An IMU perception resource runtime performance evaluation method based on space-time prediction

Abstract

The application discloses an IMU sensing resource runtime performance evaluation method based on space-time prediction, which comprises collecting IMU sensor data; preprocessing the IMU sensor data; constructing a hybrid space-time prediction model composed of a long short-term memory network and a convolutional autoencoder; predicting the IMU sensing resource demand and performance at future time; dynamically monitoring the resource usage and performance indicators of the system; according to the real-time performance evaluation results, optimizing the allocation and use strategy of the IMU sensing resource by dynamically adjusting the resource scheduling scheme; applying optimization algorithms to ensure efficient use of resources and avoid resource waste and performance bottlenecks; outputting the final evaluation results and optimization suggestions to guide the subsequent operation of the system and the management of the IMU sensing resource. The application combines space-time prediction with IMU sensing resource scheduling, significantly improves the operation performance of the IMU sensing resource, and is suitable for application requirements in complex dynamic environments.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of sensors, and specifically relates to an IMU perception resource runtime performance evaluation method based on space-time prediction. BACKGROUND

[0002] According to the existing technical background, IMU (Inertial Measurement Unit) is widely used in navigation, robots and unmanned aerial vehicles and other fields. IMU sensors measure the motion state and attitude of the device through accelerometers, gyroscopes and magnetometers. However, with the complication of application scenarios, the existing technology faces many challenges in the runtime performance evaluation of IMU perception resources.

[0003] Firstly, the traditional IMU data processing method mainly focuses on static data analysis, which cannot effectively cope with complex and dynamic environmental changes. This leads to low utilization rate and accuracy of performance evaluation of IMU perception resources in real-time applications. Traditional methods usually rely on fixed model parameters, lack flexible adjustment mechanism, and cannot dynamically update according to actual scene, thereby affecting the overall system performance.

[0004] Secondly, existing space-time prediction techniques are mostly based on single data source or simple linear model, failing to fully utilize the potential information of multi-sensor data. This method ignores the spatio-temporal correlation between sensor data, resulting in limited accuracy and reliability of prediction results. Especially in complex environments, the data of a single sensor is easily affected by noise and outliers, affecting the effect of the prediction model.

[0005] In addition, in terms of IMU perception resource optimization and scheduling, traditional methods usually adopt static or predefined IMU perception resource allocation strategies, which cannot respond to actual IMU perception resource demand fluctuations in time. This static allocation method may lead to waste of IMU perception resources or performance bottleneck, reducing the overall performance of the system. The existing technology lacks accurate prediction and real-time adjustment mechanism for IMU perception resource demand, and cannot effectively solve the problem of low utilization rate of IMU perception resources.

[0006] At the same time, the fusion and processing of multi-sensor data is also a big difficulty in the existing technology. Due to the diversity of sensor types and different data characteristics, how to extract effective information from multi-sensor data, reduce noise interference and improve the accuracy of space-time prediction has become a big challenge. Existing methods usually simply average process multi-sensor data, failing to fully utilize the complementarity between sensors.

[0007] The above defects can lead to inefficient use of IMU sensing resources in practical applications, not only wasting system resources, but also possibly causing insufficient performance of the system at critical moments. Therefore, an innovative method is urgently needed to improve the overall operating efficiency of IMU sensing resources by combining spatio-temporal prediction, real-time performance evaluation, resource optimization, and multi-sensor data fusion to meet the application requirements in complex dynamic environments. SUMMARY

[0008] In order to overcome the shortcomings of the prior art, the present application provides a spatio-temporal prediction-based IMU sensing resource runtime performance evaluation method, which comprises collecting IMU sensor data; preprocessing the IMU sensor data; constructing a hybrid spatio-temporal prediction model composed of a long short-term memory network and a convolutional autoencoder; predicting the IMU sensing resource demand and performance at future time; dynamically monitoring the resource usage and performance indicators of the system; optimizing the allocation and use strategy of the IMU sensing resource by dynamically adjusting the resource scheduling scheme according to the real-time performance evaluation result; applying optimization algorithms to ensure efficient use of resources and avoid resource waste and performance bottlenecks; outputting the final evaluation result and optimization suggestions for guiding the subsequent operation of the system and the management of the IMU sensing resource. The present application significantly improves the operating efficiency of the IMU sensing resource by combining spatio-temporal prediction with IMU sensing resource scheduling, and is suitable for application requirements in complex dynamic environments.

[0009] The technical solution adopted by the present application to solve its technical problems is as follows:

[0010] Step 1: Collecting IMU sensor data, the IMU sensor including an accelerometer, a gyroscope and a magnetometer;

[0011] Step 2: Data cleaning, filtering and noise removal are performed on the IMU sensor data;

[0012] Step 3: A hybrid spatio-temporal prediction model composed of a long short-term memory network and a convolutional autoencoder is constructed based on historical IMU sensor data and real-time IMU sensor data, and a transfer learning algorithm based on attention mechanism is used to train the hybrid spatio-temporal prediction model;

[0013] Step 4: The hybrid spatio-temporal prediction model is used to predict the IMU sensing resource demand and performance at future time, and the spatio-temporal prediction result is obtained;

[0014] Step 5: During operation, the performance of the IMU sensing resource is evaluated in real time based on the spatio-temporal prediction result, the resource usage and performance indicators of the system are dynamically monitored, and the real-time performance evaluation result is obtained;

[0015] Step 6: According to the real-time performance evaluation result, the allocation and use strategy of the IMU sensing resource is optimized by dynamically adjusting the resource scheduling scheme;

[0016] Step 7: Output the final evaluation results and optimization suggestions.

[0017] Further, the step 1 is as follows:

[0018] Step 1-1: Obtain IMU sensor data, including accelerometer, gyroscope and magnetometer data, collected at fixed time intervals;

[0019] Step 1-2: Collect accelerometer data A(t) and record linear acceleration in three-axis direction ; respectively represent the acceleration components on X, Y, Z axes;

[0020] Step 1-3: Collect gyroscope data G(t) and record angular velocity in three-axis direction ; respectively represent the angular velocity components on X, Y, Z axes;

[0021] Step 1-4: Collect magnetometer data M(t) and record magnetic field intensity in three-axis direction , respectively represent the angular velocity components on X, Y, Z axes.

[0022] Further, the step 2 is as follows:

[0023] Step 2-1: Data cleaning of accelerometer data A(t), gyroscope data G(t) and magnetometer data M(t) obtained by IMU sensor, using statistical analysis-based anomaly detection algorithm to remove outliers and missing data;

[0024] Step 2-2: Apply adaptive Kalman filter to filtered accelerometer data A(t), gyroscope data G(t) and magnetometer data M(t), and dynamically adjust the filter gain combined with real-time environmental variables;

[0025] Step 2-3: Multi-sensor data fusion of filtered accelerometer data A(t), gyroscope data G(t) and magnetometer data M(t), using self-attention mechanism to enhance signal features, and the fused data is:

[0026] ;

[0027] ;

[0028] wherein, are weights based on attention mechanism; represents the fused data; , 、 Represent the filtered accelerometer, gyroscope, and magnetometer data respectively; represent query vector, key vector and value vector respectively, represents the attention score calculated by the self-attention mechanism, represents the exponential function;

[0029] Step 2-4: Fusion data For denoising, an adaptive denoising algorithm based on a deep convolutional neural network is applied to extract multi-scale features and obtain the final feature representation after convolution and activation operations:

[0030] ;

[0031] in, (.) is the convolution operation, is the convolutional layer weight; represents the final feature representation after convolution and activation operations, represents the total number of convolutional layers, Indicates the layer of convolutional layers, represents the convolution operation, represents the activation function;

[0032] Steps 2-5: Output IMU sensor data after data cleaning, filtering, data fusion and denoising , calculated by multi-scale feature enhancement method:

[0033] ;

[0034] in, and Represent the multi-scale analysis function and feature extraction function respectively; Indicates the number of dimensions for multi-scale analysis, Indicates the mth dimension of multi-scale analysis.

[0035] Furthermore, the step 3 is specifically as follows:

[0036] Step 3-1: Preprocessed IMU sensor data The input is fed into the enhanced long short-term memory network, and the temporal features and spatial correlation are extracted by combining the spatiotemporal gating unit. The output sequence feature H(t) is expressed as:

[0037] ;

[0038] in, , , denotes a weight matrix, respectively, for the current input, the hidden state at the previous time, and the output of the spatio-temporal gating unit, denotes an activation function, denotes a bias vector; STGU(S(t)) denotes a spatio-temporal gating unit, and the calculation formula is as follows:

[0039] ;

[0040] wherein, denotes a weight matrix of H(t), denotes a bias vector, denotes a hyperbolic tangent activation function;

[0041] The enhanced long short-term memory network refers to introducing a spatio-temporal gating unit STGU in a long short-term memory network;

[0042] Step 3-2: input H(t) into a multi-scale convolutional autoencoder to capture features of different scales and generate a compressed representation C(t):

[0043] ;

[0044] wherein, denotes a convolution operation with different convolution kernel sizes , and denotes a pooling operation for reducing the feature dimension;

[0045] Step 3-3: combine a dynamic attention mechanism to adjust the attention weight according to the historical context :

[0046] ;

[0047] wherein, is a dynamic adjustment function, which denotes the dot product of the query vector Q and the key vector K; R(t) denotes a time-dependent bias, which captures the historical context information; denotes the dimension size of K;

[0048] Step 3-4: apply the attention weight to the compressed representation to generate a spatio-temporal feature representation :

[0049] ;

[0050] wherein, denotes a balance parameter for fusing the original feature and the weighted feature;

[0051] Step 3-5: establish a hybrid spatio-temporal prediction model M and introduce a multi-objective optimization loss function:

[0052] ;

[0053] in, represents the multi-objective optimization loss function, represents the predicted value of the model output, Y(t) represents the actual observed value, 、 、 Represents the loss function balance parameters, corresponding to the error term, L2 regularization and L1 regularization respectively; Represents the L2 and L1 norm regularization terms of the model weights respectively; represents the total number of time steps, Indicates the number of weights involved in L2 regularization, Indicates the number of weights involved in L1 regularization;

[0054] Step 3-6: Output the optimized hybrid spatiotemporal prediction model M.

[0055] Furthermore, the step 4 is specifically as follows:

[0056] Step 4-1: Representing spatiotemporal features Perform full connection layer transformation and output the potential feature representation:

[0057] ;

[0058] in, is the weight matrix, is the bias vector; represents the latent feature representation;

[0059] Step 4-2: Input the potential feature Z(t) into the prediction module, which is divided into IMU perception resource demand prediction and performance prediction :

[0060] ;

[0061] ;

[0062] in, and is the weight matrix of each module, and is bias; 、 They represent the amount of resources required at future times and the performance at future times, respectively. 、 They represent mapping functions respectively;

[0063] The prediction module is constructed on the basis of a hybrid space-time prediction model plus a fully connected layer;

[0064] Step 4-3: Introduce an adaptive learning rate optimization algorithm to minimize the loss function:

[0065] ;

[0066] Where R(t) and P(t) are the actual IMU perception resource demand and performance, respectively; and is a weight parameter for balancing the prediction error;

[0067] Step 4-4: Optimize model parameters through iterative training to ensure the prediction accuracy of model M, and finally output the predicted IMU perception resource demand and performance .

[0068] Further, the step 5 is specifically as follows:

[0069] Step 5-1: During operation, real-time collection of current IMU perception resource usage data and performance index data ;

[0070] Step 5-2: Combine the predicted IMU perception resource demand and performance , to construct a real-time monitoring matrix M(t):

[0071] ;

[0072] Where, represents the resource utilization rate:

[0073] ;

[0074] Where, is the resource fluctuation coefficient, is the fluctuation frequency, is the phase shift; represents the actual resource amount used at time t, represents the maximum available resource amount at time t;

[0075] represents the IMU perception resource demand error, plus the time derivative term for capturing the rate of change:

[0076]

[0077] is the performance error, is the variance influence coefficient; represents the actual performance value; represents the variance of performance;

[0078] Step 5-3: Using the adaptive weight distribution algorithm, analyze the real-time monitoring matrix, and the weight adjustment formula is:

[0079] ;

[0080] wherein, and are adaptive adjustment parameters, which are dynamically updated according to historical data; is the weight adjustment coefficient; represents the dynamic weight, represents the historical value of the real-time monitoring matrix;

[0081] Step 5-4: According to the real-time monitoring matrix and the weight, generate the performance evaluation report :

[0082] ;

[0083] wherein, is the time influence coefficient, used to adjust the historical weight influence; is the historical time point;

[0084] Step 5-5: According to the result of the performance evaluation report, trigger the intelligent optimization module to automatically adjust the system resource configuration;

[0085] Step 5-6: Output the performance evaluation report and the result of step 5-5 to the user interface.

[0086] Further, the step 7 is specifically as follows:

[0087] Step 7-1: According to the performance evaluation report determine the IMU sensing resource utilization rate and the bottleneck index, and define the optimization objective function:

[0088] ;

[0089] wherein, , , is the balance coefficient, adjusting the weight of different objectives;

[0090] Step 7-2: Apply the particle swarm optimization algorithm to initialize the particle swarm, and each particle represents an IMU sensing resource configuration scheme, and the particle position is represented as:

[0091] ;

[0092] wherein, The allocated amount of the jth IMU sensing resource is denoted as:

[0093] ;

[0094] wherein, denotes the predicted demand amount of the jth resource at time t, denotes the predicted demand amount of the kth resource at time t;

[0095] Step 7-3: Calculate the position and velocity update of each particle:

[0096] ;

[0097] ;

[0098] wherein, is an inertia weight, adjusting the influence of the previous velocity of the particle, , is a learning factor, affecting the dependence of the particle on its own experience and the group experience; , is a random number between 0 and 1, increasing the randomness of the particle search; is the historical best position of the particle, is the global best position, is a derivative coefficient, combining the gradient information of the optimization objective function; denotes the velocity of particle i at time t+1, denotes the velocity of particle i at time t, denotes the position of particle i at time t, denotes the objective function value corresponding to each particle, denotes the position of particle i at time t+1;

[0099] Step 7-4: According to the updated particle position, calculate the optimization objective function value, and update the historical optimal position of the particle and the global optimal position :

[0100] ;

[0101] ;

[0102] Step 7-5: Determine whether the stopping condition is met, if met, output the optimal IMU sensing resource configuration scheme ; otherwise, return to step 7-3 to continue iteration:

[0103] ;

[0104] wherein, is the convergence threshold, ensuring that the optimization process reaches stability; represents the objective function value corresponding to the global optimal position found in the previous time t-1 system iteration process;

[0105] Step 7-6: Apply the optimal IMU sensing resource configuration scheme , dynamically adjust the system IMU sensing resource allocation, ensure the utilization and performance optimization of IMU sensing resource:

[0106] ;

[0107] wherein, is the jth item of IMU sensing resource allocation after optimization;

[0108] Step 7-7: Output the results of step 7-6 to the user interface, including the optimal IMU sensing resource configuration and error analysis.

[0109] The beneficial effects of the present application are as follows:

[0110] (1) The present application constructs a hybrid spatio-temporal prediction model, combines long short-term memory network and convolutional autoencoder, fully utilizes historical and real-time IMU sensor data, improves the accuracy of IMU sensing resource demand and performance prediction, based on real-time efficiency evaluation of spatio-temporal prediction results, dynamically monitors system IMU sensing resource usage and performance indicators, so that the system can respond to environmental changes in time, and improves the overall adaptive ability.

[0111] (2) The present application uses particle swarm optimization algorithm combined with adaptive learning rate to adjust the IMU sensing resource configuration scheme in real time, ensures the efficient use of IMU sensing resource, reduces the waste of IMU sensing resource, avoids performance bottleneck, uses adaptive Kalman filter and multi-sensor data fusion method based on attention mechanism, reduces the error and uncertainty of single sensor data, and improves the reliability of spatio-temporal prediction.

[0112] (3) The present application realizes intelligent adjustment of system resources through analysis of real-time monitoring matrix and triggering of intelligent optimization module, significantly improves the overall performance of IMU sensing resource, provides user interface of optimization results and adjustment suggestions, so that users can intuitively understand the improvement of system performance and IMU sensing resource configuration, and facilitate subsequent management and optimization. BRIEF DESCRIPTION OF DRAWINGS

[0113] Figure 1 is the flow chart of the method of the present application. DETAILED DESCRIPTION

[0114] The present application will be further described below in conjunction with the drawings and examples.

[0115] One object of the present application is to provide a spatiotemporal prediction-based IMU sensing resource runtime performance evaluation method. The present application combines spatiotemporal prediction with IMU sensing resource scheduling, significantly improving the operational performance of IMU sensing resources, and is suitable for application requirements in complex dynamic environments.

[0116] Reference Figure 1 A spatiotemporal prediction-based IMU sensing resource runtime performance evaluation method is applied to a complex human behavior recognition scenario. In this scenario, multiple intelligent devices (such as wearable devices, smartphones, fitness devices, etc.) are equipped with IMU sensors to monitor real-time human posture, motion, and location parameters. Due to the diversity of actions and environmental interference, traditional methods have many shortcomings in real-time IMU sensing resource scheduling and performance evaluation.

[0117] In a human behavior recognition task, multiple intelligent devices are deployed to collect user behavior data and transmit it to a central processing system. These intelligent devices are equipped with IMU sensors, including accelerometers, gyroscopes, and magnetometers. The recognition scenario has complex and varied action characteristics, which puts high demands on the accuracy and real-time performance of device data collection. Under traditional methods, due to the lack of response capability to real-time action changes, it is easy to cause IMU sensing resource waste and performance bottlenecks.

[0118] In the method of the present application, intelligent devices collect data in real time through IMU sensors, including linear acceleration, angular velocity, and magnetic field strength. Using adaptive Kalman filters and self-attention mechanisms, multi-sensor data is fused to eliminate noise and outliers, improving data accuracy and reliability. Based on historical and real-time data, a hybrid spatiotemporal prediction model is constructed to accurately predict future IMU sensing resource demand and performance.

[0119] During the task, the particle swarm optimization algorithm is used to adjust the IMU sensing resource configuration of intelligent devices in real time to respond to rapid changes in actions. The model dynamically monitors IMU sensing resource utilization and performance indicators based on real-time performance evaluation results to ensure stable performance of intelligent devices in complex action scenarios.

[0120] In the behavior recognition task, intelligent devices not only effectively respond to complex action changes, but also maintain efficient resource utilization and stable performance throughout the task duration. Traditional methods are prone to improper IMU sensing resource scheduling and performance bottlenecks when action changes are large, while the present method effectively avoids these problems through real-time prediction and optimization, improving the overall performance of the behavior recognition task.

[0121] Overall, the method of the application makes full use of the sensing ability of the IMU sensor in a complex dynamic environment, realizes efficient use of IMU sensing resources and significant improvement of system performance through space-time prediction, performance evaluation and intelligent optimization, and provides a solid technical guarantee for human behavior recognition tasks.

[0122] In summary, the application combines space-time prediction with IMU sensing resource scheduling to significantly improve the operating efficiency of IMU sensing resources and is suitable for application requirements in a complex dynamic environment.

Claims

1. A method for evaluating the runtime performance of IMU perception resources based on spatiotemporal prediction, characterized in that: The steps include: Step 1: Collect IMU sensor data, which includes an accelerometer, gyroscope, and magnetometer; Step 2: Clean, filter and remove noise from the IMU sensor data; Step 3: Based on historical IMU sensor data and real-time IMU sensor data, a hybrid spatiotemporal prediction model consisting of a long short-term memory network and a convolutional autoencoder is constructed. The hybrid spatiotemporal prediction model is trained using a transfer learning algorithm based on an attention mechanism. Step 4: Use the constructed hybrid spatiotemporal prediction model to predict the IMU perception resource requirements and performance at future moments and obtain spatiotemporal prediction results; Step 5: During operation, the efficiency of IMU sensing resources is evaluated in real time based on the spatiotemporal prediction results. The resource usage and performance indicators of the system are dynamically monitored to obtain real-time efficiency evaluation results. Step 5-1: During operation, collect the current IMU perception resource usage data R in real time through the sensor network a (t) and performance index data P a (t); Step 5-2: Combine predicted IMU sensing resource requirements and performance Construct real-time monitoring matrix M(t): Among them, U(t) represents the resource utilization rate: Among them, λ is the resource fluctuation coefficient, ω is the fluctuation frequency, is the phase shift; R a (t) represents the actual amount of resources used at time t, R m (t) represents the maximum available resources at time t; E R (t) represents the IMU perception resource requirement error, plus the time derivative term to capture the rate of change: is the performance error, γ1 is the variance influence coefficient; P a (t) represents the actual performance value; Var(.) represents the variance of the performance; Step 5-3: Use the adaptive weight allocation algorithm to analyze the real-time monitoring matrix. The weight adjustment formula is: Among them, α2 and β2 are adaptive adjustment parameters, which are dynamically updated according to historical data; k is the weight adjustment coefficient; W(t) represents the dynamic weight, and M(τ) represents the historical value of the real-time monitoring matrix; Step 5-4: Generate the performance evaluation report E(t) based on the real-time monitoring matrix and weights: Among them, θ is the time influence coefficient, which is used to adjust the historical weight influence; τ i For historical time points; Step 5-5: Based on the results of the performance evaluation report, trigger the intelligent optimization module to automatically adjust the system resource configuration; Step 5-6: Output the performance evaluation report and the result of step 5-5 to the user interface; Step 6: Based on the real-time performance evaluation results, optimize the allocation and usage strategy of IMU sensing resources by dynamically adjusting the resource scheduling plan; Step 7: Output the final evaluation results and optimization suggestions.

2. The method for evaluating the runtime performance of IMU sensing resources based on spatiotemporal prediction according to claim 1, characterized in that: The step 1 is specifically as follows: Step 1-1: Obtain IMU sensor data, including accelerometer, gyroscope, and magnetometer data, which are collected at fixed time intervals; Step 1-2: Collect accelerometer data A(t) and record the linear acceleration (a x (t), a y (t), a z (t)); a x (t), a y (t), a z (t) represents the acceleration components on the X, Y, and Z axes respectively; Step 1-3: Collect gyroscope data G(t) and record the angular velocity (ω) in the three-axis direction x (t),ω y (t),ω z (t));ω x (t),ω y (t),ω z (t) represents the angular velocity components on the X, Y, and Z axes respectively; Step 1-4: Collect the magnetometer data M(t) and record the magnetic field strength (m x (t), m y (t), m z (t)),ω x (t),ω y (t),ω z (t) represents the angular velocity components on the X, Y, and Z axes, respectively.

3. The method for evaluating the runtime performance of IMU sensing resources based on spatiotemporal prediction according to claim 2, characterized in that: The step 2 is specifically as follows: Step 2-1: Clean the accelerometer data A(t), gyroscope data G(t), and magnetometer data M(t) obtained by the IMU sensor and use an anomaly detection algorithm based on statistical analysis to remove outliers and missing data. Step 2-2: Apply an adaptive Kalman filter to the cleaned accelerometer data A(t), gyroscope data G(t), and magnetometer data M(t), and dynamically adjust the filter gain based on real-time environmental variables; Step 2-3: Perform multi-sensor data fusion on the filtered accelerometer data A(t), gyroscope data G(t), and magnetometer data M(t), using the self-attention mechanism to enhance signal features. The fused data is: D f (t)=αA f (t)+βG f (t)+γM f (t); Among them, α, β, and γ are weights based on the attention mechanism; D f (t) represents the fused data; A f (t), G f (t), M f (t) represents the filtered accelerometer, gyroscope, and magnetometer data, respectively; Q, K, V represent the query vector, key vector, and value vector, respectively; Attention(Q, K, V) represents the attention score calculated by the self-attention mechanism; exp(.) represents the exponential function; Step 2-4: Fusion data D f (t) Denoising: Apply an adaptive denoising algorithm based on a deep convolutional neural network to extract multi-scale features and obtain the final feature representation after convolution and activation operations: Among them, Conv(.) is the convolution operation, w l is the convolution layer weight; D c (t) represents the final feature representation after convolution and activation operations, L represents the total number of convolution layers, l represents the lth convolution layer, Conv l (.) represents the convolution operation, and ReLU(.) represents the activation function; Step 2-5: Output the IMU sensor data D after data cleaning, filtering, data fusion and denoising e (t), calculated by multi-scale feature enhancement method: Among them, Scale m (.) and Feature m (.) represent the multi-scale analysis function and feature extraction function respectively; M represents the number of dimensions of multi-scale analysis, and m represents the m-th dimension of multi-scale analysis.

4. The method for evaluating the runtime performance of IMU sensing resources based on spatiotemporal prediction according to claim 3, characterized in that: The step 3 is specifically as follows: Step 3-1: The pre-processed IMU sensor data D e (t) is input into the enhanced long short-term memory network, combined with the spatiotemporal gating unit to extract temporal features and spatial correlation, and the output sequence feature H(t) is expressed as: H(t)=σ(W h ·D e (t)+U h ·H(t-1)+V h ·STGU(S(t))+b h ); Among them, W h ,U h ,V h represents the weight matrix, which is used for the current input, the hidden state of the previous moment, and the output of the spatiotemporal gating unit, σ(.) represents the activation function, b h represents the bias vector; STGU(S(t)) represents the spatiotemporal gating unit, and the calculation formula is as follows: STGU(S(t))=tham(W s ·D e (t)+b s ); Among them, W s Indicates D e The weight matrix of (t), b s represents the bias vector, tanh(.) represents the hyperbolic tangent activation function; The enhanced long short-term memory network refers to the introduction of a spatiotemporal gating unit STGU into the long short-term memory network; Step 3-2: Input H(t) into the multi-scale convolutional autoencoder to capture features of different scales and generate a compressed representation C(t): in, Indicates different convolution kernel sizes k i Convolution operation, Pooling (.) represents the pooling operation, which is used to reduce the feature dimension; Step 3-3: Combine the dynamic attention mechanism and adjust the attention weight W according to the historical context a : in, is a dynamic adjustment function, which represents the calculation of the dot product of the query vector Q and the key vector K; R(t) represents the time-related bias, which captures the historical context information; d k Indicates the dimension size of K; Step 3-4: Apply attention weights W to the compressed representation C(t) a , generate spatiotemporal feature representation C′(t): C'(t)=W a ·C(t)+β1·tanh(C(t)); Among them, β1 represents the balance parameter, which is used to fuse the original features and weighted features; Step 3-5: Establish a hybrid spatiotemporal prediction model M and introduce a multi-objective optimization loss function: in, Denotes the multi-objective optimization loss function, M(D e (t)) represents the predicted value of the model output, Y(t) represents the actual observed value, λ1, λ2, and λ3 represent the loss function balance parameters, corresponding to the error term, L2 regularization, and L1 regularization, respectively; ||W i || 2 ,||V j || 1 Represents the L2 and L1 norm regularization terms of the model weights respectively; T represents the total number of time steps, n represents the number of weights participating in L2 regularization, and m1 represents the number of weights participating in L1 regularization; Step 3-6: Output the optimized hybrid spatiotemporal prediction model M.

5. The method for evaluating the runtime performance of IMU sensing resources based on spatiotemporal prediction according to claim 4, characterized in that: The step 4 is specifically as follows: Step 4-1: Perform a fully connected layer transformation on the spatiotemporal feature representation C′(t), and output it as a potential feature representation: Z(t)=W z ·C′(t)+b xz ; Among them, W z is the weight matrix, b z is the bias vector; Z(t) represents the potential feature representation; Step 4-2: Input the potential feature Z(t) into the prediction module, which is divided into IMU perception resource demand prediction M R and performance prediction M P : Among them, W r and W p is the weight matrix of each module, b r and b p is bias; They represent the amount of resources required at the future time and the performance at the future time, respectively, R (.), f P (.) represent mapping functions respectively; The prediction module is constructed by adding a fully connected layer on the basis of the hybrid spatiotemporal prediction model; Step 4-3: Introduce the adaptive learning rate optimization algorithm to minimize the loss function: Among them, R(t) and P(t) are the actual IMU perception resource requirements and performance respectively; λ3 and λ4 are weight parameters used to balance the prediction error; Step 4-4: Optimize model parameters through iterative training to ensure the prediction accuracy of model M, and finally output the predicted IMU perception resource requirements and performance 6. The method for evaluating the runtime performance of IMU sensing resources based on spatiotemporal prediction according to claim 5, characterized in that: The step 7 is specifically as follows: Step 7-1: Determine the IMU perception resource utilization and bottleneck indicators based on the performance evaluation report E(t), and define the optimization objective function: Among them, α3, β3, and γ3 are balance coefficients, which adjust the weights of different objectives; Step 7-2: Apply the particle swarm optimization algorithm to initialize the particle swarm. Each particle represents an IMU sensing resource configuration scheme, and the particle position is expressed as: X i (t)=[r1,r2,...r j ,...,r n ]; Among them, r j is the allocation of the j-th IMU perception resource, expressed as: in, represents the predicted demand for the jth resource at time t, represents the predicted demand for the kth resource at time t; Step 7-3: Calculate the position and velocity update of each particle: Among them, ω1 is the inertia weight, which adjusts the influence of the particle's previous speed; c1 and c2 are learning factors, which affect the particle's dependence on its own experience and group experience; rand1 and rand2 are random numbers between [0,1], which increase the randomness of the particle search; P best，i is the best historical position of the particle, G best is the global optimal position, ξ is the derivative coefficient, combined with the gradient information of the optimization objective function; V i (t+1) represents the velocity of particle i at time t+1, V i (t) represents the velocity of particle i at time t, X i (t) represents the position of particle i at time t, Represents the objective function value corresponding to each particle, X i (t+1) represents the position of particle i at time t+1; Step 7-4: Calculate the optimization objective function value based on the updated particle position and update the particle's historical optimal position P best，i (t) and the global optimal position G best (t): G best (t)=argmin{P best,i (t)}; Step 7-5: Determine whether the stopping condition is met. If so, output the optimal IMU perception resource configuration plan. Otherwise, return to step 7-3 and continue iterating: Stop condition: |G best (t)-G best (t-1)|<∈; Among them, ∈ is the convergence threshold, which ensures that the optimization process reaches stability; G best (t-1) represents the objective function value corresponding to the global optimal position found during the system iteration at the previous time t-1; Step 7-6: Apply the optimal IMU perception resource configuration solution Dynamically adjust the system IMU perception resource allocation to ensure IMU perception resource utilization and performance optimization: in, is the optimized j-th IMU perception resource allocation; Step 7-7: Output the results of step 7-6 to the user interface, including the optimal IMU perception resource configuration and error analysis.

Images