69results about How to "Avoid underfitting" patented technology

The invention discloses a pedestrian re-identification model, method and system for adaptive difficulty mining. The identification method comprises the steps of: randomly dividing sample pictures intoa training set used for each iteration, inputting the training set into a convolutional neural network, obtaining the probability that each sample pair belongs to a positive or negative sample pair by using a softmax function, and then obtaining the loss of each sample pair by using a multinomial logistic function; obtaining a difficult sample pair by using the loss of each sample pair; and training the convolutional neural network by using the difficult sample pair until the current number of iterations reaches the upper limit of the number of iterations, thus obtaining the pedestrian re-identification model. The pedestrian re-identification model is used to extract features of each picture in a picture set to be identified, and then a similarity order of the sample pairs in the pictureset to be identified is obtained. The pedestrian re-identification model, method and system avoid over-fitting and under-fitting, and have high recognition accuracy.

The invention provides an abnormal behavior detection method and an abnormal behavior detection device based fused characteristics. The method comprises steps that, according to a detection tracking processing result of a motion target in a to-be-detected video, a behavior type of the motion target is determined; multi-dimensional characteristics of the motion target are extracted, including a pixel point change degree, a pixel point arrangement tightness degree, an integral shape, a frame image similarity degree, motion characteristics, position characteristics and form characteristics; the multi-dimensional characteristics are analyzed and processed according to a characteristic fusion model corresponding to the behavior type, whether the motion target has abnormal behaviors is determined according to the processing result; according to innovative characteristics of the multiple abnormal behaviors, algorithm robustness and stability can be effectively improved, according to the characteristic fusion model acquired through learning and training large-scale abnormal behaviors, the multi-dimensional characteristics are analyzed and processed, problems of algorithm overfitting and insufficient fitting can be effectively avoided, the method is suitable for multiple types of complex application scenes, time cost and manpower cost are greatly saved, and the method has high popularization values.

The invention discloses a semi-supervised intrusion detection method combining the improved grey wolf algorithm, and belongs to the technical field of network information security. The method can effectively avoid the defect that the basic GWO algorithm is prone to premature convergence, and balance the global exploration and local development ability of the GWO algorithm; and the cloud GWO algorithm is used to optimize the K-means algorithm to mark data having similar features, and realization of generation of the large-scale accurate mark training data set is realized on the basis of manually marking the data in a small amount. The proportion of unmarked data and marked data is provided, the phenomenon of "under-fitting" and "over-fitting" of the model is avoided, and the detection accuracy of the model is ensured. The optimized semi-supervised learning method is combined with the cloud GWO algorithm to optimize parameters of a single-class support vector machine model. Compared withthe unilateral optimization, the invention achieves higher detection precision.

The invention discloses a logging lithology recognition method based on convolutional neural network learning. The method comprises the following steps: 1, taking a data curve acquired for drilling coring as an input feature; taking a drilling lithology result as an input feature label, cleaning the sample data, and establishing a learning data sample; 2, sequentially arranging the three-porositycurve, the three-resistivity curve and the three-lithology curve, dividing drilling lithology into four types, and dividing learning data samples into a training set and a test set; 3, extracting feature parameters through one-time convolution and one-time pooling, linking a Softmax regression layer, and establishing a convolutional neural network model; 4, training the convolutional neural network model, testing the accuracy of the convolutional neural network model by using the test set; if the required accuracy is met, putting the convolutional neural network model into use, and if the required accuracy is not met, increasing the training amount; and 5, identifying the lithology of the new well by using the trained convolutional neural network model. Rock stratum information can be identified more accurately, and the convergence speed is high.

The invention relates to the field of credit card finance, in particular to a cash-out detection method and apparatus. The method comprises the steps of obtaining attribute information and transactionrecord information of a to-be-detected merchant, wherein the attribute information comprises a merchant type of the to-be-detected merchant; then according to the merchant type of the to-be-detectedmerchant, determining a normal sample merchant with the same merchant type as the to-be-detected merchant in a preset database and obtaining attribute information and transaction record information ofthe determined normal sample merchant; and finally, according to the attribute information and the transaction record information of the to-be-detected merchant and the normal sample merchant, detecting whether the to-be-detected merchant is a cash-out merchant or not. The cash-out detection is performed from a merchant dimension, so that the detection method is more scientific and effective.

The invention discloses a convolutional neural network-based music signal multi-instrument identification method. The method comprises the following steps of S1, extracting two features from an inputaudio, wherein the two features comprise a pitch feature matrix and a constant Q transformation matrix based on tone; S2, carrying out classification according to musical instrument groups, includingtubes, strings and percussion music, inputting the constant Q transformation matrix into a primary convolutional neural network to obtain a classification matrix, and inputting the classification matrix into a classifier to obtain a coarse classification result, namely the musical instrument group type; and S3, on the basis of the classification matrix, inputting the classification matrix into a secondary convolutional neural network with an attention network in combination with a pitch matrix to obtain a subdivision result, namely a specific musical instrument, wherein the attention network allocates weights to different harmonic waves. The method is suitable for musical instrument identification tasks in music information retrieval and can be used for the musical instrument identification method in music automatic transcription.

The invention discloses a photovoltaic power station short-term power prediction method, and the method comprises the steps: selecting meteorological factors such as season types, weather types and irradiation intensity / temperature as an input data set according to the historical output power data and numerical weather prediction data of a photovoltaic power station; preprocessing the input data set, extracting a day type feature vector, and clustering the day type feature vector by adopting a K-Means clustering method to obtain K different day type results; according to the numerical weatherprediction data of the prediction day, determining a day type to which the prediction day belongs, obtaining a data sample set of the most similar day of n days in the day type to which the predictionday belongs based on a similar day theory, taking the data sample set as prediction model training data, performing training modeling on the training data set by adopting a random forest regression prediction algorithm, and establishing a photovoltaic power station short-term power prediction model; and calling a photovoltaic power station short-term power prediction model based on the predictionday numerical weather prediction data, and obtaining a short-term power prediction result of the photovoltaic power station in the prediction day.

The invention relates to a target recognition training method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring video stream data, wherein the video stream data comprises multiple frames of images; reading the image, detecting a corresponding target in the image, and generating a plurality of types of sample image sets by utilizing the target;training the sample image sets of multiple categories by using a recognition model to obtain recognition probabilities corresponding to the categories; adjusting samples in the corresponding sample image set according to the identification probability; and optimizing the identification model by using the adjusted sample image set, and training the optimized identification model through the adjusted sample image set. By adopting the method, the problem of poor model learning effect caused by unbalanced samples can be solved, so that the trained recognition model can perform accurate target recognition, and the problems of under-fitting and over-fitting of the training model are avoided.

The invention discloses an automatic electrocardiogram classification method, system and device based on deep learning. The method comprises the following steps: dividing acquired labeled original electrocardiogram data into a training set and a verification set according to a preset proportion; constructing a convolutional neural network with residual connection, and substituting the convolutional neural network into the training set and the verification set for training and verification to obtain a trained convolutional neural network; evaluating the trained convolutional neural network by using the labeled test set, and acquiring an automatic electrocardiogram classification model passing the test in combination with evaluation indexes; and inputting an electrocardiogram to be tested into the automatic electrocardiogram classification model to obtain an electrocardiogram classification result. The automatic electrocardiogram classification method based on deep learning can carry outcomprehensive feature extraction, complete judgment task of multi-label classification, comprehensively extract information in the electrocardiogram, and complete classification.

The invention relates to a smoke detection method based on multi-network model fusion, and two network models of VGG16 and ResNet50 are fused to realize reliable detection of smoke. According to the fusion network provided by the invention, more abundant smoke image detail features can be extracted, and the distinguishing capability of the features on the smoke image and the smoke-like image is enhanced. By adopting a feature transfer learning method based on an isomorphic space, feature extraction layers of pre-trained VGG16 and ResNet50 models can be well migrated to a target data set classification task in a smoke scene, and meanwhile, the generalization capability of the model is improved. By fusing the multi-network model, the distinguishing capability of the characteristics is enhanced, the false alarm phenomenon caused by targets such as cloud and fog similar to smoke is reduced, and the reliability of smoke detection is further improved.

The invention belongs to the field of wireless equipment identification, and discloses a wireless equipment fingerprint identification method, a wireless equipment fingerprint identification system, equipment and a readable storage medium. The wireless equipment fingerprint identification method comprises the following steps of: collecting a network data frame of wireless equipment in a communication process; extracting characteristic parameters of the wireless equipment capable of reflecting the characteristics of the wireless equipment from the network data frame; constructing a characteristic fingerprint of the wireless equipment according to the characteristic parameters of the wireless equipment; training the characteristic fingerprint of the wireless equipment through using a classifier and then determining classifier parameters to obtain a wireless equipment fingerprint identification model; and identifying the current wireless equipment through using the wireless equipment fingerprint identification model and a characteristic fingerprint preset in a fingerprint library. The wireless equipment fingerprint identification method overcomes the defect that a traditional equipment identification method based on a certain characteristic parameter of the wireless equipment is prone to be forged and tampered, accurate identification of the wireless equipment is achieved by implicitly collecting the network traffic of the wireless equipment, the whole identification process does not interfere with normal operation of the wireless equipment, participation of a user is not needed, and good user experience is achieved.

The invention discloses a welding device for industrial machinery production. The welding device comprises a base; an arc groove is formed in the outer wall of the top of the base; a first ring is mounted on one side of the outer wall of the top of the arc groove; the inner wall of the first ring is rotationally connected with a first fixed ring; a tooth groove is formed in one side of the peripheral outer wall of the first fixed ring; a fixed plate is mounted on the outer wall of the top of the first ring; a motor is mounted on the outer wall of the top of the fixed plate; a gear sleeves theperipheral outer wall of an output shaft of the motor; and the gear is engaged with the tooth groove on the first fixed ring. In the welding device, through arrangement of a lead screw, a slide blockand a servo motor, a second ring can be driven to move, and two round pipes to be welded can be tightly bonded to achieve welding; and through arrangement of the motor, the gear and the tooth groove on the first fixed ring, the round pipes on the second fixed ring and the first fixed ring can be driven to rotate for total welding of the peripheral outer walls thereof, so that the working efficiency is improved.

The invention discloses an agricultural product growth environment big data detection system which is composed of a soil environment parameter acquisition and evaluation platform and a farmland soil environment big data processing subsystem, and realizes farmland soil environment parameter detection and management and farmland parameter evaluation. The system effectively solves the problems that an existing farmland soil environment detection system does not influence the production yield of farmland plants according to nonlinearity and large lag of changes of farmland plant growth environment parameters, large and complex farmland soil environment areas and the like, and therefore farmland soil yield prediction and production management are greatly influenced.

The invention discloses a self-learning image super-resolution reconstruction method and system based on a convolutional neural network. The method comprises the steps of 1, obtaining a training sample of a to-be-reconstructed image; 2, constructing a convolutional neural network; the convolutional neural network comprises a feature extraction unit, a feature enhancement unit, a residual error unit and a reconstruction unit. 3, training the convolutional neural network constructed in the step 2 based on the training sample obtained in the step 1 to obtain a trained reconstructed convolutionalneural network; and 4, performing super-resolution reconstruction on a to-be-reconstructed image based on the reconstructed convolutional neural network trained in the step 3. According to the method,the problem of insufficient training samples of the self-learning algorithm can be effectively solved, and the over-fitting phenomenon of the network can be avoided; meanwhile, a high-resolution image with a higher peak signal-to-noise ratio and a better visual effect can be obtained.

The invention discloses an intrusion detection method, which comprises the following steps: classifying data in an obtained data set, with the category of the data comprising a large sample and a small sample; carrying out data expansion on the data of which the category is a small sample in the data set; dividing the expanded data set into a training set and a test set, training a preset networkmodel by using the training set, and performing performance evaluation on the trained preset network model by using the test set; and determining the preset network model with the optimal performanceas an intrusion detection model, and performing intrusion detection on the obtained perception data by using the intrusion detection model. Data of which the category is a small sample in the data setis subjected to data expansion, so that the underfitting phenomenon of a learner on the small sample is avoided, the learning efficiency and generalization ability of the model are improved, and theintrusion detection accuracy is further improved. The invention further provides an intrusion detection system and device and a readable storage medium which have the above beneficial effects.

The invention discloses a moisture detection system based on a cloud platform. The moisture detection system is composed of a granary environment parameter acquisition and control platform and a granary environment big data processing subsystem, wherein the granary environment parameter acquisition platform detects, adjusts and monitors granary environment parameters, and the granary environment parameter big data processing subsystem comprises a plurality of moisture detection models and a detection parameter fusion model, so that moisture of grains is predicted, and the granary environment production management efficiency and benefits are improved. According to the invention, the system effectively solves the problems that an existing granary environment parameter detection system does not consider influence of nonlinearity and large lag of granary environment parameter change, large and complex granary environment area and like on grain safety, and does not predict granary environment moisture and accurately adjust granary environment parameters so as to greatly affect granary safety and grain management.

The invention discloses an aerial handwriting inertia sensing signal generation method based on deep adversarial learning, which comprises the following steps: carrying out filtering, smoothing and noise reduction on an obtained aerial handwriting inertia sensing signal sequence to serve as a training sample set; designing a deep convolution condition confrontation generation network based on timesequence feature map position coding; using the training sample set to train the generation network; and inputting the formulated sample length, the sample category label and the random noise vectorinto the trained adversarial generation network, and taking the input of the generator as a generated air handwriting inertial sensing signal sequence. According to the invention, aerial handwriting inertial sensing signal samples with certain diversity and good quality can be generated, and the effective lengths and types of the generated samples are controllable.

The invention discloses a pipe network big data detection system which is composed of a pipe network parameter acquisition platform and a pipe network safety big data processing subsystem, the pipe network parameter acquisition platform detects pipe network parameters, the pipe network safety big data processing subsystem classifies pipe network safety, and pipe network operation benefits and reliability are improved; the method effectively solves the problems that an existing pipe network has no influence on pipe network safety according to nonlinearity and large lag of pipe network parameter changes, large pipe network area, complex parameter changes and the like, pipe network parameters are not predicted, and pipe network safety is not pre-warned, so that reliable operation and intelligent management of the pipe network are greatly influenced.

The invention relates to the field of automatic production, and discloses an air injection control system. An NARX neural network model 1 and an NARX neural network model 2 are used for controlling air flow and predicting an air flow value respectively, and an NARX neural network establishes a dynamic recursive network of the models by introducing a time delay module and outputting feedback, input and output vector delay feedback is introduced into network training to form a new input vector, good nonlinear mapping capability is achieved, the input of the NARX neural network not only comprises the error of original gas flow, the control quantity and the input data of actual gas flow, but also comprises the trained corresponding output data; the generalization ability of the network is improved, so that the system has better prediction precision and adaptive ability in prediction of corresponding parameters of the gas flow compared with a traditional static neural network, and the single-chip microcomputer controller improves the precision, robustness and reliability of the control system.

The invention discloses an extreme random tree furnace temperature prediction control method based on longicorn beard search, and belongs to the field of furnace temperature prediction and control inthe industrial combustion process. By establishing the regression relation between the key variables influencing the temperature and the temperature of the combustion chamber, the predicted value of the furnace temperature in the combustion chamber is obtained in real time, and the prediction precision of the furnace temperature reaches + / -1.5 DEG C. In consideration of low control precision of anoriginal system, a longicorn beard search algorithm is designed, a secondary performance index function commonly used in predictive control is selected as a fitness function of longicorn, and an optimal control quantity is searched through an olfaction search mechanism of the algorithm, so that the control effect of the whole system is better.

The invention discloses a livestock feed detection system, and the system comprises a livestock breeding environment parameter collection and control platform and a feed ratio big data processing subsystem, and the livestock breeding environment parameter collection and control platform achieves the detection, adjustment and monitoring of livestock environment parameters, and the feed ratio big data processing subsystem is used for predicting the feed-to-weight ratio of the livestock feed ratio. The system effectively solves the problems that an existing livestock breeding environment parameter detection system has no influence on livestock breeding economic benefits according to nonlinearity, large lag, large breeding environment and the like of livestock breeding environment parameter changes, and does not predict the feed weight ratio of livestock feed and accurately adjust the livestock feed ratio, therefore, the economic benefit of livestock breeding and breeding management are greatly influenced.

The invention discloses a dust concentration intelligent detection system which is composed of a dust concentration environmental parameter acquisition platform based on a CAN bus and a dust concentration intelligent prediction module. The dust concentration environmental parameter acquisition platform based on the CAN bus realizes detection and adjustment of dust concentration environmental factor parameters. The dust concentration intelligent prediction module is composed of a dust concentration interval number neural network model, an interval number prediction model and an interval numberCMAC cerebellar neural network dust concentration fusion model. The system effectively solves the problems that the existing industrial and agricultural production environment does not have the characteristics of nonlinearity, large lag, complex dynamic change and the like according to the change of the dust concentration, and the dust concentration cannot be accurately detected and predicted, sothat the effective management of the dust concentration of the industrial and agricultural production environment is greatly influenced.

The invention provides a fan noise prediction method based on an optimal neural network. The fan noise prediction precision and the generalization ability of the neural network are improved mainly through joint control on the number of input neurons and the number of hidden layer neurons. Based on correlation analysis, the influencing importance of input parameters in fan samples towards output parameters is sorted, and according to the training precision and the prediction precision, an input layer neuron number range and the maximum input layer neuron number are determined. The correlation analysis is used to effectively reduce the input neuron number, and the building difficulty of the optimal neural network structure is reduced. The best hidden layer neuron number is used to determinethe optimal neural network structure, over fitting and under fitting can be effectively avoided, and while the training precision is improved, the prediction precision and the generalization ability are also improved.

The invention discloses a bearing fault diagnosis method based on a semi-supervised adversarial network. The method comprises the following steps: S100, collecting a vibration signal xf when a bearing has a true fault, a vibration signal xh when the bearing normally operates, and a vibration signal of a to-be-detected bearing; S200, constructing a semi-supervised generative adversarial network composed of a generator g, a feature network f, a fault classifier fc, a discriminator d, an auxiliary classifier ac and a diagnosis network diag, and training the semi-supervised generative adversarial network: S201, training the generator g to generate a fault state and a pseudo bearing vibration signal under normal operation; S202, training the feature network f, the fault classifier fc, the discriminator d and the auxiliary classifier ac according to the vibration signals xf and xh and the pseudo bearing vibration signal; and S203, after training convergence of the step S201 and the step S202, training the diagnosis network diag by using the vibration signals xh and xf and the pseudo bearing vibration signal; and S300, inputting the vibration signal of the to-be-detected bearing into the trained diagnosis network diag for fault diagnosis.

The invention belongs to the technical field of data classification, and particularly relates to a gesture classification recognition method and application thereof. However, some current gesture classification recognition algorithms about sEMG signals are low in recognition accuracy, and over-fitting and under-fitting, gradient disappearance, poor robustness and long training time also exist in a model training process. The invention provides a gesture classification recognition method. The method comprises the following steps: acquiring a surface electromyogram signal; performing feature extraction on the surface electromyogram signal to obtain a gesture feature sequence and a gesture type; and inputting the gesture feature sequence and the gesture type into a circulation gate circuit neural network for training to obtain a classification model, and adopting the classification model to realize gesture classification recognition. And the prediction classification accuracy is improved.

The invention discloses a substance mixing intelligent system, which is characterized in that the system is composed of a parameter acquisition and control platform and a substance mixing adjustment subsystem, the system realizes the parameter detection of a mixed object and the adjustment of a substance mixing process, and reliability and accuracy of the process of adjusting the mixed substance are improved; according to the influence of problems of inaccuracy, poor robustness and the like of an existing gas or liquid mixing ratio control system on the performance of the gas or liquid mixing ratio system, problems of poor robustness, slow response speed and the like of the gas or liquid mixing ratio system are effectively solved.

The invention discloses an ICA sparse representation and SOM-based distorted image quality evaluation method. The method comprises the steps of 1) performing ICA sparse representation on a reference image and a to-be-tested image to obtain sparse representation information of the images; 2) calculating an SSIM value between the reference image and the to-be-tested image after the sparse representation; 3) drawing a scatter diagram of the SSIM value and a DMOS value; 4) clustering scatter diagram data by using an SOM algorithm: classifying the data distributed more intensively into one category, and classifying other data into the other category; 5) using a cross validation regression algorithm for each category of the data, and performing regression mapping on the SSIM value in the step 4)to obtain the DMOS value; 6) calculating an error value between the DMOS value of the two categories of the data and a DMOS value in an actual database; and 7) performing weighted average on the obtained error value to serve as an index value of final image quality evaluation.

The invention belongs to the related technical field of laser probe element analysis, and discloses a method for improving the classification precision of a laser probe by utilizing spectral characteristic expansion. The method comprises the following steps: S1, collecting a plasma spectrum by utilizing a laser probe spectrum collection device; S2, averaging the plasma spectrum, and selecting an analysis line and corresponding start and stop wavelengths in the obtained flat spectrum; S3, extracting spectral intensity, spectral peak area, spectral peak full width at half maximum, spectral peakstandard deviation, spectral peak signal-to-noise ratio and spectral peak signal-to-noise ratio characteristics from the original spectrum; S4, performing feature expansion on the input feature vectorby utilizing the features to obtain an expanded mixed spectrum feature vector; S5, training the expanded mixed spectral features in combination with a classification algorithm to obtain a classification model based on the mixed spectral features; and S6, inputting the mixed spectral characteristics of a test set into the classification model, and outputting a classification result by the classification model to finish classification. According to the method, traditional spectral feature vectors taking the spectral intensity as main components are effectively expanded, and the characterizationcapability and the classification accuracy of the spectral feature vectors are improved.

The invention belongs to the field of battery design, and particularly relates to a laminated solar battery structure optimization method, which comprises the steps of taking structure information ofa to-be-optimized laminated solar battery as population information of a differential evolution algorithm, taking a battery performance index as an optimization target of the differential evolution algorithm, and initializing the structure information; controlling a differential evolution algorithm to perform iterative evolution on the initial structure information for multiple times by adaptivelyadjusting a scaling factor and a crossover probability required by each iteration, wherein each iterative evolution is to jointly adjust each layer of structure in the laminated solar cell to obtaina new population and predict a cell performance index according to the new population by adopting a pre-constructed cell performance prediction neural network, and finally optimal structure information is obtained. According to the self-adaptive differential evolution algorithm, the structures of all layers can be jointly adjusted, the problem of local optimization is avoided, the differential evolution algorithm is combined with the battery performance prediction neural network, the battery structure can be designed in a high-efficiency and time-saving self-adaptive reverse optimization mode,and the optimization efficiency is improved.