76 results about "Wavelet filter" patented technology

The invention relates to the field of data processing, in particular to a composite insulator vulcanization uniformity detection method and device. The method comprises the following steps: acquiring an NMR spectrogram of a vulcanized insulator; performing first-layer decomposition processing on the NMR spectrogram by using a wavelet filtering algorithm to obtain a detail coefficient; a noise detail coefficient and a signal detail coefficient in the approximation coefficient are distinguished; performing second-layer decomposition processing on the approximation coefficient to obtain a detail coefficient; calculating the signal purity; in response to the fact that the signal purity is smaller than a preset purity threshold value, noise detail coefficients in the detail coefficients obtained through second-layer decomposition processing are screened out; continuing decomposition and calculating the signal purity, and stopping decomposition when the signal purity is not less than a preset purity threshold value; after the noise detail coefficient is corrected and adjusted, a denoised NMR spectrogram is reconstructed; therefore, vulcanization uniformity detection is realized. Therefore, the accuracy of vulcanization uniformity detection is improved by improving the accuracy of denoising.

The invention discloses a photovoltaic power ultra-short-term prediction method, system and device and a storage medium, and relates to the technical field of new energy prediction. The method comprises the following steps: constructing an improved ensemble empirical mode decomposition algorithm to decompose original photovoltaic power data, and distinguishing and reconstructing a plurality of decomposed intrinsic mode functions to obtain reconstruction sequences of high, medium and low frequency components; for high-frequency and intermediate-frequency components, carrying out dynamic denoising by adopting an extended Kalman filtering algorithm; and inputting the original data, the low-frequency component, the denoised high-frequency component and the denoised intermediate-frequency component into a WTConv1d module, extracting multi-scale features through a wavelet filter and convolution operation, and inputting a multi-dimensional input matrix constructed by the multi-scale features and meteorological data into a Transform model for prediction. According to the method, noise interference can be effectively suppressed, multi-time scale feature expression is enhanced, and a self-attention mechanism is fully utilized to capture a cross-scale dependency relationship, so that the prediction precision and stability are improved.

The invention relates to the field of data processing, in particular to a data processing method and system of a thulium laser treatment device. The method comprises the following steps: acquiring a laser data time sequence and a biological reflection data time sequence; dividing the laser data time sequence and the biological reflection data time sequence into a plurality of laser data segments and a plurality of biological reflection data segments respectively; setting a plurality of preset layer numbers, setting the decomposition layer number of a wavelet filtering algorithm as the preset layer number, and carrying out filtering processing on the laser data segment to obtain a denoised data segment; the denoising credibility of the preset number of layers is calculated, and the denoising credibility is positively correlated with the signal-to-noise splitting degree and negatively correlated with the information loss degree; and selecting the denoised data segments obtained by the preset number of layers with the maximum denoising credibility as final denoised data segments, and splicing all the final denoised data segments into a final denoised time sequence. The denoising accuracy is improved by setting a proper number of decomposition layers.

The invention provides a hot rolled steel strip surface defect detection method based on target detection, and aims to solve the problems of difficulty in multi-scale small target recognition, insufficient feature fusion and complex background interference in hot rolled steel strip surface defect detection. Extracting multi-scale defect features by using wavelet convolution through decomposition, convolution and reconstruction processes; multi-scale features are adaptively fused based on channels and space attention through a dynamic feature fusion module (DFF); a dynamic hyperbolic tangent function (DyT) is used to replace a normalization layer to dynamically adjust the input range and compress extreme values to reduce the amount of calculation. The method has the effects of high detection average precision (mAP), low missing report rate and remarkably reduced calculation complexity, effectively improves the efficiency and reliability of identifying the surface defects of the hot rolled steel strip in a complex industrial environment, and is suitable for intelligent steel manufacturing quality control.

The invention provides an attention detection method and system based on a brain-computer interface, and particularly relates to the technical field of brain-computer interfaces, and the system integrates a hardware unit (including TGAM electroencephalogram acquisition, Bluetooth connection and LED illumination control), a software unit (including login, detection, display and data export) and an algorithm unit. According to the method, electroencephalogram collection is carried out, wavelet filtering is utilized to decompose EEG signals, an Attention value and a Meditation value are extracted after calculation, a fusion value AM of the Attention value and the Meditation value is further provided, and control over the LED illumination brightness is achieved according to the fusion value. Besides, the functions of user verification, signal processing, result display and data recording are achieved through the software unit, the problems of extraction and fusion of attention features and relaxation features of electroencephalogram signals in the man-machine interaction process are solved through the algorithm unit, the accuracy and reliability of collection and processing of electroencephalogram data are improved, and the system is suitable for popularization and application. Effective attention feature extraction and feature fusion based on the electroencephalogram signals and adjustment of the LED illumination brightness are achieved, and development of the human-computer interaction technology is promoted.

The invention relates to the technical field of data processing, in particular to an equipment operation state monitoring method and system based on the Internet of Things. The method comprises the following steps: acquiring an oxygen concentration time sequence and a temperature time sequence of equipment; acquiring a plurality of preset local ranges of the oxygen concentration data at each moment in the oxygen concentration time sequence, calculating the noise expression degree of the oxygen concentration data, and calculating the corrected noise expression degree of the oxygen concentration data according to the noise expression degree, the number of decomposition layers in a wavelet filtering algorithm is set according to the corrected noise expression degree, and denoising processing is achieved so as to assist equipment operation state monitoring. Therefore, efficient and accurate de-noising processing is realized through adaptive decomposition of the number of layers.

The invention discloses a method for simplifying data points of a piecewise linear excitation source, a time sequence and a data value sequence corresponding to waveform data points needing to be simplified are set as tk and vk respectively, k = 1, 2,..., n, and the method comprises the following steps: S1, carrying out wavelet filtering; s2, screening out special extreme points; s3, data points are screened through difference values; and S4, screening the data points through the slope. According to the simplification method for the piecewise linear excitation source data points, the core data points in the actual waveform original data points can be extracted, the main features of the original data points are reserved by the core data points, and the number of the core data points is far smaller than that of the original data points. Core data points are used as input of a piecewise linear excitation source, so that simulation time is remarkably shortened, and simulation convergence is facilitated.

The present invention relates to the field of wind turbine monitoring technology, and specifically to a monitoring method, device, equipment and medium for a wind turbine pitch system. The method comprises: collecting image data of the wind turbine pitch system; grayscale processing the image data based on an adaptive grayscale algorithm adjusted by parameters; filtering the grayscale image data using median filtering, Gaussian filtering and wavelet filtering to obtain filtered image data; edge detection and feature and feature point recognition of the filtered image data, wherein the feature point recognition adopts SIFT and SURF fusion algorithms; based on the recognized feature points, the angle of the wind turbine pitch system is calculated according to a geometric relationship algorithm, or, based on the recognized features, the angle of the wind turbine pitch system is calculated according to a template matching algorithm. This method overcomes many drawbacks of traditional sensor monitoring methods, and can timely and accurately grasp the operating status of the pitch system, effectively reducing the probability of wind turbine shutdown due to pitch system failure.

The invention discloses a physical information neural network and a classification model construction and classification method for micro-seismic monitoring data. The constructed classification model comprises a signal embedding block, a Transform encoder and a multi-layer sensor. A signal embedding block maps a microseismic signal to query, key and value vector spaces, and a Transform encoder captures physical information features in an input query feature vector, a key feature vector and a value feature vector. The multi-layer sensor carries out classification according to the physical information characteristics captured by the Transform encoder, and determines the category of the input signal. The Transform encoder adopts a discrete wavelet filter to replace a linear transformation layer of a self-attention mechanism. Wavelet physical information is introduced through a discrete wavelet filter to restrain the physical information neural network, so that the network can learn features which have clear physical significance and can be used for classification from microseismic signals, the interpretability and reliability of the model are improved, effective extraction and fusion of the features are realized, and the classification accuracy is improved. And thus, the reliability and accuracy of road slope micro-seismic early warning are improved.

The embodiment of the invention provides a bearing fault diagnosis method and system for optimizing filtering MCKD, and the method comprises the steps: collecting a vibration signal, and carrying out the Fourier transform of a frequency spectrum of the vibration signal, and obtaining a key function; setting an initial interception length, performing inverse Fourier transform on the key function based on the initial interception length to obtain a frequency spectrum trend, dividing filtering frequency bands with different bandwidths and center frequencies by taking a minimum value of a trend spectrum as a boundary of frequency spectrum segmentation, increasing the interception length, repeating the steps, and recording all results; designing a Mayer wavelet filter based on the obtained filtering frequency band, and taking a designed Mayer wavelet filter coefficient as an initial filter coefficient of MCKD; mCKD filtering results of different initial filter coefficients are calculated, the fault characteristics of all filtering signals are quantized through harmonic spectrum kurtosis HSK, the filtering result corresponding to the maximum HSK value is selected as the global optimal solution of MCKD, and fault diagnosis is completed.

The embodiment of the invention provides a self-adaptive target threshold determination method and device, and the method comprises the steps: obtaining a detection cavity sonar signal, and enabling the detection cavity sonar signal to be collected by a detection cavity sonar; performing multi-layer wavelet decomposition on the cavity measurement sonar signal to obtain a low-frequency coefficient of each layer and a high-frequency coefficient of each layer after decomposition; performing threshold processing on the high-frequency coefficient of each layer according to an initial threshold, and performing signal reconstruction based on the low-frequency coefficient of each layer and the high-frequency coefficient of each layer after the threshold processing to obtain a reconstruction signal corresponding to the initial threshold; and according to the noise degrees of the reconstructed signals corresponding to the plurality of initial thresholds, determining a target threshold from the plurality of initial thresholds, the target threshold being used for performing wavelet filtering on a target signal collected by the cavity measurement sonar. The target threshold value obtained through the method can be used for removing periodic noise in signals collected by the cavity measuring sonar, and the quality of the signals is improved.

The application discloses a kind of based on adaptive wavelet decomposition's rotating machinery fault analysis method and system, wherein, method includes the vibration signal of the rotating machinery to be analyzed is collected and data preprocessing is carried out, obtains data sample;Depth feature extraction processing is carried out to the data sample by feature extraction module, and depth feature is obtained;The depth feature is input to adaptive wavelet decomposition model, and target decomposition feature is output and obtained;The target decomposition feature is subjected to fault type discrimination processing, and obtains fault analysis result.The adaptive wavelet decomposition model is carried out in the embodiment of the application wavelet filter and the automatic learning of wavelet decomposition route, and then separates fault component and interference component in rotating machinery fault analysis task, effectively improves the anti-noise performance of method, and can be widely applied in intelligent monitoring technical field.

The invention belongs to the technical field of inertial navigation, and provides an inertial instrument anti-disturbance alignment method based on sliding window wavelet filtering to solve the problems that measurement information cannot be fully utilized, an alignment result cannot be globally optimal and is seriously influenced by sudden change interference in an existing alignment method. The anti-disturbance alignment method comprises the following steps: extracting a speed error of an inertial instrument; based on a sliding window wavelet filtering method, carrying out noise reduction processing on system noise of the inertial instrument; and carrying out anti-disturbance self-alignment on the inertial instrument according to the speed error and the noise-reduced output data of the inertial instrument. According to the method, the output data of the inertial instrument can be subjected to noise reduction processing, the holding time of the alignment precision of the inertial measurement unit is prolonged, the dispersion of a convergence value is reduced, and the navigation precision is improved.

The application discloses a kind of traffic image segmentation control method and system, it is related to image processing field, wherein the method comprises: the camera of interactive target vehicle is obtained first traffic image;According to wavelet filtering algorithm, first traffic image is carried out denoising processing, and second traffic image is obtained;Based on the second traffic image, ORB feature recognition is carried out, and ORB feature recognition result is obtained;Based on the ORB feature recognition result, clustering is carried out, and ORB feature clustering result is obtained;Based on the ORB feature clustering result, the second traffic image is carried out background isolation, and third traffic image is obtained, and, third traffic image has ORB feature mark;Traffic image segmentation channel is built;Based on the ORB feature mark, according to the traffic image segmentation channel, third traffic image is carried out segmentation control, and target traffic image segmentation result is obtained.The technical problem that present technique is low in accuracy for traffic image segmentation is solved.

The embodiment of the specification provides a bearing fault diagnosis method and system of optimized filtering MCKD, wherein the method comprises the following steps: collecting a vibration signal, performing Fourier transform on a frequency spectrum of the vibration signal to obtain a key function; setting an initial truncation length, performing inverse Fourier transform on the key function based on the initial truncation length to obtain a frequency spectrum trend, taking a minimum value of the trend spectrum as a boundary of spectrum segmentation to divide a filtering frequency band with different bandwidths and center frequencies, increasing the truncation length, repeating the above steps, and recording all results; designing a Mayer wave filter based on the obtained filtering frequency band, and taking the designed Mayer wave filter coefficient as an initial filter coefficient of the MCKD; calculating MCKD filtering results of different initial filter coefficients, quantifying fault characteristics of all filtering signals by a harmonic spectrum kurtosis (HSK), selecting a filtering result corresponding to a maximum value of the HSK as a global optimal solution of the MCKD, and completing fault diagnosis.

The invention discloses a GPS (Global Positioning System) and inertial navigation fused high-precision navigation method, aiming at the problems that a GPS signal is easily shielded and interfered and an inertial navigation error is accumulated along with time, high-precision navigation is realized through data preprocessing, a self-adaptive fusion algorithm, GPS signal interruption processing and fusion result optimization. In the data preprocessing, multi-path effect detection and correction are performed on GPS data, and zero offset error compensation is performed on inertial navigation data; the adaptive fusion algorithm constructs a model based on a fuzzy logic system, and updates a fusion weight in real time; when a GPS signal is interrupted, map matching is adopted to assist navigation, and the prediction model is utilized to compensate an inertial navigation error; and a fusion result is subjected to wavelet filtering and smoothing processing, and the accuracy is ensured by using consistency check. The method can effectively overcome the limitation of independent use of GPS and inertial navigation, realizes high-precision, continuous and reliable navigation in a complex environment, and improves the adaptability and stability of a navigation system.

This invention belongs to the field of video coding and decoding technology, and relates to a three-dimensional wavelet scalable video bitrate control method and system based on video content-aware characteristics. This method fully considers the impact of wavelet filter type, subband coupling phenomenon, and temporal subband content-aware characteristics on reconstruction error propagation, and proposes an adaptive Lagrange multiplier selection algorithm suitable for the wavelet domain. Then, an R-λ model based on video content-aware characteristics is established. Finally, based on the proposed bitrate control model, the coding parameters are adaptively adjusted to achieve the bitrate control target. Experimental results show that, compared with the bitrate control methods used in current classic three-dimensional wavelet video coding and decoding schemes, the method proposed in this invention achieves accurate bitrate control without significantly increasing complexity, while significantly improving the rate-distortion performance of the codec, better maintaining the continuity of video quality, and obtaining a more stable reconstructed video quality.

According to the switch cabinet vibration signal denoising method, the switch cabinet vibration signal denoising device, the storage medium and the computer equipment provided by the invention, before the vibration signal is denoised, direct current removal processing and normalization processing can be firstly performed on the vibration signal, so that the signal precision is improved; in order to accurately identify a key part carrying fault information in a signal, the vibration signal variation mode can be decomposed into a plurality of connotation mode components, and each connotation mode component is divided into an effective component and a background component according to the kurtosis value of each connotation mode component; according to the method, a target wavelet filter is used for carrying out targeted denoising on effective components, and the filter is constructed by adopting a golden proportion optimization algorithm to carry out iterative optimization on parameter vectors by taking minimization of a root mean square error as a target, so that subjectivity and blindness of parameter selection can be avoided, and the denoising effect is further improved; and finally, performing signal reconstruction on the background component and the filtered effective component to ensure the integrity and high signal-to-noise ratio of the denoised vibration signal.

The invention provides an easily-confused emotion recognition method based on multi-path feature extraction and gating fusion. The method can be applied to scenes such as human-computer interaction, emotion calculation and psychological health auxiliary diagnosis. According to the method, a feature path is constructed for each group of emotions aiming at three groups of easily-confused emotions of happiness-angry, fear-surprising and sadness-neutral, each path adopts different wavelet parameter sets to generate a multi-scale wavelet filtering kernel, and multi-path time-frequency features are extracted from the same voice signal. And inputting each path output into a gating attention mechanism module, calculating a path weight and performing weighted summation to obtain a fusion feature for easily confused emotion discrimination, and inputting the fusion feature into a classifier to output an emotion category. According to the method, a wavelet filtering kernel group emphasizing different emotion categories is constructed, a multi-path feature extraction structure is generated, and fusion is performed by using a gating mechanism, so that the model can realize easy-to-confusion emotion recognition according to feature contributions of different paths, and the easy-to-confusion emotion recognition accuracy of the model is improved.

The present invention discloses a radar communication modulation recognition method based on hierarchical classification, comprising the following steps: S1, acquiring modulation signal data; S2, normalizing the modulation signal data; S3, performing wavelet filtering on the normalized data; S4, performing feature extraction on the filtered data; S5, using the extracted features as input to a support vector machine, and constructing a hierarchical modulation recognition classifier based on the support vector machine. The present invention utilizes wavelet filtering to reduce the impact of the noise environment on the characteristics of the received signal, and employs time domain statistical features to mitigate frequency bias problems caused by carrier frequency estimation deviations. By hierarchically classifying the signal through a support vector machine trained with corresponding features, the accuracy and stability of signal recognition are improved, and good robustness can be achieved under noise and frequency bias conditions.

The invention discloses a method for extracting and classifying terahertz time-domain spectral features of hidden articles in an atmospheric environment. Comprising the following steps: carrying out wavelet filtering on collected terahertz time-domain spectral data of the hidden article to obtain a denoised spectral data set; performing data enhancement on the denoised spectral data set by using a deep learning model to form an expanded data set; preprocessing the expanded data set to form a multi-channel graph feature data set; carrying out transfer learning on the multi-channel graph feature data set by utilizing a neural network model to obtain a hidden article real-time detection model; and detecting and classifying the hidden articles in real time by using the detection model. According to the invention, through adoption of advanced data enhancement and multi-channel processing technologies, the identification capability of hidden articles is obviously improved. Besides, transfer learning is carried out through the improved neural network model, key features in the terahertz spectrum are fully extracted and distinguished, and the feature extraction capability and classification performance of the model on terahertz time-domain spectrum data are remarkably improved.

The invention relates to the technical field of radiation detection, in particular to radiation monitoring equipment and a signal processing method, device and equipment based on combination of wavelet filtering and differential filtering, and the signal processing method based on combination of wavelet filtering and differential filtering comprises the following steps: acquiring a signal to be processed; preprocessing the signal to obtain a processed signal; and processing the processed signal by combining a wavelet filtering algorithm with a delay differential filtering algorithm to obtain a pulse signal peak value. According to the signal processing method based on combination of wavelet filtering and differential filtering, the obtained signal is preprocessed, and then the processed signal is processed by combining a wavelet filtering algorithm and a delay differential filtering algorithm to obtain a pulse signal peak value, so that the signal filtering precision is improved; the technical problem that an existing signal processing filtering mode is low in filtering precision is solved.

The invention belongs to the field of fault diagnosis, and particularly relates to a bearing single-source-domain generalization fault diagnosis method based on wavelet adaptive causal decoupling, and the method comprises the steps: carrying out the time domain decomposition and adaptive feature extraction of a source domain sample based on a training set and a test set, and obtaining a global feature; in the second convolutional neural network, introducing a causal decoupling strategy based on a binary mask, and displaying and decoupling the extracted global features into causal features and non-causal features; and performing second convolutional neural network training, establishing a fault classifier, and determining causal purification loss based on the fault classifier. In the feature extraction stage, multi-scale time-frequency decomposition of mechanical vibration signals is achieved by adaptively optimizing wavelet filter parameters, low-frequency stable information and high-frequency impact features can be effectively separated, the stability and discrimination of feature expression can still be kept even in a strong-noise and multi-component coupling environment, and the method is suitable for large-scale popularization and application. Therefore, the accuracy and robustness of fault feature extraction are remarkably improved.

The invention relates to the technical field of crossing of artificial intelligence and tobacco processing, discloses a cut tobacco feeding moisture regulation and control method and system based on LightGBM, and aims to solve the problems of moisture signal distortion and low regulation and control precision caused by dust interference. The method comprises the following steps: acquiring dust concentration and noise-containing moisture signals of a production line in real time; dynamically generating an optimal wavelet filtering parameter through a nonlinear mapping model based on the dust concentration; performing adaptive wavelet filtering on the moisture signal to reconstruct a pure signal; inputting the pure signal into a pre-trained LightGBM model to predict a moisture value; and generating a regulation and control instruction according to the deviation between the predicted value and the target value to realize closed-loop control. The system comprises a working condition sensing module, a signal acquisition module, a filtering parameter dynamic generation module, an adaptive filtering module, a moisture prediction module and a regulation and control instruction generation module. Through decoupling integration of dynamic filtering and an existing LightGBM model, on the premise that an original model is not changed, the moisture regulation and control precision and the response speed are remarkably improved, and the engineering transformation cost is reduced.

The invention provides an intelligent tire lateral force estimation method and system based on an improved iTransform network, and the method is realized through the following steps: S1, carrying out the acceleration signal collection of a tire, and obtaining the three-axis acceleration data; s2, performing IEEEEMDAN-permutation entropy-wavelet filtering noise reduction on the x-axis acceleration signal and the y-axis acceleration signal, and performing 200Hz low-pass filtering on the z-axis acceleration signal; s3, identifying starting and ending feature points of the z-axis acceleration signal based on an adaptive threshold method, and segmenting three-axis acceleration data; s4, constructing a hybrid network model, and replacing a linear mapping layer and a regression prediction layer in an iTransform model by using a KAN network; and S5, predicting the lateral force of the tire by using the trained model. According to the method, by combining the advantages of the iTransform in time sequence modeling and the KAN in nonlinear representation, the precision and stability of tire lateral force estimation are remarkably improved, and the method is suitable for vehicle dynamics analysis and control scenes and has great significance in improving the maneuverability and safety of vehicle driving.

Techniques in wavelet filtering for advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency. An array of processing elements comprising a portion of a neural network accelerator performs flow-based computations on wavelets of data. Each processing element comprises a compute element to execute programmed instructions using the data and a router to route the wavelets in accordance with virtual channel specifiers. Each processing element is enabled to perform local filtering of wavelets received at the processing element, selectively, conditionally, and / or optionally discarding zero or more of the received wavelets, thereby preventing further processing of the discarded wavelets. The wavelet filtering is performed by one or more configurable wavelet filters operable in various modes, such as counter, sparse, and range modes.

This invention relates to a hyperspectral image demixing method based on a graph wavelet transform convolutional network, belonging to the field of hyperspectral image demixing technology. The method includes: using a Mexican Hat graph wavelet filter to perform frequency domain decomposition on the undirected graph of the original hyperspectral image, obtaining low-frequency and high-frequency components, which are then input into a parallel dual-branch feature encoder network for feature learning, extraction, and adaptive recalibration. Enhanced low-frequency and high-frequency features are obtained and fused element-wise to obtain multi-frequency comprehensive features, which are then input into an endmember constraint and reconstruction module for abundance estimation. This yields a reconstructed hyperspectral image, which is compared with the original hyperspectral image. A multi-dimensional joint loss function is constructed and optimized to obtain a hyperspectral demixing model for hyperspectral image demixing. This method aims to solve the technical problem of existing technologies' inability to effectively distinguish the mixing characteristics of high-frequency boundary regions and low-frequency homogeneous regions, leading to boundary blurring and spectral distortion.

The invention discloses a method for accurately calculating passive discharge time, and the method comprises the steps: S1, carrying out the wavelet filtering processing, obtaining a time sequence tk and a voltage sequence vk of a passive discharge curve of a passive discharge resistor collected by an oscilloscope, carrying out the wavelet filtering of the voltage sequence vk, and obtaining a voltage sequence vfk after the wavelet filtering; s2, determining a passive discharge starting point, and obtaining a time point tstart corresponding to the starting point; s3, determining a passive discharge termination point and a time point tend corresponding to the termination point; and S4, calculating the passive discharge time. According to the method for accurately calculating the passive discharge time, the defects that in the prior art, manual point taking errors are large, and the accuracy of actually measured passive discharge time cannot be guaranteed are overcome.

The present invention relates to the technical field of power system electricity consumption monitoring and analysis, and discloses a load identification method based on power signal similarity clustering. A collector or a special transformer terminal deployed at the user side incoming line is used to perform denoising on the collected voltage and current waveform data using a wavelet filtering algorithm. Time domain features, frequency domain features and time-frequency domain features are extracted from the pre-processed voltage and current waveform data respectively, and a measurement method combining Euclidean distance and cosine similarity is used. The present invention uses the characteristic similarity of voltage and current waveforms to perform non-invasive electrical equipment identification, which is deployed on a collector or a special transformer terminal to realize energy consumption structure analysis on the user side. With the energy consumption structure analysis, users can clearly know the energy consumption status and usage patterns of each electrical equipment, thereby formulating a more scientific and reasonable electricity consumption plan, avoiding the use of high-energy-consuming equipment during peak hours, reducing unnecessary electricity waste, and thus significantly reducing electricity costs.

Techniques in wavelet filtering for advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency. An array of processing elements comprising a portion of a neural network accelerator performs flow-based computations on wavelets of data. Each processing element comprises a compute element to execute programmed instructions using the data and a router to route the wavelets in accordance with virtual channel specifiers. Each processing element is enabled to perform local filtering of wavelets received at the processing element, selectively, conditionally, and / or optionally discarding zero or more of the received wavelets, thereby preventing further processing of the discarded wavelets. The wavelet filtering is performed by one or more configurable wavelet filters operable in various modes, such as counter, sparse, and range modes.