Search

CN-121981173-A - Multi-variable measurement sensor state lightweight evaluation method based on knowledge distillation

CN121981173ACN 121981173 ACN121981173 ACN 121981173ACN-121981173-A

Abstract

A multivariable measurement sensor state light-weight evaluation method based on knowledge distillation belongs to the technical field of electric digital data processing and multi-sensor data fusion. The method comprises the steps of firstly carrying out time synchronization and physical consistency constraint modeling on raw data of a plurality of sensors, extracting feature representation, then training a high-precision teacher model in a cloud, learning a high-dimensional mapping relation of sensor states, guiding light-weight chemical raw model training by extracting middle features, soft output and uncertainty information of the teacher model as distillation knowledge, combining soft label constraint, feature alignment and uncertainty guiding mechanisms to achieve effective migration of discrimination knowledge, and finally compressing, quantifying and optimizing the student model and then deploying the student model to a vehicle-mounted end to achieve real-time evaluation and dynamic update of the multi-sensor states. According to the invention, the knowledge distillation frame is used for guaranteeing the evaluation accuracy and greatly reducing the complexity of the model, so that efficient and reliable state sensing and fault-tolerant control support is provided for the intelligent driving system.

Inventors

  • CHEN YANFENG
  • LIU ZIHAN
  • WANG YAN
  • LIU HAOQIAN
  • BAI YITONG

Assignees

  • 辽宁大学

Dates

Publication Date
20260505
Application Date
20260409

Claims (10)

  1. 1. The method for evaluating the state of the multivariable measurement sensor based on knowledge distillation in a light-weight manner is characterized by comprising the following steps of: Step 1) modeling multi-source perception data acquisition and time synchronization: collecting multi-variable measurement data, performing time synchronization and alignment processing on different data sources, realizing time registration of multi-source data by adopting a unified time stamp and interpolation algorithm, obtaining a preprocessing data set with unified time sequence resolution through noise filtering, missing value compensation and scale normalization processing, and constructing multi-variable joint observation representation; Step 2) modeling the multivariate state feature construction and physical consistency constraint: Defining a multidimensional state feature space containing physical variables of position, speed, acceleration and posture based on a kinematic and dynamic model of an intelligent driving system, establishing a mapping relation between measured quantities of each sensor and the actual physical state of a vehicle, constructing a physical consistency error function, performing constraint optimization on multisource observation data, extracting common mode features meeting a physical coupling rule, and forming an informationized state feature vector containing time sequence information and physical constraint; step 3) sensor state evaluation labels and prior constraint construction: Constructing a state evaluation label for describing the running state of each sensor according to the characteristics of the multivariable state and the physical consistency deviation thereof, introducing historical statistical information and priori knowledge, and carrying out weighted correction on the state label to form a supervision label with data driving characteristics and physical constraints, thereby providing an objective function foundation for the subsequent neural network model training; Step 4), constructing a high-complexity teacher neural network model and performing sensor state discrimination training: Constructing and training a high-complexity teacher neural network model by taking multivariable state characteristics as input and state evaluation labels as supervision targets, and enabling the teacher model to learn a high-dimensional discrimination relation between multivariable measurement data and the health state, abnormal mode and credibility of the sensor through cascading and nonlinear mapping of the multi-layer characteristic extraction modules; Step 5) extracting middle characteristics and soft output distillation knowledge of the teacher model: Extracting middle layer characteristic representation, soft output probability distribution and state uncertainty information for knowledge distillation from a trained teacher model to form a structured distillation knowledge representation, regulating the smoothness of a soft label through temperature parameters, and enhancing knowledge migration efficiency among different state categories; step 6) designing and constraint modeling of a lightweight chemical neural network model structure: Constructing a lightweight chemical neural network model meeting the constraint of vehicle-mounted computing resources, wherein the number of network layers, the parameter scale and the computing complexity of the model are smaller than corresponding indexes of a teacher model, the parameter scale is not higher than the preset proportional upper limit of the parameter scale of the teacher model, the computing complexity is not higher than the preset proportional upper limit of the computing complexity of the teacher model, the model complexity is measured by adopting the sum of the non-zero parameter numbers in each network layer, and the student model is adapted to the vehicle-mounted real-time operation requirement through the constraint of the model complexity index and the reasoning delay threshold; Step 7) knowledge distillation combined training method based on multiple constraints: based on the extracted distillation knowledge representation, carrying out joint distillation training on the student model, introducing soft label constraint loss, intermediate feature alignment loss and uncertainty-guided hard supervision loss, and realizing effective migration of the teacher model discrimination capability to the student model through multi-objective optimization; Step 8) compression, quantification and vehicle-mounted reasoning optimization of the lightweight chemo-biological model: carrying out parameter importance assessment and low contribution parameter cutting on the trained student model, reducing model storage and calculation cost by adopting a fixed-point quantization or low-bit quantization technology, and ensuring real-time operation performance of the model on a vehicle-mounted platform through inference structure optimization; Step 9) real-time evaluation and dynamic update of the state of the vehicle-mounted terminal multivariable measurement sensor: The optimized student model is deployed to an intelligent driving automobile vehicle-mounted computing platform, the running state of the multivariable measurement sensor is evaluated on line in real time, a sliding window filtering and on-line self-adaptive updating mechanism is introduced, smooth output of a state evaluation result and dynamic adjustment of model parameters are achieved, and high-state input is provided for a perception fusion, fault-tolerant control and decision-making module.
  2. 2. The method for evaluating the state of a multivariate measurement sensor based on knowledge distillation according to claim 1, wherein in the step 1): In the intelligent driving system, the multi-type sensors are utilized to collect multi-variable measurement data, including laser Radar LiDAR, camera, millimeter wave Radar Radar and inertial measurement unit IMU, and the multi-variable measurement data are respectively recorded as: LiDAR point cloud data Image data Radar range-speed data IMU inertial data ; To achieve multi-sensor time synchronization and data alignment, uniform time stamping is adopted for different data sources Time interpolation operator And (3) completing time synchronization processing: Wherein, the The synchronization data representing the ith sensor at time t, A time offset representing the sensor relative to a system reference time; Stitching synchronized data of all sensors into a unified multivariate joint observation vector : T is the vector transpose symbol; by noise modeling and missing value compensation, a normalized multivariate observation representation is formed that can be used for neural network input.
  3. 3. The method for evaluating the state of a multivariate measurement sensor based on knowledge distillation according to claim 1, wherein in the step 2): Based on multivariate joint observation vectors Defining a feature mapping function Extracting original state features: is shown at the moment Constructing the obtained multivariable state feature vector; Representing a feature mapping function formed by normalization, differential operation, statistic extraction or nonlinear transformation; introducing a historical time window Constructing a time sequence expansion feature: The transposition of the original feature vector at the time t is carried out; representing a time-series state feature vector containing history information; establishing physical consistency constraint functions based on vehicle kinematic relationships Calculating a consistency residual: representing a bias vector between the multivariate state feature and the physical consistency constraint; representing a physical consistency mapping function formed by a vehicle kinematic relation, a sensor physical characteristic or a constraint rule among variables; Defining a physical consistency metric: The greater the value, the higher the degree of deviation of the measured state from the physical model, the metric being used to quantify the degree of deviation between the multivariate state feature and the physical model.
  4. 4. The method for evaluating the state of a multivariate measurement sensor based on knowledge distillation according to claim 1, wherein in the step 3): let the state variable of the ith sensor be In fact For a preset state set, through a monotonic mapping function Converting the physical consistency metric into a state confidence label: A state evaluation tag indicating that the i-th sensor is at time t; Representing a physical consistency metric or a characteristic deviation indicator associated with the sensor; representing a monotonic mapping function for converting the consistency deviation into a state confidence value; Introducing historical statistical prior probability And carrying out weighted correction on the state label: Constructing a multi-sensor joint state label vector: Indicating that the system is at time The multi-sensor state evaluation label vector of (2), N represents the number of sensors, will As a supervision target for subsequent teacher model training.
  5. 5. The method for the light-weighted evaluation of the state of a multivariate measurement sensor based on knowledge distillation of claim 1, wherein in said step 4): characterised by time sequence state For input, constructing a high-complexity teacher neural network model : Wherein the method comprises the steps of Model for teacher Training by supervising the loss function: further outputting a state probability distribution: the probability distribution of states is output for the teacher neural network model, Is an activation function that normalizes the original prediction to the [0,1] interval.
  6. 6. The method for evaluating the state of a multivariate measurement sensor based on knowledge distillation according to claim 1, wherein in the step 5): Selecting teacher models Extracting an intermediate feature set by the key hidden layers: Wherein the method comprises the steps of A key middle characteristic set of a teacher model at the moment t, Model No. for teacher Hidden layer feature vector of layer at time t, introducing temperature parameter Constructing a soft output probability distribution : Calculating output entropy as a measure of state uncertainty : The total number of categories for which the time series states are classified, And finally, forming a structured distillation knowledge representation for guiding the training of the student model.
  7. 7. The method for evaluating the state of a multivariate measurement sensor based on knowledge distillation according to claim 1, wherein in the step 6): construction of light-weight chemonerve network model The vehicle-mounted calculation constraint is satisfied: Wherein the method comprises the steps of Model for teacher Is a set of all learnable parameters; defining model complexity index : Wherein the method comprises the steps of For the number of layers of the student model, For the first layer weight parameter set of the student model, Representing a non-zero parameter number; introducing a computational complexity constraint: therefore, the student model meets the vehicle-mounted real-time reasoning requirement.
  8. 8. The method for evaluating the state of a multivariate measurement sensor based on knowledge distillation according to claim 1, wherein in the step 7): constructing a distillation training objective function based on multiple constraints: Wherein: Super parameter weight coefficient for balancing soft distillation loss, characteristic alignment loss, hard supervision loss, and soft output distillation loss : As a function of the degree of divergence, Is a teacher model soft probability distribution with temperature coefficient, Synchronizing soft probability distribution after temperature scaling for student model, feature alignment loss : Model number one for students The hidden layer feature vector of the layer at time t, Is the mth projection function, uncertainty weighted hard supervision loss : Wherein the method comprises the steps of As the weight coefficient of the light-emitting diode, Is an exponential function of the natural constant e, In order to reduce the coefficient of the coefficient, Is a measure of state uncertainty in step 5).
  9. 9. The method for evaluating the state of a multivariate measurement sensor based on knowledge distillation according to claim 1, wherein in the step 8): Importance assessment is carried out on student model parameters: scoring the importance of the j-th parameter of the first layer of the student model, For the jth learnable parameter of the first layer of the student model, the learning parameter is based on a threshold value Parameter cutting is carried out : Quantifying the parameters after clipping : Wherein the method comprises the steps of Quantization functions for fixed point or low bits; evaluating the calculation complexity of the model after optimization: the floating point operation times corresponding to the quantization parameter of the first layer are satisfied at the same time 。
  10. 10. The method for evaluating the state of a multivariate measurement sensor based on knowledge distillation according to claim 1, wherein in the step 9): deploying the optimized student model to a vehicle-mounted end, and carrying out real-time state evaluation: the original state evaluation result output by the vehicle-mounted end student model at the time t is output smoothly through sliding window filtering : Defining a state decision function: Wherein the method comprises the steps of As a result of the binary decision of the i-th class of state, For the state determination threshold value, To indicate the function, the condition output 1 is satisfied, otherwise, the output 0 is output, and an online self-adaptive update mechanism is introduced: Wherein the method comprises the steps of In order for the rate of on-line learning, Gradient of hard supervision loss to student model parameters at the present moment, Is the current supervision loss.

Description

Multi-variable measurement sensor state lightweight evaluation method based on knowledge distillation Technical Field The invention belongs to the technical field of electric digital data processing and multi-sensor data fusion, and particularly relates to a multi-variable measurement sensor state light-weight evaluation method based on knowledge distillation, which relates to the technologies of time synchronization, physical consistency modeling, teacher-student model collaborative training, knowledge distillation migration, model compression optimization, vehicle-mounted real-time deployment and the like of multi-source heterogeneous sensor data, and is suitable for the scenes of multi-sensor health state monitoring, anomaly detection, reliability evaluation, fault-tolerant control and the like in intelligent driving automobiles. Background Along with the development of intelligent driving technology, the vehicle sensing system integrates various types of sensors such as a laser Radar (LiDAR), a Camera (Camera), a millimeter wave Radar (Radar), an Inertial Measurement Unit (IMU) and the like, and is used for acquiring environmental information and vehicle states in real time. The sensors have the characteristics of sampling frequency difference, time dissynchrony, strong data heterogeneity and the like in multivariate measurement, so that the sensor state evaluation faces the problems of high model complexity, large calculation resource requirement, difficult guarantee of real-time performance and the like. The current sensor state evaluation method mainly comprises residual analysis based on a physical model, anomaly detection based on statistics, end-to-end state recognition based on deep learning and the like. Although the deep learning method can effectively learn complex nonlinear relations and improve state evaluation accuracy, the model parameters are large, reasoning delay is high, and the method is difficult to directly deploy to a vehicle-mounted computing platform with limited resources. Although the traditional lightweight method such as network pruning, quantification and the like can reduce the model scale, the requirement of intelligent driving on high-reliability state perception is difficult to meet at the expense of evaluation precision. Therefore, how to realize the light deployment and real-time operation of the complex neural network model at the vehicle-mounted end on the premise of ensuring the state evaluation precision and the robustness of the sensor becomes a key problem to be solved urgently for the intelligent driving system. Aiming at the problems, the invention provides a multivariable measurement sensor state light-weight evaluation method based on knowledge distillation, which transfers the discrimination knowledge of a high-complexity teacher model to a light-weight chemical model by constructing a knowledge distillation frame of 'teacher-student' cooperation, and remarkably reduces the model scale and calculation cost while maintaining the evaluation precision, thereby realizing the efficient deployment and real-time operation of a complex neural network on an intelligent driving automobile. Disclosure of Invention The invention aims to solve the problems of complex sensor state evaluation model, heavy calculation load, difficult vehicle-mounted deployment and the like in the existing intelligent driving system, and provides a multivariable measurement sensor state lightweight evaluation method based on knowledge distillation. The technical scheme of the invention is that the method for evaluating the state of the multivariable measurement sensor based on knowledge distillation comprises the following steps: Step 1) modeling multi-source perception data acquisition and time synchronization: And collecting multi-variable measurement data, performing time synchronization and alignment processing on different data sources, realizing time registration of multi-source data by adopting a unified time stamp and interpolation algorithm, obtaining a preprocessing data set with unified time sequence resolution through noise filtering, missing value compensation and scale normalization processing, and constructing multi-variable joint observation representation. In the step 1): In the intelligent driving system, the multi-type sensors are utilized to collect multi-variable measurement data, including laser Radar LiDAR, camera, millimeter wave Radar Radar and inertial measurement unit IMU, and the multi-variable measurement data are respectively recorded as: LiDAR point cloud data Image dataRadar range-speed dataIMU inertial data; To achieve multi-sensor time synchronization and data alignment, uniform time stamping is adopted for different data sourcesTime interpolation operatorAnd (3) completing time synchronization processing: Wherein, the The synchronization data representing the ith sensor at time t,A time offset representing the sensor relative to a system reference time; St