55 results about "Eigenvalue distribution" patented technology

The invention discloses a commodity purchase prediction modeling method. The method comprises the steps that a purchase record marking training sample is used to predict whether to purchase or not; a sliding window commodity purchase sample is constructed; commodity purchase features are designed based on a time preference; a gradient improvement decision tree algorithm is used for training prediction; after the sample and the features are constructed, feature processing and selection need to be performed, and then the features are input into the gradient improvement decision tree algorithm for training prediction; and feature selection indicators include feature value distribution and relevancy, feature information gains, feature calling frequency, influences of feature knockout, etc. Ordering is performed on feature importance by integrating the indicators, and redundant features with low importance are eliminated. According to the method, a sliding window sample construction method and a feature system based on the time preference are proposed, the accuracy of a commodity purchase prediction model is effectively improved, and the method is used for realizing commodity personalized recommendation in a big data background to precisely recommend proper commodities to a user at a proper time and a proper place.

The invention belongs to the technical field of situation awareness for a dynamic electric power system, and discloses a recursive state estimation fused anomaly detection method for the dynamic electric power system. The method comprises the following steps: firstly establishing a simplified non-linear recursive model of system node voltage according to the dynamic characteristic of the electricpower system, and reasonably representing the influence of the dynamic change of electric power system load on system node voltage; then realizing system node voltage dynamic estimation based on the non-linear recursive model based on a recursive state estimation filtering algorithm, and on this basis, further constructing a residual error random matrix of system node voltage; finally constructinga system residual error dynamic performance index based on characteristic spectrum mean value-square deviation statistical analysis so as to effectively reflect the influence of electric power systemanomaly on the residual error matrix eigenvalue distribution of system node voltage, and then judging according to a self-adaptive statistical threshold value to finally realize effective state estimation and anomaly detection of the dynamic electric power system.

There is provided a specimen information acquisition apparatus that can obtain an image in which an influence of a local fluence distribution in a specimen is reduced. A signal processing apparatus includes a region extraction unit and an optical constant estimation unit. The region extraction unit extracts two or more regions that have a known absorption coefficient, from an optical characteristic value distribution of the specimen. The optical constant estimation unit estimates an optical constant of a background of the specimen by using optical characteristic values of the two or more regions extracted by the region extraction unit and a local fluence distribution obtained from an optical constant which is set in advance for the specimen background.

The invention discloses a near-field accompanying optimization method for far-field sonic boom suppression of a supersonic aircraft, and the method comprises the steps: giving a parameterized initialnear-field waveform, obtaining near-field and far-field overvoltage signals, and carrying out the modal decomposition; extracting each decomposition mode and characteristic value distribution thereof,and reversely solving a near-field overvoltage signal; on the basis of CFD grid division, according to the reversely solved near-field overvoltage signals and a given line segment at the lower position of the fuselage, completing target function variation formed by near-field overvoltage and target overvoltage distribution, and constructing and solving an adjoint equation; and performing optimization design based on a sequential quadratic programming algorithm, and performing forward calculation check on a design result. According to the method, the defects of traditional near-field accompanying optimization can be overcome, meanwhile, the tedious operation of far-field-near-field-body variation of far-field signal design is avoided, and the efficiency of accompanying optimization can becontinuously and fully utilized.

The invention discloses a multivariate flame monitor-based on-line judgment method for fuel type, which comprises the following steps: when a known fuel and a new fuel exist in burning, extracting characteristic values in a time domain and a frequency domain from flame radiation signals of the known fuel, and mathematically transforming the characteristic values to acquire orthogonal characteristic values; establishing a joint probability density model and a neural network model of characteristic value distribution of the known fuel based on the orthogonal characteristic values; extracting characteristic values in the time domain and the frequency domain from burning flame radiation signals of a fuel to be identified, mathematically transforming the characteristic values to acquire orthogonal characteristic values and inputting the orthogonal characteristic values into the joint probability density model of various known fuels to judge, and if the fuel to be identified in burning is the new fuel, storing the orthogonal characteristic values of the flame radiation of the new fuel to establish the joint probability density model of the new fuel and update the neural network model; and if the fuel to be identified in burning is not the new fuel, judging the type of the fuel through the neural network model.

A resistor array is provided in an element array. A mean value of a characteristic-value distribution is associated with a median of combined resistance values obtained by the element array. An array of trimming information pieces corresponding to combined resistance values larger than the median is set in a descending order of ‘15’ to ‘8’ in decimal number, and an array of trimming information pieces corresponding to combined resistance values less than the median is set in an ascending order of ‘0’ to ‘7’ in decimal number. A circuit converts trimming information derived from the trimming information generation circuit to generate element selection information for selecting turn-off resistors to obtain combined resistance values from the resistor array. Thus, the number of melted-and-cut fuses involved in generation of trimming information associated within the range of “mean value±2σ” in the distribution is reduced.

The invention discloses a power grid mountain fire potential fire point identification method based on a two-dimensional OTSU. The method comprises the following steps: firstly, dividing a satellite image data plane into a plurality of rectangular sub-regions according to latitude and longitude; Selecting a sub-region as a to-be-identified region, calculating an average channel brightness temperature characteristic value of pixel points in a neighborhood of each pixel point in the to-be-identified region, and calculating two-dimensional probability distribution of the pixel brightness temperature characteristic values in the to-be-identified region; And obtaining a fire point determination optimization threshold based on a two-dimensional OTSU method, and determining whether each pixel point in the to-be-identified area is a potential fire point according to the optimization threshold. According to the method, the judgment threshold of the potential fire point is calculated only according to the current image; According to the method, the characteristic value distribution information of the whole pixels in the region is fully utilized, the adaptive adjustment of the mountain fire judgment threshold can be carried out according to different time and different regions without the help of historical threshold values, the alarm accuracy is high, the calculation efficiency is high,the practicability is high, and the method plays an important role in maintaining the safety and stability of a power grid and preventing and reducing disasters.

The invention relates to a data processing method and device based on artificial intelligence, computer equipment and a storage medium. The method comprises the steps of obtaining a preset number of samples, wherein the samples comprise user features and user behavior sequences; performing sequence pattern mining based on the user behavior sequence to obtain a plurality of user behavior sequence patterns; identifying the relevancy between the user characteristics and the user behavior sequence mode, and extracting a strong correlation sequence mode corresponding to the user characteristics according to the relevancy; counting characteristic value distribution corresponding to each user characteristic in the strong correlation sequence mode; and when one sample has the user feature with themissing feature value, filling the missing feature value according to the feature value distribution to obtain a sample filled with the feature value. A sample containing accurate and effective userfeatures can be provided for model learning, and then the accuracy of machine learning is improved.

A resistor array is provided in an element array. A mean value of a characteristic-value distribution is associated with a median of combined resistance values obtained by the element array. An array of trimming information pieces corresponding to combined resistance values larger than the median is set in a descending order of ‘15’ to ‘8’ in decimal number, and an array of trimming information pieces corresponding to combined resistance values less than the median is set in an ascending order of ‘0’ to ‘7’ in decimal number. A circuit converts trimming information derived from the trimming information generation circuit to generate element selection information for selecting turn-off resistors to obtain combined resistance values from the resistor array. Thus, the number of melted-and-cut fuses involved in generation of trimming information associated within the range of “mean value±2σ” in the distribution is reduced.

The invention discloses an evaluation method and device for face image similarity, and the method comprises the steps: collecting face image information, obtaining a face multi-dimensional feature vector, and carrying out the In < t > quantification of the face multi-dimensional feature vector; carrying out dimensionality reduction on the quantized face multi-dimensional feature vectors, and taking d-dimensional feature values with high contribution degrees; training face image data, counting feature value distribution of each dimension and segmenting to divide a value domain of the feature value of each dimension into k sections with unequal widths; for the two face image feature vectors of which the similarity is to be recognized, obtaining the similarity of each dimension according to the segment where the feature value of each dimension is located, and obtaining an evaluation value of the similarity of the face image vectors according to the similarity of each dimension. Accordingto the embodiment of the invention, the method can reduce the consumption of system resources, achieves the quick and objective calculation of the similarity of the feature vectors of different humanface images, improves the efficiency and applicability of a clustering algorithm, and can achieve the quick and accurate judgment of the similarity of the feature vectors of the human face.