327 results about "Short-time Fourier transform" patented technology

A scheme for stereo and multi-channel synthesis of inter-channel correlation (ICC) (normalized cross-correlation) cues for parametric stereo and multi-channel coding. The scheme synthesizes ICC cues such that they approximate those of the original. For that purpose, diffuse audio channels are generated and mixed with the transmitted combined (e.g., sum) signal(s). The diffuse audio channels are preferably generated using relatively long filters with exponentially decaying Gaussian impulse responses. Such impulse responses generate diffuse sound similar to late reverberation. An alternative implementation for reduced computational complexity is proposed, where inter-channel level difference (ICLD), inter-channel time difference (ICTD), and ICC synthesis are all carried out in the domain of a single short-time Fourier transform (STFT), including the filtering for diffuse sound generation.

The invention discloses a clustering-based blind source separation method for synchronous orthogonal frequency hopping signals. The method comprises the following steps of: acquiring M sampled paths of discrete time-domain mixed signals; obtaining M time-frequency domain matrixes of the mixed signals; preprocessing the time-frequency domain matrixes of the frequency hopping mixed signals; estimating frequency hopping moments, normalized mixed matrix column vectors and frequency hopping frequency; estimating time-frequency domain frequency hopping source signals by utilizing the estimated normalized mixed matrix column vectors; splicing the time-frequency domain frequency hopping source signals between different frequency hopping points; and recovering time-domain source signals according to time-frequency domain estimate values of the source signals. According to the method, the frequency hopping source signals are estimated only according to the received mixed signals of a plurality of frequency hopping signals under the condition of unknown channel information, and the frequency hopping signals can be subjected to blind estimation under the condition that the number of receiving antennae is smaller than that of the source signals; short-time Fourier transform is utilized, so that the method is low in computation amount; and the frequency hopping signals are subjected to blind separation, and meanwhile, a part of parameters can also be estimated, so that the method is high in practicability.

The invention discloses a multi-target detection method based on short-time Fourier transform and fractional Fourier transform, which belongs to the technical field of the radar target detection. The method comprises the following steps that the short-time Fourier transform is firstly used for conducting the primary detection on a signal, then a binaryzation method is used for processing a primary detection result, phase position of the signal is kept in the processing, the fractional Fourier transform is used for detecting a signal after being restored by the short-time Fourier transform, by adopting multiple methods for combined processing, advantages of overcoming phenomenon that a strong signal side lobe presses a weak signal main lobe, improving the signal-to-noise ratio of the signal to be detected, and solving the problem of the large false alarm possibility which is caused by adopting traditional method to detect the signal at the low signal-to-noise ratio can be realized; and meanwhile, an image contrast method and a gradual elimination method are adopted, multiple strong signals and weak signals with different or identical frequency modulation rates can be detected by utilizing the space and power strength information of the signal, so that the detection probability and the calculation efficiency can be further improved, easiness in project realization is realized, and the method is worth of being adopted and popularized.

The invention provides a voice noise reduction method for a conference terminal based on the neural network model. The method comprises steps that S1, an audio file is collected by the conference terminal device to generate a digital audio signal in the time domain; S2, the digital audio signal is framed, and short-time Fourier transform is performed; S3, the amplitude spectrum of the frequency domain is mapped into a frequency band, and a Mel-frequency cepstral coefficient is further solved; S4, first-order and second-order differential coefficients are calculated through utilizing the Mel-frequency cepstral coefficient, a pitch correlation coefficient is calculated on each frequency band, and pitch period features and VAD features are further extracted; S5, input characteristic parameters of an audio are used as the input of the neural network model, the neural network is trained offline, the frequency band gain generating the noise reduction speech is learned, and the trained weightis solidified; S6, the neural network model is utilized to learn, the frequency band gain is generated, the outputted frequency band gain is mapped to the spectrum, the phase information is added, and a noise reduction speech signal is reduced through inverse Fourier transform. The method is advantaged in that real-time noise reduction can be achieved.

The invention provides a CEEMD (Complementary Empirical Mode Decomposition)-STFT (Short-Time Fourier Transform) time-frequency information entropy and multi-SVM (Support Vector Machine) based fault diagnosis method for a centrifugal pump. The method comprises the following steps: 1, preprocessing the fault data of the centrifugal pump; 2, extracting fault features; 3, performing dimensionality reduction for the fault features; 4, automatically recognizing a fault mode through a multi-SVM classifier. Vibration signals of the centrifugal pump have the characteristics of being non-stable and low in repeatability and reproducibility, so that some traditional time domain or frequency domain based analysis methods cannot timely reflect the running conditions of a system. The CEEMD is a self-adaptive signal decomposition method and can decompose the signals into a series of intrinsic mode functions; the STFT is a time-frequency analysis method and can analyze non-stable signals; the time-frequency information entropy is a metric of the signal time-frequency distribution complexity and can reflect the fault information hidden in the signals. According to the method, the CEEMD, the STFT and the information entropy method are combined; the method is applied to the actual diagnosis experiment, and the data analysis result shows that the method is high in diagnosis accuracy.

The invention discloses a variable-speed rolling bearing fault recognition method based on an order enveloping time-frequency energy spectrum. Short time Fourier transform is carried out on low-band signals in acquired rolling bearing vibration signals under a variable speed to acquire a short time spectrum, and in combination of a peak search method, instantaneous rotating frequency estimation for a bearing reference axis harmonic component is acquired; then, through a fast kurtosis spectrum technique, the optimum resonance frequency band pass parameters of the vibration signals are determined, the band pass signals after band pass filtering are subjected to Hilbert transform to acquire enveloping signals, and through short time Fourier transform, an enveloping time-frequency energy matrix is obtained; and an extraction interval is set, an order enveloping energy calculation path is obtained, order enveloping time-frequency energy in different orders is calculated through the energy calculation path and the enveloping time-frequency energy matrix, an order enveloping time-frequency energy spectrum is acquired, and through comparison between the order corresponding to the maximum order enveloping time-frequency energy and a fault order, bearing fault recognition is realized. The method has the advantages that the calculation efficiency is high, and requirements of real-time monitoring in actual engineering application can be met.

The invention discloses an intrusion detecting and positioning method for a distributed optical fiber fence. The method comprises the following steps of: performing short-time Fourier transform on a monitoring signal to acquire instantaneous energy distribution of the monitoring signal along a laid optical cable; removing large-range background noise energy distributed along space; extracting information and a position of a mutation point which is relatively greatly changed and is locally abnormal from an energy sequence from which the background noise energy is removed; and judging whether the fence is actually intruded or not, and determining an intrusion position. In the method, mutation information at the mutation point is extracted by a short-time Fourier transform method, and the background noise energy is reduced while random noise is removed, so that the influences of interference to fluctuation noise caused by an unstable optical system and other factors are effectively eliminated, the problems of long-distance intrusion detection and positioning of the distributed optical fiber fence in a complicated application environment under a large noise background are solved, the accurate detection rate of an optical fiber fence intrusion monitoring system can be effectively improved, a false alarm rate can be reduced, and the confidence degree of a security alarming system during actual application can be improved.

The invention discloses a multi-speaker voice separation method based on a convolutional neural network and depth clustering. The method comprises the following steps: 1, a training stage: respectively performing framing, windowing and short-time Fourier transform on single-channel multi-speaker mixed voice and corresponding single-speaker voice; and training mixed voice amplitude frequency spectrum and single-speaker voice amplitude frequency spectrum as an input of a neural network model; 2, a testing stage: taking the mixed voice amplitude frequency spectrum as an input of a threshold expansion convolutional depth clustering model to obtain a high-dimensional embedded vector of each time-frequency unit in the mixed frequency spectrum; using a K-means clustering algorithm to classify thevectors according to a preset number of speakers, obtaining a time-frequency masking matrix of each sound source by means of the time-frequency unit corresponding to each vector, and multiplying thematrixes with the mixed voice amplitude frequency spectrum respectively to obtain a speaker frequency spectrum; and combining a mixed voice phase frequency spectrum according to the speaker frequencyspectrum, and obtaining a plurality of separate voice time domain waveform signals by adopting short-time Fourier inverse transform.

The invention provides an underwater target feature extraction method based on a convolutional neural network (CNN). 1, a sampling sequence of an original radiation noise signal is divided into 25 consecutive parts, and each part is set with 25 sampling points; 2, normalization and centralization processing are carried out on a sampling sample of the j-th segment of the data signal; 3, short-time Fourier transform is carried out to obtain a LoFAR graph; 4, a vector is assigned to an existing 3-dimensional tensor; 5, an obtained feature vector is input to a fully-connected layer for classification and calculation of error with label data, whether the loss error is below an error threshold is tested, if the loss error is below the error threshold, network training is stopped, and otherwise, step 6 is entered; and 6, a gradient descent method is used to carry out parameter adjustment layer by layer on the network from back to front, and shifting to the step 2 is carried out. Compared with the traditional convolutional neural network algorithms, the method of the invention carries out a weighted operation of spatial information multi-dimensions on a feature graph layer to compensate for a defect of spatial information losses caused by one-dimensional vectorization of the fully-connected layer.

Observed signals x₁(t) to x_j(t) from a plurality of sensors disposed in two dimensions are subject to a short-time Fourier transform, from which signals X₁(ω₁) to X₁(ω_N), . . . X_J(ω₁) to X_J(ω_N) are derived. Using the independent component analysis process, separation matrices W(ω₁) to W(ω_N) are produced. Their inverse matrices H(ω₁) to H(ω_N) are calculated, and for each ω_n (n=1, . . . , N), using a pair of elements from each column of H(ω_n), H_ji(ω_n) and H_j′i(ω_n), an angle {circumflex over (θ)}_i,jj′(ω_n)=cos⁻¹(arg(|H_ji(ω_n)/H_j′i(ω_n))/(ω_nc⁻¹∥d_j−d_j′∥)) is calculated where arg(α) represents an argument of α, c the propagation velocity of a signal, and ∥d_j−d_j′∥ represents a spacing between sensors i and j′. Columns are permuted so that {circumflex over (θ)}_i,jj′(ω_n) obtained from each column of H(ω₁) to H(ω_N) assume an ascending order. For columns which cannot be permuted, the equation ∥q_i−d_j′∥/∥q_i−d_j∥=|H_ji(ω_n)/H_j′i(ω_n)|=DR_i,jj′(ω_n) is solved for q_i to calculate R_i,jj′(ω_n)=∥DR_i,jj′(ω_n)·(d_j−d_j′)/(DR²_i,jj′(ω_n)−1)|. Columns of H(ω_n) is permuted so that R_i,jj′(ω_n) assume an ascending order. H(ω_n) is used to solve the permutation problem of W(ω).

The invention provides a speech enhancement method applied to a dual-microphone system. The dual-microphone system comprises a microphone array module, a control module, a processing module and an output module, wherein the microphone array module consists of a first microphone and a second microphone; the control module is used for receiving a signal transmitted by the microphone array module and controlling a microphone array; the processing module is used for receiving data transmitted by the control module; and the output module is used for outputting the processed data output by the processing module. The speech enhancement method comprises the following step that: 1, the first microphone and the second microphone receive time domain speech signals with noise (S1) and (S2) and transmit the signals to the control module, and the control module performs framing and pre-emphasis processing on the time domain speech signals with noise (S1) and (S2) and converts the time domain speech signals with noise (S1) and (S2) into frequency domain signals with noise (X1) and (X2) through short-time Fourier transform. Therefore, the attenuation on related noise is improved.

A system performs local feature extraction. The system includes a processing device that performs a Short Time Fourier Transform to obtain a spectrogram for a discrete-time speech signal sample. The spectrogram is subdivided based on natural divisions of frequency to humans. Time-frequency-energy is then quantized using information obtained from the spectrogram. And, feature vectors are determined based on the quantized time-frequency-energy information.

The invention provides a method for extracting a time-frequency domain spectrum of mixed signals of rotating machinery. The method comprises the following steps of: carrying out normalization pre-processing on collected vibration signals at first, and carrying out time-frequency analysis to obtain instantaneous parameters representing non-steady signal characteristics, namely short-time Fourier transform coefficients; then, calculating a spectrogram, and carrying out non-negative matrix factorization on the spectrogram with the order number r of 1; and finally, obtaining a characteristic spectrum of original vibration signals from a factorized basis matrix. The method disclosed by the invention is capable of extracting the characteristic spectrum rapidly and accurately and effectively reducing interference of outside noise at the same time.

The present invention provides a gear case fault diagnosis method based on non-timescale short-time phase demodulation. An acceleration or speed sensor is attached to a gear case to be detected for collecting a vibration signal of the gear case to be detected. The vibration signal is subjected to short-time Fourier transform, and an instant frequency signal of characteristic vibration is obtained. The instant frequency signal is plug into generalized Fourier transform for time frequency reverse of the vibration signal. Then, a stable projection of a characteristic frequency component is abstracted by means of band-pass filtering, and subjected to generalized Fourier inverse transform, and an engagement frequency component is obtained. Thereafter, the engagement frequency component is subjected to Hilbert transform, and an instantaneous phase of the engagement frequency component is obtained. By the instant phase, angular domain re-sampling is performed on the original vibration signal, and the re-sampled signal is subjected to phase demodulation after reasonable fragmentation and angular domain average. Thus, the short-time phase demodulation is completed. The method provided by the invention can reduce the hardware cost of the detection system, and improve the market competitiveness of the detection system.

The invention discloses a convolutional neural network optimization method for underwater target recognition, and belongings to the technical field of underwater target recognition. The original sounddata is converted into gray-scale sound spectrum images through short-time Fourier transform and the like, and then the single-layer SAE discriminant classification method and the multi-layer SAE reconstruction classification method are added to the Alexnet model respectively. The gray-scale sound spectrum images are tagged and trained, 70% of the gray-scale sound spectrum images are used as thetraining set and the remaining 30% are used as the test set. The training set and the test set are applied to the Alexnet model before improvement and the improved Alexnet model for the contrast experiment of accuracy and training time and the test results are analyzed. The target classification layer in the convolutional neural network is optimized, the problem that the classification accuracy ofthe current convolutional neural network in the field of underwater target recognition field is low can be solved, the method is more suitable for the field of underwater target recognition than thatbefore improvement and thus the better classification effect in the underwater target field can be obtained.

The invention discloses a method for suppressing harmonic-wave interference existed in an original seismic record of controllable land seismic source excitation acquisition. Method is characterized by: taking a related previous earthquake data to carry out short-time Fourier transform and transforming a time domain signal to a time-frequency domain; taking a time sampling sequence number and a frequency sampling sequence number in a related previous time-frequency domain signal as two coordinate axes of a rectangular coordinate, carrying out coordinate transformation so that a fundamental wave is parallel to a time axis; separating the harmonic wave interference from the signal, carrying out coordinate inverse transformation of the time-frequency domain and carrying out short-time Fourier inverse transform so as to acquire time domain data which can remove the harmonic wave interference; after all unrelated seismic data collected from a filed is processed, acquiring the related seismic data after the harmonic wave interference is suppressed. By using the method of the invention, disadvantages of a filtering method in a frequency wave number domain can be avoided. Through the controllable seismic source, the seismic data with high qualities can be collected.

The invention provides a gas pipeline leakage identification method based on a convolution neural network. The method comprise the following steps that after a leakage sound signal and a background sound signal of a typical leakage type are collected, framing processing and short-time Fourier transform are carried out to obtain a time-frequency diagram representing the original leakage sound signal; then a convolution neural network classification model aiming at the leakage sound signal is built, a traditional square convolution kernel is changed into a specific strip-shaped rectangular convolution kernel, so that the line spectrum characteristics in the time-frequency diagram are better extracted; and the time-frequency diagram of the leakage sound and the time-frequency diagram of the background sound are mixed and sent to the built convolution neural network for training, K-fold cross validation is adopted for training, and a network model superparameter is optimized, so that the optimal model superparameter is selected and the robustness and universality of the model are enhanced. Compared with the pipeline leakage identification method in the prior art, the method has the advantages that the identification rate is further improved, and the problem of feature screening which is most difficult to process in the prior art can be effectively solved.

In an automotive system (300), a knock detection scheme is provided for detecting knock events in an internal combustion engine that are sensed by a transducer structure, such as a non-intrusive acoustic accelerometer sensor (310) to generate sensor signal information which is processed by a signal processing structure (312, 316, 318, 326) which extracts digital signal parameters from the sensor signal information to identify a predetermined pitch frequency (330) and any short-term energy increase (328) in the digital signal information which in combination are used to provide a positive indication of engine knock behavior. When a short term Fourier transform (324) is used to extract the digital signal parameters, time frequency resolution may be improved by appropriately windowing the digital signal being transformed.

The invention discloses a single-channel monitor-free voice and noise separating method based on low-rank and sparse matrix decomposition. The method includes the steps of converting a time domain waveform of noise-contained voice to a time frequency domain via short-time Fourier transform to obtain a magnitude spectra with noises; decomposing the magnitude spectra with the noises into a sum of a noise magnitude spectra, a voice magnitude spectra and a residual noise magnitude spectra by a low-rank and sparse matrix decomposition algorithm; and finally reconstructing a voice time-domain waveform from the voice magnitude spectra via short-time Fourier transform. Without any priori information about voice and noise, the method is the single-channel monitor-free voice and noise separating method, pure voice can be separated from noise-contained voice by the aid of an algorithm, and the single-channel monitor-free voice and noise separating method is simple, effective, and particularly suitable for voice extraction in strong-noise environment.

The invention provides an underwater target identification method based on a semi-tensor product neural network, and the method comprises the steps: receiving an underwater sound signal through an underwater sonar sensor, and enabling the time domain and frequency domain information of the sound signal to be presented in a LOFAR map through short-time Fourier transform; constructing a data sample semi-tensor product neural network by taking a LOFAR map sample as an input characteristic matrix; dividing the received underwater acoustic signals into a training set and a verification set, and inputting the training set and the verification set into a semi-tensor product neural network for training and verification; selecting different hyper-parameters, carrying out model training on a semi-tensor product neural network by using a training set, comparing test effects of a verification set, and determining hyper-parameters with high test accuracy; and finally inputting the currently acquired sound signal of the underwater target into the trained semi-tensor product neural network after model, and giving a judgment result. The underwater target recognition rate can be improved, the application scene is expanded, and the method is suitable for recognizing the underwater target in complex marine environment noise.

The invention provides a laser additional material manufacturing defect online diagnosis method based on visual sensing. The method comprises the following steps of collecting a molten pool analog image signal by a CCD camera in real time; converting the molten pool analog image signal into a digital image signal by an image collecting card and importing the digital image signal into a computer; performing real-time image processing on the digital image signal through the computer so as to obtain a time domain picture of the molten pool area; performing short-time Fourier transform on the time domain picture so as to obtain a frequency domain picture of the molten pool area; judging whether the molten pool area suddenly fluctuates or changes or not on the basis of the time domain picture of the molten pool area; if not, judging that no manufacturing defect exists; if so, judging whether the obvious abnormal fluctuation occurs in the frequency domain picture of the molten pool area or not in the sudden fluctuation or change time period of the molten pool area; if so, judging that the macroscopic manufacturing defect exists; if not, judging that the non-macroscopic manufacturing defect exists. The method can be used for judging the generation of the defects, the occurrence moment and the defect types in the laser additional material manufacturing process.

A technique is disclosed for accurate measurement of the distance and detection of an object such as a human body behaving in a complicated way and having a complicated reflection surface using the two-frequency Doppler method. Each of the two-frequency Doppler signals is subjected to the short time Fourier transform to produce a spectrum in a short time window. Based on restraints, it is determined whether the waveforms of the two spectra are analogous to each other or not, and only in the case where the waveforms are analogous, the Doppler signals of the particular time window are used for distance calculation. As a result, the Doppler signals having the phase information low in reliability are eliminated, and therefore high-accuracy distance measurement and object detection are possible.

The invention discloses a microphone array postfiltering sound enhancement method based on multi-models and hearing characteristic, aiming at two important factors influencing the postfiltering sound enhancement performance of a microphone array, i.e. accurate estimation for signal parameters and suitable compromise between increasing noise reduction performance and reducing voice distortion. Thescheme of the invention comprises the following steps of carrying out time domain alignment on signals collected by the microphone array, and carrying out short-time Fourier transform and characteristic value analysis based of power spectrum; determining the dimensionality of a signal subspace through the existence probability of target voice signal in maximation noise-carried voice signals; self-adaptively selecting a distribution model of a noise power spectrum in the noise-carried voice signals; estimating noise power spectrum by utilizing a conditional probability; estimating an auditory masking threshold value based on the signal subspace; and estimating a postfilter by combining Lagrange multipliers according to the auditory sensing characteristics.