156results about How to "Avoid precision loss" patented technology

The invention discloses a prediction method of the forecabin thermal protection system whole trajectory temperature boundary of a hypersonic velocity aircraft. The prediction method comprises the following steps: 1) according to the layout features of the forecabin thermal protection system of the hypersonic velocity aircraft, establishing a three-dimensional geometric model of the thermal protection system; 2) extracting the thermal conductivity coefficient and the emissivity of each material as feature parameters, and finishing the parametric establishment of a forecabin thermal protection system finite element model which takes the feature parameters as driving; 3) on the basis of the whole trajectory reentry process flight working condition of the aircraft, comprehensively considering thermal conductivity and thermal radiation effects to realize the transient thermal analysis of the thermal protection system in a whole trajectory process; and 4) considering material dispersibility, taking the thermal conduction coefficient and the emissivity of the material as uncertain input parameters, and finishing the analysis of the thermal protection system whole trajectory temperature boundary in the trajectory process on the basis of an interval vertex analysis method. The prediction method accurately and efficiently predicts the forecabin thermal protection system temperature boundary in the harsh reentry environment of the hypersonic velocity aircraft, and lays a foundation for the subsequent reliability rational evaluation and the uncertain optimization design of the structure.

The invention relates to a digital closed loop servo circuit applied to a quartz flexible accelerometer, which comprises a differential capacitor detection circuit, a triangular wave generating circuit, a pre-filtering amplifying circuit, an A / D conversion circuit, a D / A conversion circuit, a drive amplifying circuit and a digital signal processing circuit, wherein the triangular wave generating circuit is connected with the input end of the differential capacitor detection circuit, the differential capacitor detection circuit is connected with a differential capacitor of a meter mechanical part of the quartz flexible accelerometer, the differential capacitor detection circuit is connected with the pre-filtering amplifying circuit, the pre-filtering amplifying circuit is connected with the A / D conversion circuit, the digital signal processing circuit is connected with the A / D conversion circuit, the digital quantity output by the digital signal processing circuit is connected with a navigation computer through an output end and is connected with the D / A conversion circuit through the output end at the same time, the D / A conversion circuit is connected with the drive amplifying circuit, and a current signal output by the drive amplifying circuit is connected with a torque coil of the meter mechanical part of the quartz flexible accelerometer through the output end of the drive amplifying circuit.

The invention discloses a dynamic MEMS (micro-electromechanical systems) inertial attitude measuring system based on a self-adaptive robust integration algorithm, and provides a sensing system for measuring an attitude angle of an inertial space at high accuracy under a dynamic acceleration environment. The dynamic MEMS inertial attitude measuring system comprises an MEMS inertial unit, a magnetometer, a data integration computer, and a synchronous sampling and data communication unit, wherein the MEMS inertial unit is used for collecting the angular speed and acceleration of a sensor in the inertial space; the magnetometer is used for collecting the component information of an earth magnetic field; the data integration computer is used for operating the self-adaptive data integration algorithm; the synchronous sampling and data communication unit is used for obtaining the data of the synchronous sensor, and transmitting to other equipment on a bus through a digital interface. The dynamic MEMS inertial attitude measuring system has the advantages that the problem of low attitude accuracy due to the easy influence by the acceleration in the existing MEMS inertial attitude measuring system is solved; a high-accuracy attitude measuring system with small size and low cost is realized.

The present invention discloses a three-dimensional print slicing method based on a T-spline surface. The method comprises: inputting a to-be-processed T-spline surface, a slice layer thickness and a slice data allowable error; in a T-spline surface parameter field, generating a uniformly distributed parameter grid, and calculating a corresponding Euclidean space grid; according to a model layer thickness, selecting a grid point whose z coordinate is closest to a current layer as a slide initial point at each layer; iterating the slide initial point that exceeds the error along a steepest change direction; by using the slide initial point of each layer as a start point, tracking along an unchanged direction of the z coordinate of the surface, and iterating a slide point that exceeds the error along the steepest change direction, and finally obtaining all slice points that satisfy a precision requirement; and connecting the slice points of each layer, and outputting a slice contour. According to the method provided by the present invention, slicing of the T-spline surface is realized by using a numerical method, and the defects of huge data volume and low precision of traditional three-dimensional printing are solved, so that the method is particularly applicable for three-dimensional print slice processing of models with a complex free surface appearance or internal holes.

The invention discloses a high-speed and high-precision SAR image ship detection method. The method is based on a target detection theory of deep learning, and a ship detection model is constructed byusing a deep separation convolutional neural network. The detection model provided by the invention mainly comprises a backbone network and five modules, wherein the backbone network is used for extracting the characteristics of a ship, the five modules are a multi-receptive-field module, a cavity convolution module, a channel and space attention module, a characteristic fusion module and a characteristic pyramid module which are used for improving the detection precision. The method is advantaged in that a model provided by the invention has relatively small depth and width, therefore, the ship detection model has fewer parameters, so the ship detection model provided by the invention has higher detection speed, and the detection speed on an SSDD data set, a Gaofen-SSDD data set and a Sentinel-SSDD data set exceeds 220FPS (FPS is a frame rate).

The invention discloses a high-precision calculating method of a random vibration dynamic stress of a mechanical structure. The random excitation is converted into certainty excitation, and calculation efficiency is improved when mode-coupling effect is completely taken into consideration. By the mode stress coefficient calculation, precision reduction during stress calculation by displacement derivation is avoided; static force is corrected, neglected influence of high-order modes in mode interception is considered, and calculation accuracy of random dynamic stress is improved. Cross-correlation terms among all vibration-involved modes are included during calculation, and exact solution is obtained. For large-scale and complex projects, the high-precision calculating method is easy to operate and implement and is higher in calculation efficiency than a conventional algorithm. In addition, calculation accuracy is improved by introducing a stress mode matrix and static force correction terms of the high-order modes.

A suppressor includes baffles with geometry and spacing minimizing sound level in the human hearing range, and overlapping tapers on consecutive elements replacing welds. The geometry includes conical baffles with approach angles between 153.7 and 163.7 degrees and at least one inch separation. The suppressor is assembled by compressing the elements between threaded end caps, thus expending the overlapping tapers against the interior of a suppressor tube to center and align the baffles. The suppressor may be attached to a rifle using a quick disconnect mount which includes an adapter fixed to the rifle barrel and having an “L” shaped slot with a first leg parallel to the barrel and a second leg turned greater than 90 degrees towards the front of the adapter. A post in the suppressor engaged the slot and a spring biases the suppressor forward and holds the post at the end of the turned leg.

The invention discloses a distributed multi-robot synergetic location algorithm. The Distributed multi-robot synergetic location algorithm aims at n robots (respectively identified as R1, R2...Rn) which are provided with locating devices and can communicate mutually and steps such as a synergetic location algorithm at time of k=1 and a synergetic location algorithm at time of k+1 (k=1,2...). The method is a distributed bayes estimation algorithm. Under the premise that a decoupling approximation is not carried out on a system model, three-type alignment sensor data such as movement metrical information of the robots, external observation information of a single robot and relative observation information between the robots are fused in robot local; the communication of the robot is point-to-point communication which comprises two parts: one of the parts is the communicated of the relative observation information between two robots which relatively observe, the other of the parts is all robots form a communication chain, and the communication content of adjacent robots is associated state estimation of robots after a sender fuses local alignment sensor data. The method has the advantages of being modularized, good in expandability, good in fault tolerance and the like.

The invention discloses a method of filtering an airborne LiDAR (Light Detection and Ranging) point cloud. The method comprises the following steps of firstly, carrying out gross error elimination and regular grid transformation on LiDAR point cloud data so as to generate a depth image; secondly, computing a segmented elevation threshold through an Otsu algorithm in an image threshold segmentation technology, and carrying out iterative rough classification of ground points and non-ground points on the point cloud data, which are obtained before regular grid transformation and resampling, through the threshold; lastly, respectively carrying out progressive triangulation network filtering on the classified ground points and non-ground points through the two different thresholds, and outputting network construction point cloud data, namely, ground point data. According to the method, the point cloud data, which participate in a filtering process, are data, which are obtained before regular grid transformation and resampling, so that the problem of accuracy loss of the point cloud due to regular grid transformation can be effectively avoided; a categorical attribute guidance is provided for the progressive triangulation network filtering, a filtering strategy is correspondingly adjusted for different terrain conditions, so that a better filtering effect is obtained.

The invention relates to a face key point recognition and tracking method and system applied to cross-platform. The face key point recognition and tracking method comprises the steps: firstly collecting face images, marking the key point of each face image, and making a face image training sample set; based on a multi-task convolutional neural network algorithm, training to obtain a multi-task convolutional neural network model; collecting a face image, preprocessing the face image, loading a multi-task convolutional neural network model, reading a current frame image, and synchronously obtaining a face region and corresponding face key point position information; adopting the key point information of the first frame of face image as the input of the current frame, calculating the face keypoint information of the current frame through a multi-task convolutional neural network model, and judging whether the key point of the current frame of face image is in a successful tracking stateor not; and finally, after accumulatively tracking the predicted number of face key point information, calculating the Euler angle of the face through a pre-trained face deflection angle calculation model to complete face posture estimation.

A suppressor includes baffles with geometry and spacing minimizing sound level in the human hearing range, and overlapping tapers on consecutive elements replacing welds. The geometry includes conical baffles with approach angles between 153.7 and 163.7 degrees and at least one inch separation. The suppressor is assembled by compressing the elements between threaded end caps, thus expending the overlapping tappers against the interior of a suppressor tube to center and align the baffles. The suppressor may be attached to a rifle using a quick disconnect mount which includes an adapter fixed to the rifle barrel and having an “L” shaped slot with a first leg parallel to the barrel and a second leg turned greater than 90 degrees towards the front of the adapter. A post in the suppressor engaged the slot and a spring biases the suppressor forward and holds the post at the end of the turned leg.

The invention relates to a structural deformation observation and data processing method at the operation and maintenance stage of a high-speed train. Though a structural deformation observation and data processing at a construction period can predict under-track settlement, the method has adaptability-related problems in the structural deformation monitoring at an operation period. The method comprises the steps of: laying reference points, deformation monitoring points and characteristic section points at two levels; conducting rigorous constraint adjustment to calculate the heights of the reference points, the deformation monitoring points and the characteristic section points; under the situation that the deformation monitoring points are broken or lost, conducting broken height repair to actually measured height difference to eliminate high-difference mutation errors caused by broken height; and calculating actually measured height difference between two adjacent monitoring points, directly adopting the actually measured height difference to conduct differential deformation analysis to the adjacent deformation monitoring points and calculating current additional gradients, accumulated additional gradients, gradient difference and accumulated gradient difference of a line. The structural deformation observation and data processing method at the operation and maintenance stage of the high-speed train has the advantages that the field observation workload is reduced, the differential deformation is accurately and intuitively reflected and comprehensive suggestions are made for the safe operation of railways.

The invention relates to a calibrating and aligning method of a CCD (Charge-coupled Device) camera. An aligning platform and two CCD cameras are used in the method; during calibration, a standard sample is fixed on the aligning platform, the aligning platform quantitatively deviates and rotates on the horizontal plane, the two CCD cameras record the coordinate of a Mark point after each motion, and the conversion relation between each coordinate system of the CCD cameras and the coordinate system of the aligning platform is obtained through calculating the change of the Mark point; after the calibration of the two CCD cameras is finished, a product to be processed can be aligned by means of the conversion relation. The method has the advantages that during calibration, the aligning platform is controlled to horizontally move and rotate, the change of the Mark point is respectively captured by using the two CCD cameras, and the conversion relation between the coordinates is obtained through calculation; during alignment, the position coordinate of the Mark point of the product is directly captured by using the two CCD cameras, the actual deviation value is obtained after a series of calculations are carried out, and then, the alignment is carried out; the CCD cameras are not required to be moved in both the calibration and alignment processes, so that the precision loss is reduced, the calibrating and aligning precisions are improved, and the calibrating and aligning speeds are increased.

The invention provides a Beidou PPP-RTK-based low earth orbit satellite space-time reference establishment method. According to the method, an observation equation is established by fully utilizing various error correction information broadcasted by Beidou PPP-RTK service, a navigation satellite precision orbit and a clock difference product; and a low earth orbit satellite space-time reference issolved by adopting a Kalman filtering parameter estimation strategy. Compared with an existing method, the method effectively avoids precision loss caused by resolving the space-time reference of a low earth orbit satellite by only adopting a navigation satellite orbit introduced by broadcast ephemeris and clock error precision difference, and also avoids precision loss caused by the fact that asatellite-borne receiver clock error absorbs various deviation items which are not accurately modeled; and the method has remarkable characteristics and advantages compared with a traditional low-orbit satellite orbit determination and time service method.

The invention belongs to the field of precision revolution body high-precision measuring and particularly relates to a sensor mounting angle selecting method for a three-point method roundness error isolation technology. On the basis of the three-point method roundness error isolation technology, the character of a weight function during an isolation process is combined, a specific value is not used as a selecting target of a sensor mounting angle any longer, under the condition of a range where certain errors are allowed, angle combination sets which accord with the preset condition are screened, and then the best sensor mounting angle combination is obtained by comparing. Compared with a traditional angle selecting method, the phenomenon of weight function harmonic suppression caused by sensor mounting errors is effectively controlled, and when a sensor mounting angle deviates from a preset angle slightly, the precision of an error isolation process can still be guaranteed. A complex data post-processing step is of no need, an angle selecting cycle is short, an implementation process is simple, a deducing process is simple, and computer programming can be well achieved.

Method and system for a direct transmission of motion path data (310) from a generating system (204) to a control system (205) for use in controlling a servo-driven machine. For example, a machine tool for machining stock material, and the data generating system includes a computer-aided-design software package (301). Tool paths for directing the tool members of the machine tool can be extracted from a design file opened by the CAD software through the operation interface of the CAD software (302). Alternatively, the design file can be used to specify motion paths rather than end product geometry. The CAD interpreter application extracts the motion path data from the design file, eliminating the need to translate the data into another or intermediate form for controlling the machine tool.

An apparatus is provided for measuring viewing angle characteristic of luminance and positional characteristic of a radiant object under measurement, by moving a condensing device and an imaging device relative to the object. The apparatus includes a first mechanism for moving the light receiving elements of the condensing device while keeping constant the solid angle subtended by the light receiving element at the radiant area. A second mechanism is provided for moving the imaging device in association with the condensing device, while maintaining thereon the image of the radiant area. A memory is provide. A circuit is provided for calculating viewing angle dependent and position dependent characteristics. A display device is provided for displaying radiance characteristic and distribution characteristic of the object and the result of evaluation of the calculated characteristic of luminance.

Method and system for a direct transmission of motion path data (310) from a generating system (204) to a control system (205) for use in controlling a servo-driven machine. For example, a machine tool for machining stock material, and the data generating system includes a computer-aided-design software package (301). Tool paths for directing the tool members of the machine tool can be extracted from a design file opened by the CAD software through the operation interface of the CAD software (302). Alternatively, the design file can be used to specify motion paths rather than end product geometry. The CAD interpreter application extracts the motion path data from the design file, eliminating the need to translate the data into another or intermediate form for controlling the machine tool.

The invention discloses an integrated processing technology of a high speed motor train unit bogie frame, comprising a magnetic dial indicator which is used to quantize and control clamping deformation and clamping force, an integrated accompanying chemical device which is used for the integrated processing of the frame, a numerical control core milling processing center which is used for the integrated processing of the frame. By using the magnetic dial indicator to quantize and control clamping deformation and clamping force, the integrated processing technology effectively controls the stress produced by clamping frocks, and furthermore guarantees the precision of the size of the frame after finishing processing and loosening the clamp. By using the method for determining the relationship between the workpiece and the machine tool coordinate system through two positioning, the integrated processing technology improves the utilization rate of the machine tool, and ensures that the improvement of the clamping precision of the frame becomes possible on the other hand. By using the numerical control core milling processing center, the integrated processing technology fully utilizes all kinds of functions of the machine tool to process the frame with a pentahedron analogously type, with the result that the precision loss which may be produced by multiple step processing can be effectively avoided, and the processing precision and quality of the frame can be greatly improved.

The invention discloses a rowing motion compensation method for fiber optic gyroscope trapdown inertial navigation, which comprises the following steps of: collecting and storing angular rate data output by a fiber optic gyroscope and specific force output by an accelerometer; updating an attitude matrix; carrying out speed updating and rowing motion compensation, segmentedly fitting a change in the angular rate and the specific force of a carrier by using a cubic parabola, and directly utilizing the angular rate output of the gyroscope and the specific force output of the accelerometer to carry out rowing motion compensation; and updating a carrier position. According to the rowing motion compensation method for fiber optic gyroscope trapdown inertial navigation provided by the invention, the change in the angular rate and the specific force of the carrier is segmentedly fit, and a rowing effect compensation term is converted into a form expressed by the angular rate and the specific force, so that the precision loss brought by angular increment conversion is avoided, and the performance of a fiber optic gyroscope strapdown inertial navigation system is improved.

The invention discloses a titling mode sampling modeling and super-resolution reconstruction method so as to achieve the reconstruction of high spatial resolution satellite remote sensing images. The method comprises the steps that a pixel coordinate system on a titling mode sampling grid is mapped onto a pixel coordinate system on a high-resolution conventional sampling grid; in a space coordinate system, triangulation is carried out on tilting mode sampling pixel points; according to an unknown pixel on the high-resolution grid, the value of the unknown pixel on the high-resolution grid is calculated by the adoption of the super-resolution reconstruction method like a polynomial interpolation through the value of the known pixel of the vertex of the triangle where the unknown pixel is located. According to the disclosed titling mode sampling modeling and super-resolution reconstruction method, single-linear-array titling mode sampling imaging is adopted, the problems that two rows of sensor linear arrays are difficult to manufacture, and the registration accuracy is hard to control are solved, meanwhile, image data collection is carried out on sampling intervals by the control over the push broom direction, the spatial sampling grid formed by tilting mode sampling is square, and the problem of accuracy losses caused by the additional geometric correction and the gray-level interpolation is solved.

The invention discloses a relay protection action delay time grading test system and method on the basis of Rogowski coil transient simulation. The system comprises an upper computer and a test host computer; the upper computer is provided with a data source generating module and a delay time test module; the test host computer comprises a master controller, a data generating module, a data acquisition module, a D / A module, an A / D module and an acquisition interface module. According to the system and method, firstly, a transient simulation model of the Rogowski coil under electrical power system failure conditions is established by the data source generating module, and then transient simulation data is sent to the test host computer in real time; differential small analog quantity signals are sent to a collector through the data generating module and the D / A module to establish whole interval dynamic analog data; at the same time, the small analog quantity signals are used as a standard source, and then serial data signals, merged unit 9-2 data, digital relay protected GOOSE data and intelligent terminal switching values are collected, and therefore, a whole set of test system is established; a delay time of those signals relative to the transient standard source and an overall action delay time can be tested.

The invention discloses a method for automatically matching characteristic lines of close-range photogrammetry in order to have noncontact measurement of contour sides, molded surface control lines and other curves playing a key role in the surface of an object. The method does not need to adopt image curve resampling technology and optimization algorithm to search an optimal pixel point of a different visual photo showing a certain point of homologous curve, directly solves the matching problem of forming the point of the homologous curve in the different visual photo, avoids the time-consuming process of solving a nonlinear optimization equation and the precision loss in the image curve resampling process and promotes the precision and efficiency of the curve matching.

The invention discloses a non-probabilistic reliability pneumatic structure coupling optimization design method for a hypersonic velocity three-dimensional wing. The method comprises the following steps: expressing beam and rib thickness size, elasticity modulus and density into interval variables, and acquiring a sample space of input parameters by utilizing an interval vertex method; finishing finite element parametric modeling of the hypersonic velocity three-dimensional wing according to sample points in the sample space; finishing pneumatic structure coupling analysis of all sample points in the sample space by utilizing an iterative algorithm; screening to obtain internal upper and lower limits of the maximum displacement and maximum stress, and finishing propagation analysis of uncertain parameters in a coupling system; introducing a volume method thought, and defining non-probabilistic reliability indexes of a pneumatic structure coupling system; taking wing structure weight as an optimization object and taking the reliability of the structural maximum displacement and maximum stress smaller than allowable values as a constraint condition, realizing the non-probabilistic reliability optimization design of the wing structure. According to the method disclosed by the invention, the wing structure mass is reduced under the condition that the hypersonic velocity wing is high in reliability, and the wing performance is improved.

Method and system for a direct transmission of motion path data (310) from a generating system (204) to a control system (205) for use in controlling a servo-driven machine. For example, a machine tool for machining stock material, and the data generating system includes a computer-aided-design software package (301). Tool paths for directing the tool members of the machine tool can be extracted from a design file opened by the CAD software through the operation interface of the CAD software (302). Alternatively, the design file can be used to specify motion paths rather than end product geometry. The CAD interpreter application extracts the motion path data from the design file, eliminating the need to translate the data into another or intermediate form for controlling the machine tool.

The invention discloses a sliding nozzle. A rotatable synchronous ring is arranged on an upper nozzle ring in a coaxial mode, and a synchronous ring inner circle connected with the upper nozzle ring in an assembling mode is of an interrupted structure in the radial direction. A first sliding block device comprises a flow guiding vane and an open shifting fork connected with the shaft end of the flow guiding vane. The shaft end of the flow guiding vane is provided with two parallel lateral planes. One end of the open shifting fork is connected with the two lateral planes in a matched mode. A second sliding block device comprises a main shifting fork assembly, one end of the main shifting fork assembly is connected with an actuator, and the other end of the main shifting fork assembly is connected with the synchronous ring through a synchronous pin. By means of the sliding nozzle, precision loss caused by fast abrasion of a line-plane friction pair in actual use is effectively avoided. Transmission precision and abrasion resistance are effectively improved. The inner circle is of an interrupted synchronous ring structure, matching area of the synchronous ring and the upper nozzle ring is reduced, and the phenomenon that the synchronous ring and the nozzle ring are clamped due to thermal deformation is effectively prevented.

An artificial intelligence guiding system for cornea crosslinking comprises the components of an ocular surface dispersed light image acquisition system which is configured for acquire a user ocular surface dispersed light image and sufficiently reflects cornea information; a cornea topographical map generating system which is pre-trained for converting the ocular surface dispersed light image toa corresponding cornea topographical map; a cornea crosslinking guiding system which is pre-trained for outputting a parameter required for a crosslinking surgery through comparison between a before-crosslinking cornea topographical map and an expected postoperative cornea topographical map; an ocular surface dispersed light image and corresponding cornea topographical map database which is used for storing an image pair of the ocular surface dispersed light image and the cornea topographical map and is used for pre-training a model and continuously correcting the model; and a cornea topographical map and cornea crosslinking parameter corresponding database which is used for recording the influence of the cornea crosslinking parameter to the cornea topographical map and performs a re-correcting function on the crosslinking guiding system. The artificial intelligence guiding system can greatly improve efficiency, accuracy and stability in a cornea crosslinking surgery process and reduces errors caused by subjective determining of a doctor.

The invention provides a single-main-shaft box one-time clamping double-end phase friction welder and belongs to the technical field of friction welders. A main shaft box is arranged on the upper portion of one side of a lathe bed. A tool locating clamping device is arranged on the upper portion of the other side of the lathe bed. A back pushing device is arranged on the upper portion of the tool locating clamping device. A main shaft is arranged in the main shaft box. A coding machine is arranged at the outer end of the main shaft. A direct-current motor component is arranged on the upper portion of the main shaft box. The direct-current motor component and the main shaft are in transmission connection through a synchronous pulley transmission mechanism. A rotating clamp is arranged at the inner end of the main shaft. During using, three workpieces are clamped at one time, accuracy loss caused by clamping again in secondary welding in a 180-degree-turning mode is avoided, welding phase accuracy is improved, and meanwhile welding production efficiency is improved. A direct-current motor is used for accurately controlling the angular velocity and the angular displacement of the main shaft, and accurate phase control of welding friction is achieved. The phase control accuracy can reach + / -0.5 degree.

The invention provides a GPS L2C signal tracking method. The unscented Kalman filter based method involves: adopting a CM code and a CL code for simultaneous tracking so as to reduce a carrier wave and code tracking threshold so that energy waste and precision loss are avoided, and arranging a weight coefficient and enabling the coefficient to be superior to any an individual CM or CL code tracking mode. The unscented Kalman filter can reduce the complexity of tracking calculating and has higher tracking sensitivity and precision. By using the method provided by the invention, not only can the tracking success rate of an L2C signal in a low signal-to-noise ratio environment be improved, the calculating complexity is also reduced, the calculating speed is faster, thus conditions are created for the application of a GPS L2C signal in a weak-signal condition, the potential economic and social values are high, and the method also takes important significance in the positioning application field of a GPS receiver in the low signal-to-noise ratio environment.

The invention discloses a compression method and a compression device of a deep neural network model, a terminal and a storage medium. The method comprises the following steps: obtaining a trained deep neural network model; quantifying parameters of each layer in the deep neural network model layer by layer, performing iterative quantification N times for the deep neural network model, until realizing that the quantified deep neural network model satisfies preset compression volume requirements, wherein, quantification treatment is executed for the i layer parameters in the trained deep neuralnetwork model as follows: performing clustering analysis for the i layer parameters in the trained deep neural network model, and determining corresponding quantification range and quantification level according to a clustering analysis result; and quantifying the parameters in the quantification range of the i layer according to the determined quantification level. By the scheme above, precisionand validity of the deep neural network model can be considered while the deep neural network model is compressed.