A signal processing method and device for a distribution transformer

Abstract

The application provides a signal processing method and device of a distribution transformer, and belongs to the technical field of signal processing. The method comprises the following steps: acquiring an actual noise signal and a load current signal in the operation process of the distribution transformer, obtaining a model noise signal by using a current wavelet neural network model and calculating the amplitude absolute value of a residual noise signal; when the amplitude absolute value reaches or exceeds an optimization trigger threshold value, determining a current load working condition type based on the load current signal, and adaptively determining the population size and the crossover probability of a genetic algorithm in combination with the amplitude absolute value and the working condition type, and performing noise reduction processing after globally optimizing the wavelet neural network model; and when the amplitude absolute value is lower than the threshold value, directly using the current model to perform noise reduction processing. The signal processing method and device of the distribution transformer provided by the application can meet the real-time requirement in the online monitoring scene of the distribution transformer.

Description

Technical Field

[0001] This application belongs to the field of signal processing technology, and more specifically, relates to a signal processing method and device for a power distribution transformer. Background Technology

[0002] Noise signals generated by distribution transformers during operation are crucial for assessing their operational status. However, the acquired noise signals often contain environmental interference, necessitating effective noise reduction to improve signal quality. Wavelet neural networks, combining the multi-resolution analysis capabilities of wavelet transform with the nonlinear mapping capabilities of neural networks, are suitable for modeling and denoising non-stationary signals such as transformer noise. In practical applications, the operating conditions of distribution transformers continuously change with load fluctuations, altering their noise characteristics. This causes the modeling accuracy of pre-trained wavelet neural network models to gradually decline over time, making it difficult to maintain ideal noise reduction performance. Existing technologies typically employ fixed-parameter optimization algorithms to periodically retrain the neural network model to restore its noise reduction performance. However, this approach has significant drawbacks: firstly, fixed-parameter optimization algorithms cannot adapt to the varying complexity of noise characteristics under different load conditions, resulting in wasted computational resources under simple conditions and difficulty in maintaining optimization accuracy under complex conditions; secondly, the strategy of performing global optimization every time incurs excessive computational overhead, failing to meet the real-time requirements of online monitoring scenarios for distribution transformers. Summary of the Invention

[0003] The purpose of this application is to provide a signal processing method and device for distribution transformers that can meet the real-time requirements of online monitoring scenarios. To achieve the above objective, the technical solution provided by this application is as follows: Firstly, a signal processing method for a distribution transformer is provided, including: Acquire the actual noise signal and load current signal of the distribution transformer during operation; Based on the actual noise signal, and through the current wavelet neural network model, the model noise signal is obtained; The residual noise signal is obtained based on the difference between the actual noise signal and the model noise signal; Calculate the absolute value of the residual noise signal amplitude and compare it with the preset optimized trigger threshold. If the absolute value of the amplitude is greater than or equal to the optimization trigger threshold, the current load condition type of the distribution transformer is determined based on the load current signal, and the population size and crossover probability of the genetic algorithm are determined based on the absolute value of the amplitude and the current load condition type. Based on the population size and crossover probability, the network parameters of the current wavelet neural network model are globally optimized using a genetic algorithm to obtain an optimized wavelet neural network model, which is then used as the updated current wavelet neural network model. The actual noise signal is denoised based on the updated wavelet neural network model to obtain the denoised transformer noise signal. If the absolute value of the amplitude is less than the optimized trigger threshold, the actual noise signal is denoised based on the current wavelet neural network model to obtain the denoised transformer noise signal.

[0004] Secondly, a signal processing device for a distribution transformer is provided, comprising: The signal acquisition module is used to acquire the actual noise signal and load current signal of the distribution transformer during operation; The model output module is used to obtain the model noise signal based on the actual noise signal and through the current wavelet neural network model; The residual calculation module is used to obtain the residual noise signal based on the difference between the actual noise signal and the model noise signal; The comparison module is used to calculate the absolute value of the amplitude of the residual noise signal and compare the absolute value of the amplitude with a preset optimized trigger threshold. The parameter determination module is used to determine the current load condition type of the distribution transformer based on the load current signal if the absolute value of the amplitude is greater than or equal to the optimization trigger threshold, and to determine the population size and crossover probability of the genetic algorithm based on the absolute value of the amplitude and the current load condition type. The model update module is used to globally optimize the network parameters of the current wavelet neural network model based on the population size and crossover probability, and obtain the optimized wavelet neural network model. The optimized wavelet neural network model is then used as the updated current wavelet neural network model. The first noise reduction module is used to perform noise reduction processing on the actual noise signal based on the updated current wavelet neural network model to obtain the noise-reduced transformer noise signal. The second noise reduction module is used to perform noise reduction processing on the actual noise signal based on the current wavelet neural network model if the absolute value of the amplitude is less than the optimized trigger threshold, so as to obtain the noise-reduced transformer noise signal.

[0005] Thirdly, embodiments of this application also provide an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement a signal processing method for a distribution transformer provided in any possible implementation of the first aspect.

[0006] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements a signal processing method for a distribution transformer provided by any possible implementation of the first aspect.

[0007] The beneficial effects of the technical solution provided in this application are as follows: This application provides a signal processing method and device for a distribution transformer. Compared with related technologies, this application calculates the absolute value of the amplitude of the residual noise signal and compares it with an optimization trigger threshold. When the absolute value of the amplitude is greater than or equal to the threshold, the current load condition type is determined based on the load current signal. The population size and crossover probability of the genetic algorithm are adaptively determined by combining the absolute value of the amplitude and the condition type, thereby globally optimizing the wavelet neural network model. When the absolute value of the amplitude is less than the threshold, the current model is directly used for noise reduction. This application avoids the huge computational overhead caused by performing global optimization of the genetic algorithm for each signal processing by introducing an optimization trigger mechanism. At the same time, based on the dual-dimensional joint decision of load condition type and residual noise amplitude, the genetic algorithm parameters can accurately match the complexity of the current condition. Adaptive optimization is triggered in a timely manner when the model accuracy decreases, ultimately achieving a dynamic balance between computational efficiency and noise reduction accuracy, meeting the real-time requirements of online monitoring scenarios for distribution transformers. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below.

[0009] Figure 1 A schematic flowchart illustrating a signal processing method for a distribution transformer provided in an embodiment of this application; Figure 2 A structural block diagram of a signal processing device for a distribution transformer provided in an embodiment of this application; Figure 3 A schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0010] The embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the embodiments described below with reference to the accompanying drawings are exemplary descriptions for explaining the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions of the embodiments of this application.

[0011] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the terms “comprising” and “including” as used in embodiments of this application mean that the corresponding feature can be implemented as the presented feature, information, data, step, operation, element, and / or component, but do not exclude implementation as other features, information, data, step, operation, element, component, and / or combinations thereof supported by the art. It should be understood that when we say that an element is “connected” or “coupled” to another element, the one element can be directly connected or coupled to the other element, or it can mean that the one element and the other element establish a connection relationship through an intermediate element. Furthermore, “connected” or “coupled” as used herein can include wireless connection or wireless coupling. The term “and / or” as used herein indicates at least one of the items defined by the term; for example, “A and / or B” can be implemented as “A,” or as “B,” or as “A and B.” When describing multiple (two or more) items, if the relationship between the multiple items is not explicitly defined, the multiple items can refer to one, several or all of the multiple items. For example, the description of "parameter A includes A1, A2, A3" can be implemented as parameter A includes A1 or A2 or A3, or it can be implemented as parameter A includes at least two of the three items A1, A2 and A3.

[0012] This application provides a signal processing method for a distribution transformer, which can be executed by an electronic device, such as... Figure 1 As shown, the method may include: S101: Acquire the actual noise signal and load current signal of the distribution transformer during operation.

[0013] In this embodiment, the actual noise signal refers to the sound wave signal generated by the distribution transformer during operation due to a combination of factors such as the magnetostrictive effect of the iron core, the electromagnetic force of the winding, and the mechanical vibration of the cooling system. It is represented as time-domain waveform data with certain spectral characteristics. The load current signal refers to the current signal output by the secondary winding of the distribution transformer. This signal can reflect the current load level and load change trend of the transformer.

[0014] In this embodiment, an industrial-grade microphone array, such as a four-channel free-field microphone, can be deployed on the outer wall or near-field area of ​​the distribution transformer tank to continuously collect transformer noise at a preset sampling rate, thereby obtaining the actual noise signal. In this embodiment, the sampling rate can be set to 10240Hz. Simultaneously, a current transformer is installed at the secondary side output terminal of the distribution transformer to collect the load current signal at the same sampling rate. Each time a 1-second segment of signal data is collected, an actual noise signal sequence containing 10240 sampling points and a load current signal sequence containing 10240 sampling points are obtained. These two signal sequences are used as input data for subsequent processing.

[0015] S102: Based on the actual noise signal, the model noise signal is obtained through the current wavelet neural network model.

[0016] In this embodiment, the wavelet neural network model refers to a nonlinear function approximation model that introduces wavelet basis functions into the hidden layer of a neural network as activation functions. It integrates the multi-resolution analysis capability of wavelet transform and the nonlinear mapping capability of neural networks. The current wavelet neural network model refers to the wavelet neural network model that is in the latest parameter state at the current moment. The model noise signal refers to the noise prediction result output by the current wavelet neural network model after performing forward inference calculation on the input actual noise signal. This model noise signal characterizes the noise component of the transformer under normal electromagnetic characteristics.

[0017] In this embodiment, the current wavelet neural network model has the following structure: 8 nodes in the input layer, 6 nodes in the hidden layer, and 1 node in the output layer. The hidden layer uses the Morlet wavelet basis function as the activation function. This embodiment uses the actual noise signal sequence with 10240 sampling points as input data and feeds it into the current wavelet neural network model for forward propagation calculation. The forward propagation calculation process is as follows: First, the actual noise signal sequence is segmented in a sliding window manner, with each window containing 8 consecutive sampling points and 4 sampling points overlapping between adjacent windows, thereby converting 10240 sampling points into multiple sets of 8-dimensional input vectors. Second, for each group of 8-dimensional input vectors, the weighted summation result from the input layer to the hidden layer is calculated, and then the output of the hidden layer is obtained through the Morlet wavelet basis function nonlinear transformation. Finally, the weighted summation result from the hidden layer to the output layer is calculated to obtain the single-point output value. Third, the single-point output values ​​of each window are spliced ​​together in chronological order to obtain the model noise signal.

[0018] The mathematical expression for the Morlet wavelet basis function is: ,in Let t be the function value (i.e., wavelet amplitude) of the Morlet wavelet basis function when the independent variable is t, where t is the normalized time variable. The term extracts the characteristic frequency components of the signal through oscillation. The term concentrates energy near t=0 through Gaussian attenuation, thereby focusing on the signal characteristics within the local time window. The product of the two terms gives the Morlet wavelet excellent time-frequency resolution, which is suitable for the complex characteristics of the noise signal of the distribution transformer, which contains both steady-state harmonics (requiring frequency domain analysis) and transient impacts of load change (requiring time domain analysis).

[0019] In this embodiment, a wavelet neural network model is used to accurately model transformer noise. This fully utilizes the multi-resolution analysis capability of wavelet transform to capture the nonlinear and non-stationary features in the noise signal, providing a high-precision noise reference benchmark for subsequent residual noise extraction and adaptive optimization.

[0020] S103: The residual noise signal is obtained based on the difference between the actual noise signal and the model noise signal.

[0021] In this embodiment, the residual noise signal refers to the sequence of amplitude differences between the actual noise signal and the model noise signal at the same sampling time. The residual noise signal reflects the degree of deviation of the current wavelet neural network model from the actual noise of the transformer. Ideally, if the wavelet neural network model can completely and accurately characterize the noise characteristics of the transformer, the residual noise signal should approach zero; conversely, the larger the amplitude of the residual noise signal, the lower the modeling accuracy, and the more necessary the optimization.

[0022] In this embodiment, the actual noise signal sequence and the model noise signal sequence are aligned point-by-sampling. Since the model noise signal is obtained through piecewise forward inference via a sliding window, its number of data points may differ slightly from the actual noise signal sequence. Therefore, during the alignment process, linear interpolation is performed on the model noise signal to ensure that its number of data points is completely consistent with the actual noise signal sequence, both being 10240 sampling points. Subsequently, the difference is calculated point-by-point according to the following formula: e(n) = x(n) - y(n); Where x(n) is the amplitude of the actual noise signal at the nth sampling point, y(n) is the amplitude of the model noise signal at the nth sampling point, and e(n) is the amplitude of the residual noise signal at the nth sampling point. This yields a residual noise signal sequence containing 10240 sampling points.

[0023] In this embodiment, by accurately calculating the residual noise signal point by point, the modeling deviation of the current wavelet neural network model for the actual noise of the transformer can be quantified. This provides a reliable decision basis for the subsequent adaptive triggering optimization mechanism based on the degree of deviation, avoiding the waste of computing resources caused by blindly triggering optimization, and also avoiding the degradation of model performance caused by not triggering optimization for a long time.

[0024] S104: Calculate the absolute value of the amplitude of the residual noise signal and compare the absolute value of the amplitude with the preset optimized trigger threshold.

[0025] In this embodiment, the absolute amplitude value refers to the statistical mean of the absolute values ​​of the residual noise signal at each sampling point. This index is used to comprehensively quantify the overall energy level of the residual noise signal. The optimization trigger threshold is a pre-set threshold for the absolute amplitude value used to determine whether the genetic algorithm optimization process needs to be triggered. When the absolute amplitude value of the residual noise signal reaches or exceeds this threshold, it indicates that the modeling accuracy of the current wavelet neural network model has decreased to the point where it needs to be re-optimized. In this embodiment, the optimization trigger threshold can be set to 0.2. The optimization trigger threshold can be determined as follows: During the debugging phase, noise data of the distribution transformer running continuously for 72 hours under typical steady-state conditions is collected. The absolute amplitude value of the residual noise signal at each time window is calculated using the wavelet neural network model after initial training. The 95th percentile value of the statistical distribution is taken as the optimization trigger threshold to ensure that optimization is not mistakenly triggered within the normal operating fluctuation range, and that optimization is triggered in a timely manner when the model performance deteriorates significantly.

[0026] In this embodiment, the absolute value of the residual noise signal amplitude is first calculated according to the following formula: ; Where N is the total number of sampling points for the residual noise signal, in this embodiment N=10240, e(n) is the residual noise amplitude at the nth sampling point, |e(n)| is its absolute value, and MAE is the absolute value of the amplitude. The calculated MAE value is compared with a preset optimization trigger threshold. If MAE is greater than or equal to the optimization trigger threshold, it is determined that the genetic algorithm optimization process needs to be triggered; if MAE is less than the optimization trigger threshold, it is determined that the current wavelet neural network model still has sufficient modeling accuracy and optimization is not required.

[0027] S105: If the absolute value of the amplitude is greater than or equal to the optimization trigger threshold, the current load condition type of the distribution transformer is determined based on the load current signal, and the population size and crossover probability of the genetic algorithm are determined based on the absolute value of the amplitude and the current load condition type.

[0028] In this embodiment, the current load condition type refers to the load state category of the distribution transformer at the current operating time. The noise characteristics of the transformer are different under different load conditions. For example, under light load, the core magnetostriction noise is the main noise, while under heavy load, the winding electromagnetic force noise increases. When the load changes suddenly, transient impact noise will also be superimposed.

[0029] In one embodiment of this application, determining the current load condition type of the distribution transformer based on the load current signal includes: Calculate the effective value and rate of change of the load current signal within a preset time window; Based on the effective value of the current and the rate of change of the current, the current load condition type of the distribution transformer is determined; the current load condition type includes one of the following: steady-state light load condition, steady-state heavy load condition, and load change condition.

[0030] In this embodiment, the effective value of the current refers to the root mean square value of the load current signal within a specified time window, used to quantify the average energy level of the load current within that time window; the rate of change of the current refers to the speed at which the effective value of the load current signal changes with time within a specified time window, used to quantify the degree of dynamic change of the load; steady-state light load condition refers to the operating state where the transformer load level is low and the load change is gradual; steady-state heavy load condition refers to the operating state where the transformer load level is high but the load change is gradual; and load change condition refers to the transient operating state where the transformer load changes rapidly and significantly.

[0031] In this embodiment, the load current signal collected in S101 is acquired, and the load current signal data within the most recent 0.5-second time window is taken, i.e., 5120 sampling points.

[0032] Then, calculate the effective value of the current within this time window using the following formula: ; Where M is the number of sampling points within the time window (M=5120 in this embodiment), and i(m) is the instantaneous value of the load current at the m-th sampling point. Assuming that the calculated value is... =325A, the rated current of the distribution transformer is 500A, then the effective value of the current accounts for 65% of the rated current.

[0033] Calculate the rate of change of current: Divide the 0.5-second time window into two equal 0.25-second sub-windows, and calculate the effective value of current in each sub-window. and The rate of change of current is: ; in, The current change rate is denoted as .

[0034] Assumption =320A, =330A, then =40A / s, which is 40 / 500=8% / s of the rated current.

[0035] The judgment is made according to the preset operating condition judgment rules: if the effective value of the current is less than 40% of the rated current and the current change rate is less than 5% / s of the rated current, it is judged as a steady-state light load condition; if the effective value of the current is greater than or equal to 40% of the rated current and the current change rate is less than 5% / s of the rated current, it is judged as a steady-state heavy load condition; if the current change rate is greater than or equal to 5% / s of the rated current, it is judged as a sudden load change condition regardless of the effective value of the current.

[0036] In this embodiment, the effective current value accounts for 65% (≥40%) of the rated current, but the current change rate accounts for 8% / s (≥5% / s) of the rated current. Therefore, the current load condition is ultimately determined to be a sudden load change condition. The determination of the above-mentioned thresholds (40% and 5% / s) is based on the following: according to the operating procedures of distribution transformers and field measurement data, 40% of the rated current is a commonly used dividing point to distinguish between light load and heavy load, and the current change rate threshold of 5% / s corresponds to the typical current change rate when the transformer load undergoes a significant change.

[0037] As can be seen from the above, this embodiment, by simultaneously considering two dimensions of indicators—the effective value of current and the rate of change of current—can accurately distinguish between three typical operating conditions: steady-state light load, steady-state heavy load, and load change. This provides a precise basis for the differentiated configuration of subsequent genetic algorithm parameters, avoids the problem of misjudgment of operating conditions caused by relying solely on a single current indicator, and significantly improves the reliability of the adaptive optimization mechanism.

[0038] In this embodiment, the population size refers to the number of candidate solution individuals maintained by the genetic algorithm in each generation of evolution. The larger the population size, the stronger the global search capability, but the greater the computational cost. The crossover probability refers to the probability that two parent individuals in the genetic algorithm will perform a crossover operation to generate offspring individuals. The higher the crossover probability, the stronger the algorithm's search space exploration capability, but it may destroy existing excellent gene combinations.

[0039] In one embodiment of this application, determining the population size and crossover probability of a genetic algorithm based on the absolute value of the amplitude and the current load condition type includes: Based on the preset amplitude range division rules, the target amplitude range to which the absolute value of the amplitude belongs is determined; Based on the preset mapping relationship between operating condition type and population size benchmark value, determine the population size benchmark value corresponding to the current load operating condition type. Based on the preset amplitude adjustment coefficient corresponding to the target amplitude range, the baseline value of the population size is corrected to obtain the population size of the genetic algorithm; Based on the preset mapping relationship between operating condition types and cross probability benchmark values, determine the cross probability benchmark value corresponding to the current load operating condition type; The crossover probability baseline value is corrected based on the preset probability adjustment coefficient corresponding to the target amplitude range to obtain the crossover probability of the genetic algorithm.

[0040] In this embodiment, the target amplitude range refers to the specific numerical range to which the absolute value of the residual noise signal amplitude belongs under the preset amplitude range division rule; the population size benchmark value refers to the reference benchmark value of the genetic algorithm population size under a specific load condition; the preset amplitude adjustment coefficient refers to the multiplier factor for scaling the population size benchmark value according to the target amplitude range; the crossover probability benchmark value refers to the reference benchmark value of the genetic algorithm crossover probability under a specific load condition; and the preset probability adjustment coefficient refers to the multiplier factor for scaling the crossover probability benchmark value according to the target amplitude range.

[0041] In this embodiment, it is assumed that the calculated absolute value of amplitude (MAE) is 0.42, and the current load condition type is a sudden load change condition.

[0042] First, the target amplitude range to which MAE belongs is determined according to the preset amplitude range division rules. In this embodiment, the preset amplitude range division rules are three consecutive ranges: the first range [0.20, 0.30), the second range [0.30, 0.50), and the third range [0.50, +∞). MAE=0.42 falls into the second range [0.30, 0.50), therefore the target amplitude range is determined to be the second range. The specific range boundaries can be determined during the debugging phase, for example: statistically analyzing the absolute amplitude distribution of a large number of historical residual noise signals, using the MAE values ​​corresponding to the 33rd and 67th percentiles as the range boundaries, thereby ensuring that the sample size in each range is roughly balanced.

[0043] Secondly, based on the preset mapping relationship between operating condition types and population size benchmark values, the population size benchmark value corresponding to the load mutation operating condition is determined. In this embodiment, the population size benchmark values ​​corresponding to the three operating condition types are: steady-state light load condition 40, steady-state heavy load condition 60, and load mutation condition 80. These benchmark values ​​can be determined through offline experiments. For example, the optimization effect is tested under each operating condition with population sizes of 20, 40, 60, 80, and 100, and the smallest population size is selected when the noise reduction signal-to-noise ratio gain reaches the saturation inflection point and the calculation time does not exceed 500 milliseconds. Thus, the population size benchmark value corresponding to the load mutation operating condition is determined to be 80.

[0044] Then, based on the preset amplitude adjustment coefficient of 1.2 corresponding to the target amplitude interval (second interval), the baseline value of the population size of 80 is corrected: 80 × 1.2 = 96. After rounding, the population size of the genetic algorithm is determined to be 96. The preset amplitude adjustment coefficients corresponding to each amplitude interval are: 0.8 for the first interval, 1.2 for the second interval, and 1.5 for the third interval. The design logic of the preset amplitude adjustment coefficients is: the larger the absolute value of the amplitude, the more serious the model bias, and the larger the population size is required to ensure global search capability.

[0045] Based on the preset mapping relationship between operating condition types and crossover probability benchmark values, the crossover probability benchmark value corresponding to the load mutation operating condition is 0.80 (0.70 for steady-state light load, 0.75 for steady-state heavy load, and 0.80 for load mutation). According to the preset probability adjustment coefficient of 1.1 corresponding to the target amplitude interval (second interval), the crossover probability benchmark value of 0.80 is corrected: 0.80 × 1.1 = 0.880, thus determining the crossover probability of the genetic algorithm to be 0.880. The preset probability adjustment coefficients corresponding to each amplitude interval are: 0.9 for the first interval, 1.1 for the second interval, and 1.15 for the third interval. It should be noted that if the corrected crossover probability exceeds 1.0, 1.0 is taken as the upper limit to ensure the legality of the probability value.

[0046] As can be seen from the above, this embodiment introduces a two-level determination mechanism by introducing the working condition type to determine the baseline value and then using the adjustment coefficient to correct it according to the amplitude range. This enables the population size and crossover probability of the genetic algorithm to accurately reflect the combined effects of the complexity of the current working condition and the severity of the model deviation, avoiding the coarseness of single-dimensional parameter adjustment. It achieves efficient utilization of computing resources while ensuring optimization accuracy.

[0047] S106: Based on the population size and crossover probability, the network parameters of the current wavelet neural network model are globally optimized using a genetic algorithm to obtain an optimized wavelet neural network model, which is then used as the updated current wavelet neural network model.

[0048] In this embodiment, network parameters refer to all adjustable parameters to be optimized in the current wavelet neural network model, including the connection weights from the input layer to the hidden layer, the connection weights from the hidden layer to the output layer, and the translation and scaling parameters of the wavelet basis functions of the hidden layer. These parameters together determine the nonlinear mapping characteristics of the wavelet neural network model. Global optimization refers to the process by which the genetic algorithm searches for the optimal or near-optimal solution in the global scope of the parameter space by performing parallel search and iterative evolution on multiple candidate solutions in the population. This is different from local optimization methods such as gradient descent, which can only find local optima near the initial value.

[0049] In one embodiment of this application, based on population size and crossover probability, and using a genetic algorithm to globally optimize the network parameters of the current wavelet neural network model, an optimized wavelet neural network model is obtained, including: The network parameters of the current wavelet neural network model are encoded as individual chromosomes of a genetic algorithm; where the network parameters include the weight coefficients of the current wavelet neural network model, the translation parameters and scaling parameters of the wavelet basis functions; An initial population is generated based on the population size, and the initial population contains multiple individual chromosomes; The fitness value of each individual's chromosome in the initial population is calculated by using the minimization of the residual noise signal energy as the fitness function; Based on the fitness value, crossover probability, and preset mutation probability, the initial population is subjected to selection, crossover, and mutation operations in sequence, and the population is iteratively updated until the preset iteration termination condition is met. The chromosome of the individual with the largest fitness value is determined as the target individual chromosome. The target individual's chromosome is decoded into corresponding network parameters, and the current wavelet neural network model is updated with the decoded network parameters to obtain the optimized wavelet neural network model.

[0050] In this embodiment, an individual chromosome refers to a data structure representing a candidate solution in a genetic algorithm. In this embodiment, each individual chromosome corresponds to a complete set of network parameters of a wavelet neural network model. The fitness value is a numerical index that measures the quality of an individual chromosome. The larger the fitness value, the better the noise reduction effect of the model is due to the combination of network parameters corresponding to that individual. The roulette wheel selection method is a genetic algorithm selection strategy that selects probabilistically according to the fitness value. The arithmetic crossover operation refers to the crossover method that generates offspring by linearly combining the corresponding gene positions of two parent individual chromosomes. The Gaussian mutation operation refers to the mutation method that superimposes a random perturbation that follows a normal distribution on the gene positions of an individual chromosome.

[0051] In this embodiment, the network structure consists of an 8-node input layer, a 6-node hidden layer, and a 1-node output layer. The network parameters include: a connection weight matrix from the input layer to the hidden layer (8×6=48 parameters), a connection weight vector from the hidden layer to the output layer (6×1=6 parameters), translation parameters (6 parameters) and scaling parameters (6 parameters) of the Morlet wavelet basis functions in the hidden layer, totaling 66 parameters. A real-number encoding method is used, with each individual chromosome represented as a 66-dimensional real-number vector, where each element corresponds to one of the 66 network parameters.

[0052] Based on the determined population size (96 in this example), 96 individual chromosomes are randomly generated in the parameter space to form the initial population. The values ​​of each gene locus on each individual chromosome are randomly generated in a uniform distribution within the interval [-2, 2]. For each individual chromosome in the initial population, the following fitness calculation process is performed: The 66 gene values ​​of the individual's chromosome were decoded into network parameters and configured into a wavelet neural network model to obtain a candidate wavelet neural network model. The actual noise signal was input into the candidate wavelet neural network model using a sliding window to obtain the candidate model noise signal. The sum of squares of the differences between the actual noise signal and the candidate model noise signal was calculated. ,in This represents the actual noise signal amplitude. Let be the amplitude of the noise signal in the candidate model; the reciprocal of this sum of squares, 1 / E, is determined as the fitness value of the chromosome for that individual. A higher fitness value indicates that the network parameters corresponding to that individual result in less residual noise energy and higher modeling accuracy.

[0053] In this embodiment, 96 individuals are selected from the initial population using a roulette wheel selection algorithm based on their fitness values. The probability of each individual being selected is... ,in Let be the fitness value of the k-th individual. This is the sum of the fitness values ​​of all individuals in the population. The same individual is allowed to be selected repeatedly to maintain a constant population size. Then, arithmetic crossover is performed on the crossover probability of the selected individuals (0.880 in this example): The 96 individuals are randomly paired. For each pair of parent individuals, a uniformly random number r in the interval [0,1] is generated. If r < 0.880, arithmetic crossover is performed on that pair of parent individuals. , ,in A uniformly random number within the interval [0,1] is applied to each of the 66 gene loci sequentially. If r ≥ 0.880, then directly copy the two parent individuals to the offspring population.

[0054] After the crossover operation, a Gaussian mutation operation is performed with a preset mutation probability of 0.05: for each gene locus on the chromosome of each individual in the offspring population, a uniformly random number in the interval [0,1] is generated. If this random number is less than 0.05, a random perturbation value following a normal distribution with a mean of 0 and a standard deviation of 0.1 is superimposed on that gene locus. The three steps of selection, crossover, and mutation constitute one iteration (i.e., generation) of the genetic algorithm. The iteration process is repeated, with each generation using the population output from the previous generation as input, until the preset iteration termination condition is met.

[0055] In this embodiment, there are two iteration termination conditions, and either one needs to be satisfied: (1) The number of iterations reaches the maximum number of iterations, 100 generations; (2) The change in the maximum fitness value over 20 consecutive generations of the population is less than 10. -6 This indicates that the algorithm has converged.

[0056] The chromosome of the individual with the highest fitness value in the population at the time of termination is identified as the target individual chromosome. Its 66-dimensional gene value is decoded into the corresponding network parameters, and all network parameters in the current wavelet neural network model are replaced to obtain the optimized wavelet neural network model. This optimized wavelet neural network model is then used as the updated current wavelet neural network model for subsequent noise reduction processing and the next signal processing cycle.

[0057] As can be seen from the above, this embodiment uses a genetic algorithm process of real number encoding, roulette wheel selection, arithmetic crossover, and Gaussian mutation to globally optimize all 66 network parameters of the wavelet neural network. This effectively overcomes the inherent defects of local optimization methods such as gradient descent, which are sensitive to initial parameters and prone to getting trapped in local extrema. Combined with the adaptively determined population size and crossover probability, the optimization process achieves a dynamic balance between search breadth and convergence speed.

[0058] S107: Based on the updated current wavelet neural network model, the actual noise signal is denoised to obtain the denoised transformer noise signal.

[0059] In this embodiment, noise reduction refers to the process of separating the inherent noise component of the transformer from the actual noise signal using the updated current wavelet neural network model, and subtracting this inherent noise component from the actual noise signal to obtain the noise-reduced signal. The core idea is that the wavelet neural network model, after global optimization by the genetic algorithm, can more accurately approximate the real noise characteristics of the transformer under the current operating conditions. Removing the output of this model from the actual noise can effectively suppress noise interference. Signal reconstruction refers to the process of restoring the processed signal data to the original sampling rate and data length.

[0060] In one embodiment of this application, the actual noise signal is denoised based on the updated current wavelet neural network model to obtain a denoised transformer noise signal, including: The actual noise signal is input into the updated wavelet neural network model to obtain the optimized model noise signal. The target noise reduction signal is obtained by subtracting the optimized model noise signal from the actual noise signal. The target noise-reduced signal is reconstructed to obtain the noise-reduced transformer noise signal.

[0061] In this embodiment, the optimized model noise signal refers to the noise prediction result obtained by forward inference of the actual noise signal using a wavelet neural network model that has been globally optimized by a genetic algorithm. Compared with the model noise signal before optimization, it more accurately represents the true noise characteristics of the transformer under the current operating conditions. The target noise reduction signal refers to the difference signal obtained by subtracting the actual noise signal from the optimized model noise signal point by point.

[0062] In this embodiment, the actual noise signal sequence with 10240 sampling points is segmented using the same sliding window method (window length of 8 sampling points, with adjacent windows overlapping by 4 sampling points), resulting in several sets of 8-dimensional input vectors. These input vectors are sequentially fed into the updated wavelet neural network model for forward inference calculation, with each input vector generating a single-point output value. All single-point output values ​​are concatenated in chronological order to obtain the preliminary optimized model noise signal.

[0063] Because there is an overlap of four sampling points between adjacent windows, two output values ​​from different windows will appear within the overlap area. To eliminate this boundary discontinuity, a linearly gradient weighted average is used to smooth the two output values ​​within the overlap area: for the k-th sampling point (k=1,2,3,4) within the overlap area, the weight of the output from the previous window is (4-k) / 4, and the weight of the output from the next window is k / 4. After smoothing, the final optimized model noise signal is obtained, which has the exact same data length as the actual noise signal.

[0064] In this embodiment, the optimized model noise signal after smoothing is subtracted from the actual noise signal at each sampling point, i.e., d(n) = x(n) - (n), where x(n) is the amplitude of the actual noise signal at the nth sampling point. d(n) represents the amplitude of the optimized model noise signal at the nth sampling point, and d(n) represents the amplitude of the target noise-reduced signal at the nth sampling point.

[0065] Finally, the target denoised signal undergoes signal reconstruction. Specifically, the Daubechies-4 wavelet basis function, corresponding to wavelet decomposition, is used to perform a three-level wavelet decomposition on the target denoised signal, obtaining the detail coefficients and approximation coefficients at each level. The first-level detail coefficients (corresponding to the highest frequency component, with a frequency range of 2560Hz to 5120Hz) are set to zero, while the second-level, third-level detail coefficients, and third-level approximation coefficients remain unchanged. Then, the Daubechies-4 wavelet basis function is used to perform a three-level inverse wavelet transform reconstruction to obtain the denoised transformer noise signal. The reason for setting the first-level detail coefficients to zero is that the high-frequency artifacts introduced during the sliding window segmentation and model prediction processes are mainly concentrated in the highest frequency band. Removing components in this frequency band can effectively eliminate artifacts without losing the effective characteristic information of the transformer noise. The denoised transformer noise signal is finally output to the downstream transformer condition monitoring module or fault diagnosis module.

[0066] As can be seen from the above, this embodiment effectively eliminates the signal boundary discontinuities and high-frequency artifacts introduced during the segmented processing and model prediction process by performing linear gradual weighted smoothing on the overlapping area of ​​the sliding window and combining it with wavelet reconstruction to remove high-frequency artifacts. This ensures the continuity and smoothness of the signal after noise reduction and provides high-quality signal input for downstream transformer condition analysis.

[0067] S108: If the absolute value of the amplitude is less than the optimized trigger threshold, the actual noise signal is denoised based on the current wavelet neural network model to obtain the denoised transformer noise signal.

[0068] In this embodiment, when the absolute value of the residual noise signal amplitude is less than the optimization trigger threshold, it indicates that the current wavelet neural network model still meets the requirements for modeling transformer noise accuracy, and there is no need to trigger the computationally expensive global optimization process of the genetic algorithm. At this time, the current wavelet neural network model is directly used to denoise the actual noise signal. The specific operation method is exactly the same as the forward inference, sliding window segmentation, signal subtraction, and wavelet reconstruction process. The only difference is that the wavelet neural network model used here is the original model that has not been updated in this optimization, that is, the current wavelet neural network model keeps its original parameters unchanged.

[0069] For example, assuming the calculated MAE is 0.15, which is less than the optimization trigger threshold of 0.2, the actual noise signal is directly input into the current wavelet neural network model to obtain the model noise signal. The actual noise signal is then subtracted from the model noise signal to obtain the target denoised signal. Finally, wavelet reconstruction is performed on the target denoised signal to obtain the denoised transformer noise signal. This process omits the determination of the operating condition type, the calculation of genetic algorithm parameters, and the global optimization process, thereby shortening the computation time of this signal processing cycle and significantly reducing computational resource consumption and system latency.

[0070] As can be seen from the above, this embodiment calculates the absolute value of the residual noise signal amplitude and compares it with the optimization trigger threshold. When the absolute value of the amplitude is greater than or equal to the threshold, the current load condition type is determined based on the load current signal. The population size and crossover probability of the genetic algorithm are adaptively determined by combining the absolute value of the amplitude and the condition type, thereby performing global optimization of the wavelet neural network model. When the absolute value of the amplitude is less than the threshold, the current model is directly used for noise reduction. This scheme avoids the huge computational overhead caused by performing global optimization of the genetic algorithm for each signal processing by introducing an optimization trigger mechanism. At the same time, based on the dual-dimensional joint decision of load condition type and residual noise amplitude, the genetic algorithm parameters can accurately match the complexity of the current condition. When the model accuracy decreases, adaptive optimization is triggered in time, ultimately achieving a dynamic balance between computational efficiency and noise reduction accuracy, meeting the real-time requirements of online monitoring scenarios for distribution transformers.

[0071] In one embodiment of this application, before acquiring the actual noise signal and load current signal of the distribution transformer during operation, the method further includes: Historical noise signal samples of distribution transformers under various historical load conditions are obtained, and wavelet decomposition is performed on the historical noise signal samples to obtain wavelet coefficients at each scale. Using wavelet coefficients as input data, an initial wavelet neural network model is constructed, comprising an input layer, a hidden layer, and an output layer; wherein, the hidden layer uses wavelet basis functions as activation functions; The initial wavelet neural network model is trained based on the input data, and the trained initial wavelet neural network model is determined as the current wavelet neural network model.

[0072] In this embodiment, historical noise signal samples refer to a set of noise signal data that was pre-collected and stored under various load conditions during the historical operation of the distribution transformer after it was put into operation, using the same acquisition method as in step S101. These data are used to train the initial model of the wavelet neural network during the initialization phase. Wavelet decomposition refers to the mathematical transformation process of decomposing a signal into coefficient components at different scales (i.e., different frequency ranges) using wavelet basis functions. Wavelet coefficients at each scale refer to the coefficient sequences obtained at each decomposition level after wavelet decomposition, including the detail coefficients at each level and the approximation coefficients at the final level.

[0073] In this embodiment, during the commissioning phase after the distribution transformer is put into operation, the transformer load is manually adjusted or natural load changes are waited for, allowing the transformer to operate under various typical conditions such as steady-state light load, steady-state heavy load, and sudden load changes. Under each condition, a microphone array is used to continuously collect noise signals at a sampling rate of 10240Hz, with at least 300 seconds of data collected for each condition, thereby obtaining historical noise signal samples under various historical load conditions. For the historical noise signal samples under each condition, a three-level wavelet decomposition is performed using the Daubechies-4 wavelet basis function, resulting in four sets of wavelet coefficients: level 1 detail coefficients, level 2 detail coefficients, level 3 detail coefficients, and level 3 approximation coefficients. The number of data points in each set of coefficients is 5120, 2560, 1280, and 1280 respectively, depending on the downsampling characteristics of the wavelet decomposition. The above sets of wavelet coefficients are concatenated into a unified wavelet coefficient vector according to time alignment, which serves as the input data for the initial model of the wavelet neural network.

[0074] In this embodiment, the normalized wavelet coefficient vector has a dimension of 8, therefore the number of input layer nodes is set to 8. The number of hidden layer nodes is calculated according to a preset formula: Number of hidden layer nodes = (Number of input layer nodes + Number of output layer nodes) / 2 + Preset constant, where the number of output layer nodes is 1 (corresponding to single-point noise amplitude prediction), and the preset constant is set to 2. Therefore, the number of hidden layer nodes is calculated to be (8+1) / 2+2 = 6.5, rounded down to 6. The hidden layer uses the Morlet wavelet basis function as the activation function. The initial wavelet neural network model is trained using the above input data. During training, gradient descent is used to update the network parameters, the learning rate is set to 0.01, and the training epochs are set to 500. When the mean square error of the training set drops to 10... -4 Training will stop when the maximum number of training epochs is reached, and the initial wavelet neural network model after training will be determined as the current wavelet neural network model. The preset constant is set to 2 based on the following: through cross-validation experiments, the generalization performance of the models with preset constants of 1, 2, and 3 was compared, and it was found that the model with a preset constant of 2 had the smallest prediction error on the validation set.

[0075] As can be seen from the above, this embodiment pre-trains the wavelet neural network using noise samples under various historical load conditions during the system initialization phase, enabling the initial model to have prior knowledge of the basic characteristics of transformer noise. This avoids the problems of slow convergence speed and poor initial noise reduction effect caused by starting optimization from random parameter states, and provides a good parameter starting point for subsequent online adaptive optimization.

[0076] In one embodiment of this application, a wavelet neural network initial model comprising an input layer, a hidden layer, and an output layer is constructed using wavelet coefficients as input data, including: The number of nodes in the input layer is determined based on the dimension of the wavelet coefficients; The number of nodes in the hidden layer is calculated based on the number of nodes in the input layer and a preset formula for determining the number of nodes in the hidden layer. The formula for determining the number of nodes in the hidden layer is that the number of nodes in the hidden layer is equal to half the sum of the number of nodes in the input layer and the number of nodes in the output layer plus a preset constant. The Morlet wavelet basis function is selected as the activation function of the hidden layer, and the initial model of the wavelet neural network is constructed according to the number of nodes in the input layer, the number of nodes in the hidden layer, and the number of nodes in the output layer.

[0077] In this embodiment, the Morlet wavelet basis function is a complex-valued wavelet basis function composed of a cosine-modulated Gaussian window. The preset hidden layer node determination formula is an empirical formula for calculating the number of hidden layer nodes based on the number of nodes in the input and output layers.

[0078] In this embodiment, the number of nodes in the input layer is determined based on the dimension of the acquired wavelet coefficients. In this embodiment, the wavelet coefficients have a dimension of 8 after normalization, therefore the number of input layer nodes is determined to be 8. Based on the input layer node count of 8 and the output layer node count of 1, the number of hidden layer nodes is calculated according to a preset formula for determining hidden layer nodes. This formula is: Number of hidden layer nodes = (Number of input layer nodes + Number of output layer nodes) / 2 + Preset constant. The preset constant typically ranges from 1 to 5; in this embodiment, the preset constant is set to 2. Substituting the values, we get: Number of hidden layer nodes = (8 + 1) / 2 + 2 = 6.5. Following the rounding down rule, the number of hidden layer nodes is determined to be 6. The design principle of this formula is that too few hidden layer nodes will lead to insufficient model expressive power, failing to fully fit the complex nonlinear characteristics of transformer noise; too many hidden layer nodes will lead to redundant model parameters, increasing the computational burden and easily causing overfitting problems. The above formula introduces a preset constant on top of the input and output dimensions, providing the hidden layer with a suitable amount of additional degrees of freedom.

[0079] Select Morlet wavelet basis functions The Morlet wavelet basis function is used as the activation function for the hidden layers. The parameter 1.75 in the Morlet wavelet basis function is the modulation frequency of the standard Morlet wavelet, which achieves the optimal balance between the time and frequency domains, satisfying the wavelet admissibility condition. Finally, based on the number of input layer nodes (8), hidden layer nodes (6), and output layer nodes (1), an initial wavelet neural network model is constructed according to the standard feedforward neural network topology. The input and hidden layers are fully connected, and the hidden and output layers are also fully connected. The initial values ​​of the network parameters are randomly initialized within the interval [-1, 1].

[0080] As can be seen from the above, this embodiment uses an empirical formula based on the input and output dimensions to determine the number of hidden layer nodes, and combines the excellent time-frequency characteristics of the Morlet wavelet basis function. This avoids excessive complexity of the network structure while ensuring that the model has sufficient nonlinear approximation capability, thus achieving a good balance between parameter efficiency and expressive power in the initial wavelet neural network model.

[0081] Based on the same principle as the signal processing method for a distribution transformer provided in the embodiments of this application, the embodiments of this application also provide a signal processing device for a distribution transformer, such as... Figure 2 As shown, the signal processing device 20 for a distribution transformer may specifically include: a signal acquisition module 21, a model output module 22, a residual calculation module 23, a comparison module 24, a parameter determination module 25, a model update module 26, a first noise reduction module 27, and a second noise reduction module 28. Signal acquisition module 21 is used to acquire the actual noise signal and load current signal of the distribution transformer during operation; The model output module 22 is used to obtain the model noise signal based on the actual noise signal and through the current wavelet neural network model; The residual calculation module 23 is used to obtain the residual noise signal based on the difference between the actual noise signal and the model noise signal; Comparison module 24 is used to calculate the absolute value of the amplitude of the residual noise signal and compare the absolute value of the amplitude with a preset optimized trigger threshold. The parameter determination module 25 is used to determine the current load condition type of the distribution transformer based on the load current signal if the absolute value of the amplitude is greater than or equal to the optimization trigger threshold, and to determine the population size and crossover probability of the genetic algorithm based on the absolute value of the amplitude and the current load condition type. The model update module 26 is used to globally optimize the network parameters of the current wavelet neural network model based on the population size and crossover probability and through a genetic algorithm to obtain the optimized wavelet neural network model, and use the optimized wavelet neural network model as the updated current wavelet neural network model. The first noise reduction module 27 is used to perform noise reduction processing on the actual noise signal based on the updated current wavelet neural network model to obtain the noise-reduced transformer noise signal. The second noise reduction module 28 is used to perform noise reduction processing on the actual noise signal based on the current wavelet neural network model if the absolute value of the amplitude is less than the optimized trigger threshold, so as to obtain the noise-reduced transformer noise signal.

[0082] In one embodiment of this application, a signal processing device 20 for a distribution transformer further includes: a model determination module, used for: Historical noise signal samples of distribution transformers under various historical load conditions are obtained, and wavelet decomposition is performed on the historical noise signal samples to obtain wavelet coefficients at each scale. Using wavelet coefficients as input data, an initial wavelet neural network model is constructed, comprising an input layer, a hidden layer, and an output layer; wherein, the hidden layer uses wavelet basis functions as activation functions; The initial wavelet neural network model is trained based on the input data, and the trained initial wavelet neural network model is determined as the current wavelet neural network model.

[0083] In one embodiment of this application, the model determination module is specifically used for: The number of nodes in the input layer is determined based on the dimension of the wavelet coefficients; The number of nodes in the hidden layer is calculated based on the number of nodes in the input layer and a preset formula for determining the number of nodes in the hidden layer. The formula for determining the number of nodes in the hidden layer is that the number of nodes in the hidden layer is equal to half the sum of the number of nodes in the input layer and the number of nodes in the output layer plus a preset constant. The Morlet wavelet basis function is selected as the activation function of the hidden layer, and the initial model of the wavelet neural network is constructed according to the number of nodes in the input layer, the number of nodes in the hidden layer, and the number of nodes in the output layer.

[0084] In one embodiment of this application, the parameter determination module 25 is specifically used for: Calculate the effective value and rate of change of the load current signal within a preset time window; Based on the effective value of the current and the rate of change of the current, the current load condition type of the distribution transformer is determined; the current load condition type includes one of the following: steady-state light load condition, steady-state heavy load condition, and load change condition.

[0085] In one embodiment of this application, the parameter determination module 25 is specifically used for: Based on the preset amplitude range division rules, the target amplitude range to which the absolute value of the amplitude belongs is determined; Based on the preset mapping relationship between operating condition type and population size benchmark value, determine the population size benchmark value corresponding to the current load operating condition type. Based on the preset amplitude adjustment coefficient corresponding to the target amplitude range, the baseline value of the population size is corrected to obtain the population size of the genetic algorithm; Based on the preset mapping relationship between operating condition types and cross probability benchmark values, determine the cross probability benchmark value corresponding to the current load operating condition type; The crossover probability baseline value is corrected based on the preset probability adjustment coefficient corresponding to the target amplitude range to obtain the crossover probability of the genetic algorithm.

[0086] In one embodiment of this application, the model update module 26 is specifically used for: The network parameters of the current wavelet neural network model are encoded as individual chromosomes of a genetic algorithm; where the network parameters include the weight coefficients of the current wavelet neural network model, the translation parameters and scaling parameters of the wavelet basis functions; An initial population is generated based on the population size, and the initial population contains multiple individual chromosomes; The fitness value of each individual's chromosome in the initial population is calculated by using the minimization of the residual noise signal energy as the fitness function; Based on the fitness value, crossover probability, and preset mutation probability, the initial population is subjected to selection, crossover, and mutation operations in sequence, and the population is iteratively updated until the preset iteration termination condition is met. The chromosome of the individual with the largest fitness value is determined as the target individual chromosome. The target individual's chromosome is decoded into corresponding network parameters, and the current wavelet neural network model is updated with the decoded network parameters to obtain the optimized wavelet neural network model.

[0087] In one embodiment of this application, the first noise reduction module 27 is specifically used for: The actual noise signal is input into the updated wavelet neural network model to obtain the optimized model noise signal. The target noise reduction signal is obtained by subtracting the optimized model noise signal from the actual noise signal. The target noise-reduced signal is reconstructed to obtain the noise-reduced transformer noise signal.

[0088] The apparatus in this application embodiment can execute the method provided in this application embodiment, and the implementation principle is similar. The actions performed by each module in the apparatus of each embodiment of this application correspond to the steps in the method of each embodiment of this application. For detailed functional descriptions of each module of the apparatus, please refer to the descriptions in the corresponding methods shown above, which will not be repeated here.

[0089] Figure 3 A schematic diagram of the structure of an electronic device to which this application embodiment applies is shown, such as... Figure 3 As shown, the electronic device can be used to implement the methods provided in any embodiment of this application.

[0090] like Figure 3 As shown, the electronic device 300 may primarily include at least one processor 301. Figure 3 The diagram shows components such as a memory 302, a communication module 303, and an input / output interface 304. Optionally, these components can be connected and communicate with each other via a bus 305. It should be noted that... Figure 3The structure of the electronic device 300 shown is merely illustrative and does not constitute a limitation on the electronic devices to which the methods provided in the embodiments of this application are applicable.

[0091] The memory 302 can be used to store operating systems and applications, etc. The applications can include computer programs that implement the methods shown in the embodiments of this application when invoked by the processor 301, and can also include programs for implementing other functions or services. The memory 302 can be ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices that can store information and computer programs, or it can be EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited thereto.

[0092] Processor 301 is connected to memory 302 via bus 305 and implements corresponding functions by calling the application programs stored in memory 302. Processor 301 can be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 301 can also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0093] Electronic device 300 can connect to a network via communication module 303 (which may include, but is not limited to, components such as a network interface) to communicate with other devices (such as user terminals or servers) through the network and achieve data interaction, such as sending data to or receiving data from other devices. Communication module 303 may include wired network interfaces and / or wireless network interfaces, meaning the communication module may include at least one of wired or wireless communication modules.

[0094] The electronic device 300 can connect to necessary input / output devices, such as a keyboard and display device, via the input / output interface 304. The electronic device 300 itself may have a display device, and other external display devices can also be connected via the input / output interface 304. Optionally, a storage device, such as a hard drive, can also be connected via the input / output interface 304 to store data from the electronic device 300, retrieve data from the storage device, or store data from the storage device in the memory 302. It is understood that the input / output interface 304 can be a wired interface or a wireless interface. Depending on the actual application scenario, the device connected to the input / output interface 304 can be a component of the electronic device 300 or an external device connected to the electronic device 300 when needed.

[0095] The bus 305 used to connect the components may include a path for transmitting information between the components. The bus 305 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Depending on its function, the bus 305 may be divided into an address bus, a data bus, a control bus, etc.

[0096] Optionally, for the solution provided in the embodiments of this application, the memory 302 can be used to store a computer program that executes the solution of this application, and the processor 301 runs the computer program. When the processor 301 runs the computer program, it implements the operation of the method or apparatus provided in the embodiments of this application.

[0097] Based on the same principle as the method provided in the embodiments of this application, the embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the corresponding content of the aforementioned method embodiments.

[0098] This application also provides a computer program product, which includes a computer program that, when executed by a processor, can implement the corresponding content of the aforementioned method embodiments.

[0099] It should be noted that the terms "first," "second," "third," "fourth," "1," "2," etc. (if present) in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in a sequence other than that shown in the figures or text descriptions.

[0100] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0101] It should be understood that although arrows indicate various operation steps in the flowcharts of this application's embodiments, the order in which these steps are implemented is not limited to the order indicated by the arrows. Unless explicitly stated herein, in some implementation scenarios of this application's embodiments, the implementation steps in each flowchart can be executed in other orders as required. Furthermore, some or all steps in each flowchart, based on the actual implementation scenario, may include multiple sub-steps or multiple stages. Some or all of these sub-steps or stages can be executed at the same time, and each sub-step or stage can also be executed at different times. In scenarios where execution times differ, the execution order of these sub-steps or stages can be flexibly configured according to requirements, and this application's embodiments do not limit this.

[0102] The above description is only an optional implementation method for some implementation scenarios of this application. It should be noted that for those skilled in the art, other similar implementation methods based on the technical concept of this application without departing from the technical concept of this application also fall within the protection scope of the embodiments of this application.

Claims

1. A signal processing method for a distribution transformer, characterized in that, include: Acquire the actual noise signal and load current signal of the distribution transformer during operation; Based on the actual noise signal, and through the current wavelet neural network model, the model noise signal is obtained; The residual noise signal is obtained based on the difference between the actual noise signal and the model noise signal; Calculate the absolute value of the amplitude of the residual noise signal and compare the absolute value of the amplitude with a preset optimized trigger threshold; If the absolute value of the amplitude is greater than or equal to the optimization trigger threshold, the current load condition type of the distribution transformer is determined based on the load current signal, and the population size and crossover probability of the genetic algorithm are determined based on the absolute value of the amplitude and the current load condition type. Based on the population size and the crossover probability, the network parameters of the current wavelet neural network model are globally optimized using the genetic algorithm to obtain an optimized wavelet neural network model, and the optimized wavelet neural network model is used as the updated current wavelet neural network model. Based on the updated wavelet neural network model, the actual noise signal is denoised to obtain the denoised transformer noise signal. If the absolute value of the amplitude is less than the optimized trigger threshold, then the actual noise signal is denoised based on the current wavelet neural network model to obtain the denoised transformer noise signal.

2. The signal processing method for a distribution transformer as described in claim 1, characterized in that, Before acquiring the actual noise signal and load current signal of the distribution transformer during operation, the method further includes: Historical noise signal samples of the distribution transformer under various historical load conditions are obtained, and wavelet decomposition is performed on the historical noise signal samples to obtain wavelet coefficients at each scale. Using the wavelet coefficients as input data, an initial wavelet neural network model comprising an input layer, a hidden layer, and an output layer is constructed; wherein, the hidden layer uses wavelet basis functions as activation functions; The initial wavelet neural network model is trained based on the input data, and the trained initial wavelet neural network model is determined as the current wavelet neural network model.

3. The signal processing method for a distribution transformer as described in claim 2, characterized in that, The step of constructing an initial wavelet neural network model containing an input layer, a hidden layer, and an output layer using the wavelet coefficients as input data includes: The number of nodes in the input layer is determined based on the dimension of the wavelet coefficients; The number of nodes in the hidden layer is calculated based on the number of nodes in the input layer and a preset formula for determining the number of nodes in the hidden layer; wherein, the formula for determining the number of nodes in the hidden layer is equal to half the sum of the number of nodes in the input layer and the number of nodes in the output layer plus a preset constant. The Morlet wavelet basis function is selected as the activation function of the hidden layer, and the initial model of the wavelet neural network is constructed according to the number of nodes in the input layer, the number of nodes in the hidden layer, and the number of nodes in the output layer.

4. The signal processing method for a distribution transformer as described in claim 1, characterized in that, Determining the current load condition type of the distribution transformer based on the load current signal includes: Calculate the effective value of the load current signal and the rate of change of the current within a preset time window; Based on the effective value of the current and the rate of change of the current, the current load condition type of the distribution transformer is determined; wherein, the current load condition type includes one of steady-state light load condition, steady-state heavy load condition, and load change condition.

5. The signal processing method for a distribution transformer as described in claim 4, characterized in that, The process of determining the population size and crossover probability of the genetic algorithm based on the absolute value of the amplitude and the current load condition type includes: Based on the preset amplitude range division rules, the target amplitude range to which the absolute value of the amplitude belongs is determined; Based on the preset mapping relationship between operating condition type and population size benchmark value, the population size benchmark value corresponding to the current load operating condition type is determined. The population size baseline value is corrected according to the preset amplitude adjustment coefficient corresponding to the target amplitude range to obtain the population size of the genetic algorithm; Based on the preset mapping relationship between operating condition types and cross probability benchmark values, determine the cross probability benchmark value corresponding to the current load operating condition type; The crossover probability benchmark value is corrected according to the preset probability adjustment coefficient corresponding to the target amplitude range to obtain the crossover probability of the genetic algorithm.

6. The signal processing method for a distribution transformer as described in claim 1, characterized in that, Based on the population size and the crossover probability, and by globally optimizing the network parameters of the current wavelet neural network model using the genetic algorithm, an optimized wavelet neural network model is obtained, including: The network parameters of the current wavelet neural network model are encoded as individual chromosomes of the genetic algorithm; wherein, the network parameters include the weight coefficients, translation parameters, and scaling parameters of the wavelet basis functions of the current wavelet neural network model; An initial population is generated based on the population size, and the initial population contains multiple individual chromosomes; Using the minimization of the residual noise signal as the fitness function, the fitness value of each individual's chromosome in the initial population is calculated; Based on the fitness value, the crossover probability, and the preset mutation probability, the initial population is subjected to selection, crossover, and mutation operations in sequence, and the population is iteratively updated until the preset iteration termination condition is met, and the individual chromosome with the largest fitness value is determined as the target individual chromosome. The target individual's chromosome is decoded into corresponding network parameters, and the current wavelet neural network model is updated with the decoded network parameters to obtain the optimized wavelet neural network model.

7. The signal processing method for a distribution transformer as described in claim 1, characterized in that, The actual noise signal is denoised based on the updated wavelet neural network model to obtain the denoised transformer noise signal, including: The actual noise signal is input into the updated wavelet neural network model to obtain the optimized model noise signal; Subtract the optimized model noise signal from the actual noise signal to obtain the target noise reduction signal; The target noise-reduced signal is reconstructed to obtain the noise-reduced transformer noise signal.

8. A signal processing device for a distribution transformer, characterized in that, include: The signal acquisition module is used to acquire the actual noise signal and load current signal of the distribution transformer during operation; The model output module is used to obtain the model noise signal based on the actual noise signal and through the current wavelet neural network model; The residual calculation module is used to obtain the residual noise signal based on the difference between the actual noise signal and the model noise signal; The comparison module is used to calculate the absolute value of the amplitude of the residual noise signal and compare the absolute value of the amplitude with a preset optimization trigger threshold. The parameter determination module is used to determine the current load condition type of the distribution transformer based on the load current signal if the absolute value of the amplitude is greater than or equal to the optimization trigger threshold, and to determine the population size and crossover probability of the genetic algorithm based on the absolute value of the amplitude and the current load condition type. The model update module is used to globally optimize the network parameters of the current wavelet neural network model based on the population size and the crossover probability, and obtain an optimized wavelet neural network model, and use the optimized wavelet neural network model as the updated current wavelet neural network model. The first noise reduction module is used to perform noise reduction processing on the actual noise signal based on the updated current wavelet neural network model to obtain the noise-reduced transformer noise signal. The second noise reduction module is used to perform noise reduction processing on the actual noise signal based on the current wavelet neural network model if the absolute value of the amplitude is less than the optimized trigger threshold, so as to obtain the noise-reduced transformer noise signal.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program, and the processor executes a signal processing method for a distribution transformer according to any one of claims 1 to 7 when running the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements a signal processing method for a distribution transformer as described in any one of claims 1 to 7.

Images