352results about How to "Strong explainability" patented technology

The present invention relates to a method for defense of attack of adversarial examples based on a convolutional denoising auto-encoder. Adversarial image examples x* (output tags identified by an image classification device is y*) are constructed by manual addition of adversarial disturbance on clean image samples x without modification (output tags identified by the image classification device is y), the fraud purpose that y* is not equal to y can be achieved, even though the image classification device classifies two images essentially showing the same meaning to two classes by mistake. Thepresent invention designs an integration defense model connected with a target image classifier based on a convolutional denoising auto-encoder (CDAE), namely input samples are subjected to coding and decoding at the internal portion of a well trained CDAE to remove most of adversarial disturbances in the input samples so as to output denoising samples close to original clean samples, and then are transmitted to the target image classifier so as to improve the classification correction of the target classifier and have an effect for defense of attack of adversarial examples.

The invention discloses a malicious encrypted flow feature analysis method, which belongs to the crossing field of network security and machine learning, and is used for analyzing the HTTPS flow and detecting the malicious threats in the HTTPS flow. The method comprises the following steps of modeling the flow data by taking a connection quaternion as a data structure, analyzing the encrypted flowfrom four levels by taking the connection quaternion as a unit, and extracting the flow-level features, TLS handshake features, X.509 certificate features and context features of the connection quaternion to obtain an initial feature set, adopting a recursive feature elimination method based on a decision tree to screen the initial feature set to obtain an optimal feature set used for the machinelearning model training to thereby obtain a malicious encrypted flow detection model.

The invention discloses a face recognition method based on weighted diagnostic sparseness constraint nonnegative matrix decomposition, and mainly aims to solve the problem that the method in the prior art is not robust to an obscured face and is of low recognition rate. According to the technical scheme, the method comprises the following steps: (1) constructing a nonnegative weight matrix according to the obscured area of a test image; (2) introducing the weight matrix into a general KL divergence objective function, applying a sparseness constraint to a basis matrix, and applying intra-class and inter-class divergence constraints to a coefficient matrix to get a weighted diagnostic sparseness constraint nonnegative matrix decomposition objective function; (3) solving the objective function, and decomposition-training a data matrix to get a basis matrix and a coefficient matrix; (4) projecting a test data matrix on the basis matrix to get a corresponding low-dimensional representation set, and taking the low-dimensional representation set as final test data; and (5) using a nearest neighbor classifier to classify the test data by taking the coefficient matrix as training data, and outputting the result. By using the method, the effect of obscured face recognition is improved. The method can be used in identity recognition and information security.

The invention discloses a random-forest-model-based power transformer fault diagnosis method, which comprises the following steps of: acquiring transformer state overhauling data, training a random forest model by utilizing the transformer state overhauling state, checking the sensitivity of the random forest model, and diagnosing a fault of a transformer by using the trained and checked random forest model. The method provided by the invention is high in adaptability and interpretability; and criticality between normality and the fault is separated by utilizing a k-means clustering method, and a system is endowed with fault early warning capability.

The invention belongs to the virtual character attitude control in the field of speech synthesis, and particularly relates to a modeling and controlling method for synchronizing the voice and the mouth shape of a virtual character. The object of the invention is to reduce the mouth shape animation data annotation amount and to achieve accurate and naturally smooth mouth motion synchronized with the voice. The method comprises: generating a phoneme sequence corresponding to the to-be-synchronized voice; converting the phoneme sequence into a phoneme category sequence; converting the phoneme category sequence into a static mouth shape configuration sequence; and converting the static mouth shape configuration sequence distributed on a time axis into dynamically changing mouth shape configuration by a dynamic model; rendering the dynamically changing mouth shape configuration into an attitude image of the head and neck of the virtual character, and displaying the attitude image in synchronization with a voice signal. The method can realize efficient and natural virtual character mouth shape synchronous control without mouth shape animation data and with a phonetic prior knowledge anddynamic model.

The invention discloses a server fault automatic detection system and a detection method based on a decision tree, which combines an expert system and an IPMI management unit to generate a historicaldata set. The server running state data, i.e. Abnormal data stream, is obtained by IPMI management unit. According to the abnormal data stream, the new fault feature vector is extracted, and the new feature vector and the fault cause relation pair are formed into a fault data set, and the fault data set is trained into a self-diagnosing decision tree model. When the server fails while running, thecorresponding fault feature vectors are extracted, the fault types are automatically judged by the self-diagnostic decision tree model, After the fault is cleared, the fault feature vector and the fault cause relation are updated and the self-diagnosing fault tree model is updated. Therefore, the fault diagnosis system will be more accurate and reliable with the improvement of the history fault set.

The invention discloses a fault diagnosis method based on adaptive manifold embedding dynamic distribution alignment. According to the method, the feature distortion of data in an original Euclidean space can be effectively avoided through the automatic calculation of the optimal subspace dimension and the calculation of the streaming kernel of a geodesic line and converted manifold feature representations; a similarity measure A-distance is introduced to define a self-adaptive factor; relative weights of condition distribution and edge distribution of sample data are dynamically adjusted, andtherefore, the distribution difference of source domain and target domain samples can be effectively reduced, the accuracy and effectiveness of rolling bearing fault diagnosis under variable workingconditions can be greatly improved. The method is high in interpretability, is lower in requirements for computer hardware resources, is higher in execution speed, and is excellent in diagnosis precision, algorithm convergence and parameter robustness. The method is especially suitable for multi-scene and multi-fault bearing fault diagnosis under variable working conditions, and can be widely applied to fault diagnosis tasks of complex systems such as machinery, electric power, chemical engineering and aviation under variable working conditions.

The invention provides a role-based access control model constructing system, which comprises a database, a weight module, a role module, a restriction module, a result display module, a user right distribution and management module and an audit management module, wherein the weight module comprises a similarity module and a weight calculation module; the role module comprises a role generation module and a role hierarchy generation module; the restriction module comprises a role restriction module, a right restriction module, a user restriction module and a mutual exclusion restriction module; and the result display module comprises a role display module, a restriction display module and a result adjustment module. In the invention, the importance of the user right is measured, the restriction to a role-based access control model is generated, and therefore the accuracy of the role and the safety of the system are improved. When the system is adopted, the construction of the role-based access control model can be realized, and the automatic and semi-automatic construction of the role-based access control model is ensured; moreover, the system has the characteristics of high safety, high interpretability and easy extendibility.

The invention provides a multi-resource coordination control method and apparatus for a direct current block fault impact alternating current channel. Static state stability limit of the alternating current channel is determined firstly; then the multi-resource coordination control sequence and coordination control quantity are determined; and finally, multi-resource coordination control is realized. By virtue of the technical scheme provided by the invention, multiple types of control resources of pumping, storing and pump switching, direct current power emergency modulation, accurate load shedding, and load shedding are taken into comprehensive consideration, so that control means are increased; the target function of a multi-resource coordination control model takes the minimum of the load shedding quantity as the optimization object, and constraint conditions are taken into consideration, so that control effectiveness is improved, control cost is lowered, and power grid defensive performance is improved; and in addition, by adoption of a heuristic algorithm, the multi-resource coordination control sequence and the multi-resource coordination control quantity are obtained, and the multi-resource coordination control quantity obtained by adopting the heuristic algorithm is high in interpretability, so that relatively high practicability can be realized.

The invention belongs to the technical field of artificial intelligence, and particularly relates to a domain adaptation method based on a machine vision scene text recognition task. The method comprises the following steps: constructing a CNN-LSTM network and an attention network; combining the CNN-LSTM network and the attention network into a scene text recognition network; inputting the scene images from the source domain and the target domain into a scene text recognition network, extracting image features from the input scene images by CNN-LSTM, recoding the image features by an attentionnetwork, extracting corresponding features of each character, and segmenting text information in the images into character level information; and finally, constructing a domain classification networkby applying a transfer learning technology based on adversarial learning, forming an adversarial generation network together with the scene text recognition network, and finally enabling the model toeffectively adapt to a target domain. According to the method, a small number of target domain calibration samples are fully utilized, the problem of sample scarcity frequently occurring in an actualscene text recognition task is solved, and the recognition effect is improved.

The invention relates to an image description text generation method based on a generative adversarial network. The method comprises the following steps: 1) constructing an encoder for realizing feature extraction of an image; 2) performing word embedding on the text, and constructing a decoder for generating an image description text; 3) pre-training a generator composed of an encoder and a decoder according to maximum likelihood estimation; 4) constructing a discriminator based on a convolutional neural network and training the discriminator; 5) jointly training a generator and a discriminator; and 6) inputting the test image data of the description text to be generated into the trained generator, and outputting the generated description text. Compared with the prior art, the method has the advantages that the objective evaluation score of the generated text is improved, the interpretability is good, and the diversity is high.

The invention discloses a sintered ore FeO content detection method and a sintered ore FeO content detection system. The method comprises: obtaining a thermal image; extracting a key frame image in combination with a dust change rule at the tail portion of a sintering machine; according to the key frame image, extracting an interested infrared thermal image by utilizing the geometrical characteristics of a trolley at the tail portion of the sintering machine so as to obtain a sintered ore cross section infrared thermal image; based on the sintered ore cross section infrared thermal image, extracting the shallow characteristics and the deep characteristics for describing the quality of the sintered ore; establishing a sintering process multiphase thermodynamic model based on Gibbs free energy theorem; according to the multiphase thermodynamic model, obtaining the FeO content classification characteristics of the sintered ore at the highest temperature, and establishing a FeO content prediction model based on multiple heterogeneous characteristics; and real-timely and onlinely predicting the FeO content of the sintered ore by utilizing the shallow characteristics, the deep characteristics and the FeO content classification characteristics. By adopting the technical scheme, the technical problem that the FeO content of the sintered ore cannot be accurately detected in real time inthe prior art is solved, the FeO content can be accurately detected in real time, and the method has the characteristics of high precision and strong interpretability.

The invention relates to a Rochester model-naive Bayesian model-based data classification system, which comprises a data processing module, a sampling module, a modeling module and a data testing module, wherein the data processing module divides an original sample set into a saturated layer and a lacking layer according to the input missing value ratio of each sample variable in the original sample set and relativity among the sample variables and sample attributes; the sampling module randomly extracts a training sample variable and a testing sample variable from the saturated layer and the lacking layer to form a training sample set and a testing sample set of which each comprises the saturated layer and the lacking layer respectively; the modeling module models training samples in the saturated layer through a Rochester regression model and models the training samples in the lacking layer through a naive Bayesian model to obtain a hybrid dynamic model with the Rochester regression model and the naive Bayesian model; and the data testing module inputs testing samples in the saturated layer into the Rochester regression model in the hybrid dynamic model, inputs the testing samples in the lacking layer into the naive Bayesian model in the hybrid dynamic model and performs a test to obtain and output scoring results. The Rochester model-naive Bayesian model-based data classification system is integrated with the functions of the Rochester regression model and the naive Bayesian model so as to have complementary advantages and can be widely applied to the financial industry, retailing and the telecommunication industry.

The invention discloses a domain knowledge pushing method based on a knowledge graph. The domain knowledge pushing method comprises the steps of collecting domain knowledge texts to construct a text knowledge base; performing semantic analysis and topic modeling on the knowledge base text; obtaining semantic distribution vectors of the nodes through domain knowledge graph embedding; establishing task context features according to the user task description and the task theme; performing entity alignment according to domain entities in the task description and the knowledge graph, performing feature extension based on graph node paths and graph node semantic distribution features, and performing task associated knowledge recall; performing text similarity calculation on the recalled text and the user task text to obtain candidate text scores; and pushing the sorting result text to the user according to the score. The text matching degree and the user experience of domain knowledge pushing are improved through node associated knowledge of the knowledge graph and a graph embedding technology.

The invention discloses a nonnegative local coordinate factorization-based clustering method, which comprises the following steps that: (1) a sample characteristic matrix is built; (2) a low-dimensional sparse matrix is iteratively outputted; (3) and the low-dimensional sparse matrix is clustered and analyzed. A sparse coding concept is introduced into the nonnegative matrix factorization (NMF) process, nonnegative local coordinate factorization is undertaken on a high-dimensional sample characteristic matrix, so the factorized coefficient matrix is used as a low-dimensional expression of the high-dimensional sample characteristic matrix, and the low-dimensional matrix is clustered to analyze, so the clustering analysis is simple and valid; and at the same time, data after the dimensional reduction has good explanatory property. Compared with the dimensional reduction method of the prior art, the judgment capacity of the clustering analysis can be further improved.

The invention relates to an electronic nose gas sensor array optimization method based on dynamic characteristic importance and belongs to the technical field of sensor array optimization. The methodincludes the steps of firstly, acquiring information of gas in a target environment through an electronic nose; secondly, screening out gas sensors having no response to gas in the target environmentby variance; thirdly, extracting characteristic parameter of each sensor of the gas sensor array obtained after primary screening; fourthly, sequencing the sensors of the current sensor array according to the dynamic characteristic importance, selecting the most important sensors and allowing the same to join in an optimized array, and judging whether the current optimized array meet requirementsor not. The method has the advantages of solving the problem of low efficiency, caused by the fact of designing the sensor array according to experiences or optimizing the arrays according to simple correlation, of the existing electronic nose system, and being simple and rapid and easy to implement, having high universality and high interpretability.

The invention discloses an image-text cross-modal feature unentanglement method based on depth mutual information constraint. The method comprises the following steps: reading text and image files ina specified data mode; secondly, respectively extracting original features of the text and the image data by utilizing ResNet and BiGRU; then, under the effect of depth mutual information constraint,mapping the original features to a mixed feature space; finally, using the generative adversarial network to reconstruct the data to different degrees. By controlling the reconstruction process, the unentanglement of the cross-modal features is realized, so that the modal common information and the modal specific information are mapped to different feature spaces respectively. According to the method, the unentanglement features can be learned on large-scale image-text data, and only the features are unentangled, so that the retrieval accuracy is improved, and the depth features have better interpretability.

The invention discloses a graph classification method of a cyclic neural network model based on Attention. The Attension idea is applied in the graph classification problem, and the graph classification problem is regarded as a decision process of interaction between a machine and a graph environment in reinforcement learning. Based on the Attention idea, the machine preferentially observes the target area of the graph classification task instead of directly processing the whole graph, so that the target area can be preferentially processed by ignoring the nodes irrelevant to the classification task, and the visual angle movement direction of the machine observation graph can be trained and determined by reinforcement learning rules. At the same time, the model can control the parameters and computational load, and get rid of the constraint on the size of graph data. The invention constructs a cyclic neural network, which integrates the local information of the graph observed before bythe machine through the hidden layer of the circulating neural network, and is used for assisting the decision of the angle of view movement and the graph classification. The invention avoids the problem of subgraph isomorphism in frequent subgraph mining and the problem that the graph kernel function method lacks scalability.

The invention discloses a dynamic process monitoring method based on a latent variable autoregression model, and aims to establish the latent variable autoregression model and implement dynamic process monitoring on the basis of the latent variable autoregression model. Specifically, the method comprises the steps of defining a least square objective function of an autoregression model of a latentvariable, inferring a corresponding feature mining algorithm, and then establishing a fault monitoring model so as to implement online fault monitoring. According to the method disclosed by the invention, the dynamic autocorrelation latent variable is mined by establishing the target of the latent variable autoregression model, and the autoregression model meeting the least square condition is given correspondingly. Through the latent variable autoregression model, autocorrelation characteristics in original training data can be mined, and the influence of latent variable autocorrelation canbe eliminated. Therefore, the method provided by the invention is obviously different from the traditional dynamic process monitoring method, and the interpretability of the model is stronger. In other words, the method provided by the invention is a more preferable dynamic process monitoring method.

The invention discloses a zero-sample target detection system and a learnable semantic and fixed semantic fusion method, a zero-sample learning mechanism is introduced into a deep target detection framework, a set of zero-sample target detection system LATNet with strong discrimination capability is established, and an end-to-end zero-sample target detection task is realized through the LATNet. Alearnable semantic feature and fixed semantic feature combined method is used, so that when a network is trained in a source domain, word vector information of a category can be fully utilized, end-to-end learning can also be utilized, a category prototype with better identification capability is discovered, and the best detection accuracy is obtained. The system is simple in framework, convenientto use, high in expandability and high in interpretability, and the results of the two tasks of zero sample detection and generalized zero sample detection of the two mainstream visual attribute datasets exceed those of an existing method. And the support of a basic framework and a method is provided for the target detection technology in the military and industrial application fields.

The invention discloses a linear model method used for simplified-Chinese readability measurement. The method includes the steps of: constructing simplified-Chinese text and a readability level corpusthereof; preprocessing the text, wherein word segmentation, sentence segmentation, part-of-speech labeling, named-entity recognition, component syntax analysis, dependency syntax analysis, clause labeling and stroke counting are included; extracting and calculating text language features; constructing a best feature combination according to the language features and a regression algorithm; and constructing a linear regression model of readability measurement. The text language features adopted by the model cover four aspects of shallow-layer features, part-of-speech label features (also called semantic or lexical features), grammatical features, textual features and the like, a readability level of simplified-Chinese text for learners of which native languages are Chinese can be automatically predicted, and a gap of readability prediction models based on the simplified-Chinese text is filled. The model of the invention is high in a fitting degree, is high in interpretability, and hasextensibility and an important reference value for evaluating readability of application text.

The invention claims a medical CT image efficient storage and retrieval method based on random forest hash learning. According to the method, a random forest hash model is trained with a medical CT image set, and the model and a hash code library corresponding to an image library are stored. When a user enters a new image that needs to be retrieved, the model maps the image to a hash code; K hashcodes nearest to the hash code are retrieved in the hash code library; and the K hash codes are decoded and reconstructed into an image through a maximum compatible rule defined by the decision pathsof a tree, and the image is returned to the user. With the method implemented, the retrieval speed of images can be effectively improved, and the storage space of the images is greatly saved; the medical diagnosis of doctors can be facilitated; the workload of the doctors can be reduced; work efficiency can be improved; and the utilization rate of hospital storage devices can be improved.

The invention discloses an automobile sale prediction method considering brand emotions. The method comprises the steps that 1, comment data of automobiles purchased by consumers is collected; 2, an emotional dictionary construction method based on pointwise mutual information is proposed, and emotional values of all emotional words in the comment data are obtained; 3, the emotions in the comment text data are quantized based on an emotional dictionary, and the emotions are further collected to obtain an emotional value of an automobile brand; 4, an automobile monthly sale prediction model of brand granularity is constructed; 5, an objective function of the automobile brand sale prediction model is defined; and 6, a grid search method is utilized to solve the objection function to obtain an optimal parameter value, and therefore the prediction model is utilized to predict the sale of the next month. Through the method, seasonal variation of a sale time sequence and a short-term prediction ability of public praise emotional tendency factors of the consumers in network big data can be fully mined, therefore, automobile sale prediction is refined to more microcosmic brand granularity, and predication precision is improved.

The training method and detection method of electric wire based on depth separable convolution neural network comprises that steps of: constructing a neural network using the deeply separable convolution; trainingconvolution kernels with uniformly distributed mini-slices, wherein the trained convolution kernel is used for feature extraction of infrared gray-scale image; binarizing the image according to the threshold value, removing the small area region, and connecting the linear region by the probabilistic hough transform. The invention trains a convolution kernel capable of extracting infrared wire features by a machine learning method, and can effectively extract wire features from infrared gray-scale images. Combined with morphological processing and linear Hough transform, wire detection can be carried out in real time.

The invention discloses a remote-sensing image decision tree classification system and method, which comprises the following parts: remote-sensing image storage unit, display, image display convergence roaming control unit, exercise domain man-machine interdefining unit, decision tree growth and pruning unit, file storage unit of decision tree growth and classification precision assessment result, remote-sensing image classification disposal unit and classification result image file storage unit. The method comprises the following steps: the starting program starts the program and classification system; the remote-sensing image display program displays the image; the exercise region definition program defines the exercise region; the sample data extraction program extracts the exercise sample data; the decision tree growth and pruning unit forms the decision tree; the precision assessment program calculates the classification precision and assessment index; the remote-sensing image classification disposal program generates the classified result image. The invention can be used in the satellite-borne or airborne sensor, which classifies and disposes kinds of received remote-sensing image.

The invention discloses a relation visual attention mechanism-based scene graph generation method and mainly aims to solve the problem that redundant relation prediction and interpretability are poorin the prior art. According to the embodiments of the invention, the method includes the following steps of: 1) obtaining the category and boundary frame of targets in images through target detection,and establishing a full-connection relation graph; 2) carrying out sparsification on the relational graph through analyzing a data set to obtain sparse relational graph representations; (3) learninga relation attention transfer function through alternate iteration, transferring subjects and objects to relation occurrence sites through union set features, and learning accurate relation representations; and 4) classifying the learned relation representations, and combining the relation representations into a final scene graph. According to the method, on the basis of the internal relationshipof the relation occurrence of two targets, a relation attention mechanism is established to accurately pay attention to a region where relations occur; a scene graph is accurately generated; the interpretability of a network is improved; and the method can be used for image description and visual question and answer tasks.

The invention discloses a surface defect detection method, device and equipment and a storage medium, and the method comprises the steps: obtaining a detection picture of the surface of a to-be-detected object under the irradiation of a plurality of light sources, wherein the number of the plurality of light sources is at least three; according to the gray value of the pixel point in each detection picture and the direction vector of each light source at the pixel point, determining normal graphs corresponding to the plurality of detection pictures, wherein the normal graphs is used for indicating the normal vector of each pixel point in the detection pictures; determining a gradient graph of the normal graph according to the normal graph, wherein the gradient graph is used for indicating a gradient value of each pixel point in the normal graph; and determining a defect area according to an area formed by the pixel points with the corresponding gradient values greater than a preset threshold in the gradient map. The surface defect detection method is high in detection accuracy.

The invention discloses an EEG signal classification method based on fast multidimensional empirical mode decomposition. The method comprises: (1) collecting several groups of EEG signals and preprocessing the signals; (2) conducting the fast multidimensional empirical mode decomposition on the preprocessed signals to obtain all intrinsic mode function signals; (3) conducting spectrum analysis oneach layer of each intrinsic mode function signal and selecting a signal layer with the average power spectrum concentrated in the frequency bands of 8-12 Hz and 18-26 Hz as a new multidimensional signal; (4) making the new multidimensional signal passing through a spatial filter to extract characteristics of the EEG signals; (5) inputting the characteristics into a classifier for classification,selecting optimal parameters of a CSP according to the classification accuracy and classifying the EEG signals under different motion imaging tasks by utilizing the EEG characteristics under the optimal characteristics. The method solves the problems of pattern aliasing and low computing efficiency of a common multidimensional empirical mode decomposition algorithm, a decomposition result is moreinterpretable, and the classification accuracy of the EEG signals is improved.

The invention discloses a multi-working-condition process monitoring method based on sparse representation and belongs to the technical field of industrial process monitoring and diagnosing. According to the method, process data are not required to obey normal distribution, and only normal operating data of a process in a certain working condition are supposed to be identical to historical data of the working condition in distribution. The method includes: firstly, building a dictionary according to historical data of each working condition; and secondly, computing sparse representation of on-line data in the dictionary, and then judging whether a process is abnormal or not according to the concentration ratio of presentation coefficients. In addition, the process can be identified to be in a certain single working condition or transition process currently according to normal data, and accordingly products are guaranteed to meet production requirements. The concept of sparse representation is used for multi-working-condition process monitoring; the method does not require the process data to obey normal distribution, thereby being wider in application range and higher in interpretability.

The invention discloses a time series data risk prediction method and system based on knowledge guidance, and application thereof. The method comprises the following steps: collecting sample time series data, converting each piece of sample time series data into an input sequence matrix, and acquiring the feature information of risk events and an event relation from a knowledge graph; inputting adistance weight matrix, and the feature information of the risk events and the event relation into a preset network model to obtain two context vectors, and performing full connection on the two context vectors to realize risk prediction; training a preset network model according to each sample time series data sequence and the two context vectors, performing supervised training, and carrying outtraining to a preset convergence condition when risk occurrence probability is obtained so as to obtain a trained risk prediction model; and performing risk prediction via the obtained risk predictionmodel. The method is more reasonable and effective in time series data representation, and can improve the accuracy of time series data risk prediction.