found within the timeout period, an exception handling process is triggered.

[0093] The circular buffer for the distance parameter set is implemented using a ring structure with pointers at both ends: the write pointer always points to the latest record, and the read pointer lags behind the write pointer by 10 positions. When the buffer is full, the oldest data is automatically overwritten, and the distance statistics (current maximum, minimum, and moving average) are updated simultaneously. Distance values ​​are stored as single-precision floating-point numbers, and the statistical calculations use an incremental update algorithm to avoid full traversal.

[0094] The dynamic update strategy for the predicted sequence is based on a feedback mechanism: every 10 new actual observation points triggers a fine-tuning of the prediction model parameters. The adjustment process includes: adding the new observation data to the training dataset, recalculating the Gaussian distribution parameters in the observation probability matrix, and keeping the transition probability matrix stable. During model updates, the prediction module switches to a backup cache sequence, and immediately switches back to the newly generated sequence after the update is complete.

[0095] Example 3:

[0096] See Figure 4 The construction of the service characteristic deviation vector begins with multi-source data acquisition: the current service duration of the transmission jumper is obtained from the equipment ledger system, accurate to the hour; the maximum value of the Euclidean distance parameter in the trajectory deviation parameters is determined by analyzing the distance value set of the most recent 100 time points, and the peak value in this set is taken; the gradient calculation of the change of the environmental interference coefficient adopts the time window difference method, selecting the environmental interference coefficient value of the current moment and the same moment of the previous hour to perform the difference operation, and then dividing by the time interval (3600 seconds) to obtain the gradient value; the aging coefficient of the jumper material is based on the year of line commissioning and the conductor material type (such as steel-cored aluminum stranded wire identified as L1), and the discretization level value is obtained by looking up the preset aging coefficient reference table. The above four components are assembled in a fixed order: the first dimension stores the service duration (unit: hours), the second dimension stores the maximum Euclidean distance (dimensionless), and the third dimension stores the environmental interference gradient (unit: The fourth dimension stores the aging coefficient level (integer values ​​from 1 to 5), forming a four-dimensional vector structure.

[0097] The historical maintenance database adopts a distributed time-series architecture. Each record includes fields such as timestamp, device identifier, and deviation modulus. Before the matching operation, the tolerance range of each dimension component needs to be set: the service duration tolerance is set to ±240 hours, which allows querying historical records with service times differing by 10 days; the maximum Euclidean distance tolerance is set to ±0.15, covering a fluctuation range of 15% above and below the current value; the environmental interference gradient tolerance is set to... The aging coefficient tolerance is set to ±1 level. Database scanning employs a multi-dimensional parallel retrieval strategy.

[0098] Service duration dimension search: Within the device identifier partition, filter records with service time in the range [T-240, T+240], where T represents the current service duration;

[0099] Euclidean distance dimension search: In a subset of records that meet the duration condition, filter for the maximum distance value within the interval. The records, among which This indicates the current maximum distance value;

[0100] Environmental gradient dimension retrieval: Further filter environmental gradient values ​​within the range The record, where G represents the current environmental gradient value;

[0101] Aging coefficient dimension search: The final filter is based on aging coefficient levels within a certain range. The record, where L represents the current aging factor.

[0102] The aggregation module extracts the deviation modulus field from historical records that meet all tolerance conditions, forming a historical deviation modulus set. This set is stored in ascending order, while retaining the original timestamp information corresponding to each record. The calculation process of the environmental disturbance gradient is defined as follows:

[0103]

[0104] in: This represents the environmental interference coefficient at the current moment. express Environmental interference coefficient before time ( =3600 seconds). This is the calculation result. This value reflects the rate of change of environmental factors; the gradient value increases significantly when the environmental disturbance coefficient fluctuates drastically.

[0105] The aging coefficient comparison table is constructed based on accelerated material testing data: common jumper conductor materials are divided into five categories (L1-L5), and each category is further divided into aging stages according to its years of operation. For example, the corresponding rules for steel-cored aluminum stranded wire (L1 category) are: 0-5 years of operation = coefficient 1, 6-10 years = coefficient 2, 11-15 years = coefficient 3, 16-20 years = coefficient 4, and over 20 years = coefficient 5. The coefficient values ​​are obtained in real time through the application programming interface of the equipment management system.

[0106] The database index structure uses a composite key design: the primary key is the device identifier (12-bit character encoding), and the secondary indexes contain quantified values ​​for four dimension fields. A B+ tree index is built for the service duration field to support range queries; a hash index is used for the Euclidean distance field to accelerate equality matching; a bitmap index is used for the environmental gradient field to improve the efficiency of multi-condition filtering; and an inverted index is built for the aging coefficient field. During query execution, the optimizer automatically selects an index merging strategy to avoid full table scans.

[0107] The storage management of the deviation modulus set includes a data validation step: checking the validity range of the values ​​in the set (0.01-10.00) and filtering out abnormal records that are outside the range; verifying the continuity of timestamps and adding data integrity markers to the record set with time discontinuities; calculating the statistical indicators of the set (mean, extreme values, quantiles) but not outputting them immediately, only as intermediate preparation for subsequent variance calculation.

[0108] Dynamic adjustment mechanism for tolerance threshold: When the number of records returned in the initial retrieval is less than the set threshold (e.g., less than 20), the tolerance expansion procedure is initiated. The tolerance for each dimension is reset to 150% of the initial value for a second retrieval. If it is still insufficient, it continues to expand to 200%, with a maximum of three iterations. The expanded tolerance range is recorded in the query log, but the original tolerance setting value remains unchanged.

[0109] Anomaly handling during vector construction: When the trajectory deviation parameter set is empty, the moving average of the distance values ​​in the last 24 hours is used to replace the maximum value; if environmental sensor data is missing, the gradient value is the average gradient of the same time period in the last three days; when the ledger system has no service duration record, the theoretical value is calculated based on the line commissioning date and a data source mark is added.

[0110] Historical record aggregation employs a streaming processing framework: database query results are transmitted in shards via a message queue, and aggregation nodes perform sorting and merging operations on each shard. The final collection is stored using a skip list data structure, supporting fast range queries and sequential access. Each record is appended with metadata: original record ID, acquisition device number, and data quality flags.

[0111] Matching result caching mechanism: A 128-bit MD5 hash value is generated for the four-dimensional vector parameters of each query as the cache key. When the Euclidean distance between the new query vector and the cache key is less than a set threshold, the cached historical deviation modulus set is directly returned to avoid duplicate database accesses. The cache validity period is set to 24 hours, and it is automatically cleared after expiration.

[0112] Example 4: See Figure 5 The calculation of the proportion of historical data storage periods uses the transmission line identifier as the index unit. When the analysis object is jumper number L-4032, the system first accesses the metadata storage area of ​​the historical operation and maintenance database. The earliest recorded timestamp of this jumper is extracted: by traversing all record header information under the device ID partition, the minimum timestamp value is obtained. For example, the search results may show the earliest record time as January 15, 2020, 08:30:25 UTC. The current analysis operation timestamp is taken from the system clock, assumed to be June 18, 2024, 14:22:10 UTC. The total time period is calculated and converted to a second-level difference: the two timestamps are converted to Unix timestamp format, and the difference is obtained by subtracting them. In this example, the total time period is 138,234,765 seconds.

[0113] The determination of effective coverage period involves a dual verification mechanism. The first verification checks data continuity: starting from the earliest recorded time, the timestamps of records in the historical deviation modulus set are scanned chronologically. The time interval between adjacent records is checked to ensure it remains constant equal to the preset sampling interval (e.g., 300 seconds). When a time interval exceeds the tolerance range (±5 seconds), that point is marked as a continuity breakpoint. The second verification checks data integrity: it checks whether each sampling point contains a valid deviation modulus record, excluding null values ​​or incorrectly marked data. The effective coverage period is defined as the longest continuous, uninterrupted interval. For example, data from March 10, 2021 to September 18, 2023 might meet the criteria. This period length, calculated after timestamp conversion, is 77,587,200 seconds. The storage period ratio is calculated by dividing the effective period length by the total period length; in this example, the result is 77,587,200 / 138,234,765≈0.5612.

[0114] The state optimization scheme generation module receives two inputs: the storage period proportion value (56.12%) and the offset direction set in the trajectory deviation parameters. The offset direction set contains an array of direction identifiers for the most recent 100 time points, and the frequency of positive and negative offsets for each dimension is counted. The decision tree model uses a predefined rule set:

[0115] The first-level branch is based on a storage period ratio threshold (70%). When the input ratio (56.12%) is below the threshold, the branch enters the left subtree. The left subtree analyzes the positive and negative proportions of the offset direction set: calculating the proportion of positive offsets in 20 dimensions. If the positive offset dimensions exceed 60% (e.g., 14 dimensions), the priority weight is adjusted to 0.7; if it is between 40% and 60% (e.g., 10 dimensions), the weight is set to 0.5; and if it is below 40%, the weight is set to 0.3.

[0116] The second-level branch associates specific offset dimension types. Image-related dimensions (1-12 dimensions) are assigned a higher priority coefficient of 1.2, deformation dimensions (13-17 dimensions) a coefficient of 1.0, and environmental dimensions (18-20 dimensions) a coefficient of 0.8. The adjustment amplitude baseline is obtained by looking up a table using the jumper type: for example, the baseline amplitude for steel-cored aluminum stranded wire is 25 units. The final adjustment amplitude calculation formula is: baseline amplitude × priority weight × dimension coefficient. In this example, if the priority weight is 0.7 and the image dimension coefficient is 1.2, then the adjustment amplitude is 25 × 0.7 × 1.2 = 21 units.

[0117] The output structure of the solution includes a sequence of adjustment items: each adjustment item corresponds to a specific dimension group, recording the dimension number range, adjustment priority (levels 1-3), and adjustment magnitude. Priority is determined by weight values: weight ≥ 0.6 is level 1, 0.4-0.6 is level 2, and ≤ 0.4 is level 3. The output in this example might contain: "Dimensions 1-12: Priority 1, magnitude 21; Dimensions 13-17: Priority 2, magnitude 17.5; Dimensions 18-20: Priority 3, magnitude 14".

[0118] Timestamp processing follows a standardized process: all timestamps are uniformly converted to UTC timezone for storage. A timeline bitmap is established for continuity verification: each sampling point corresponds to one bit, set to 1 for valid data and 0 for missing data. The longest consecutive sequence of 1s is achieved by scanning the bitmap, using two pointers to record the start and end positions of the current consecutive segment. When a 0-bit is detected, the length of the current consecutive segment is compared with the historical maximum value, and the record for the longest valid time period is updated.

[0119] Data integrity verification includes outlier filtering rules: records with a deviation modulus exceeding a preset reasonable range (0.05-8.00) are marked as invalid; records with a timestamp deviating from the system clock by more than 24 hours are marked as suspicious data; for outliers in the sampling interval, three adjacent records need to be verified to confirm whether they are isolated outliers caused by equipment failure.

[0120] The decision tree dimension coefficients are set based on physical characteristics: the image dimension corresponds to the spatial location sensitivity of the conductor and is assigned a higher coefficient; the environment dimension is greatly affected by external factors and is assigned a lower coefficient. The weight allocation adopts discretization: every 10% increase in the proportion of the positive offset dimension corresponds to a 0.2 increase in weight, avoiding frequent changes in the scheme caused by continuous values.

[0121] Adjustment Amplitude Unit Conversion Mechanism: Amplitude values ​​in the output scheme need to be converted into specific execution parameters. For example, an image dimension adjustment of 21 units corresponds to a pixel offset threshold adjustment of 42 pixels in the video analysis system (unit conversion factor of 2); a deformation dimension adjustment of 17.5 units corresponds to a strain sensor alarm threshold change of 0.035% (conversion factor of 0.002). The conversion factor table is stored in the device configuration database, categorized by jumper model.

[0122] Scheme generation anomaly handling: When there are all-zero records (no offset) in the offset direction set, the nearest non-zero record is used as the replacement; if the storage period ratio is less than 40%, a data shortage warning is triggered, and a default conservative scheme is output (priority 3 for all dimensions, amplitude is taken as the baseline value of 50%); when the decision tree branch conditions conflict, the conflict resolution rule is activated: the higher priority branch is given priority, and the maximum amplitude value is taken for conflicts at the same level.

[0123] The output interface uses a structured data format: each adjustment item includes four fields: dimension range, priority value, original amplitude value, and engineered amplitude value. The engineered value is calculated in real time through unit conversion, with an additional source identifier for the conversion coefficient. The solution package includes additional metadata: generation timestamp, input parameter summary, and checksum.

[0124] Example 5: The dynamic adjustment instruction execution process begins with the parsing operation of the state optimization scheme. The scheme text is stored using Structured Markup Language and contains several adjustment item entries. The parsing engine identifies the dimension interval descriptor, priority value, and original amplitude value fields for each entry. The priority value is converted into an execution sequence index: priority 1 corresponds to index 0 (highest execution level), priority 2 corresponds to index 1, and priority 3 corresponds to index 2. The original amplitude value is converted through the device configuration library: accessing the device type registry and matching the jumper model with the actuator mapping relationship. For example, when processing the steel-cored aluminum stranded wire model, the image dimension adjustment amplitude is converted into a displacement instruction for the tension regulator, with the conversion rule being that the amplitude value is multiplied by a coefficient of 0.8 to generate a millimeter-level displacement; the deformation dimension adjustment amplitude is converted into a damper pressure valve opening instruction, with a conversion coefficient of 0.05 to generate a kPa pressure value. The converted device control instruction is formatted as a standard control frame: the frame header contains the actuator physical address (6-byte MAC address), and the frame body contains the action type code (1 byte) and parameter value (4-byte floating-point number).

[0125] The environmental compatibility verification module monitors the meteorological sensor data stream in real time. The verification rule set contains three layers of judgment logic: The first layer checks the latest reading of the wind speed sensor; if three consecutive sampled values ​​exceed a set threshold (e.g., 15 m / s), a wind speed exceeding limit flag is generated; the second layer analyzes the humidity sensor data, calculates the moving average value over the past ten minutes, and sets a humidity exceeding limit flag when it exceeds a threshold (e.g., 85%); the third layer evaluates the combined conditions, and outputs an incompatibility status code when both the wind speed exceeding limit flag and the humidity exceeding limit flag are activated. The verification process performs a polling every 200 milliseconds, and the results are cached in the shared memory area.

[0126] The instruction step-by-step transmission mechanism employs priority queue management. Three priority queues (Q0-Q2) are initialized, and the parsed instruction frames are stored in their corresponding queues according to their index numbers. The send scheduler retrieves the first instruction from the head of queue Q0, appends a timestamp and checksum, and encapsulates it into a transmission message. The message is sent to the target execution terminal via industrial Ethernet, and a response timer (with a timeout set to 2 seconds) is started after transmission. Upon receiving the message, the execution terminal verifies the integrity of the frame structure and performs physical operations, subsequently returning an acknowledgment frame containing the execution result code. Upon receiving a valid acknowledgment frame, the scheduler records the execution timestamp and terminal return code, waits 500 milliseconds, and then triggers the next instruction transmission. If the response times out or returns an error code, the instruction is reinserted into the original queue, the retry counter is incremented, and transmission of that queue is suspended if the maximum number of retries (3) is exceeded.

[0127] The state characteristic change trajectory recording system is activated synchronously with the command execution process. Recording starts 100 milliseconds before the first command is sent and continues until 300 milliseconds after the last command is confirmed. Trajectory recording parameters include: raw sampled values ​​of multimodal sensor data (image pixel blocks, deformation micro-strain values, ambient temperature and humidity), real-time calculated values ​​of the state feature matrix (20-dimensional feature vector), and updated values ​​of trajectory deviation parameters. The sampling frequency is increased to 50 times per second, and high-precision timestamps (microsecond level) are added to the data packets. After recording, the change analysis module performs difference calculation: extracting the feature matrix of the last stable state before command execution (taking the average of the first 10 seconds) as a baseline, comparing it with the feature matrix of the stable state after execution (taking the average of the last 10 seconds), calculating the absolute changes in 20 dimensions, and generating a difference report.

[0128] The terminal communication protocol employs an application-layer confirmation mechanism. Control frame transmission uses UDP for accelerated transmission, while confirmation frames use TCP to ensure reliability. The terminal's built-in instruction interpreter parses the action type code field: 0x01 represents displacement commands, activating the stepper motor drive circuit; 0x02 represents pressure commands, controlling the opening and closing of the piezoelectric valve. Parameter values ​​are converted into pulse signal counts or analog voltage quantities, maintaining a conversion accuracy within a 0.1% error range. The terminal status register is updated in real time, containing parameters such as execution progress flags, equipment temperature values, and current load; these parameters are returned to the central system with the confirmation frames.

[0129] The trajectory recording storage employs a circular buffer structure. A 100MB memory space is allocated, storing data packets in chronological order. Each data packet contains: an 8-byte timestamp, a 4-byte block of raw sensor data (compressed), a 20-dimensional single-precision floating-point feature vector, and an 8-byte deviation parameter structure. The buffer write pointer moves cyclically, overwriting the oldest data packet. During difference calculation, the system extracts data segments from the buffer based on the start and end timestamps of the records, generates a comparative analysis report, and stores it in the historical database.

[0130] Environmental verification threshold dynamic adjustment mechanism: Threshold parameters are automatically switched according to seasonal type. In summer mode (June-August), the humidity threshold is increased to 90%, and the wind speed threshold is decreased to 12 m / s; in winter mode (December-February), the humidity threshold is decreased to 80%, and the wind speed threshold is increased to 18 m / s. Threshold parameters are stored in a configuration file and are automatically updated at the beginning of each month. A time decay factor is added to the composite condition judgment: when the wind speed exceedance flag is activated, if no new exceedance event occurs within the next 5 minutes, the flag is automatically reset.

[0131] Command retry management includes a backoff strategy: the first retry is delayed by 1 second, the second by 3 seconds, and the third by 5 seconds. The retry queue is independent of the main priority queue and is monitored by a dedicated thread. All retried failed commands generate a fault event log, including the terminal address, command content, and failure reason code, triggering the operation and maintenance alarm system.

[0132] The time synchronization mechanism for the changing trajectory adopts the PTP precision clock protocol. The execution terminal, sensor nodes, and central server are all connected to the clock synchronization network, with time deviation controlled within 100 microseconds. Upon startup of trajectory recording, a synchronization pulse signal is broadcast, and each node aligns its sampling clock with the rising edge of the pulse as a reference. The difference report includes an additional time synchronization accuracy indicator, marking the maximum clock deviation value.

[0133] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for measuring the state deviation of power transmission jumpers based on multimodal sensing, characterized in that, Includes the following steps: Acquire multimodal sensor data sets from power transmission jumpers; Integrate the multimodal sensing data set to generate a power transmission jumper state feature matrix; Based on a preset standard feature template, the power transmission jumper state feature matrix is ​​subjected to time-series alignment processing, and the state feature alignment result is output. Based on the alignment results of the aforementioned state features, the future operating state sequence of the power transmission jumper can be deduced; The calculation of the trajectory deviation parameter between the power transmission jumper state feature matrix and the future operating state sequence of the power transmission jumper specifically includes: extracting the real-time feature values ​​of the power transmission jumper state feature matrix, obtaining the predicted feature values ​​of the future operating state sequence of the power transmission jumper, and calculating the Euclidean distance parameter between the real-time feature values ​​and the predicted feature values; calibrating the offset direction of the real-time feature values ​​relative to the predicted feature values, and generating a trajectory deviation parameter containing the Euclidean distance parameter and the offset direction; constructing a service feature deviation vector containing the deviation between the current state features and the predicted state, specifically including: collecting the current service duration of the power transmission jumper, obtaining the maximum value of the Euclidean distance parameter in the trajectory deviation parameter; integrating the current service duration and the maximum value of the distance parameter; adding the gradient of the change in the environmental interference coefficient, and constructing a four-dimensional service feature deviation vector; The process of retrieving historical deviation modulus values ​​from the historical operation and maintenance database and obtaining a set of historical deviation modulus values ​​that match the service characteristic deviation vector includes: decomposing the dimensional components of the service characteristic deviation vector and setting a matching tolerance threshold for each dimensional component; scanning the index tags of the historical operation and maintenance database and extracting historical records that meet the tolerance range of each dimensional component; and aggregating the deviation modulus values ​​in the historical records to form a set of historical deviation modulus values. When the variance of the historical deviation modulus set exceeds a preset variance threshold, the proportion of historical data storage periods is calculated, specifically including: determining the storage start timestamp of historical data; obtaining the timestamp of the current analysis operation; calculating the total time period from the storage start timestamp to the current analysis operation timestamp; calibrating the effective coverage period of the historical deviation modulus set; and calculating the percentage of the effective coverage period to the total time period. Based on the proportion of the historical data storage periods, a power transmission jumper status optimization scheme is generated, specifically including: parsing the percentage of the effective coverage period to the total period; associating the offset direction in the trajectory deviation parameter to construct a status optimization decision tree; matching the branch weight parameters in the status optimization decision tree to output a power transmission jumper status optimization scheme that includes adjustment priority and adjustment magnitude; Execute dynamic adjustment instructions based on the power transmission jumper status optimization scheme.

2. The method for measuring the state deviation of power transmission jumpers based on multimodal sensing according to claim 1, characterized in that, The multimodal sensing data set includes image sensing data, deformation sensing data, and environmental sensing data; The generation of the power transmission jumper state feature matrix specifically includes: Decompose the spatial feature vector of the image sensing data; Extract the time-varying fluctuation characteristics of the deformation sensing data; Environmental interference coefficient that integrates the environmental sensor data; By integrating the aforementioned spatial feature vectors, time-varying fluctuation features, and environmental disturbance coefficients, a multidimensional feature tensor is constructed. A dimensionality reduction operation is performed on the multidimensional feature tensor to generate the power transmission jumper state feature matrix.

3. The method for measuring the state deviation of power transmission jumpers based on multimodal sensing according to claim 1, characterized in that, The output state feature alignment result specifically includes: Load standard feature templates; Match the timestamp sequence of the power transmission jumper status feature matrix; The dimensional differences between the power transmission jumper state feature matrix and the standard feature template are determined; Adjust the sampling frequency and feature dimension of the power transmission jumper state feature matrix; Generate state feature alignment results that include time alignment parameters and dimension correction parameters.

4. The method for measuring the state deviation of power transmission jumpers based on multimodal sensing according to claim 1, characterized in that, The predicted sequence of future operating states of the transmission jumper specifically includes: Parse the time alignment parameters in the state feature alignment result; Construct a state transition probability model, and input the current observation window of the power transmission jumper state feature matrix; Iteratively calculate the state transition path for future time windows and output a sequence of future operating states of the power transmission jumper containing multiple predicted time points.

5. The method for measuring the state deviation of power transmission jumpers based on multimodal sensing according to claim 1, characterized in that, Executing dynamic adjustment instructions based on the power transmission jumper status optimization scheme specifically includes: The adjustment priority of the power transmission jumper status optimization scheme is decomposed, and the adjustment range is converted into a sequence of equipment control commands; Verify the compatibility between the device control command sequence and the current environmental sensor data. If the verification is successful, send the verified device control commands to the execution terminal step by step and record the trajectory of state characteristic changes after the command is executed.

Images