A free weight training monitoring method using a wearable PPG sensor

Abstract

The present application relates to a kind of free weight training monitoring method using wearable PPG sensor, belong to mobile computing application and human training auxiliary technical field.The present application uses infrared light and green light on PPG sensor to irradiate wrist skin, and collects the optical density change data containing muscle movement information.By analyzing change data, the load weight currently used by exerciser, the number of repetitions and the type of movement are obtained, and further predict the adjusted load weight required in next step.The present application proposes a novel arterial noise elimination method, effectively reduces the influence of pulse noise in artery, designs a novel feature extraction method based on recurrence plot, maps each action segment to two-dimensional phase space to extract space-time features, and constructs a novel multi-task learning model based on convolutional neural network.The present application can effectively identify load weight and movement type, realize the monitoring of free weight training.

Description

TECHNICAL FIELD

[0001] The present application relates to a free weight training monitoring method, in particular to a free weight training monitoring method using a wearable PPG (Photo Plethysmo Graphy) sensor, and belongs to the technical field of mobile computing application and human training assistance. BACKGROUND

[0002] Free weight training usually uses fitness equipment such as dumbbells, barbells and kettlebells to improve or maintain human muscles by controlling the selection of load weight, repetition number and movement type. Because of its significant role in enhancing muscle strength and improving physical and mental health, free weight training has become widely popular. However, most exercisers have difficulty in choosing the appropriate load weight without professional guidance when participating in free weight training. Generally speaking, too large load weight may cause sports injuries, and insufficient load weight may affect the training effect and hinder the progress of exercisers.

[0003] At present, various fitness assistance solutions are developing rapidly. These solutions can provide precise and personalized training monitoring and feedback for exercises such as running, bodyweight training and high-intensity interval training. However, the research progress in the field of free weight training assistance is relatively slow. Most exercisers still need to manually record, calculate and adjust the load weight with various load weight adjustment models (such as pyramid model, linear model, etc.) during training. The calculation process is not only tedious and time-consuming, but also prone to errors.

[0004] In order to solve the above problems, researchers have tried different devices to achieve automatic load weight estimation. According to the sensing object, these works can be divided into two categories: one based on special devices and the other based on wearable devices. The work based on special devices relies on intelligent adjustable dumbbells, passive RFID tags, etc. to sense the fitness equipment used by exercisers. However, the high cost of intelligent dumbbells and RFID readers hinders the large-scale deployment of such work. The method based on wearable devices uses surface electromyography, inertial measurement unit and PPG (Photoplethysmography) sensor to estimate the load weight. By sensing the exerciser rather than the fitness equipment, this kind of work can achieve monitoring of all types of movements. However, these methods are too coarse-grained and lack robustness, making it difficult to achieve high-precision load weight estimation. In addition, there is no work that can provide recommendations for load weight adjustment.

[0005] In summary, the existing technology has various deficiencies, and there is an urgent need to develop a new free weight training monitoring technology that is easy to deploy, low-cost and fine-grained. SUMMARY

[0006] The purpose of the present application is to overcome the defects and deficiencies of high cost and poor robustness of existing load weight estimation techniques in human free weight training, and creatively propose a free weight training monitoring method using wearable PPG sensor. The present application uses infrared light and green light on the PPG sensor to irradiate the wrist skin, and collects the optical density change data containing muscle movement information. By analyzing the change data, the current load weight used by the exerciser, the number of repetitions and the movement type are obtained, and the adjusted load weight required for the next step is further predicted.

[0007] The innovation of the present application includes: in the process of free weight training, the main muscle groups involved in the movement are activated, the muscles contract and produce force output to overcome the given resistance. The force generated in the activated muscle tissue increases the resistance of the surrounding blood vessels and compresses the geometry of the arteries to varying degrees. PPG is an optical technology for non-invasive detection of blood volume changes during the heartbeat cycle. Therefore, the changes in arterial blood flow caused by muscle contraction and vascular deformation can change the pattern of PPG waveform. In short, the deformation of specific muscle groups under different load weights and movement types will cause changes in the geometry of the arteries, and further produce subtle and unique change patterns in the PPG signal. By analyzing this change pattern, the present application can realize the recognition of load weight and movement type, and further monitor the free weight training.

[0008] The purpose of the present application is realized by the following technical solutions.

[0009] A free weight training monitoring method using a wearable PPG sensor, comprising the following steps:

[0010] Step 1: using a dual-wavelength PPG sensor of infrared light and green light, collecting two PPG signals at the wrist during human exercise, removing high-frequency noise and arterial noise in the PPG signal to obtain pure motion-derived signals.

[0011] Specifically, it includes the following steps:

[0012] Step 1.1: process the two collected PPG signals respectively (low-pass filter can be used), eliminate high-frequency interference that does not overlap with the frequency of PPG signal.

[0013] Step 1.2: subtract the two PPG signals to obtain an arterial noise reference signal, further eliminate arterial noise using adaptive filtering technology, and extract pure motion-derived signals.

[0014] Step 2: segment the motion-derived signals and estimate the number of movement repetitions, and generate a recurrence plot that can effectively represent the spatiotemporal features for each motion segment.

[0015] Specifically, the method comprises the following steps:

[0016] Step 2.1: The starting point and the ending point of each action segment on the motion-derived signal are detected by using a moving window method, so that action segmentation and motion repetition number estimation are realized.

[0017] Step 2.2: A recurrence plot is generated for each action segment obtained in step 2.1.

[0018] Step 3: The recurrence plot and the personal information of the human body are input into a multi-task convolutional neural network, so that load weight estimation and motion type recognition are realized, and personalized load weight adjustment suggestions are proposed according to the state of the human body and the exercise target, and free weight training monitoring is realized.

[0019] Specifically, the method comprises the following steps:

[0020] Step 3.1: A multi-task convolutional neural network model is constructed, and the recurrence plot representing the spatiotemporal features obtained in step 2.2 and the personal information of the human body are taken as inputs, and the load weight estimation and the motion type recognition results are taken as outputs.

[0021] Step 3.2: In combination with the load weight, the motion type and the repetition number obtained in step 3.1, a most suitable load weight adjustment model is selected according to the state and the target of the user, and load weight adjustment suggestions are given to the human body.

[0022] Up to now, from step 1 to step 3, the free weight training monitoring by using the wearable PPG sensor is realized.

[0023] Beneficial effects

[0024] Compared with the prior art, the method has the following advantages:

[0025] 1. The present application can realize continuous and passive free weight training monitoring only by using a low-cost PPG sensor integrated in a commercial wearable device, without interrupting the training of the exerciser, and has strong popularization. In addition, the present application redefines the load weight estimation as a regression problem, is no longer limited to discrete load weight categories, and provides personalized load weight adjustment feedback suggestions.

[0026] 2. The present application proposes a novel arterial noise elimination method, which effectively reduces the influence of pulse noise in the artery; designs a novel feature extraction method based on a recurrence plot, which maps each action segment to a two-dimensional phase space to extract spatiotemporal features; and constructs a novel multi-task learning model based on a convolutional neural network, which realizes the monitoring of load weight and motion type at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1Six free weight training types identified by the embodiment of the present invention, wherein (1) is a lateral raise, (2) is a dumbbell curl, (3) is a dumbbell external rotation, (4) is a dumbbell deadlift, (5) is a dumbbell bench press, and (6) is a seated dumbbell press;

[0028] Figure 2 Schematic diagram of a free weight training monitoring method using a wearable PPG sensor;

[0029] Figure 3 The process of generating a recurrence graph from motion-derived signals designed for the present invention;

[0030] Figure 4 The structure diagram of the multi-task convolutional neural network developed by the present invention;

[0031] Figure 5 is the root mean square error result of the load weight estimation of each participant in the embodiment of the present invention;

[0032] Figure 6 The root mean square error result of the estimated number of repetitions for each participant in the embodiment of the present invention;

[0033] Figure 7 The accuracy, recall, and F1 score results of each action in the motion type recognition task of the embodiment of the present invention;

[0034] Figure 8 The results are the root mean square error of load weight estimation and the accuracy, recall rate and F1 score of motion type recognition in experiments with different sampling rates according to the embodiment of the present invention. DETAILED DESCRIPTION

[0035] The method of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0036] Example

[0037] Will Figure 1 As shown, the present invention uses six free weight training types (F1: Lateral Raise; F2: Dumbbell Curl; F3: Dumbbell External Rotation; F4: Dumbbell Deadlift; F5: Dumbbell Bench Press; F6: Seated Dumbbell Press) as specific detection targets. The six training types selected by the present invention can cover various muscle groups and movement patterns of the human upper body.

[0038] like Figure 2 As shown, a free weight training monitoring method using a wearable PPG sensor includes the following steps:

[0039] Step 1: Use a dual-wavelength PPG sensor with infrared and green light to collect two PPG signals from the wrist during exercise. Remove high-frequency noise and arterial noise from the PPG signals to obtain a pure motion-derived signal.

[0040] Specifically include:

[0041] Step 1.1: The two collected PPG signals are processed using low-pass filters to eliminate high-frequency interference that does not overlap with the PPG signal frequency.

[0042] Specifically, since the PPG signal is mainly concentrated below 5 Hz, a fourth-order Butterworth low-pass filter with a cutoff frequency of 5 Hz can be used to remove high-frequency noise in the two PPG signals. Subsequently, the range of both signals is scaled to [-1, 1].

[0043] Step 1.2: Subtract the two PPG signals to obtain the arterial noise reference signal, and use adaptive filtering technology to further eliminate the arterial noise and extract the pure motion-derived signal.

[0044] Specifically, the infrared and green PPG signals obtained in step 1.1 are denoted as X(k) and Y(k), respectively, where k denotes the signal sequence. The arterial noise reference signal is then defined as a(k), where a(k) = X(k) - Y(k).

[0045] Furthermore, the present invention proposes an arterial noise elimination method based on adaptive filtering to eliminate highly dynamic periodic pulse signals:

[0046] The least mean square (LMS) adaptive filtering algorithm, with its simple principle and stable performance, has been widely used in the field of physiological signal denoising. However, there is a contradiction between its steady-state error and the convergence speed after convergence. Although the normalized LMS adaptive filtering algorithm resolves this contradiction by normalizing the convergence step size factor to the signal energy, a fixed normalized step size factor is not conducive to accelerating convergence.

[0047] To this end, the present invention proposes a variable-step-size normalized LMS adaptive filtering algorithm, which allows the step-size factor to change with the number of iterations. This not only speeds up the convergence speed but also more effectively balances the convergence speed and steady-state error. The details are as follows:

[0048] The variable step-size normalized LMS adaptive filter comprises two input ends, wherein the basic input end is a noisy infrared light PPG signal, i.e. X(k); the reference input end is an arterial noise reference signal a(k) which is strongly related to pulse noise. The filtered signal is estimated by subtracting the noise reference signal from the noise signal of the basic input. The adaptive filter can automatically adjust its coefficients, so that the noise reference signal estimation value finally approximates to the actual noise signal, and the filtered PPG signal is the pure motion-derived signal. The variable step-size normalized LMS algorithm iteratively converges the error between the system output and the expected response, so that the system output finally stabilizes at the expected response. After each iteration, the error between the system output and the expected response is used as a feedback value to participate in the next iteration process. The coefficients are continuously adjusted, so that the system output converges to the vicinity of the expected response. During the iteration process, the error gradually decreases. When the error is less than a set threshold, the iteration ends.

[0049] Specifically, the following steps are included:

[0050] First, the input signal weight coefficient W(k) is input, which is initially set as W(k) = 0.

[0051] Then, the output value C(k) corresponding to each input signal is calculated, C(k) = W T (k)a(k), wherein T represents matrix transposition.

[0052] After that, the error e(k) is calculated using the corresponding output signal, e(k) = X(k)-C(k).

[0053] Finally, the input signal weight coefficient W(k) is iteratively updated:

[0054] W(k+1) = W(k) + 2μ(n)e(k)X(k) (1)

[0055] Wherein n is the iteration number, which is equal to the signal length minus the filter order (in the embodiment of the present application, the filter order is preferably 256, but other values within the range of [32, 1024] are also within the scope of the present application). e represents the error. μ(n) is the step-size factor, and its expression is:

[0056]

[0057] Wherein X(k) is the input signal; β is a constant for controlling the convergence speed, and satisfies β = 2 / λ max , λ maxis the maximum eigenvalue of the input signal autocorrelation matrix; α is a constant that affects the steady-state error of the algorithm; γ is a constant that controls the step size to prevent it from being too large. The value range of α is [2, 10], and the value range of γ is [0, 1]. In this embodiment, the values ​​of α and γ are 2 and 0.05, respectively.

[0058] Considering that infrared light can penetrate the skin and reach the arteries in the subcutaneous tissue, while green light can only reach the superficial capillaries, infrared light PPG signals can better capture muscle tissue movement information during exercise. This paper focuses on the denoising analysis of infrared light PPG signals, that is, sending X(k) and a(k) into the filter to obtain the motion-derived signal.

[0059] Step 2: Segment the motion-derived signals and estimate the number of motion repetitions, generating a recurrence graph for each action segment that can effectively represent the spatiotemporal features.

[0060] The specific steps include:

[0061] Step 2.1: Use the moving window method to detect the start and end points of each action segment on the motion-derived signal to achieve action segmentation and estimation of the number of motion repetitions.

[0062] First, we can use the findpeaks algorithm provided by MATLAB software to find the position v1 of the first trough of the motion-derived signal. Then, we can use the autocorrelation algorithm to calculate the period p1 of the signal in the first 15 seconds and set the initial value of the jump step length t to p1.

[0063] Then, the minimum value of the data points whose horizontal coordinates are within the range of t / 3 before and after v1+t is found to detect the position of the second trough v2, and then the jump step size t=v2-v1 is updated, and so on.

[0064] The moving window method uses jump detection to reduce computational complexity. At the same time, by updating the jump step size and length of the detection window in real time based on the current cycle, the method becomes more stable and accurate.

[0065] Step 2.2: Generate a recurrence graph for each action segment obtained in step 2.1.

[0066] Recurrence plots are an innovative visualization technique that reveals hidden recurrence patterns and complex dynamics in time series data. Recurrence plots encode the temporal dependence and spatial fluctuations of PPG waveforms, thereby preserving critical spatiotemporal information for estimating load weight and identifying exercise type. Figure 3 The generation process of the recursion graph is shown as follows:

[0067] First, transform each point pm in the time series into the corresponding state in the phase space Reconstruct phase space;

[0068] Then, the distance between each two states, i.e., vector norm, is calculated, and threshold binarization is performed to obtain the features between two corresponding states in the recurrence plot, which is denoted as:

[0069]

[0070] where Q is the total number of states; R m，n is a QxQ matrix; ∈ is a threshold value for determining the neighborhood of states, ∈ is in the range of [-1, 1], and in this embodiment, ∈ is 0; θ is a Heaviside step function, which takes the value of 0 when is less than 0, and takes the value of 1 when is greater than or equal to 0. represents the state vector in space.

[0071] Step 3: input the recurrence plot and personal information of the human body into the multi-task convolutional neural network, simultaneously realize load weight estimation and movement type recognition, and propose personalized load weight adjustment suggestions according to the state of the human body and the exercise target, so as to realize free weight training monitoring.

[0072] Specifically, the following steps are included:

[0073] Step 3.1: construct a multi-task convolutional neural network model, input the recurrence plot representing the spatio-temporal features obtained in step 2.2 and the personal information of the human body as input, and output the load weight estimation and movement type recognition results.

[0074] Further, the present application designs a novel multi-task convolutional neural network framework for load weight regression and movement type classification, and the network structure is as shown in Figure 4 The network does not rely on pre-defined artificial features or expert knowledge, but automatically extracts robust feature representations from the input recurrence plot, and explicitly includes personal information (including age, gender, height, and weight, etc.) into the learning model, so as to consider individual physiological characteristics to obtain accurate and reliable results. In addition to Figure 1 the six training types shown, the present application also defines a NULL class to represent unknown exercise types other than the defined classes, such as additional actions caused by switching between actions. The output includes class labels and estimated load weights, and the load weight estimation result of the NULL class will be discarded.

[0075] Considering that the recursive graph is usually globally similar and locally different, the global and local structural information is very important for simultaneously achieving the classification and regression tasks. To capture the local structural information of the recursive graph, the invention uses six convolutional blocks, each of which contains a two-dimensional convolutional layer using a rectified linear unit (ReLU) activation function. The six convolutional layers use 32, 32, 64, 64, 128, and 128 kernels, respectively, with a kernel size of 3x3. After Conv2, Conv4, and Conv6, there are 2x2x2 max-pooling layers for downsampling to learn local significant information. Two fully connected layers, FC7 and FC8, have 128 and 64 units, respectively. Then, two additional FC layers, FC9 and FC10, with 64 and 68 units, respectively, are added to simulate the global structural information of the recursive graph. In addition, the concatenated representation of the FC10 output and the personal information (including age, gender, height, weight, etc.) are fed into two FC layers, FC11 and FC12, each with 64 units. Finally, two 32-unit FC13 layers are used to make motion type classification probability prediction and load weight estimation through Softmax, respectively.

[0076] Step 3.2: Based on the user's status and goals, select the most suitable load weight adjustment model according to the load weight, exercise type, and repetition number obtained in step 3.1, and give the exerciser a load weight adjustment suggestion.

[0077] Optimizing the management of load weight is a necessary condition for continuously inducing physiological adaptation and thus avoiding poor training or injury. However, due to individual differences, it is still challenging to find an ideal load weight adjustment model. To solve this problem, the invention proposes a decision tree scheme that automatically selects one of the four commonly used load weight adjustment models based on the user's status and goals, including the linear model, the two-to-two model, the pyramid model, and the repetition maximum zone model. The basic information of the four models is as follows:

[0078] Linear model: Adjust the same load weight increment for each set of training, with an increment of 1 kg.

[0079] Two-to-two model: In consecutive two sets of training, if the individual completes more than 2 times the specified repetition goal, increase the weight of the given exercise by 1 kg.

[0080] Pyramid model: Start with a lighter load weight and more repetitions, then gradually reduce the number of repetitions and increase the load weight, with a reduction of 1 repetition and an increase of 1 kg.

[0081] Repetition Maximum Zone Model: exercisers choose the heaviest weight they can lift for a given range of repetitions (e.g., 3 to 5 repetitions), with the goal of reaching muscle failure on the last set of the exercise.

[0082] Specifically, the decision nodes include the user's experience level (beginner or advanced), the recent training effect (progression or plateau), and the training goal (hypertrophy or strength), each node contains a binary category. The decision tree is constructed by a recursive partitioning process based on these nodes and values. For beginners (defined as people who have trained for less than three months), a linear model with a fixed increment is recommended. If an advanced exerciser is in the progression phase, a two-versus-two model is selected. When entering the plateau phase, a pyramid model can be used to achieve the hypertrophy goal, and a repetition maximum zone model is used to achieve the strength increase goal.

[0083] Embodiment validation

[0084] To verify the performance of the present application, the present application constructed a wrist-worn device prototype with a commercial PPG sensor to record infrared and green light PPG signals. A total of 15 participants (8 males and 7 females, aged 19 to 42 years old, training age from 0 to 6 years) were recruited to participate in the experiment. After a brief standard motion training, the participants wore the prototype tightly on the inner side of the right wrist. For each participant, the present application first determined their maximum weight (1RM) that can be lifted once for the six types of movements through a standardized strength test method. Based on these results, during the data collection period, the present application defined a personalized 60-80% 1RM range (incrementing by 0.5kg) as the individual's load weight. In addition, 80% of the data was used for training, and the remaining 20% of the data was used for testing, and five-fold cross-validation was performed.

[0085] The root mean square error (RMSE) was used to evaluate the performance of the present application's load weight and repetition number estimation, defined as where n is the number of samples, yi is the true value, is the predicted value of the load weight or the number of repetitions. The smaller the RMSE value, the higher the estimation accuracy.

[0086] The accuracy (Precision), recall, and F1 score were used to evaluate the performance of the present application's movement type identification. Among them, the recall is defined as: the percentage of correctly identified movement type F segments among all segments of movement type F; the accuracy is defined as: the percentage of correctly belonging to type F character segments among all segments predicted to be movement type F; the F1 score is defined as:

[0087] First, the overall performance of the present application's load weight estimation was tested. Figure 5The RMSE results of load weight estimation for all 15 participants are shown. The participants' RMSE results range from 0.32 kg to 0.99 kg, with an average error of 0.59 kg. The results show that the present invention can achieve accurate load weight estimation. In particular, participants 5 and 15 have the lowest estimation error, which can be due to their relatively rich fitness experience and standard exercise forms, which allow the sensor to capture more stable and reliable information during exercise.

[0088] Secondly, the overall performance of the present invention's repetition count estimation is tested. Figure 6 The RMSE results of repetition estimation for all 15 participants are given. It can be seen that the RMSE of repetition count estimation is between 0.68 and 1.2 times, with an average error of 0.96 times, and the RMSE of most participants is less than 1 time. Overall, the above results show that the present invention can reliably track the number of repetitions during free weight training.

[0089] Then, the overall performance of the present invention's motion type recognition is tested. Figure 7 The Precision, Recall and F1 score results of identifying six motion types are reported. For each type, the average value between different participants is calculated. Overall, the average Precision of the present invention is 91.55%, the Recall is 91.59%, and the F1 score is 91.57%, and the F1 score of all motion types is above 85%, which shows that the present invention is relatively accurate in recognizing motion types. In addition, the F1 scores of F3 (dumbbell external rotation) and F6 (dumbbell sitting press) are below 90%, which can be due to their more similar motion patterns, making it more difficult to distinguish.

[0090] Finally, the impact of sampling rate on the present invention is tested. The present invention is evaluated at several different sampling rates, including 50 Hz, 100 Hz, 200 Hz and 400 Hz. Figure 8 The RMSE results of load weight estimation and the Precision, Recall and F1 score results of motion type recognition under the above several sampling rates are shown. The results show that the performance of the present invention improves continuously when the sampling rate increases from 50 Hz to 200 Hz. However, when the sampling rate reaches around 200 Hz, the improvement is no longer significant, at which point the RMSE of load weight estimation is 0.59 kg, and the Precision, Recall and F1 of motion type recognition are 91.55%, 91.59% and 91.57%, respectively, achieving quite excellent performance.

[0091] The above detailed description of the specific description, the purpose, technical scheme and beneficial effects of the application are further described in detail, it should be understood that the above description is only a specific embodiment of the present application, for explaining the present application, and is not used to limit the protection scope of the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A free weight training monitoring method using a wearable PPG sensor, characterized in that: The following steps are involved: Step 1: Use a dual-wavelength PPG sensor with infrared and green light to collect two PPG signals from the wrist during exercise. Remove high-frequency noise and arterial noise from the PPG signals to obtain a pure motion-derived signal. Step 1.1: Process the two collected PPG signals separately to eliminate high-frequency interference that does not overlap with the PPG signal frequency; Step 1.2: Subtract the two PPG signals to obtain the arterial noise reference signal, and use adaptive filtering technology to further eliminate the arterial noise and extract the pure motion-derived signal; Step 2: Segment the motion-derived signals and estimate the number of motion repetitions, generating a recursive graph for each action segment that can effectively represent spatiotemporal features; Step 2.1: Use the moving window method to detect the starting and ending points of each action segment on the motion-derived signal to achieve action segmentation and estimate the number of motion repetitions; Step 2.2: Generate a recurrence graph for each action segment obtained in step 2.1; Step 3: The recurrence graph and the individual's personal information are fed into a multi-task convolutional neural network to simultaneously estimate load weight and identify exercise type. Based on the individual's condition and exercise goals, personalized load weight adjustment suggestions are made to monitor free weight training. Step 3.1: Construct a multi-task convolutional neural network model, taking the recursive image representing the spatiotemporal features obtained in step 2.2 and the human body personal information as input, and outputting the load weight estimation and movement type recognition results; Step 3.2: Based on the load weight, exercise type, and number of repetitions obtained in step 3.1, the most suitable load weight adjustment model is selected according to the user's status and goals, and load weight adjustment suggestions are given to the human body.

2. A free weight training monitoring method using a wearable PPG sensor as claimed in claim 1, characterized in that: In step 1.1, a fourth-order Butterworth low-pass filter with a cutoff frequency of 5 Hz is used to remove high-frequency noise from the two PPG signals; then, the range of both signals is scaled to [-1, 1].

3. A free weight training monitoring method using a wearable PPG sensor as claimed in claim 1, characterized in that: In step 1.2, the infrared and green PPG signals obtained in step 1.1 are denoted as X(k) and Y(k), respectively, where k denotes the signal sequence. Next, the arterial noise reference signal is defined as a(k), where a(k) = X(k) - Y(k). Adopting the arterial noise elimination method based on adaptive filtering, a normalized LMS adaptive filter with variable step size is used to eliminate the highly dynamic periodic pulse signal. The variable-step-size normalized LMS adaptive filter includes two inputs: a base input, a noisy infrared PPG signal, X(k); and a reference input, an arterial noise reference signal a(k) that is strongly correlated with pulse noise. The filtered signal is estimated by subtracting the noise reference signal from the base input noise signal. The adaptive filter can automatically adjust its coefficients so that the noise reference signal estimate is ultimately close to the actual noise signal, and the filtered PPG signal is a pure motion-derived signal. The variable step-size normalized LMS algorithm iteratively converges the error between the system output and the desired response, so that the system output is finally stabilized at the desired response. After each iteration, the error between the system output and the expected response is used as a feedback value to participate in the next iteration process; The coefficients are continuously adjusted to make the system output converge to the expected response. During the iteration process, the error gradually becomes smaller. When the error is less than the set threshold, the iteration ends. X(k) and a(k) are fed into the filter to obtain the motion-derived signal.

4. The free weight training monitoring method using a wearable PPG sensor according to claim 3, wherein: First, the input signal weight coefficient W(k) is initially set to W(k) = 0; Then, calculate the output value C(k) corresponding to each input signal, in, Represents matrix transpose; Afterwards, the error e(k) is calculated using the corresponding output signal, e(k) = X(k) - C(k); Finally, the input signal weight coefficient W(k) is iteratively updated: W(k+1)=W(k)+2μ(n)e(k)X(k) (1) Where n is the number of iterations, which is the signal length minus the filter order; e represents the error; μ(n) is the step size factor, which is expressed as: Where X(k) is the input signal; β is a constant that controls the convergence rate, satisfying β = 2 / λ max ,λ max is the maximum eigenvalue of the autocorrelation matrix of the input signal; α is a constant that affects the steady-state error of the algorithm, and the value range of α is [2, 10]; γ is a constant that controls the step size to prevent it from being too large, and the value range of γ is [0, 1].

5. The free weight training monitoring method using a wearable PPG sensor according to claim 1, wherein: In step 2.1, first find the position v1 of the first trough of the motion-derived signal, use the autocorrelation algorithm to calculate the period p1 of the signal in the first 15 seconds, and set the initial value of the jump step t to p1; Then, the minimum value of the data points whose horizontal coordinates are within the range of t / 3 before and after v1+t is found to detect the position of the second trough v2, and then the jump step size t=v2-v1 is updated, and so on.

6. The free weight training monitoring method using a wearable PPG sensor according to claim 1, wherein: In step 2.2, first each point p in the time series m Transformed into the corresponding state in phase space Reconstruct phase space; Then, the distance between each two states, i.e., the vector norm, is calculated and thresholded to obtain the features between the two corresponding states in the recursive graph; the recursive graph reveals the trajectory movement from the current state to the previous state, which is expressed as: Where Q is the total number of states; R m,n is a Q×Q square matrix; ∈ is the threshold for determining the state neighborhood, and the value range of ∈ is [-1, 1]; θ is the Heaviside step function. When it is less than 0, θ takes the value of 0. When it is greater than or equal to 0, θ takes the value of 1; Represents the state vector in space.

7. The free weight training monitoring method using a wearable PPG sensor according to claim 1, wherein: The multi-task convolutional neural network model constructed in step 3.1 automatically extracts robust feature representations from the input recurrence graph and explicitly incorporates personal information into the learning model; Define the NULL class to represent unknown exercise types outside the defined class. The output contains the class label and the estimated load weight. The load weight estimation result of the NULL class will be discarded. Six convolutional blocks are used, each of which contains a two-dimensional convolutional layer using a linear rectifier unit activation function. The six convolutional layers use 32, 32, 64, 64, 128 and 128 kernels with a kernel size of 3×3 respectively; Conv2, Conv4 and Conv6 are followed by a 2×2×2 maximum pooling layer for downsampling to learn local salient information; two fully connected layers FC - FC7 and FC8 have 128 and 64 units respectively; then, two additional FC layers, FC9 and FC10 with 64 and 68 units respectively, are added to simulate the global structural information of the recurrent graph; in addition, the cascade representation and personal information output by FC10 are fed into two FC layers, namely FC11 and FC12, each with 64 units; finally, two 32-unit FC13 layers are used for motion type classification probability prediction and load weight estimation, respectively.

8. The free weight training monitoring method using a wearable PPG sensor according to claim 1, wherein: In step 3.2, a decision tree approach is used to automatically select one of four load weight adjustment models based on the user's state and goals. These models include a linear model, a two-to-two model, a pyramid model, and a repeated maximum area model. The information for these four models is as follows: Linear model: Each training session was adjusted with the same load weight increment of 1 kg. Two-for-Two Model: If an individual exceeds the specified repetition goal by more than 2 repetitions during two consecutive training sets, the weight of a given exercise is increased by 1 kg. Pyramid model: Start with a lighter load and more repetitions, then gradually reduce the number of repetitions and increase the load, with the number of repetitions decreasing by 1 and the load increasing by 1 kg; Repetition Max Zone Model: The exerciser chooses the heaviest load they can lift within a given repetition range, with the goal of reaching muscle failure on the last set of the exercise. Decision nodes include the user's experience level, recent training results, and training goals. Each node contains a binary category, and a decision tree is constructed through a recursive partitioning process based on these nodes and values. For beginners, use a linear model with fixed increments. If advanced exercisers are in the progression phase, choose a two-to-two model. When entering the plateau phase, use a pyramid model to achieve muscle growth goals and a repetition maximum zone model to achieve strength gains.

Images