Neural network model parameter compression method for fitness equipment
By employing a multi-module collaborative optimization method, the limitations of the neural network model of fitness equipment in handling the collaborative relationship between human joints and equipment trajectories, as well as the issues of user adaptability, were resolved. This approach enabled efficient motion recognition and storage optimization, thereby enhancing the intelligence level of fitness equipment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-31
- Publication Date
- 2026-03-27
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing neural network models for fitness equipment are one-sided in handling the collaborative relationship between human joints and equipment trajectories, cannot adapt to individual differences among different users, and are prone to losing key features during pruning, resulting in a decrease in motion recognition accuracy.
A multi-module collaborative optimization method is adopted, including data preprocessing, model pruning, parameter quantization and encoding storage. Through multi-source data analysis and dynamic fault tolerance threshold of Gaussian mixture model, adaptive adjustment is made in combination with user historical data. Unimportant parameters are pruned, and sparse matrix encoding and hierarchical storage are used to optimize the model.
It improves the accuracy and robustness of fitness movement recognition, reduces computation and storage requirements, and provides more accurate movement feedback and a smoother user experience.
Smart Images
Figure CN121745174A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer technology and relates to a method for compressing neural network model parameters for fitness equipment. Background Technology
[0002] In today's booming era of digital fitness, fitness equipment is rapidly advancing towards intelligence and precision. Neural network models, with their powerful data processing and pattern recognition capabilities, are widely used in fitness equipment to achieve core functions such as exercise movement recognition, motion status monitoring, and personalized exercise program development. However, the hardware resources of fitness equipment, such as memory and computing power, have significant limitations, posing a huge challenge to the efficient operation of neural network models. In neural network model optimization techniques, model pruning is a key method, aiming to reduce the number of model parameters, lower computational and storage costs, while maintaining model performance as much as possible. However, existing model pruning techniques have many unavoidable problems. On the one hand, traditional pruning techniques can only process single data points of human joints or equipment trajectories, failing to capture the synergistic relationship between the two. For example, when analyzing the bench press movement using a barbell, it cannot comprehensively consider the interaction between the arm joint movement and the barbell's trajectory, resulting in a one-sided understanding and analysis of the movement, making it difficult to accurately judge the correctness and effectiveness of the movement. On the other hand, traditional pruning uses a static judgment method, relying on fixed joint angle thresholds. Because different users have individual differences in body structure and exercise habits, fixed thresholds are difficult to adapt to diverse user needs, reducing the applicability of the model. Furthermore, traditional pruning processes are prone to feature loss. When removing model parameters, crucial movement features essential for fitness movement recognition and motion analysis may be mistakenly deleted, leading to a significant decrease in model recognition accuracy. For example, when recognizing squat movements, important parameters related to the trend of knee flexion angle changes may be mistakenly deleted, making it difficult for the model to accurately determine whether the squat movement is standard.
[0003] Therefore, this application provides a method for compressing neural network model parameters for fitness equipment. It utilizes multi-source data for comprehensive analysis and can dynamically adapt to individual differences among different users while avoiding the loss of key features. This significantly reduces the model size and improves the accuracy and robustness of fitness movement recognition, thus better serving the intelligent development of fitness equipment. Summary of the Invention
[0004] To address the technical problems mentioned above, this invention provides a method for compressing neural network model parameters for fitness equipment. The design goal of this invention is to overcome the shortcomings of traditional neural network model parameter processing techniques by using multi-module collaborative optimization, addressing the limitations of the memory and computing power of fitness equipment.
[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for compressing neural network model parameters for fitness equipment includes a data preprocessing module, a model pruning module, a parameter quantization module, and an encoding and storage module. The data preprocessing module collects multi-source data, including motion images, movement data, and physiological indicator data, generated by fitness equipment. It uses cameras to capture motion images, accelerometers to acquire movement data, and physiological monitoring devices to collect physiological indicator data. The multi-source data is then aggregated via communication technologies such as Bluetooth and Wi-Fi. Addressing the differences in numerical range and distribution among the multi-source data, the module normalizes pixel values in image data to a specific range and uses max-min normalization to unify the scale of movement and physiological indicator data. The normalized data is then divided into a learning set, a calibration set, and an evaluation set in a 7:2:1 ratio. A dataset view is constructed for each data sample, and pseudo-labels are assigned based on data characteristics and the application scenario of the fitness equipment. The K-Means algorithm is used to cluster and group the fitness data according to the multi-source data. Motion images are clustered based on joint coordinates and contour feature vectors of human posture; movement data is clustered by movement type and intensity; and physiological indicator data is clustered based on health status and trends. The model pruning module, based on the multi-source data processed by the data preprocessing module, generates simulated data using a Gaussian Mixture Model (GMM). The simulated data is similar in distribution to the multi-source data but does not contain motion identification information. The multi-source and simulated data are input into the neural network model, and through each processing layer, corresponding multi-source and simulated features are obtained. Simultaneously, bimodal feature modeling is performed, including joint-level modeling, defining the parent-child joint hierarchy and calculating the rotation matrix of the child joint relative to the parent joint, and equipment trajectory modeling, analyzing dumbbell acceleration data to extract features and calculating the association weights between human joints and equipment data. Based on the multi-source data, simulated data, motion identification, and modulation factor, an error function is calculated, and a Gaussian Mixture Model (GMM) is used to dynamically calculate the fault tolerance threshold for the user's joint error data. When the output of the error function meets preset conditions, the importance index of each processing unit in the neural network model is calculated based on the modulation factor at that time. This index consists of joint importance, equipment importance, and fusion importance. Unimportant processing units are removed based on the importance index to obtain the pruned neural network model. Finally, the model is fine-tuned. The parameter quantization module employs a threshold-based quantization strategy. By analyzing a large number of fitness data samples and combining model training and practical application performance, a suitable quantization threshold is determined to convert model parameters from high precision to 8-bit integer values. Before applying the quantized model parameters to the neural network model of the fitness device, the model is fine-tuned using calibration data stored on the fitness device. K-means scalar quantization is used for parameter values, and network performance is tested and optimized under different bit parameter quantization conditions. Simultaneously, the application of 16-bit fixed-point representation based on stochastic rounding in CNN training is investigated. The encoding and storage module employs sparse matrix encoding methods such as COO (CoordinateList) to encode the numerous sparse parameters resulting from model pruning. For the 8-bit integer parameters converted by the parameter quantization module, Delta encoding is used to further compress the data. The model parameters are divided into a core parameter layer and an auxiliary parameter layer for hierarchical storage. The core parameter layer stores key parameters closely related to fitness movement recognition and analysis in the high-speed storage area of the device, while the auxiliary parameter layer stores relatively minor parameters in a larger but slower-access area. Huffman coding is used on the quantized weights and codebook. When retraining the network to learn the final weights of the remaining sparse connections, the relationship between compression rate and accuracy is optimized to ensure stable and efficient model operation on fitness equipment.
[0006] The beneficial effects of this invention are as follows: The model pruning module of this system effectively solves the problems currently encountered in the application of neural network models in fitness equipment. Its dual-modal collaborative mechanism comprehensively analyzes the key joint points of the human body and the motion trajectory information of the equipment, avoiding the one-sidedness of single-modal processing, improving the accuracy of fitness movement recognition, providing users with more accurate movement feedback, and reducing the risk of sports injuries. It uses a Gaussian mixture model to dynamically determine the fault tolerance threshold and makes adaptive adjustments based on the user's historical movement data, overcoming the problem that traditional static judgment relies on fixed thresholds and is difficult to adapt to individual differences among different users. During the pruning process, it accurately removes unimportant parameters based on the importance index of joints, equipment, and fusion importance calculation units, reducing the model's computational load and storage space requirements, and expanding its application range on various fitness equipment. At the same time, based on dual-modal feature modeling and comprehensive consideration of error functions, pruning avoids the loss of key features and ensures that the model maintains high recognition accuracy and stability even when the size is compressed. Finally, by fine-tuning the model and working in conjunction with other modules, it comprehensively improves the operating efficiency of the neural network model on fitness equipment, bringing users a smoother and more intelligent fitness experience.
[0007] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0008] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0009] Figure 1 This is a block diagram of a neural network model parameter compression method for fitness equipment according to the present invention; Figure 2 This is a flowchart of a neural network model parameter compression method for fitness equipment according to the present invention; Figure 3 This is a schematic diagram illustrating the compression principle of a neural network model parameter compression method for fitness equipment according to the present invention. Detailed Implementation
[0010] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0011] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in these embodiments have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0012] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0013] In this invention, terms such as "upper," "lower," "left," "right," "front," "back," "vertical," "horizontal," "side," and "bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are only used to facilitate the description of the structural relationships of the various components or elements of this invention and do not specifically refer to any component or element in this invention. They should not be construed as limiting the invention.
[0014] In this invention, terms such as "fixed connection," "connected," and "linked" should be interpreted broadly, indicating a fixed connection, an integral connection, or a detachable connection; a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can determine the specific meaning of these terms in this invention based on the specific circumstances, and they should not be construed as limitations on the invention.
[0015] Example 1, as Figure 1 As shown, this embodiment provides a method for compressing neural network model parameters for fitness equipment, including a data preprocessing module, a model pruning module, a parameter quantization module, and an encoding and storage module; The data preprocessing module is responsible for collecting multi-source data such as motion images, exercise data, and physiological index data from fitness equipment. After normalization, the data scale is unified, and the dataset is divided into training set, calibration set, and evaluation set according to the proportion. The data structure is optimized, a dataset view is constructed and pseudo-labels are assigned, and the multi-source data is clustered and grouped to improve training efficiency, model generalization ability, and adaptability to different fitness scenarios.
[0016] Specifically, firstly, multi-source data generated by fitness equipment is collected, including motion images, exercise data, and physiological indicator data. Motion images are captured using cameras, exercise data is obtained using accelerometers, and physiological indicator data is collected using physiological monitoring devices. The multi-source data is then aggregated into the data preprocessing module via communication technologies such as Bluetooth and Wi-Fi.
[0017] Secondly, a unified scaling process is performed to address the differences in numerical range and distribution among multi-source data.
[0018] For image data, normalize the pixel values to The interval, the formula is: ; For exercise data and physiological index data, the maximum-minimum normalization method is used, and the formula is:
[0019] in, The original data, and These are the minimum and maximum values of the feature, respectively.
[0020] The normalized data is then divided into a training set, a calibration set, and an evaluation set in a 7:2:1 ratio. A dataset view is constructed for each data sample. Simultaneously, based on the characteristics of the data and the application scenario of the fitness equipment, a decision tree algorithm is used to fuse posture features, motion data, and physiological indicator data to generate pseudo-labels. For example, by combining the posture features of the motion images, the resistance magnitude in the motion data, and the heart rate zone in the physiological indicator data, the type and intensity level of the movement are determined. This information is used as pseudo-labels to provide more valuable supervision information for subsequent model training.
[0021] After partitioning the dataset, the K-Means algorithm was used to cluster and group the fitness data according to multi-source data. Specifically, motion images were clustered based on joint coordinates and contour feature vectors of human posture, grouping images with similar postures into one category. Exercise data was clustered by exercise type (running, swimming, strength training) and intensity (e.g., pace, resistance level) to distinguish the feature distribution of different exercise modes. Physiological index data were clustered based on health status (e.g., resting heart rate range) and trends (slope of heart rate increase during exercise), providing a basis for personalized exercise prescriptions. Grouping similar data into groups allows the model to learn from the characteristics of different groups during training, improving training efficiency and the model's adaptability to different fitness scenarios.
[0022] The model pruning module generates simulated data based on multi-source data from the data preprocessing module. Then, it inputs the multi-source data and simulated data into the neural network to perform dual-modal feature modeling. Next, it calculates the error function based on relevant data and modulation factors, and dynamically determines the fault tolerance threshold using a Gaussian mixture model. Once the error function meets the threshold, it calculates the importance of processing units according to the modulation factor and prunes them to reduce computation and storage overhead. Finally, it fine-tunes the model to improve performance and ensure its efficient operation on fitness equipment.
[0023] Current model pruning techniques have significant limitations: they can only perform single-modal processing, pruning only single data points of human joints or equipment trajectories, failing to capture the collaborative relationship between the two; they employ static judgment, relying on fixed joint angle thresholds, making it difficult to adapt to individual differences among users; and they are prone to feature loss, easily deleting key motion features during pruning, leading to decreased recognition accuracy. This model pruning module, through a dual-modal collaborative mechanism, analyzes and compares information on key human joints and equipment motion trajectories to determine whether the movements are standard, significantly improving the accuracy and robustness of fitness movement recognition while greatly reducing model size.
[0024] Specifically, the data preprocessing module collects multi-source data, encompassing information on key human joints and equipment motion trajectories. Human joint data is obtained by capturing the three-dimensional coordinates of 18 joints at 30 frames per second using motion capture equipment. Equipment trajectory data includes the three-axis acceleration data of a smart dumbbell and the speed and incline data of a treadmill. After collecting this data, the preprocessing module performs preprocessing operations: standardizing the joint data by normalizing the coordinates to the [0, 1] interval; using Kalman filtering to remove noise from the equipment data; and synchronizing both types of data to the same timestamp. Based on the preprocessed multi-source data, a Gaussian Mixture Model (GMM) is used to generate simulated data. The initial parameters of the GMM are initialized using K-Means clustering results to ensure that the distribution of the simulated data is consistent with the clustering grouping. The simulated data is similar in distribution to the multi-source data but does not contain motion identification information.
[0025] Multi-source data and simulated data are input into a neural network model. After passing through each processing layer of the model, multi-source features and simulated features corresponding to each layer are obtained. Each processing unit of the neural network model is assigned a modulation factor. Defined as joint importance Importance of equipment Linear combination: = + ,in and Modality fusion weights are parameters used to adjust the relative importance of different modalities during the fusion process. Essentially, they are a set of coefficients that are adjusted... and The value of can alter the contribution of different modalities in the comprehensive analysis. During feature extraction, bimodal feature modeling is performed, including joint hierarchy modeling. Parent-child joint hierarchical relationships are defined, such as "trunk → upper limb: shoulder → elbow → wrist; trunk → lower limb: hip → knee → ankle," and the rotation matrix of child joints relative to their parent joints is calculated. Equipment trajectory modeling is also performed, and time-domain and frequency-domain analyses are conducted on dumbbell acceleration data to extract features such as peak acceleration and frequency components. An attention mechanism is used to calculate the association weights between human joints and equipment data.
[0026] Based on multi-source data, simulated data, motion identifiers, and modulation factors, the error function is calculated. .
[0027]
[0028] in, ; The task loss function ensures that the model can correctly recognize actions such as squatting and running; To compare loss functions, prevent the model from remembering noise, and learn the essential features of actions; For regularization terms, sparsification parameters, and marking of invalid units.
[0029] Specifically, , in, Number of training samples in a batch : The true label of the i-th sample : The probability distribution of actions predicted by the model; = , in, Number of simulated data samples : The fusion feature of the j-th real sample : The fusion feature of the j-th simulated sample; = , in, Number of layers in a neural network The height and width of the feature map in layer c. The modulation factor of the i-th processing unit in the c-th layer. Regularization coefficient.
[0030] The task loss function ensures core functionality, the contrast loss function enhances generalization ability, and the regularization term marks invalid parameters. The three work together to achieve "precise pruning," maintaining high accuracy and robustness while compressing the model.
[0031] The error function comprehensively considers the model's fit to multi-source data and the impact of simulated data on the model's generalization ability, reflecting the model's performance under the current modulation factor settings. For example, in fitness movement recognition, the error function can measure the model's accuracy in recognizing different movements such as squats and running. Simultaneously, a Gaussian mixture model (GMM) is used to dynamically calculate the fault tolerance threshold for user joint error data to determine whether the model needs pruning adjustments. The steps are as follows: S1: Collect user's historical motion data and calculate joint rotation error; S2: Use GMM to fit the error distribution and obtain the mean. and standard deviation ; S3: Dynamically calculate fault tolerance threshold When the error function value exceeds At that time, the pruning operation is triggered.
[0032] Specifically, the calculation of joint rotation error requires combining the three-dimensional coordinate data of human joints, and is achieved by comparing the differences in joint rotation between the actual movement and the standard movement; the formula for calculating the tolerance threshold is as follows: , where the mean This represents the average joint rotation error in a user's historical motion data, reflecting the average deviation of the user's joint movements during a specific action. The standard deviation is... This indicates the dispersion of user joint rotation errors, measuring the range of fluctuation of error data relative to the mean. A larger standard deviation indicates poorer consistency of joint movement across different actions (e.g., differences in movement stability between different users or before and after fatigue for the same user). Due to differences in body structure and movement habits among different users (e.g., joint range of motion, muscle strength), the distribution of joint errors for the same movement varies significantly. By fitting a user-specific error distribution using a Geometric Matrix (GMM) instead of a fixed threshold, the model can automatically adapt to the movement characteristics of different users, avoiding the universality limitations of traditional static thresholds.
[0033] The error function integrates the model's fitting error and generalization ability to multi-source data. When the output of the error function meets preset conditions, the importance index of each processing unit in each processing layer of the neural network model is calculated based on the modulation factor at that time. The importance index of each network layer consists of three parts: joint importance, equipment importance, and fusion importance. Based on the calculated importance index of each processing unit in each processing layer, unimportant processing units are removed, resulting in a pruned neural network model. This pruning strategy can remove parameters that are unimportant to fitness movement recognition and motion analysis, reducing the computational load and storage space occupied by the model.
[0034] The parameter quantization module employs a threshold-based quantization strategy to convert model parameters from high precision to 8-bit integer values, thereby reducing storage and computational complexity.
[0035] In fitness equipment, due to limitations in the device's memory and computing power, this module converts model parameters from high-precision numerical forms to low-precision forms. Threshold-based quantization is an effective strategy. Fitness data contains various types, such as joint angle changes, speed, and acceleration during user movement. This data is crucial for the model to accurately identify fitness movements and analyze movement states. By analyzing a large number of fitness data samples and combining the model's performance in training and practical applications, a suitable quantization threshold can be determined. The quantization threshold is linked to the normalized dynamic range. When the value of the stimulus data exceeds the set threshold, it is quantized into an 8-bit integer value. This method can reduce the precision requirements of the data while preserving key features, thereby reducing the complexity of the computation process. For example, when analyzing running movements, for speed as the stimulus data, a threshold is set based on the common running speed range, quantizing different speed values into corresponding 8-bit integer values. This ensures that the model can capture key information about speed changes while reducing the computational load.
[0036] Before applying the quantized model parameters to the neural network model of the fitness equipment, the model is fine-tuned using calibration data stored on the equipment. Calibration data typically consists of representative data samples collected during the equipment's production or use. Fine-tuning the model using calibration data allows it to better adapt to the quantized accuracy requirements, further optimizing its performance. For example, fine-tuning enables the model to more accurately identify various fitness movements when processing quantized fitness data, improving its stability and reliability.
[0037] Furthermore, the parameter quantization module incorporates optimization methods similar to those used in network quantization. K-means scalar quantization is applied to parameter values, and network performance is tested and optimized under different bit-level quantization conditions. An "8-bit master quantization + 16-bit auxiliary fine-tuning" strategy is employed to ensure significant speedup with minimal accuracy loss. Simultaneously, the application of a 16-bit fixed-point representation based on stochastic rounding in CNN training is investigated to reduce memory usage and floating-point operations, minimize network error, and optimize network stability in deep network processing chips.
[0038] The encoding and storage module is responsible for encoding and storing the pruned and quantized model to meet the limited storage resources and fast retrieval requirements of fitness equipment.
[0039] Because model pruning results in a large number of sparse parameters, and the parameter quantization module converts these parameters into 8-bit integer values, the encoding and storage module employs various encoding techniques to address these characteristics. For sparse parameters, when the sparsity exceeds 50%, the encoding method used for sparse matrices, such as the COO (Coordinate List) format, is adopted. For example, after pruning, the weight matrix of a convolutional layer may contain a large number of zero values; using the COO format can significantly reduce the amount of data stored. For the quantized 8-bit parameters, Delta encoding is used. The differences between adjacent data are stored first, and then Huffman coding is applied to the Delta encoding result to further compress the data. For example, for a series of consecutive weight parameters [10, 12, 15, 14], Delta encoding might store [10, 2, 3, -1], thus reducing the amount of data stored while maintaining accuracy.
[0040] Considering the storage characteristics of fitness equipment, a hierarchical storage approach was adopted. Model parameters were divided into a core parameter layer and an auxiliary parameter layer. The core parameter layer stores key parameters closely related to fitness movement recognition and analysis, such as parameters related to joint motion characteristics and key feature parameters of equipment trajectory. These parameters are directly related to the importance indicators of the pruning module and are preferentially stored in the device's high-speed storage area (such as SRAM) to ensure fast retrieval during model runtime and improve model inference speed. The auxiliary parameter layer stores some relatively minor parameters, such as some initialization parameters of the model and some intermediate calculation results. These parameters are stored in a larger but slower-access storage area (such as Flash). This hierarchical storage approach satisfies the model's need for fast retrieval of key parameters while making full use of the device's storage resources.
[0041] In addition, the encoding and storage module uses Huffman coding on the quantized weights and codebook to further reduce the compression ratio and improve storage efficiency when retraining the network to learn the final weights of the remaining sparse connections. However, given that binary networks based on parameter quantization and binary optimization experience significant performance degradation when processing large networks, and that existing binaryization methods are based on simple matrix approximations that ignore the impact of binaryization on accuracy loss, the encoding and storage module needs continuous optimization in practical applications to balance the relationship between compression ratio and accuracy, ensuring stable and efficient model operation on fitness equipment.
[0042] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for compressing parameters of a neural network model for fitness equipment, characterized in that: It includes a data preprocessing module, a model pruning module, a parameter quantization module, and an encoding and storage module. Data preprocessing module: Collects multi-source data generated by fitness equipment, including motion images, exercise data, and physiological index data. Normalizes the data to unify the data scale, divides the dataset into learning set, calibration set, and evaluation set according to the proportion, optimizes the data structure, constructs a dataset view and assigns pseudo labels, and clusters and groups the multi-source data. The model pruning module utilizes preprocessed multi-source data to generate simulated data without motion markers via a Gaussian mixture model. The multi-source and simulated data are then input into a neural network for bimodal feature modeling. Based on relevant data and modulation factors, an error function is calculated. The dynamic fault tolerance threshold for user joint errors is determined using a Gaussian mixture model. When the error function meets the threshold, the importance index of processing units, including the importance of joints, equipment, and fusion, is calculated according to the modulation factor. Unimportant units are removed to obtain the pruned model, which is then fine-tuned. Finally, model parameters are provided to the parameter quantization module. Specifically, error function , in, ; The task loss function ensures that the model can correctly recognize actions such as squatting and running; To compare loss functions, prevent the model from remembering noise, and learn the essential features of actions; For regularization terms, sparsification parameters, and marking of invalid units; The parameter quantization module uses a threshold-based quantization strategy to convert the model parameters obtained from the model pruning module from high precision to 8-bit integer values, thereby reducing storage and computational complexity. The encoding and storage module encodes and further compresses the large number of sparse parameters that appear after model pruning, as well as the 8-bit integer parameters converted by the parameter quantization module. The model parameters are then divided into a core parameter layer and an auxiliary parameter layer for hierarchical storage.
2. The method for compressing neural network model parameters for fitness equipment according to claim 1, characterized in that: In the data preprocessing module, the specific interval for normalizing image data pixel values is [0, 1]. The maximum-minimum normalization formula for motion data and physiological indicator data is: in, The original data, and These are the minimum and maximum values of the feature, respectively.
3. The method for compressing neural network model parameters for fitness equipment according to claim 1, characterized in that: The model pruning module uses a Gaussian mixture model to dynamically calculate the fault tolerance threshold for user joint error data, including: collecting historical user action data and calculating joint rotation error; fitting the error distribution using a Gaussian mixture model to obtain the mean and standard deviation; and dynamically calculating the fault tolerance threshold based on the mean and standard deviation.
4. The method for compressing neural network model parameters for fitness equipment according to claim 1, characterized in that: Before applying the quantized model parameters to the neural network model of the fitness equipment, the parameter quantization module fine-tunes the model using calibration data stored on the fitness equipment. We used K-means scalar quantization for the parameter values and tested and optimized the network performance under different bit parameter quantization conditions. We also studied the application of 16-bit specific point representation based on random rounding in CNN training.
5. The method for compressing neural network model parameters for fitness equipment according to claim 1, characterized in that: The encoding and storage module uses sparse matrix encoding methods such as COO format to encode the large number of sparse parameters that appear after model pruning. For the 8-bit integer parameters converted by the parameter quantization module, Delta encoding technology is used.