A main transformer monitoring signal filtering method, device, equipment and medium

Abstract

The application discloses a kind of main transformer monitoring signal filtering method, device, equipment and medium.The application tracks the steady-state periodic law of main transformer monitoring signal by autoregressive prediction model and generates prediction reference signal, simultaneously, the random fluctuation and noise interference of signal are handled by variable step size adaptive least mean square filter, the prediction output of autoregressive prediction model is used as the reference signal of adaptive filter, according to the error of reference signal and filter output signal, step size is dynamically adjusted and weight vector is iteratively updated, so that filter maintains faster convergence speed when error is larger, quickly tends to be stable when error is smaller, so as to give consideration to convergence speed and steady-state accuracy, effectively extract non-stationary monitoring signal with low-frequency positive and negative alternating, strong random fluctuation and high noise interference, avoid transient impact information loss and effective low-frequency alternating information is submerged.

Description

Technical Field

[0001] This invention relates to the technical field of power systems, specifically to a method, apparatus, equipment, and medium for filtering main transformer monitoring signals. Background Technology

[0002] In my country, urban rail transit systems commonly employ DC traction power supply systems, with the running rails also serving as return conductors. Under these conditions, some traction current inevitably leaks through the rails into the surrounding soil, forming stray currents. As my country's power grid structure becomes increasingly complex and electrical connections become closer, the phenomenon of stray currents intruding into the power grid is becoming more frequent, further leading to serious consequences such as accelerated corrosion of buried metals and increased DC bias in main transformers, posing a serious threat to the safe and stable operation of urban infrastructure.

[0003] Real-time and accurate online monitoring of electrical quantities such as transformer neutral point current, vibration, and noise is fundamental to obtaining transformer health status and implementing corresponding magnetic field suppression measures. However, existing transformer monitoring signal processing technologies mainly target two engineering scenarios: ground current and geomagnetic induced current in DC transmission projects, where the monitored signals change slowly and their patterns are relatively predictable. In contrast, stray currents in urban rail transit are affected by various factors such as train start-stop frequency, traction load power, and track insulation status, exhibiting non-stationary characteristics such as low-frequency alternating positive and negative currents and strong random fluctuations. When stray currents intrude into a transformer, the monitoring signals of transformer neutral point current, vibration, and noise exhibit complex characteristics such as low-frequency alternating positive and negative currents, strong random fluctuations, and high noise interference. This makes it easy for existing online transformer monitoring and signal processing methods to suffer from signal distortion when applied in the field. On the one hand, methods such as low-pass filtering are difficult to capture transient impact signals caused by random fluctuations in subway stray currents; on the other hand, methods such as wavelet transform are easily overwhelmed by random interference, resulting in the loss of effective low-frequency alternating information. The aforementioned signal distortion problem will directly lead to the distortion of the main transformer's magnetic bias assessment results based on the monitoring signal, which in turn will cause misjudgment and omission, making the assessment and suppression of the main transformer's magnetic bias lose its accurate basis, and seriously affecting the safe and stable operation of the main transformer in the power grid.

[0004] In related technologies, in scenarios where stray currents from urban rail transit intrude into the main transformer, there is a technical problem that existing transformer monitoring signal filtering methods are unable to effectively filter non-stationary monitoring signals that exhibit low-frequency alternating positive and negative currents, strong random fluctuations, and high noise interference. Summary of the Invention

[0005] The technical problem this invention aims to solve is that, in related technologies, existing transformer monitoring signal filtering methods struggle to effectively filter non-stationary monitoring signals exhibiting low-frequency alternating positive and negative values, strong random fluctuations, and high noise interference in scenarios involving stray current intrusion into the main transformer of urban rail transit. The objective is to provide a main transformer monitoring signal filtering method, device, equipment, and medium that solves the technical problem of effectively filtering non-stationary monitoring signals with low-frequency alternating positive and negative values, strong random fluctuations, and high noise interference.

[0006] This invention is achieved through the following technical solution:

[0007] In a first aspect, the present invention provides a method for filtering main transformer monitoring signals, comprising:

[0008] Acquire the current and historical sampling data of the main transformer monitoring signals;

[0009] The model parameters of the autoregressive prediction model are solved based on historical sampling data, and the reference signal is calculated based on the model parameters and historical sampling data.

[0010] The input vector of the adaptive filter is constructed based on the current sampled data, and the output signal of the filter is calculated based on the input vector and the current weight vector.

[0011] Calculate the difference between the reference signal and the filter output signal as the error signal, determine the current step size based on the error signal, and update the weight vector based on the current step size;

[0012] The updated weight vector is used to iteratively filter the subsequent sampled data, and the filtered main transformer monitoring signal is output.

[0013] Furthermore, the step of acquiring the current sampling data and historical sampling data of the main transformer monitoring signal includes:

[0014] Collect the transformer grounding neutral point current signal, the tank surface vibration signal, and the noise signal around the transformer;

[0015] Each acquired channel signal is marked with a unified timestamp, and the time of each channel signal is aligned based on the unified timestamp;

[0016] Based on the time-aligned signals of each channel, the current sampling data and historical sampling data of each channel signal are extracted.

[0017] Furthermore, the step of solving the model parameters of the autoregressive prediction model based on historical sampling data, and calculating the reference signal based on the model parameters and historical sampling data, includes:

[0018] Calculate the sample autocorrelation coefficients corresponding to different lag orders based on historical sampling data;

[0019] Based on the sample autocorrelation coefficient, establish a set of autoregressive parameter equations corresponding to the order of candidate models, and solve for the model parameters corresponding to the order of each candidate model.

[0020] The target model order is determined based on the residual variance and model complexity corresponding to the order of each candidate model, and the model parameters corresponding to the target model order are used as the model parameters of the current autoregressive prediction model.

[0021] The reference signal is calculated based on the model parameters of the current autoregressive prediction model and the historical sampling data prior to the current time.

[0022] Further, the step of using the model parameters corresponding to the target model order as the model parameters of the current autoregressive prediction model includes:

[0023] Determine the current operating segment based on the train timetable;

[0024] Call the historical sampling data corresponding to the runtime segment to solve for the target model order and model parameters corresponding to the runtime segment;

[0025] The target model order and model parameters corresponding to the runtime segment are used as the model order and model parameters of the current autoregressive prediction model.

[0026] Further, the step of constructing the input vector of the adaptive filter based on the current sampled data, and calculating the filter output signal based on the input vector and the current weight vector, includes:

[0027] Obtain the current sampled data at the current moment and the historical sampled data at a preset number of moments prior to the current moment;

[0028] Arrange the current sampled data and historical sampled data in chronological order to form the input vector of the adaptive filter;

[0029] Call the current weight vector corresponding to the input vector;

[0030] The filter output signal is calculated based on the product of the input vector and the current weight vector.

[0031] Further, the step of calculating the difference between the reference signal and the filter output signal as an error signal, determining the current step size based on the error signal, and updating the weight vector based on the current step size includes:

[0032] Calculate the difference between the reference signal and the filter output signal at the current moment, and use it as the error signal at the current moment;

[0033] The error signal is input into a preset nonlinear mapping function to obtain a mapping value; wherein, the nonlinear mapping function increases the mapping value and slows down the change when the absolute value of the error signal increases, and decreases the mapping value more rapidly when the absolute value of the error signal decreases; wherein, the nonlinear mapping function is constructed based on the sigmoid function, and the absolute value of the error signal acts on the nonlinear mapping function in the form of a cube power;

[0034] Obtain the first adjustment coefficient and the second adjustment coefficient; wherein, the first adjustment coefficient is the square of the ratio of the error signal to the error signal at the previous time, and the second adjustment coefficient is the product of the second adjustment coefficient at the previous time and the weighting coefficient, plus the sum of the product of the difference between the absolute values ​​of the error signal and the error signal at the previous time and the complementary weighting coefficient.

[0035] The first and second adjustment coefficients are calculated with the mapping value to obtain the current step size;

[0036] Based on the current step size, error signal, and input vector, calculate the weight vector update amount, add the current weight vector to the weight vector update amount, and obtain the updated weight vector.

[0037] Furthermore, the step of iteratively filtering subsequent sampled data based on the updated weight vector and outputting the filtered main transformer monitoring signal includes:

[0038] Set the updated weight vector as the current weight vector at the next time step;

[0039] The input vector of the adaptive filter is reconstructed based on the current sampling data and historical sampling data at the next time step, and the filter output signal at the next time step is calculated based on the input vector and the current weight vector at the next time step.

[0040] The reference signal for the next time step is calculated based on the historical sampling data for the next time step and the model parameters of the current autoregressive prediction model.

[0041] Calculate the difference between the reference signal and the filter output signal at the next time step, use it as the error signal at the next time step, and determine the step size and update the weight vector at the next time step based on the error signal at the next time step.

[0042] Repeat the above steps until all the sampled data to be processed has been traversed, and output the filtered main transformer monitoring signal;

[0043] Specifically, iterative filtering is performed on the transformer grounding neutral point current signal, the tank surface vibration signal, and the noise signal around the transformer, and the corresponding filtering results are output respectively.

[0044] Secondly, the present invention provides a main transformer monitoring signal filtering device, comprising:

[0045] The acquisition module is used to acquire the current and historical sampling data of the main transformer monitoring signals;

[0046] The reference signal calculation module is used to solve the model parameters of the autoregressive prediction model based on historical sampling data, and to calculate the reference signal based on the model parameters and historical sampling data.

[0047] The filter output signal calculation module is used to construct the input vector of the adaptive filter based on the current sampled data, and to calculate the filter output signal based on the input vector and the current weight vector.

[0048] The weight vector update module is used to calculate the difference between the reference signal and the filter output signal as an error signal, determine the current step size based on the error signal, and update the weight vector based on the current step size.

[0049] The output module is used to iteratively filter subsequent sampled data based on the updated weight vector and output the filtered main transformer monitoring signal.

[0050] Thirdly, the present invention provides an electronic device, comprising: a memory, and one or more processors communicatively connected to the memory; the memory stores instructions executable by the one or more processors, the instructions being executed by the one or more processors to cause the one or more processors to implement the method described above.

[0051] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0052] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0053] This invention tracks the steady-state periodicity of the main transformer monitoring signal using an autoregressive prediction model and generates a prediction reference signal. Simultaneously, it uses a variable-step-size adaptive minimum mean square filter to handle random fluctuations and noise interference. The prediction output of the autoregressive prediction model is used as the reference signal for the adaptive filter. The step size is dynamically adjusted and the weight vector is iteratively updated based on the error between the reference signal and the filter output signal. This allows the filter to maintain a fast convergence speed when the error is large and quickly stabilize when the error is small, thus balancing convergence speed and steady-state accuracy. It effectively extracts non-stationary monitoring signals with low-frequency alternating positive and negative frequencies, strong random fluctuations, and high noise interference, avoiding the loss of transient impact information and the suppression of effective low-frequency alternating information. Attached Figure Description

[0054] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0055] Figure 1 A flowchart illustrating a method for filtering main transformer monitoring signals provided in the embodiments of this specification;

[0056] Figure 2 This is a structural diagram of the adaptive filter provided in the embodiments of this specification;

[0057] Figure 3 A flowchart illustrating a method for filtering main transformer monitoring signals based on autoregressive prediction and variable step-size adaptive LMS filtering, as provided in the embodiments of this specification;

[0058] Figure 4 This is a block diagram of a main transformer monitoring signal filtering device provided in the embodiments of this specification;

[0059] Figure 5 This is a block diagram of an electronic device provided in the embodiments of this specification. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0061] In urban rail transit power supply systems, trains often use DC traction power supply, with the rails serving as return current channels. Under these operating conditions, the traction current is affected by factors such as track insulation status, train operating conditions, grounding conditions, and the conductivity of the surrounding soil during its return flow. Some current deviates from the intended return path and diffuses into the surrounding soil and nearby power facilities, forming stray currents. When stray currents intrude into the main transformer, they cause changes in monitoring signals such as the transformer's grounding neutral point current, transformer surface vibration, and surrounding noise. Therefore, related technologies often analyze the main transformer's operating status by collecting these monitoring signals and then perform subsequent bias assessment, status identification, or anomaly monitoring based on the collected signals.

[0062] In related technologies, transformer monitoring signal processing techniques are mostly applied to scenarios where ground current changes are relatively slow and disturbance types are relatively simple. Possible processing methods include low-pass filtering, wavelet transform, and other filtering methods based on fixed parameters. These methods often aim to suppress high-frequency noise or extract features of a specific frequency band, achieving certain results when the signal statistical characteristics are relatively stable. However, in scenarios where stray currents from urban rail transit intrude into transformers, the mechanism of transformer monitoring signal changes differs significantly from the aforementioned scenarios. On the one hand, train starts, stops, accelerations, decelerations, and load changes cause stray currents to exhibit obvious time-varying characteristics, resulting in low-frequency positive and negative alternation in the transformer monitoring signal. On the other hand, changes in track discharge paths, grounding conditions, and operating conditions cause strong random fluctuations to superimpose on the monitoring signal. Simultaneously, mechanical vibrations, electromagnetic interference, and environmental noise exist in the field acquisition environment, giving the monitoring signal a high noise level. Therefore, the relevant monitoring signal exhibits non-stationary characteristics with low-frequency alternation, random fluctuations, and noise interference coexisting.

[0063] In the aforementioned non-stationary scenarios, conventional filtering methods in related technologies suffer from insufficient adaptability. While low-pass filtering can suppress high-frequency noise to some extent, its fixed parameters easily weaken transient components caused by stray currents while filtering out interference, resulting in the loss of transient impact information. Decomposition methods such as wavelet transform can process components at different scales, but when the signal exhibits strong random fluctuations and low-frequency alternating components overlap with interference components, effective low-frequency information is easily masked by noise or the reconstructed signal becomes distorted. Furthermore, because related technologies often lack dynamic modeling mechanisms for the cyclical characteristics of rail transit operations and mechanisms that can adaptively adjust filtering parameters based on error changes, it is difficult to simultaneously achieve both steady-state tracking and suppression of random disturbances.

[0064] Further analysis reveals that the aforementioned technical problems arise because the main transformer monitoring signal under the influence of stray currents from urban rail transit is not a simple superposition of noise, but simultaneously includes regular components related to train operation rhythms, random fluctuations influenced by on-site operating conditions, and environmental noise components. If only fixed-parameter filtering methods are used, it is difficult to adjust the processing strategy in a timely manner according to changes in signal state. If only the extraction of a certain frequency band is emphasized, it is easy to overlook other aspects, leading to a difficulty in balancing regular low-frequency alternating information, transient response information, and noise suppression requirements. Therefore, in related technologies, in scenarios where stray currents from urban rail transit intrude into the main transformer, existing transformer monitoring signal filtering methods face the technical challenge of effectively filtering non-stationary monitoring signals that coexist with low-frequency positive and negative alternations, strong random fluctuations, and high noise interference.

[0065] To address the aforementioned problems, the inventive concept of this invention lies in combining an autoregressive prediction mechanism that generates reference signals based on historical sampling data with a filtering mechanism that adaptively adjusts the step size based on error signals to collaboratively process the regular and random disturbance components in the main transformer monitoring signal, thereby achieving effective filtering of the non-stationary monitoring signal.

[0066] like Figure 1 As shown, this embodiment provides a method for filtering main transformer monitoring signals. The execution subject of this method can be a main transformer online monitoring device, a main transformer monitoring terminal device, an industrial control device, an edge computing device, a server, or a main transformer monitoring system composed of signal acquisition device and signal processing device. The execution subject is used to acquire the current sampling data and historical sampling data of the main transformer monitoring signal, and perform autoregressive prediction, reference signal calculation, adaptive filtering, error signal calculation, step size update, weight update, and iterative filtering output.

[0067] The method may include:

[0068] Step S10: Obtain the current sampling data and historical sampling data of the main transformer monitoring signal.

[0069] In this embodiment, the acquisition action can be represented as the executing entity obtaining data from the monitored signal source that can be used for subsequent processing, and organizing the acquired data into a data format suitable for prediction and filtering. The acquisition can be either direct collection or reading from a buffer, database, message queue, or other storage medium.

[0070] In this embodiment, the main transformer monitoring signal can be an electrical and physical quantity signal reflecting the operating status of the main transformer, specifically including the transformer grounding neutral point current signal, the transformer tank surface vibration signal, and the noise signal around the transformer. Specifically, the current sampled data can be the data value obtained at the latest sampling time after discretizing the main transformer monitoring signal. The historical sampled data can be a sequence of data values ​​from multiple consecutive moments collected and stored sequentially at fixed sampling intervals before the current sampling time.

[0071] In one possible and specific implementation, the transformer grounding neutral point current signal can be acquired using a current sensor nested within the transformer grounding neutral point. This current sensor can be a Hall effect current sensor, with an electromagnetic shield on its outer side to suppress the influence of external electromagnetic fields on the current measurement. The transformer tank surface vibration signal can be acquired using an accelerometer fixed to the surface of the transformer tank. The accelerometer's installation position can avoid the tank's reinforcing structure to ensure that the collected vibration signal reflects the mechanical vibration characteristics of the transformer core and windings. The noise signal around the transformer can be acquired using a noise sensor positioned within one to two meters of the transformer and facing the transformer tank. This noise sensor can be equipped with a windproof cover to reduce interference from ambient airflow on the acoustic measurements.

[0072] In one possible and specific implementation, the sampling process for the three types of signals can be synchronized using the same clock source and the same trigger source. After analog-to-digital conversion, the analog input signals of each channel are marked with a high-precision Coordinated Universal Time (UTC) data timestamp by the BeiDou time synchronization module, achieving strict alignment of the multi-channel signals in the time dimension. Based on the time-aligned signals of each channel, the data value of the latest sampling moment is extracted from the data buffer as the current sampling data, and the data values ​​of multiple consecutive sampling moments before the latest sampling moment are extracted as historical sampling data. The data length of the historical sampling data covers at least three cycles of the signal; in one implementation, sampling data of ten minutes in length can be used.

[0073] Step S12: Solve for the model parameters of the autoregressive prediction model based on historical sampling data, and calculate the reference signal based on the model parameters and historical sampling data.

[0074] In this embodiment, the model parameters for solving the autoregressive prediction model can be represented as a set of parameters determined based on the temporal correlation between adjacent sampled values ​​in historical sampled data to characterize the mapping relationship between the current signal and past signals.

[0075] In this embodiment, the autoregressive prediction model can be represented as a model that uses the sampled values ​​of the same signal at multiple previous times to predict the signal value at the current time or the next time.

[0076] In this embodiment, the model parameters can be a set of parameters used to represent the degree of influence of each historical sample value on the prediction result.

[0077] In this embodiment, the autoregressive prediction model may specifically include models of different orders, i.e., using different numbers of historical sampled values ​​for prediction. It may also include models with different update frequencies, for example, updating model parameters at a fixed period or updating model parameters when changes occur during runtime. The model parameters may be a single set of parameters or multiple sets of parameters stored separately for different runtime periods. The reference signal may correspond to the predicted value at the current moment, or to the predicted value at the next or subsequent moment, as long as it maintains consistency with the temporal relationship of subsequent error calculations.

[0078] In one possible and specific implementation, the sample autocorrelation coefficients corresponding to different lag orders can be calculated based on historical sampling data. Specifically, the lag order can be the number of sampling intervals between two data points. The sample autocorrelation coefficient can be obtained by multiplying the pairs of data points in the historical sampling data that are separated by that lag order and taking the average. Based on the obtained sample autocorrelation coefficients, a system of autoregressive parameter equations corresponding to the candidate model orders is established. This system of equations uses the sample autocorrelation coefficients to form a coefficient matrix and the sample autocorrelation coefficients corresponding to the lag orders to form a right-hand vector. By solving this system of linear equations, the model parameters corresponding to each candidate model order are obtained.

[0079] In one possible and specific implementation, candidate model orders can be sequentially assigned integers from one to ten. The model parameters corresponding to each order are then calculated, and the residual variance for each order is also calculated. The residual variance characterizes the dispersion of the difference between the model's predicted values ​​and the actual historical sampled data. The target model order is determined based on the residual variance and model complexity corresponding to each candidate model order. Model complexity can be represented by the logarithmic product of the model order and the amount of historical sampled data. The fitting term representing the residual variance is added to the model complexity to obtain an evaluation index. The order corresponding to the minimum of this evaluation index is selected as the target model order, and the model parameters corresponding to this target model order are used as the model parameters of the current autoregressive prediction model.

[0080] In one possible and specific implementation, the monitoring period can be divided into multiple operating phases according to the train timetable. For each operating phase, the corresponding target model order and model parameters are calculated using historical sampling data within that phase and saved. At the start time of each phase specified in the timetable, the target model order and model parameters for that phase are periodically triggered and loaded as the model order and model parameters of the current autoregressive prediction model. The reference signal for the current moment is obtained by weighted summing of the model parameters of the current autoregressive prediction model and the historical sampling data before the current moment.

[0081] Step S14: Construct the input vector of the adaptive filter based on the current sampled data, and calculate the filter output signal based on the input vector and the current weight vector.

[0082] In this embodiment, the input vector for constructing the adaptive filter can be a data structure that organizes the current sampled data and its related historical sampled data in a preset order for filter operation. Specifically, the input vector can be an ordered data set composed of sampled values ​​from the current time and multiple time points prior to the current time in chronological order. The current weight vector can be a set of weights that corresponds one-to-one with the data at each position of the input vector and is used to participate in the filtering operation.

[0083] Specifically, the input vector can be a vector of different lengths, or its dimensions can be set according to different signal types. It can consist of time-series sampled values ​​from a single channel, or it can be expanded to an input set including multiple feature components. The current weight vector can be the default weights obtained from initialization, or it can be the weights updated after multiple iterations in the preceding rounds. The filter output signal can be the output at a single current time step, or it can be the output results calculated for multiple time steps in a batch processing scenario.

[0084] In one possible and specific implementation, after obtaining the current sampled data at the current moment, the executing entity can extract the sampled values ​​from a preset number of time points prior to the current moment from the historical sampled data window, and arrange the current sampled data and the historical sampled values ​​in chronological order to form the input vector of the adaptive filter. Then, the executing entity obtains the current weight vector corresponding to the input vector. The current weight vector can be set to a preset initial value at the beginning of the method execution, or it can be directly used as the current weight vector after the processing at the previous time point is completed. Afterwards, the executing entity can perform vector operations based on the input vector and the current weight vector to obtain the filter output signal.

[0085] In one possible and specific implementation, for different types of monitoring signals, separate input vectors can be constructed and separate weight vectors can be called to obtain the filter output signal corresponding to each signal. For the same type of signal, the input vector at different times changes with the update of current and historical sampling data, and the current weight vector is also continuously updated with the iteration process. Therefore, the filter output signal will adaptively change with the signal state.

[0086] Step S16: Calculate the difference between the reference signal and the filter output signal as the error signal, determine the current step size based on the error signal, and update the weight vector based on the current step size.

[0087] In this embodiment, the error signal can be the numerical difference between the reference signal and the filter output signal, used to measure the degree of deviation of the filter's current output from the steady-state predicted value. The current step size can be a positive coefficient used to control the magnitude of the weight vector update.

[0088] In one possible and specific implementation, the executing entity first calculates the difference between the reference signal and the filter output signal at the current moment, using this difference as the error signal at the current moment. Then, the executing entity inputs the error signal into a preset nonlinear mapping function to obtain a mapping value corresponding to the magnitude of the error. The nonlinear mapping function can adopt an S-shaped variation characteristic, such that when the absolute value of the error signal is large, the mapping value increases with the increase of the error, but the rate of increase gradually slows down. When the absolute value of the error signal decreases, the mapping value decreases accordingly, and the rate of decrease accelerates. In this way, the step size can be kept at a high level during the period of large error and rapidly reduced during the period of small error.

[0089] Furthermore, the executing entity can also determine a first adjustment coefficient and a second adjustment coefficient. The first adjustment coefficient reflects the degree of change of the current error signal relative to the error signal at the previous moment, and is used to adjust the response speed of the nonlinear mapping function to changes in the error signal. The second adjustment coefficient reflects the change between the absolute value of the current error signal and the absolute value of the error signal at the previous moment, and is used to adjust the range of the step size. The executing entity can apply the first and second adjustment coefficients to the mapping value to obtain the current step size. The application method can be scaling, weighting, amplifying, compressing, or other equivalent processing of the mapping value, as long as it can achieve the adjustment of the rate of change and range of the current step size.

[0090] Finally, the executing entity can calculate the weight vector update amount based on the current step size, error signal, and input vector, and combine this update amount with the current weight vector to obtain the updated weight vector. The updated weight vector is used for filtering calculation at the next time step. In one embodiment, the first adjustment coefficient can be determined by the ratio of the error signal at the current time step to the error signal at the previous time step, and the second adjustment coefficient can be recursively determined by the second adjustment coefficient at the previous time step and the change in the absolute value of the error signals at the current and previous time steps.

[0091] Step S18: Iteratively filter the subsequent sampled data based on the updated weight vector, and output the filtered main transformer monitoring signal.

[0092] In this embodiment, the iterative filtering can use the updated weight vector at the current time as the current weight vector at subsequent time points. For subsequent sampled data, the processes of input vector construction, output signal calculation, reference signal calculation, error signal calculation, step size determination, and weight update are repeatedly executed to continuously generate the filtering results corresponding to each time point. The filtered main transformer monitoring signal can be the output signal obtained after the above iterative processing, and whose noise and random disturbances are suppressed compared to the original monitoring signal. The output can be a real-time signal output at each time point, or a continuous monitoring signal sequence composed of multiple time points.

[0093] In this embodiment, the iterative filtering can include online real-time processing and offline batch processing. In online real-time processing, at each new sampling time, the execution entity uses the currently updated weight vector as the current weight vector for the next time step and performs a complete filtering process on the newly arrived sampled data, outputting the filter output signal corresponding to the current time step in real time. In offline batch processing, the execution entity can iterate through a continuous range of sampled data time-by-time, completing the filtering process for each time step sequentially, and outputting a monitoring signal sequence composed of the filter output signals for each time step after the iteration is complete.

[0094] In one possible and specific implementation, the executing entity can set the updated weight vector as the current weight vector for the next time step and obtain the current sampled data and its corresponding historical sampled data for the next time step. Then, the executing entity can reconstruct the input vector for the next time step, calculate the reference signal for the next time step based on the model parameters of the current autoregressive prediction model and the historical sampled data before the next time step, calculate the filter output signal for the next time step based on the input vector and the current weight vector, and form the error signal for the next time step based on the difference between the two. Next, the executing entity can determine the step size for the next time step based on the error signal and update the weight vector for the next time step. By continuously repeating the above process until all the sampled data to be processed has been traversed, the filtering result in the continuous time domain is obtained.

[0095] In one possible and specific implementation, the above iterative filtering process can be performed separately for the transformer grounding neutral point current signal, the tank surface vibration signal, and the noise signal around the transformer to obtain the corresponding filtering results. The executing entity can output the filtered signals of each channel separately, or it can store the filtering results of each channel together for subsequent bias magnetization assessment, status identification, or anomaly detection.

[0096] In some implementations, the step of acquiring the current sampling data and historical sampling data of the main transformer monitoring signal includes:

[0097] Step S102: Collect the transformer grounding neutral point current signal, the tank surface vibration signal, and the noise signal around the transformer.

[0098] In this embodiment, the data acquisition can be achieved by obtaining raw monitoring data that characterizes the operating status of the main transformer through a corresponding signal acquisition channel. Specifically, the transformer grounding neutral point current signal can be a signal characterizing the change in the main transformer grounding neutral point current. The transformer enclosure surface vibration signal can be a signal characterizing the mechanical vibration state of the main transformer enclosure. The transformer surrounding noise signal can be a signal characterizing the acoustic changes during the operation of the main transformer. The above three types of signals can respectively reflect the electrical response, mechanical response, and acoustic response of the main transformer after being affected by stray currents.

[0099] In this embodiment, the data acquisition can include different forms such as analog signal acquisition, digital signal acquisition, continuous acquisition, periodic acquisition, and triggered acquisition. The acquisition can be performed by a single monitoring device or by multiple acquisition units performing the acquisition separately before uploading to the same processing device. For signals from different channels, independent sampling links or integrated multi-channel sampling links can be used.

[0100] In this embodiment, the executing entity can acquire the transformer grounding neutral point current signal through the current acquisition channel, the vibration signal of the tank surface through the vibration acquisition channel, and the noise signal around the transformer through the sound acquisition channel. Each acquisition channel can be connected to a corresponding sensor or signal conversion module to convert the measured physical quantity into processable data. The acquired data can be raw waveform data or sampled data after front-end conditioning. The front-end conditioning may include amplification, isolation, analog-to-digital conversion, and preliminary removal of outliers, as long as it does not affect the subsequent filtering method's recognition of the original trend.

[0101] In this embodiment, the three types of signals can be acquired synchronously or separately according to a unified scheduling cycle and then processed uniformly. Regarding the sampling frequency, the same sampling frequency can be used for all three types of signals, or different sampling frequencies can be used based on the changing characteristics of each type of signal, followed by alignment processing under a unified time reference in subsequent steps. As for the acquisition duration, it can be long-term continuous acquisition or acquisition within a preset monitoring period.

[0102] Step S104: Mark each channel signal with a unified timestamp, and perform time alignment of each channel signal based on the unified timestamp.

[0103] In this embodiment, the unified timestamp can be a time identifier that represents the sampling time added to signal data acquired from different channels based on the same time reference. Specifically, the unified timestamp may include an absolute time identifier, a relative time identifier, a time identifier corresponding to the sampling sequence number, or other identifier forms that can uniformly represent the chronological relationship of sampling times.

[0104] In this embodiment, the time alignment can be achieved by establishing a correspondence between different channel signals according to a unified time base, so that the data of each channel corresponds to the same actual sampling time or the same target time window at the same processing time. Specifically, it can include directly matching according to the same timestamp, merging according to a preset time window, performing interpolation to fill in missing times, performing downsampling on high-frequency channels, performing resampling on low-frequency channels, or other equivalent time unification processing methods.

[0105] In one possible and specific implementation, after the signal acquisition for each channel is completed, the executing entity can read the sampling time information corresponding to each sampling point and convert this sampling time information into a timestamp in a unified format. Then, the executing entity can match the sampling data from the current signal channel, vibration signal channel, and noise signal channel according to the timestamps to determine the corresponding sampling points for each channel at the same time. When a channel lacks data at the target time, the executing entity can select the sampling value closest to the target time from the data of nearby times, or estimate adjacent sampling values ​​to form the data corresponding to the target time.

[0106] In one possible and specific implementation, if there are differences in the sampling frequencies of different channels, the implementing entity can first map the data of each channel to a unified time scale. For example, for channels with higher sampling frequencies, target sampling points can be selected according to a unified time interval. For channels with lower sampling frequencies, the corresponding values ​​under the unified time interval can be estimated based on neighboring sampling points. After completing the above processing, each channel signal can be mapped to the same set of time positions under a unified time series, thereby obtaining a time-aligned multi-channel signal.

[0107] Step S106: Based on the time-aligned signals of each channel, extract the current sampling data and historical sampling data of each channel signal.

[0108] In one possible and specific implementation, the executing entity determines the current processing moment on a unified timeline and reads the corresponding sampled values ​​from the current signal channel, vibration signal channel, and noise signal channel at that moment, respectively, as the current sampled data for each channel. Then, the executing entity can trace back a preset number of sampling moments along the unified timeline, reading the corresponding sampled values ​​from each channel and assembling the historical sampled data for each channel in chronological order. For each channel, a set of current sampled data and a set of historical sampled data can be generated for subsequent prediction and filtering processing of the signals from each channel.

[0109] In one possible and specific implementation, historical sampling data can be maintained using a sliding window approach. Specifically, when a new sampling time arrives, the executing entity incorporates the new current sampled value into the data window of the corresponding channel and removes the data from the window at the earliest time, ensuring that the historical sampling data always maintains a preset length. In this way, a set of historical sampling data corresponding to the current sampled data can be formed at each processing time. The preset length can be set according to the order requirements of the subsequent prediction model, the length requirements of the adaptive filter input vector, or the characteristics of changes in the field monitored signal.

[0110] In one possible and specific implementation, the length of the historical sampling data can be the same or different for different channel signals. For example, a first historical window length can be set for current signals, a second historical window length for vibration signals, and a third historical window length for noise signals to adapt to the changing characteristics of different signals.

[0111] In some implementations, the step of solving for the model parameters of the autoregressive prediction model based on historical sampling data, and calculating the reference signal based on the model parameters and historical sampling data, includes:

[0112] Step S122: Calculate the sample autocorrelation coefficients corresponding to different lag orders based on historical sampling data.

[0113] In this embodiment, the calculation of the sample autocorrelation coefficient can be based on the correlation between the current sampled value and the sampled values ​​at previous multiple time points in historical sampling data, determining the temporal correlation of the signal under different lag conditions. The sample autocorrelation coefficient can reflect the degree of correlation between the same signal sequence at different time points.

[0114] In this embodiment, the different lag orders can refer to the time interval levels between the current time and the previous time, the two previous times, the three previous times, and even earlier times.

[0115] In one possible and specific implementation, the executing entity can read multiple consecutive sample values ​​from historical sampling data prior to the current moment in chronological order, and sequentially calculate the correlation of these historical sample values ​​under different lags. For example, the executing entity can first determine a candidate lag order range, and then separately calculate the correlation between the historical sample data and the data at the corresponding lag positions, thereby obtaining the sample autocorrelation coefficients corresponding to multiple lag orders. In this way, a correlation data foundation can be formed for subsequently establishing an autoregressive parametric equation system.

[0116] In one possible and specific implementation, the sample autocorrelation coefficients can be calculated separately for different types of main transformer monitoring signals. For example, for the transformer grounding neutral point current signal, a set of sample autocorrelation coefficients can be calculated based on its historical sampling sequence. The corresponding sample autocorrelation coefficients for the tank surface vibration signal and the noise signal around the transformer can also be calculated in the same way. The range of hysteresis orders used can be the same for different signals, or it can be set separately according to the variation characteristics of each signal.

[0117] Step S124: Establish a set of autoregressive parameter equations corresponding to the order of candidate models based on the sample autocorrelation coefficients, and solve for the model parameters corresponding to the order of each candidate model.

[0118] In this embodiment, the establishment of the autoregressive parameter equation set can be based on the correlation of historical sampling data at multiple lag orders, constructing a set of data relationships for solving the autoregressive model parameters.

[0119] In this embodiment, the autoregressive parametric equation system can be a set of linear equations with the unknown model parameters as unknowns, the sample autocorrelation coefficients as known coefficients, and the right-hand side as terms. Specifically, in this equation system, the left-hand side of the i-th equation is the sum of the product of each unknown model parameter and the corresponding sample autocorrelation coefficient determined by the equation number and the unknown parameter number, and the right-hand side is the sample autocorrelation coefficient for a specific lag order corresponding to the equation number. The coefficient matrix of the equation system has a Toplitz structure, that is, the elements on the same diagonal of the matrix have the same value, which corresponds to the sample autocorrelation coefficient for a specific lag order, and the elements on the main diagonal are the sample autocorrelation coefficients when the lag order is zero.

[0120] In this embodiment, the candidate model order can be any possible autoregressive model order value within a preset search range. The lower limit of the preset search range is one, and the upper limit is a preset maximum order. The executing entity can process each integer order value within this range as a candidate model order.

[0121] In one possible and specific implementation, the executing entity can divide the preset search range into multiple candidate orders. For each candidate order p, the executing entity operates as follows: First, it extracts the sample autocorrelation coefficient values ​​with lag orders from zero to p from the sample autocorrelation array obtained in step S122. Second, it constructs a system of linear equations consisting of p equations using the extracted sample autocorrelation coefficients according to the aforementioned Toplitz structure, where the unknowns of the system are p model parameters. In this system of equations, the element in the i-th row and j-th column of the coefficient matrix is ​​the sample autocorrelation coefficient corresponding to the absolute value of the lag order i minus j, and the i-th element of the right-hand vector is the sample autocorrelation coefficient corresponding to the lag order i. Third, the executing entity calls a linear equations solver to solve the system of equations. The solving algorithm can employ direct methods such as the Levinson-Druping recursive algorithm, the Doolittle decomposition method, or the Cholisky decomposition method. The Levinson-Durbin recursive algorithm utilizes the Toplitz structure of the coefficient matrix to progressively derive the model parameters of each order from low to high order in a recursive manner. Its computational complexity is proportional to the square of the order. After solving, the executing entity obtains a set of model parameters corresponding to the candidate order p, including p coefficient values. Simultaneously, the executing entity calculates the residual variance corresponding to this set of model parameters. Specifically, the residual variance is calculated by subtracting the sum of the products of the sample autocorrelation coefficients for each model parameter and the corresponding lag order from the sample autocorrelation coefficients when the lag order is zero; the difference is the estimated residual variance. The executing entity can save the model parameters and residual variances corresponding to each candidate order separately, forming a candidate model list.

[0122] Step S126: Determine the target model order based on the residual variance and model complexity corresponding to each candidate model order, and use the model parameters corresponding to the target model order as the model parameters of the current autoregressive prediction model.

[0123] In this embodiment, the target model order can be the one that is finally selected from the candidate model orders and is the optimal order used for autoregressive prediction at the current time.

[0124] In this embodiment, the residual variance can be expressed as the degree of dispersion of the deviation between the model prediction result and the actual historical sampling data, given the order of the candidate model and the corresponding model parameters.

[0125] In this implementation, it is understood that model complexity reflects the structural complexity of the model; a higher value indicates a more complex model. If the model order is chosen solely to minimize residual variance, it may result in an overly high order, causing the model to overfit the random noise components in historical data and thus reducing the accuracy of future predictions. Conversely, if the goal is solely to minimize model complexity, it may result in an underly low order, preventing the model from fully capturing the temporal dependencies in the signal. Therefore, determining the target model order requires a comprehensive consideration of both factors.

[0126] In one possible and specific implementation, the executing entity can iterate through each candidate order in the candidate model list generated in step S124 and calculate a comprehensive evaluation index for each candidate order. Specifically, the product of the natural logarithm of the number of historical sampling data points N and the model order p can be taken as a penalty term, and the product of N and the natural logarithm of the residual variance can be taken as a fitting term. The fitting term and the penalty term are added together to obtain the information criterion value corresponding to the candidate order. When the candidate order is small, the large residual variance leads to the fitting term dominating. When the candidate order is large, the increasing model order leads to a gradual increase in the penalty term. During the traversal from low to high order, the information criterion value often shows a trend of first decreasing and then increasing, and the candidate order corresponding to its minimum value is the target model order. The executing entity can compare the information criterion values ​​corresponding to all candidate orders and select the candidate order with the smallest information criterion value as the target model order.

[0127] Step S128: Calculate the reference signal based on the model parameters of the current autoregressive prediction model and the historical sampling data before the current time.

[0128] In this embodiment, the reference signal can be a linear prediction estimate made by an autoregressive prediction model based on the solved model parameters and historical sampling data at corresponding lag times, for the components of the signal value at the current time that conform to a periodic steady-state law. This prediction estimate can be an inference of the expected value of the current signal formed by weighted combination of model parameters based on historical observations.

[0129] In one possible and specific implementation, after determining the model parameters of the current autoregressive prediction model at the current moment, the executing entity can read a number of historical sampled values ​​corresponding to the order of the target model from the historical sampled data prior to the current moment, and perform combination operations on these historical sampled values ​​according to the model parameters to obtain the reference signal at the current moment. This reference signal can be used as the prediction result of the signal change trend at the current moment, and compared with the filter output signal in subsequent steps to form an error signal.

[0130] In one possible and specific implementation, for different types of main transformer monitoring signals, reference signals can be calculated separately according to their respective current autoregressive prediction models and corresponding historical sampling data. For example, the executing entity can calculate a first reference signal for the transformer grounding neutral point current signal, a second reference signal for the tank surface vibration signal, and a third reference signal for the noise signal around the transformer. Each reference signal can then be fed into a subsequent adaptive filtering process to obtain the filtering result for the corresponding signal.

[0131] In one possible and specific implementation, if the model parameters of the current autoregressive prediction model are managed in a time-segmented manner, the executing entity first calls the corresponding model parameters according to the current runtime segment before calculating the reference signal, and then uses these model parameters and historical sampling data before the current moment to calculate the reference signal. This method allows the generation process of the reference signal to match the signal variation patterns under the current runtime segment.

[0132] In some implementations, the step of using the model parameters corresponding to the target model order as the model parameters of the current autoregressive prediction model includes:

[0133] Step S1262: Determine the current operating segment based on the train timetable.

[0134] In this embodiment, the train timetable can be a sequence of information representing the train's operation within a preset time range. The train timetable is a data set reflecting train departure time, stop time, interval travel time, operating density, operating cycle, or other time arrangement information related to the train's operating rhythm. The train timetable can be a pre-defined planned timetable or an adjusted timetable formed based on actual operating conditions.

[0135] In one possible and specific implementation, the executing entity can pre-store the train timetable and its corresponding running segment division information. The running segment division information can be stored as a time period list, a time interval table, or other data format that can characterize the start and end relationships of each time period. When the executing entity enters the model parameter determination process, it first obtains the current processing time and compares it with the start and end times of each running segment in the train timetable to determine the target running segment in which the current time falls. After determination, the executing entity can output the identifier of the target running segment for subsequent retrieval of historical sampling data and model parameters corresponding to that running segment.

[0136] Step S1264: Call the historical sampling data corresponding to the runtime segment to solve for the target model order and model parameters corresponding to the runtime segment.

[0137] In one possible and specific implementation, after determining the current operating segment, the executing entity can retrieve the historical sampling data corresponding to that operating segment from the historical data management module. For the transformer grounding neutral point current signal, the tank surface vibration signal, and the noise signal around the transformer, separate datasets corresponding to each operating segment can be established. The executing entity can then use the historical sampling data corresponding to the operating segment as input, and following the processing methods described in steps S122 to S126, calculate the sample autocorrelation information, model parameters, residual variance, and model complexity for different candidate model orders, and determine the target model order and model parameters for that operating segment.

[0138] In one possible and specific implementation, the invocation can be either a real-time invocation or a pre-invocation. A real-time invocation means that upon arrival of the current moment, the executing entity reads the corresponding historical sampling data from the storage area based on the current runtime segment and immediately completes the calculation of the model order and model parameters. A pre-invocation means that before system operation or during periodic maintenance, the executing entity pre-trains and stores the target model order and model parameters for each runtime segment, and directly reads the corresponding results upon arrival of the current moment.

[0139] Step S1266: Use the target model order and model parameters corresponding to the runtime segment as the model order and model parameters of the current autoregressive prediction model.

[0140] In this embodiment, the model order and model parameters of the current autoregressive prediction model can be the model order value and model parameter vector obtained by solving or looking up a table and written into the data area of ​​the device memory to store the working parameters of the current autoregressive model, so that it immediately takes effect on the reference signal calculation of the subsequent step S128.

[0141] In one possible and specific implementation, the execution entity can read the target model order value and the corresponding model parameter vector from the calculation result or lookup result of step S1264. The execution entity first writes the target model order value into the current value field of the model order register, and then writes each component of the model parameter vector sequentially into the corresponding address of the model parameter storage area, overwriting the previously used model parameter values. After writing, the execution entity sets a model update completion flag. Before each reference signal calculation, step S128 first checks whether this flag is set. If the flag is set, step S128 reads the latest values ​​from the model order register and the model parameter storage area to calculate the reference signal. If the flag is not set, the previous model configuration continues to be used. This mechanism ensures coordination and synchronization between the model parameter switching operation and the original filtering process, avoiding incomplete or inconsistent parameter values ​​read by step S128 during the model parameter writing process.

[0142] In one possible and specific implementation, the execution entity can re-execute the model result loading process when a runtime segment switches. While the current moment is still within the same runtime segment, the execution entity can continue to use the model order and model parameters corresponding to that runtime segment until it detects that the current moment has entered a new runtime segment, and then switch to the target model order and model parameters corresponding to the new runtime segment.

[0143] In one possible and specific implementation scheme, separate runtime model tables can be maintained for different types of main transformer monitoring signals. After determining the current runtime segment, the executing entity can retrieve the target model order and model parameters corresponding to the grounding neutral point current signal, the target model order and model parameters corresponding to the tank surface vibration signal, and the target model order and model parameters corresponding to the transformer surrounding noise signal, and use them as the model order and model parameters of the current autoregressive prediction model for the corresponding channel, respectively.

[0144] In some implementations, the step of constructing the input vector of the adaptive filter based on the current sampled data, and calculating the filter output signal based on the input vector and the current weight vector, includes:

[0145] Step S142: Obtain the current sampling data at the current moment and the historical sampling data at a preset number of moments prior to the current moment.

[0146] In this embodiment, the historical sampling data from a preset number of moments prior to the current moment can be a number of sample values ​​selected in chronological order before the current processing moment. The preset number is the number of historical sampling points used to determine the length of the input vector, and it can be a fixed value or a value dynamically set according to different signal types, different filtering configurations, or different operating conditions.

[0147] In one possible and specific implementation, when the execution entity enters the filtering process at the current moment, it can read the sampled value corresponding to the current moment from the sampling data buffer corresponding to the current channel as the current sampled data. Simultaneously, it reads the sampled values ​​corresponding to a preset number of sampling moments backward from this buffer as historical sampled data. The current sampled data and historical sampled data can originate from the time-series sampling sequence of the same channel. For example, when the current processing object is the transformer grounding neutral point current signal, the execution entity extracts the current sampled data and historical sampled data from this current signal sequence; when the current processing object is the vibration signal of the tank surface or the noise signal around the transformer, the data of the corresponding channel can be extracted separately in the same way.

[0148] In one possible and specific implementation, the preset number of historical sampling data points can be maintained by a sliding window. Specifically, after reaching a new sampling time, the executing entity writes the new sampling value into the data window and removes the data from the earliest time point in the window, keeping the window length at the preset value. Thus, at any given time, the executing entity can directly read the current sampling data and its corresponding historical sampling data from the data window. For different channel signals, the preset number can be the same or set separately to adapt to the temporal variation characteristics of different monitoring signals.

[0149] Step S144: Arrange the current sampled data and historical sampled data into the input vector of the adaptive filter according to the chronological order.

[0150] In one possible and specific implementation, the executing entity can arrange the current sampled data at the current moment and the historical sampled data from a preset number of moments prior to the current moment in a preset order to form the input vector corresponding to the current moment. If the arrangement method starts from the current moment, the current sampled data can be placed at one end of the input vector, and the remaining historical sampled data can be arranged in chronological order from most recent to oldest. If the arrangement method starts from an earlier moment, the earliest historical sampled data can be placed at one end of the input vector, and the current sampled data can be arranged at the other end.

[0151] In one possible and specific implementation, the components of the input vector can be the sampled values ​​themselves, or they can be preprocessed sampled values. For example, outliers can be removed, the units can be standardized, and the amplitude of the channel data can be normalized before constructing the input vector.

[0152] Step S146: Call the current weight vector corresponding to the input vector.

[0153] In this embodiment, the current weight vector can be a set of weights that actually participate in the filter output calculation at the current processing moment. Specifically, the current weight vector may include the initial weight vector set during the initialization phase, or it may include the weight vector obtained after iterative updates from the previous moment or several previous moments. It can be a set of weights corresponding to a single channel, or it can be multiple sets of weights maintained separately for multiple channels.

[0154] In one possible and specific implementation, after constructing the input vector for the current moment, the executing entity can read the current weight vector from the weight storage area, cache area, or model state management unit. If the current moment is the initial execution moment of the method, the preset initial weight vector is used as the current weight vector. If the current moment is not the initial execution moment, the weight vector updated in the previous moment is used as the current weight vector for the current moment.

[0155] Step S148: Calculate the filter output signal based on the product of the input vector and the current weight vector.

[0156] In one possible and specific implementation, the executing entity performs product processing on each data item in the input vector and the corresponding weight in the current weight vector, and combines all product results to obtain the filter output signal at the current time. This filter output signal represents the estimation result of the adaptive filter for the current input data under the current weight configuration. Then, this filter output signal can be compared with the reference signal in subsequent steps to form the error signal at the current time.

[0157] In one possible and specific implementation, for different types of main transformer monitoring signals, the respective filter output signals can be calculated according to their respective input vectors and current weight vectors. For example, the executing entity can calculate the first filter output signal based on the input vector and current weight vector of the transformer grounding neutral point current signal, the second filter output signal based on the input vector and current weight vector of the tank surface vibration signal, and the third filter output signal based on the input vector and current weight vector of the transformer surrounding noise signal. Each filter output signal can then be fed into the subsequent error calculation and weight update process.

[0158] In some implementations, the steps of calculating the difference between the reference signal and the filter output signal as an error signal, determining the current step size based on the error signal, and updating the weight vector based on the current step size include:

[0159] Step S162: Calculate the difference between the reference signal and the filter output signal at the current moment, and use it as the error signal at the current moment.

[0160] Understandably, when the filter output signal has not yet tracked the steady-state law represented by the reference signal, the error signal value is relatively large, indicating that the weight vector is still far from its optimal configuration. When the filter output signal is close to the reference signal, the error signal tends to fluctuate slightly near zero, indicating that the weight vector is close to convergence. When sudden random interference is superimposed on the input signal, this interference component will be reflected in the filter output signal, causing the error signal to suddenly increase, providing a transient response trigger for the step size adjustment mechanism.

[0161] In one possible and specific implementation, after obtaining the reference signal and the filter output signal at the current moment, the executing entity can perform a difference operation on the two to obtain the error signal at the current moment. This error signal can be stored in a temporary register area for subsequent nonlinear mapping, adjustment coefficient calculation, and weight updates. For continuous processing, a corresponding error signal can be obtained at each moment, and the error signal sequence can be used to characterize the output deviation changes of the filter at each processing moment.

[0162] Step S164: Input the error signal into a preset nonlinear mapping function to obtain a mapping value; wherein, the nonlinear mapping function increases the mapping value and slows down the change when the absolute value of the error signal increases, and accelerates the decrease of the mapping value when the absolute value of the error signal decreases; wherein, the nonlinear mapping function is constructed based on the sigmoid function, and the absolute value of the error signal acts on the nonlinear mapping function in the form of a cube power.

[0163] In this embodiment, the nonlinear mapping function can be a preset function with the error signal as the independent variable and the mapping value as the dependent variable, and its function curve generally exhibits an S-shaped characteristic. Specifically, when the absolute value of the error signal increases from zero, the mapping value gradually increases from a value approaching zero. The rate of increase is relatively slow when the absolute value of the error is small, accelerates when the absolute value of the error is in the middle range, and when the absolute value of the error continues to increase to a larger value, the mapping value continues to increase, but the rate of increase gradually slows down, and finally the mapping value approaches a preset upper limit value. Conversely, when the absolute value of the error signal gradually decreases from a larger value, the mapping value decreases accordingly, and accelerates its decrease when the absolute value of the error is small, rapidly approaching zero. This S-shaped response characteristic allows the step size to remain large when the error is large to provide sufficient convergence driving force, and rapidly decays to near zero when the error is small to ensure steady-state accuracy, avoiding the problem of traditional fixed-step-size filters being unable to balance convergence speed and steady-state error.

[0164] In this embodiment, the nonlinear mapping function can be constructed based on the sigmoid function. This embodiment modifies the standard sigmoid function to construct a nonlinear mapping function suitable for step size adjustment. More specifically, the modification involves using the absolute value of the error signal as a cube power as the input to the sigmoid function. The purpose of using a cube power is to give the function a steeper rate of change in the region where the absolute value of the error is close to zero. That is, as the error decreases further from a small value, the decay rate of the mapped value is faster, allowing the filter to shrink the step size to a minimum more quickly near the convergence stage, thus improving steady-state accuracy. Simultaneously, the cube power has a compression effect when the absolute value of the error is large, making the change of the mapped value more gradual in the region of large error, avoiding drastic step size jumps caused by small fluctuations in the error.

[0165] In one possible and specific implementation, the executing entity first reads the error signal at the current moment and extracts its absolute value information. Then, it performs a cube power operation on the absolute value and inputs the processing result into a nonlinear mapping function constructed based on the sigmoid function to obtain the mapping value corresponding to the current moment. This mapping value can be stored in the intermediate variable storage area corresponding to the current processing moment for use in subsequent step S168. For different channel signals, the executing entity can use the same nonlinear mapping function to process the corresponding error signal, or it can configure nonlinear mapping functions with different parameters for different channel signals.

[0166] Step S166: Obtain the first adjustment coefficient and the second adjustment coefficient; wherein, the first adjustment coefficient is the square of the ratio of the error signal to the error signal at the previous time, and the second adjustment coefficient is the product of the second adjustment coefficient at the previous time and the weighting coefficient, plus the sum of the product of the difference between the absolute values ​​of the error signal and the error signal at the previous time and the complementary weighting coefficient.

[0167] In this embodiment, the first adjustment coefficient can be used to adjust the rate of change of the nonlinear mapping function, that is, to control the steepness of the change of the mapping value as the error signal changes. The first adjustment coefficient can be used to quantify the relative rate of change of the error signal. When the error signal increases significantly relative to the previous moment, it indicates that the filter has encountered a sudden change or sudden interference in the signal's statistical characteristics, requiring accelerated convergence. At this time, the first adjustment coefficient increases accordingly, making the curve of the nonlinear mapping function steeper at the current position, and the same error increment will produce a larger increment in the mapping value, thereby accelerating the response speed of the step size. When the error signal decreases relative to the previous moment, it indicates that the filter is tending to converge. At this time, the first adjustment coefficient decreases accordingly, making the curve of the nonlinear mapping function flatter at the current position, and the step size shrinks smoothly.

[0168] In one possible and specific implementation, the first adjustment coefficient can be specifically determined by taking the ratio of the error signal at the current moment to the error signal at the previous moment, and then calculating the square of that ratio. It is understandable that the ratio is used because it reflects the relative magnitude of the error signal change, eliminating the influence of the absolute amplitude of the signal on the rate of change judgment, thus giving the adjustment coefficient uniform adaptability to signals with different amplitude levels. It is also understandable that the squaring operation is used because it further amplifies the increase in the adjustment coefficient when the error increases, making the filter respond more quickly to signal abrupt changes, while further compressing the adjustment coefficient value when the error decreases, making the filter's changes during the convergence phase smoother.

[0169] In this embodiment, the second adjustment coefficient is used to adjust the range of values ​​of the nonlinear mapping function, that is, to control the upper limit of the maximum amplitude that the mapping value can reach. Specifically, when a random burst of interference occurs in the signal, causing a sudden increase in the absolute value of the error signal, the second adjustment coefficient increases smoothly, appropriately expanding the upper limit of the step size, enabling the filter to make a larger-scale weight correction for sudden changes. During the stable operation phase of the signal, the absolute value of the error signal remains small and stable, and the second adjustment coefficient converges to a smaller value, compressing the range of values ​​of the step size, preventing unnecessary weight adjustments caused by residual small error fluctuations, and maintaining the steady-state accuracy of the filter.

[0170] In one possible and specific implementation, the second adjustment coefficient can be determined using a smooth recursive mechanism. Specifically, the second adjustment coefficient at the current moment is obtained by weighting the difference between the second adjustment coefficient at the previous moment and the absolute value of the error at the current moment using a first smoothing weight and a second smoothing weight, respectively. The second adjustment coefficient at the previous moment is multiplied by a weighting coefficient close to but slightly less than one as a continuation term, and the difference between the absolute value of the current error signal and the absolute value of the error signal at the previous moment is multiplied by a complementary weighting coefficient much less than one as an update term. The continuation term and the update term are added together to obtain the second adjustment coefficient at the current moment. This smooth recursive mechanism can be a first-order low-pass filter, where the weighting coefficient determines the speed at which the second adjustment coefficient tracks the error fluctuation amplitude: the closer the weighting coefficient is to one, the smoother and more delayed the response of the second adjustment coefficient to error fluctuations. The smaller the complementary weighting coefficient, the weaker the impact of a single error mutation on the second adjustment coefficient, effectively suppressing the drastic fluctuations in the step size range caused by random sudden interference.

[0171] Step S168: Calculate the first adjustment coefficient and the second adjustment coefficient with the mapping value to obtain the current step size.

[0172] In this embodiment, the executing entity can read the current mapping value, the first adjustment coefficient, and the second adjustment coefficient into the step size calculation module, process them according to a preset combination, and output the step size result for the current moment. The combination can involve applying rate-of-change adjustment and range adjustment to the mapping value, so that the first adjustment coefficient mainly affects the sensitivity of the step size change, and the second adjustment coefficient mainly affects the value range or adjustment amplitude of the step size. In this way, the current step size is affected not only by the magnitude of the error, but also by the change in the error relative to the previous moment and the historical trend of error fluctuations.

[0173] In one possible and specific implementation, the executing entity can read the mapping value obtained in step S164 and the first and second adjustment coefficients obtained in step S166 from the register. The executing entity can multiply the first adjustment coefficient by the cube of the absolute value of the error signal, take the negative of the product, and then perform an exponential operation. Next, using the sum of the exponential operation result plus one as the denominator and twice the second adjustment coefficient as the numerator, a division operation is performed, and then one is subtracted from the quotient to obtain the current step size. This current step size value is temporarily stored in the register for use in step S1610.

[0174] Step S1610: Calculate the weight vector update amount based on the current step size, error signal, and input vector. Add the current weight vector to the weight vector update amount to obtain the updated weight vector.

[0175] In one possible and specific implementation, the executing entity can read the current step size, the current error signal value, and the input vector constructed in step S14 from the register. The executing entity can multiply the current step size, the error signal, and the input vector together; the product is the weight vector update amount. Specifically, the current step size is multiplied by the error signal to obtain a scalar coefficient, and then this scalar coefficient is multiplied by each component of the input vector to obtain an update vector with the same dimension as the input vector. Each component in the update vector is the correction value of the component at the same index position in the corresponding weight vector.

[0176] Then, the executing entity can read the component values ​​of the current weight vector from the weight vector storage area of ​​the memory, add each component of the updated weight vector to the corresponding component of the current weight vector one by one, and the result is the updated weight vector. The executing entity can also write the updated weight vector component values ​​to the weight vector storage area, overwriting the original current weight vector, and complete this weight update operation.

[0177] In some implementations, the step of iteratively filtering subsequent sampled data based on the updated weight vector and outputting the filtered main transformer monitoring signal includes:

[0178] Step S181: Set the updated weight vector as the current weight vector for the next time step.

[0179] Step S182: Reconstruct the input vector of the adaptive filter based on the current sampling data and historical sampling data at the next time step, and calculate the filter output signal at the next time step according to the input vector and the current weight vector at the next time step.

[0180] Step S183: Calculate the reference signal for the next time step based on the historical sampling data for the next time step and the model parameters of the current autoregressive prediction model.

[0181] Step S184: Calculate the difference between the reference signal and the filter output signal at the next time step, use it as the error signal at the next time step, and determine the step size and update the weight vector at the next time step based on the error signal at the next time step.

[0182] Step S185: Repeat the above steps until all sampled data to be processed has been traversed, and output the filtered main transformer monitoring signal. Specifically, iterative filtering is performed on the transformer grounding neutral point current signal, the tank surface vibration signal, and the noise signal around the transformer, and the corresponding filtering results are output respectively.

[0183] In this embodiment, repeatedly executing the above steps can mean continuously executing the processing flow corresponding to steps S181 to S184 on the sampled data at subsequent time points, so that each time point can complete a new filtering calculation based on the updated weight state of the previous time point. Traversing all sampled data to be processed can mean sequentially performing iterative processing on each sampling time point within a predetermined processing range until the data processing for all time points is completed.

[0184] In this embodiment, the traversal can include online real-time point-by-point traversal or offline point-by-point traversal of a continuous sampled data segment. Online real-time point-by-point traversal involves performing a complete iterative process each time new sampled data is obtained, and outputting the filtering result corresponding to the current time. Offline point-by-point traversal involves performing iterative processing sequentially from the start time to the end time of a completed continuous sampled data segment, and outputting the entire filtering result after the traversal is completed.

[0185] In one possible and specific implementation, the executing entity can sequentially call steps S181 to S184 within the range of all unprocessed sampled data corresponding to the current processing object, in order of sampling time. Upon completion of processing at each time point, the filter output signal corresponding to that time point is obtained and written into the result buffer. After all unprocessed sampling times have been processed, the executing entity combines the output results from each time point in the result buffer into a filtered main transformer monitoring signal according to chronological order, and outputs it to the display unit, storage unit, or subsequent status analysis unit.

[0186] like Figure 2 and Figure 3 As shown, this invention provides a method for filtering main transformer monitoring signals based on autoregressive prediction and variable step-size adaptive LMS filtering, as detailed below:

[0187] Step 1: Sample and preprocess three signals: transformer grounding neutral point current, tank surface vibration, and ambient noise.

[0188] S11. Transformer neutral point current signal acquisition; acquired based on a Hall current sensor nested within the transformer grounding neutral point, preferably, the current sensor is provided with an electromagnetic shielding cover on its outer side;

[0189] S12. Transformer vibration signal acquisition; acquired from an accelerometer fixed to the surface of the transformer housing (avoiding the housing reinforcement structure);

[0190] S13. Transformer noise signal acquisition; acquired using a noise sensor located 1-2m away from the transformer and facing the transformer housing. Preferably, the noise sensor is provided with a windproof cover.

[0191] S14. All analog input channels use the same clock and trigger source, and use the Beidou time synchronization module to mark high-precision UTC data timestamps to achieve time synchronization under long-term monitoring.

[0192] Step 2: Based on the periodicity of urban rail transit operation, establish a dynamically updatable autoregressive prediction model for each type of signal;

[0193] S21. Construct an autoregressive prediction model based on current signals; Let the current signal u(n) at time n be the input signal corresponding to the model, and construct the autoregressive prediction model:

[0194] (1)

[0195] In formula (1), This is the predicted value at time n; The model parameters to be estimated represent the degree of influence of the current value at this historical moment on the current predicted value. The observed value at time ni; The residuals represent the unexplained portion of the model; p is the model order. These are the model parameters to be estimated.

[0196] S22. Calculate the parameters of the p-order autoregressive prediction model; construct the system of equations as shown in equation (2), where , … The sample autocorrelation coefficient ( Solving the system of equations yields the parameters corresponding to the p-order autoregressive prediction model. :

[0197] (2)

[0198] In formula (2), the sample autocorrelation coefficient ( )for:

[0199] (3)

[0200] In the above formulas (2) and (3), Let p be the parameters (weight coefficients) of the autoregressive model to be determined, where the subscript p represents the number of lag steps at a historical time. This represents the similarity between the current moment and the p-th past moment. N is the number of historical data points required to calculate the parameters of the p-th order autoregressive prediction model, which generally needs to cover at least 3 cycles of the signal, and is taken as 10 minutes. This represents the signal amplitude acquired at time j; Representative at The amplitude of the signal collected at any time.

[0201] S23. Determine the optimal order and optimal parameters of the autoregressive prediction model using the BIC criterion; let the order of the autoregressive prediction model corresponding to the signal be 1, 2, ..., 10, and calculate the parameters of the autoregressive prediction model at that order according to step S22. and residual ;

[0202] After completing the search, the optimal order is determined using the BIC criterion:

[0203] (4)

[0204] In formula (4), For p-order autoregressive prediction model residuals The variance reflects the degree to which the model fits the data; This serves as a penalty term to limit the complexity of the model;

[0205] Compare the BIC values ​​corresponding to all orders p. The order corresponding to the minimum value is the optimal order of the autoregressive prediction model, and the corresponding coefficients are the optimal coefficients of the autoregressive model.

[0206] S24. Similarly, an autoregressive prediction model is established for transformer vibration and noise signals.

[0207] S25. Dynamically update the autoregressive model according to the train schedule; according to the train schedule, the day is divided into multiple stages, and the corresponding optimal autoregressive model is pre-trained and saved using historical data in each time period. At the beginning of each stage specified in the schedule, the autoregressive prediction model order and parameters corresponding to the signal are updated.

[0208] Step 3: Build a variable step-size adaptive least mean square (LMS) filter for each signal;

[0209] S31. For current signals, let Let u(n) be the input vector formed by the current time u(n) and the previous times u(n-1), ..., u(n-M+1) (where M is the filter order); y(n) is the output signal corresponding to time n; define e(n) as the error signal at time n, whose value is the difference between the reference signal d(n) and the output signal y(n), and establish the following... Figure 2 The adaptive LMS filter shown:

[0210] (5)

[0211] In formula (5), Let be the tap weight vector of the adaptive filter at time n; n is the discrete-time index, representing the current sampling time; T represents the transpose. Let n be the input vector of the adaptive filter at time n; Let n be the output signal of the adaptive filter at time n, and its value is obtained by operating on the input vector and the weight vector. The weight vector at time n The i-th component, i.e., the i-th tap weight. The subscript i indicates the position index of this weight in the filter, corresponding to the i-th input data in the input vector. The input vector at time n The i-th component. M is the filter order (number of taps), representing the dimension of the input vector and the weight vector, which is the sum of the number of historical data points participating in the filtering operation at the current time.

[0212] The iterative formula for the weight vector of the adaptive LMS filter is as follows:

[0213] (6)

[0214] In formula (6), The updated tap weight vector at time n+1 (the next sampling processing time) is the output of this iteration and will be used for filtering calculation at the next time step. Let n be the tap weight vector at time n (the current time), and let n be the initial value for this weight update operation. For the objective function For weight vectors The gradient vector. Specifically, Let e(n) be the mean square error of the error signal. Therefore, the adaptive LMS filter essentially calculates the step size factor using the error signal e(n). Then update the weight vector. It outputs y(n) at the new time step and iterates to perform filtering.

[0215] S32. Introducing a variable step-size function to improve the traditional adaptive LMS filtering algorithm; since the traditional adaptive filter cannot simultaneously satisfy both steady-state error and convergence speed performance indicators, a sigmoid function with continuous derivative characteristics is introduced as the mapping benchmark for the step-size adjustment function:

[0216] (7)

[0217] In formula (7), The output value of the sigmoid function is a scalar result obtained after nonlinear mapping based on the input variable x, with a value range of (0,1). In this invention, this function serves as the mapping benchmark for constructing the step size factor adjustment function. x is the input variable of the sigmoid function. e is the base of the natural logarithm, a mathematical constant.

[0218] Based on the variant of the sigmoid function shown in equation (7), we obtain The nonlinear function of e(n), i.e., the step size factor adjustment function:

[0219] (8)

[0220] In formula (8), is the step size factor (output value of the step size adjustment function) at time n, whose value is determined by the error signal after nonlinear mapping, and is used to control the update amplitude of the weight vector at time n. n is the discrete-time index, representing the current sampling processing time. e(n) is the error signal at time n, defined as the difference between the reference signal and the filter output signal.

[0221] By introducing the sigmoid function to construct a step size factor adjustment function, the step size of the filter can be adjusted when the error signal e(n) is large. The error signal e(n) is relatively large and changes slowly; while when the error signal e(n) is small... It decreases rapidly.

[0222] S33, Introducing Adaptive Coefficients and The rate of change of the step size is varied; to further improve the filter's adaptability to strong random fluctuation signals, an adaptive coefficient is introduced. and An improved step size adjustment function is obtained. The nonlinear function of e(n), i.e., the improved step size adjustment function:

[0223] (9)

[0224] In formula (9), the adaptive coefficient It directly affects the rate of change of the step size. When the error signal e(n) starts to increase, the step size... Maintaining a large value accelerates algorithm convergence; when the error signal e(n) begins to decrease, the step size... The value decreases exponentially, causing the algorithm to stabilize. Specifically, This is the improved step size factor at time n, used to control the update magnitude of the weight vector at time n. The subscript n indicates the processing time to which this step size factor belongs. This is the second adaptive coefficient at time n, used to control the range (upper limit) of the step size factor, avoiding excessively large step sizes due to random sudden disturbances. The subscript n indicates the value of this coefficient at the current time. is the first adaptive coefficient at time n, used to control the rate of change of the step size factor (the steepness of the nonlinear mapping function).

[0225] Based on the above principles, define the adaptive coefficients. for:

[0226] (10)

[0227] In formula (10), It is the absolute value of the ratio of the error signal at the current time to the error signal at the previous time, reflecting the relative change in the error signal.

[0228] Adaptive coefficients This directly affects the range of step size values. To avoid the influence of random, sudden interference signals, its variation range should be appropriately reduced.

[0229] Based on the above principles, we define the adaptive coefficients. for:

[0230] (11)

[0231] In formula (11), This is a smoothing coefficient used to weight the second adaptive coefficient from the previous time step. These are the weighting coefficients for the update term, used to weight the difference term of the absolute value of the error.

[0232] S34. Similarly, an adaptive LMS filter is established for the transformer vibration signal and noise signal;

[0233] Step 4: Using the prediction output of the autoregressive prediction model as the reference signal, calculate the error signal between the reference signal and the output signal of the adaptive LMS filter, iteratively change the weights according to the step size adjustment function, and output the filtering result in real time.

[0234] S41. Initialize the adaptive LMS filter parameters; set the filter order M for the three signals to 128; set the weight vector. Initial value is 0; set adaptive coefficient. The initial value is 0.5, and the coefficient is... The initial value is 0.01;

[0235] S42, Based on Equation (5) and Weight Vector Calculate the output signal y(n) of the adaptive LMS filter at time n.

[0236] S43. The output of the autoregressive prediction model is calculated based on equation (1) and used as the reference signal d(n) for the adaptive LMS filter at time n.

[0237] S44. Calculate the difference between the reference signal d(n) and the output signal y(n) as the error signal e(n) at time n.

[0238] S45. Calculate the adaptive coefficient at time n based on equations (8) and (10). and And obtain the step size .

[0239] S46, Based on Equation (6) and Step Size Update the weight vector at the next time step .

[0240] S47. Repeat steps S42-S46 to output the signals of transformer neutral point stray current, vibration and noise after passing through the adaptive LMS filter in real time.

[0241] like Figure 4 As shown, according to an embodiment of the present invention, a main transformer monitoring signal filtering device is provided, comprising:

[0242] The acquisition module is used to acquire the current and historical sampling data of the main transformer monitoring signals;

[0243] The reference signal calculation module is used to solve the model parameters of the autoregressive prediction model based on historical sampling data, and to calculate the reference signal based on the model parameters and historical sampling data.

[0244] The filter output signal calculation module is used to construct the input vector of the adaptive filter based on the current sampled data, and to calculate the filter output signal based on the input vector and the current weight vector.

[0245] The weight vector update module is used to calculate the difference between the reference signal and the filter output signal as an error signal, determine the current step size based on the error signal, and update the weight vector based on the current step size.

[0246] The output module is used to iteratively filter subsequent sampled data based on the updated weight vector and output the filtered main transformer monitoring signal.

[0247] According to an embodiment of the present invention, an electronic device is provided; please refer to... Figure 5 The electronic device in this embodiment may include one or more of the following components: a processor, a network interface, memory, non-volatile memory, and one or more application programs, wherein the one or more application programs may be stored in non-volatile memory and configured to be executed by one or more processors, and the one or more programs are configured to perform the methods as described in the foregoing method embodiments.

[0248] According to embodiments of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a computer, causes the computer to perform the method described in any of the above embodiments.

[0249] According to embodiments of the present invention, a computer program product comprising instructions is also provided, which, when executed by a computer, cause the computer to perform a method in any of the above embodiments.

[0250] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for filtering monitoring signals of a main transformer, characterized in that, include: Acquire the current and historical sampling data of the main transformer monitoring signals; The model parameters of the autoregressive prediction model are solved based on historical sampling data, and the reference signal is calculated based on the model parameters and historical sampling data. The input vector of the adaptive filter is constructed based on the current sampled data, and the output signal of the filter is calculated based on the input vector and the current weight vector. Calculate the difference between the reference signal and the filter output signal as the error signal, determine the current step size based on the error signal, and update the weight vector based on the current step size; The updated weight vector is used to iteratively filter the subsequent sampled data, and the filtered main transformer monitoring signal is output.

2. The method according to claim 1, characterized in that, The steps for acquiring the current and historical sampling data of the main transformer monitoring signal include: Collect the transformer grounding neutral point current signal, the tank surface vibration signal, and the noise signal around the transformer; Each acquired channel signal is marked with a unified timestamp, and the time of each channel signal is aligned based on the unified timestamp; Based on the time-aligned signals of each channel, the current sampling data and historical sampling data of each channel signal are extracted.

3. The method according to claim 1, characterized in that, The steps of solving the model parameters of the autoregressive prediction model based on historical sampling data, and calculating the reference signal based on the model parameters and historical sampling data, include: Calculate the sample autocorrelation coefficients corresponding to different lag orders based on historical sampling data; Based on the sample autocorrelation coefficient, establish a set of autoregressive parameter equations corresponding to the order of candidate models, and solve for the model parameters corresponding to the order of each candidate model. The target model order is determined based on the residual variance and model complexity corresponding to the order of each candidate model, and the model parameters corresponding to the target model order are used as the model parameters of the current autoregressive prediction model. The reference signal is calculated based on the model parameters of the current autoregressive prediction model and the historical sampling data prior to the current time.

4. The method according to claim 3, characterized in that, The step of using the model parameters corresponding to the target model order as the model parameters of the current autoregressive prediction model includes: Determine the current operating segment based on the train timetable; Call the historical sampling data corresponding to the runtime segment to solve for the target model order and model parameters corresponding to the runtime segment; The target model order and model parameters corresponding to the runtime segment are used as the model order and model parameters of the current autoregressive prediction model.

5. The method according to claim 4, characterized in that, The steps of constructing the input vector of the adaptive filter based on the current sampled data, and calculating the filter output signal based on the input vector and the current weight vector, include: Obtain the current sampled data at the current moment and the historical sampled data at a preset number of moments prior to the current moment; Arrange the current sampled data and historical sampled data in chronological order to form the input vector of the adaptive filter; Call the current weight vector corresponding to the input vector; The filter output signal is calculated based on the product of the input vector and the current weight vector.

6. The method according to claim 5, characterized in that, The steps of calculating the difference between the reference signal and the filter output signal as an error signal, determining the current step size based on the error signal, and updating the weight vector based on the current step size include: Calculate the difference between the reference signal and the filter output signal at the current moment, and use it as the error signal at the current moment; The error signal is input into a preset nonlinear mapping function to obtain a mapping value; wherein, the nonlinear mapping function increases the mapping value and slows down the change when the absolute value of the error signal increases, and decreases the mapping value more rapidly when the absolute value of the error signal decreases; wherein, the nonlinear mapping function is constructed based on the sigmoid function, and the absolute value of the error signal acts on the nonlinear mapping function in the form of a cube power; Obtain the first adjustment coefficient and the second adjustment coefficient; wherein, the first adjustment coefficient is the square of the ratio of the error signal to the error signal at the previous time, and the second adjustment coefficient is the product of the second adjustment coefficient at the previous time and the weighting coefficient, plus the sum of the product of the difference between the absolute values ​​of the error signal and the error signal at the previous time and the complementary weighting coefficient. The first and second adjustment coefficients are calculated with the mapping value to obtain the current step size; Based on the current step size, error signal, and input vector, calculate the weight vector update amount, add the current weight vector to the weight vector update amount, and obtain the updated weight vector.

7. The method according to claim 6, characterized in that, The step of iteratively filtering subsequent sampled data based on the updated weight vector and outputting the filtered main transformer monitoring signal includes: Set the updated weight vector as the current weight vector at the next time step; The input vector of the adaptive filter is reconstructed based on the current sampling data and historical sampling data at the next time step, and the filter output signal at the next time step is calculated based on the input vector and the current weight vector at the next time step. The reference signal for the next time step is calculated based on the historical sampling data for the next time step and the model parameters of the current autoregressive prediction model. Calculate the difference between the reference signal and the filter output signal at the next time step, use it as the error signal at the next time step, and determine the step size and update the weight vector at the next time step based on the error signal at the next time step. Repeat the above steps until all the sampled data to be processed has been traversed, and output the filtered main transformer monitoring signal; Specifically, iterative filtering is performed on the transformer grounding neutral point current signal, the tank surface vibration signal, and the noise signal around the transformer, and the corresponding filtering results are output respectively.

8. A main transformer monitoring signal filtering device, characterized in that, include: The acquisition module is used to acquire the current and historical sampling data of the main transformer monitoring signals; The reference signal calculation module is used to solve the model parameters of the autoregressive prediction model based on historical sampling data, and to calculate the reference signal based on the model parameters and historical sampling data. The filter output signal calculation module is used to construct the input vector of the adaptive filter based on the current sampled data, and to calculate the filter output signal based on the input vector and the current weight vector. The weight vector update module is used to calculate the difference between the reference signal and the filter output signal as an error signal, determine the current step size based on the error signal, and update the weight vector based on the current step size. The output module is used to iteratively filter subsequent sampled data based on the updated weight vector and output the filtered main transformer monitoring signal.

9. An electronic device, characterized in that, include: A memory, and one or more processors communicatively connected to the memory; The memory stores instructions that can be executed by the one or more processors to cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 7.

Images