Deep learning-based badminton flying strength detection method

Abstract

The invention discloses a badminton flying strength detection method based on deep learning, and relates to the technical field of intelligent detection of kinematics, and the method comprises the steps: converting ball hitting window data containing speed information into four recurrence plots and short-time spectrograms, splicing the four recurrence plots and short-time spectrograms into an unchecked strength image tensor, carrying out the phase consistency and edge checking, and carrying out the detection of the strength image tensor; obtaining a force image tensor; and inputting the force image tensor into a convolution and attention combined deep learning network to perform feature extraction, fitting air resistance and effective quality parameters in combination with a speed change curve, and performing constraint in a micro physical layer according to a momentum conservation relationship to obtain a physically corrected badminton hitting force result and generate a detection report. According to the invention, visual representation and depth feature extraction of the multi-source motion signal are realized, the accuracy of ball hitting force detection is improved, and the physical consistency and stability of a force result are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of kinematics intelligent detection, and particularly relates to a badminton flight force detection method based on deep learning. BACKGROUND

[0002] With the rapid development of intelligent sports monitoring and artificial intelligence technology, sports detection based on wearable sensors has become an important direction of sports science and sports equipment research. In recent years, inertial measurement units (IMU) and multi-axis acceleration sensors have been widely used in the dynamics analysis of high-speed racket sports such as badminton, tennis, and baseball. By capturing the acceleration, angular velocity, and attitude changes of the athlete's limbs and equipment, the action type and the exercise intensity can be identified. At the same time, the introduction of deep learning technology has significantly improved the feature extraction and pattern recognition capabilities of sports data, making it possible to model complex nonlinear motion processes.

[0003] The existing badminton force detection methods mainly have two deficiencies. First, the sensor signals are mostly analyzed in a single source form, without fully considering the time deviation and coupling characteristics of the ball end and the wearing end signals, resulting in accumulated alignment errors of the instantaneous data of the ball hitting, affecting the reliability of the force calculation. Second, most existing deep learning algorithms use pure data-driven fitting, lack of aerodynamics and momentum conservation constraints, and the model can give numerical prediction results, but it is difficult to ensure that the results are consistent with the real physical process, resulting in precision drift in cross-scene and cross-device applications. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a badminton flight force detection method based on deep learning to solve the problem of lack of physical consistency and precision stability in traditional detection methods in badminton hitting force estimation.

[0006] To solve the above technical problems, the present application provides the following technical solutions: The present application provides a badminton flight force detection method based on deep learning, which comprises, Collecting ball-end three-axis acceleration and wearing-end six-degree-of-freedom motion data, time aligning through Bluetooth time marking and impact peak cross-correlation, and generating multi-source synchronous data; Detecting the ball hitting impact starting point in the multi-source synchronous data and cutting the ball hitting window, integrating the ball-end three-axis acceleration, combining with the zero-speed constraint to obtain the velocity change curve, and calculating the hand arm end linear velocity according to the wearing-end six-degree-of-freedom motion data to generate the ball hitting window data containing velocity information; The ball hitting window data containing speed information is converted into four recursive graphs and short-time spectrum graphs, and spliced into an unverified force image tensor, and after phase consistency and edge verification, a force image tensor is obtained; The force image tensor is input into a deep learning network combining convolution and attention for feature extraction, the fitting of the speed change curve is combined with the air resistance and effective mass parameters, and the momentum conservation relationship is used to constrain in the differentiable physical layer, so that the badminton hitting force result is obtained after physical correction, and a detection report is generated.

[0007] As a preferred scheme of the badminton flight force detection method based on deep learning, wherein: the three-axis acceleration of the ball end and the six-degree-of-freedom motion data of the wearing end are collected, time alignment is performed through Bluetooth time marking and impact peak cross-correlation, and multi-source synchronous data is generated, and the specific steps are as follows, The three-axis acceleration of the ball end and the six-degree-of-freedom motion data of the wearing end are collected, and the sampling time stamp and the Bluetooth receiving time stamp are recorded to generate a double-time-marked data sequence; According to the Bluetooth receiving time stamp, the double-time-marked data sequence is roughly aligned to generate preliminary synchronous data; The impact peak position is monitored from the three-axis acceleration of the ball end, and the swing peak position is monitored from the six-degree-of-freedom motion data of the wearing end, and the cross-correlation function between the impact peak and the swing peak is calculated; The time offset corresponding to the maximum value of the cross-correlation function between the impact peak and the swing peak is used as the fine alignment parameter; The fine alignment parameter is used to perform time translation correction on the preliminary synchronous data to generate time-aligned multi-source synchronous data.

[0008] As a preferred scheme of the badminton flight force detection method based on deep learning, wherein: the hitting impact starting point is detected in the multi-source synchronous data, and the hitting window is cut, and the specific steps are as follows, The three-axis acceleration of the ball end and the six-degree-of-freedom motion data of the wearing end are extracted from the multi-source synchronous data; The synthetic acceleration is calculated according to the three-axis acceleration of the ball end, and the sliding median filter is performed to obtain a robust threshold; The time point when the synthetic acceleration exceeds the robust threshold for the first time is taken as the hitting impact starting point; The multi-source synchronous data is divided into a preparation segment and an early flight segment forward and backward respectively with the time point of the hitting impact starting point as the center, a hitting window is generated, and the hitting window data and the hitting window time range are obtained.

[0009] As a preferred scheme of the badminton flight force detection method based on deep learning, wherein: the three-axis acceleration of the ball end is integrated, the velocity change curve is obtained by combining the zero speed constraint, the linear velocity of the arm end is calculated according to the six-degree-of-freedom motion data of the wearing end, and the hitting window data containing speed information is generated, and the specific steps are as follows, The three-axis acceleration of the ball end is offset corrected and numerically integrated in the hitting window, and the zero speed constraint is performed by combining the zero speed baseline of the average speed of the preparation segment, the synthesized acceleration after the constraint is calculated, and the velocity change curve is generated; The linear velocity of the arm end is calculated according to the six-degree-of-freedom motion data of the wearing end, and is aligned with the velocity change curve sample by sample to generate the aligned linear velocity time sequence of the arm end; The hitting window data, the velocity change curve and the linear velocity time sequence of the arm end are uniformly packaged in the hitting window time range to generate the hitting window data containing speed information.

[0010] As a preferred scheme of the badminton flight force detection method based on deep learning, wherein: the hitting window data containing speed information is converted into four recursive graphs and short-time spectrum graphs, and the specific steps are as follows, The synthesized acceleration, the instantaneous amplitude of the wearing end angular velocity, the velocity change curve and the linear velocity time sequence of the arm end are extracted from the hitting window data containing speed information to generate the hitting dynamics time sequence; The hitting dynamics time sequence is resampled and amplitude normalized, the distance matrix of time points is calculated, and four recursive graphs are obtained by converting distance to gray scale.

[0011] As a preferred scheme of the badminton flight force detection method based on deep learning, wherein: the unverified force image tensor is spliced, and the specific steps are as follows, The resampled and amplitude normalized hitting dynamics time sequence is overlapped using the Hanning window, and the time-varying energy distribution is calculated to generate a short-time spectrum graph; The four recursive graphs and the short-time spectrum graph are spliced to generate a preliminary force image tensor.

[0012] As a preferred scheme of the badminton flight force detection method based on deep learning, wherein: after phase consistency and edge checking, the force image tensor is obtained, and the specific steps are as follows, Based on the unverified force image tensor, multi-scale and multi-directional phase consistency measurement is performed, and a phase consistency graph is calculated; According to the phase consistency graph, low consistency areas are excluded, and after performing edge connectivity statistics based on connected domain, pixel filling and amplitude stretching are performed to generate a force image tensor.

[0013] As a preferred scheme of the badminton flight force detection method based on deep learning, the force image tensor is input into the deep learning network combining convolution and attention for feature extraction, and the specific steps are as follows, The force image tensor is uniformly sized and stretched in amplitude, and the speed change curve is aligned to the time axis of the force image tensor. After amplitude stretching, a uniform force image tensor and speed change curve are generated. The uniform force image tensor is input into the deep learning network combining convolution and attention, and the deep feature map is output. The global representation vector is generated by globally averaging the deep feature map. The global representation vector is input into the regression head, and the preliminary badminton hitting force result is output.

[0014] As a preferred scheme of the badminton flight force detection method based on deep learning, the speed change curve is combined to fit the air resistance and effective mass parameters, and the momentum conservation relationship is constrained in the differentiable physical layer to obtain the physically modified badminton hitting force result and generate a detection report, and the specific steps are as follows, The fixed time range of the deceleration segment after impact in the early stage is intercepted from the uniform speed change curve, and the minimum residual criterion is combined for joint fitting to generate air resistance and effective mass parameters. The preliminary badminton hitting force result, air resistance and effective mass parameters, and uniform speed change curve are input into the differentiable physical layer, and the physical residual is calculated according to the momentum conservation relationship. The preliminary badminton hitting force result is physically modified using the physical residual to generate the physically modified badminton hitting force result. The physically modified badminton hitting force result, corresponding timestamp, speed change curve summary index, and air resistance and effective mass parameters are combined into a detection report.

[0015] As a preferred scheme of the badminton flight force detection method based on deep learning, the deep learning network combining convolution and attention is composed of shallow convolution blocks, middle convolution blocks and deep convolution blocks, and channel attention and spatial attention are added at the tail of the convolution block. The regression head is composed of two layers of fully connected structures, the first layer of fully connected structure is used for nonlinear mapping, and the second layer of fully connected structure is used for numerical regression.

[0016] The present application has the beneficial effects that: by converting the ball hitting window data containing speed information into a recurrence plot and a short-time spectrum plot and splicing into a force image tensor, visual representation and deep feature extraction of multi-source motion signals are realized, and the accuracy of the ball hitting force detection is improved; and the air resistance and effective mass parameters are fitted in combination with the speed change curve, and the network output is constrained in the differentiable physical layer in the momentum conservation relationship, and the physical consistency and stability of the force result are realized. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Fig. 1 The flowchart of the badminton flight force detection method based on deep learning.

[0019] Fig. 2 The flowchart of the ball hitting window clipping and the ball hitting window data containing speed information generation.

[0020] Fig. 3 The flowchart of the force image tensor construction and the phase consistency and edge checking.

[0021] Fig. 4 The flowchart of the deep learning network combined with convolution and attention and the differentiable physical layer constraint. DETAILED DESCRIPTION

[0022] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification.

[0023] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0024] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0025] REFERENCE Figs. 1-4For an embodiment of the present application, the embodiment provides a badminton flight strength detection method based on deep learning, comprising the following steps: S1: Collect ball-end three-axis acceleration and wearing-end six-degree-of-freedom motion data, time align through Bluetooth time marking and impact peak cross-correlation, and generate multi-source synchronous data; S1.1: Collect ball-end three-axis acceleration and wearing-end six-degree-of-freedom motion data, and record sampling time stamp and Bluetooth receiving time stamp, and generate double-time-marked data sequence; Further, the ball-end sensor is fixedly installed in the bottom of the badminton holder, for collecting ball-end three-axis acceleration, and the sampling directions correspond to the longitudinal direction, the transverse direction and the vertical direction of the badminton. The wearing-end sensor is fixed on the forearm of the athlete close to the wrist position, for collecting wearing-end six-degree-of-freedom motion data, which includes three-axis acceleration and three-axis angular velocity. The ball-end sensor and the wearing-end sensor record the sampling time stamp of each frame of data at a fixed sampling frequency, and transmit to the application end through the Bluetooth channel. The application end records the Bluetooth receiving time stamp when receiving each frame of data, and binds the sampling time stamp with the Bluetooth receiving time stamp, to generate a double-time-marked data sequence containing the sampling time stamp and the Bluetooth receiving time stamp.

[0026] S1.2: Coarsely align the double-time-marked data sequence according to the Bluetooth receiving time stamp, to generate preliminary synchronous data; Further, the application end reads the double-time-marked data sequence containing the ball-end three-axis acceleration and the double-time-marked data sequence containing the wearing-end six-degree-of-freedom motion data, respectively, sorts them in the order of the Bluetooth receiving time stamp, and eliminates repeated or reverse frames. The application end takes the Bluetooth receiving time stamp as a unified reference, establishes a time corresponding relationship between the ball-end three-axis acceleration and the wearing-end six-degree-of-freedom motion data, matches the frames closest in time according to the nearest neighbor principle, and fills in the time gaps through linear interpolation. The ball-end three-axis acceleration and the wearing-end six-degree-of-freedom motion data are corresponded on the unified time axis, to form preliminary synchronous data.

[0027] S1.3: Monitor the impact peak position from the ball-end three-axis acceleration, and monitor the swing peak position from the wearing-end six-degree-of-freedom motion data, and calculate the cross-correlation function between the impact peak and the swing peak; Further, the application end extracts the ball end three-axis acceleration from the preliminary synchronization data, calculates the ball end synthetic acceleration, and performs sliding median filtering to weaken high-frequency noise, detects the maximum amplitude time point in the filtered ball end synthetic acceleration sequence as the impact peak position. The application end extracts the three-axis angular velocity in the six-degree-of-freedom motion data of the wearing end from the preliminary synchronization data, calculates the wearing end angular velocity instantaneous amplitude, and detects the time point with the maximum change rate as the swing peak position. The application end intercepts the ball end synthetic acceleration sequence containing the impact peak position and the wearing end angular velocity instantaneous amplitude sequence containing the swing peak position, performs mean value processing, and then calculates the cross-correlation function, the expression is: ; Wherein, is the cross-correlation function between the impact peak and the swing peak at a time offset , is the value of the ball end synthetic acceleration at the th sampling point on the unified time axis, is the value of the wearing end angular velocity instantaneous amplitude at the th sampling point after a time offset , is the mean value of the ball end synthetic acceleration sequence, is the mean value of the wearing end angular velocity instantaneous amplitude sequence, is the total length of the sequence samples used for cross-correlation calculation, is the time offset.

[0028] S1.4: Take the time offset corresponding to the maximum value of the cross-correlation function between the impact peak and the swing peak as the fine alignment parameter; Further, the application end reads the cross-correlation function of the ball end synthetic acceleration and the wearing end angular velocity instantaneous amplitude, obtains the time offset corresponding to the maximum value of the cross-correlation function, and determines the time offset as the fine alignment parameter. The positive sign of the fine alignment parameter indicates that the six-degree-of-freedom motion data of the wearing end is delayed relative to the ball end three-axis acceleration, and the negative sign indicates that the six-degree-of-freedom motion data of the wearing end is advanced relative to the ball end three-axis acceleration. The application end records the fine alignment parameter, the impact peak position timestamp and the swing peak position timestamp, and generates the fine alignment control field.

[0029] S1.5: Use the fine alignment parameter to perform time translation correction on the preliminary synchronization data to generate time-aligned multi-source synchronization data; Further, the application end reads the preliminary synchronization data and the fine alignment control field, and adjusts the time index of the six-degree-of-freedom motion data of the wearable end on the unified time axis according to the fine alignment parameter. When the fine alignment parameter is a positive sign, the six-degree-of-freedom motion data of the wearable end is translated backward as a whole; when the fine alignment parameter is a negative sign, the six-degree-of-freedom motion data of the wearable end is translated forward as a whole. After translation, the application end performs linear resampling on the six-degree-of-freedom motion data of the wearable end, so that the sampling interval is consistent with the three-axis acceleration of the ball end. The application end re-pairs the six-degree-of-freedom motion data of the wearable end and the three-axis acceleration of the ball end after time translation correction on the unified time axis, records a new unified time stamp, and generates multi-source synchronization data after time alignment.

[0030] S2: Detecting the ball impact starting point in the multi-source synchronization data and cutting the ball window, integrating the three-axis acceleration of the ball end, obtaining the velocity change curve combined with the zero speed constraint, and calculating the linear velocity of the hand end according to the six-degree-of-freedom motion data of the wearable end to generate the ball window data containing velocity information; S2.1: Extracting the three-axis acceleration of the ball end and the six-degree-of-freedom motion data of the wearable end from the multi-source synchronization data; Further, the three-axis acceleration of the ball end is stored independently in three directions and retains the time index. The six-degree-of-freedom motion data of the wearable end includes three-axis acceleration and three-axis angular velocity, which are arranged in continuous sequence according to the channel and time index.

[0031] S2.2: Calculating the synthetic acceleration according to the three-axis acceleration of the ball end, and performing sliding median filtering to obtain a robust threshold; Further, the synthetic acceleration is calculated sample by sample according to the three-axis acceleration of the ball end, and the square sum of the acceleration in three directions is squared to obtain the synthetic acceleration sequence. The sliding median filtering is performed on the synthetic acceleration sequence to suppress impulse noise and short-time spikes. The median and the absolute median difference of the near-time period are calculated on the synthetic acceleration sequence after sliding median filtering, and the robust threshold is generated according to the median and the absolute median difference, with a value range of one to three times the absolute median difference.

[0032] S2.3: Taking the time point when the synthetic acceleration exceeds the robust threshold for the first time as the ball impact starting point; Further, the time point when the synthetic acceleration exceeds the robust threshold for the first time is detected in time sequence on the synthetic acceleration sequence. In order to reduce the influence of incidental noise, it is required that the synthetic acceleration remains higher than the robust threshold for a plurality of continuous sampling time points to determine the ball impact starting point.

[0033] S2.4: Centering on the time point of the start of the impact, the multi-source synchronous data is divided into a preparation segment and an early flight segment respectively, a hitting window is generated, and hitting window data and hitting window time range are obtained; Further, centering on the time point of the start of the impact, the preparation segment is cut forward and the early flight segment is cut backward on the unified time axis. The preparation segment and the early flight segment are spliced to form a continuous time range covering the preparation stage, the impact stage and the early flight stage, which is defined as the hitting window time range. The ball-end three-axis acceleration and the wearing-end six-degree-of-freedom motion data are synchronized and cropped according to the hitting window time range to form the hitting window data.

[0034] S2.5: In the hitting window, the ball-end three-axis acceleration is corrected and numerically integrated, and the zero-speed constraint is performed with the average speed of the preparation segment as the zero-speed baseline to calculate the constrained synthetic acceleration and generate the speed change curve; Further, in the hitting window, the ball-end three-axis acceleration is corrected, specifically, the average values of the three-direction accelerations are calculated within the preparation segment, and the corresponding average values are subtracted from the three-direction accelerations of the hitting window sample by sample. The three-direction accelerations after bias correction are numerically integrated to obtain three-direction velocity sequences and calculate the synthetic velocity to obtain the velocity sequence. The average synthetic velocity of the preparation segment is taken as the zero-speed baseline to perform the zero-speed constraint, so that the synthetic velocity of the preparation segment returns to zero level. The constrained synthetic velocity is time-differentiated to obtain the constrained synthetic acceleration, and the zero-speed constrained synthetic velocity is taken as the speed change curve.

[0035] S2.6: Calculate the hand end linear velocity according to the wearing-end six-degree-of-freedom motion data, and align it with the speed change curve sample by sample to generate the aligned hand end linear velocity time sequence; Further, the product of the angular velocity instantaneous amplitude calculated from the wearing-end three-axis angular velocity and the equivalent radius of the wrist joint to the racket force point in the registered human individual parameters at the application end is taken as the hand end linear velocity time sequence. The hand end linear velocity time sequence is aligned with the speed change curve sample by sample according to the unified time axis, and the aligned hand end linear velocity time sequence is output.

[0036] S2.7: The hitting window data, the speed change curve and the hand end linear velocity time sequence are unified and encapsulated within the hitting window time range to generate the hitting window data containing speed information; Further, the hitting window data, the speed change curve and the aligned hand end linear velocity time sequence are unified and encapsulated according to the hitting window time range. The encapsulation content includes unified time axis index and corresponding channel data, and keeps the field name and order consistent, forming a single hitting entry containing complete kinematic information, and outputting the hitting window data containing speed information.

[0037] S3: Convert the ball hitting window data containing speed information into four recursive graphs and short-time frequency spectrum graphs, and splice them into an unverified force image tensor, and obtain the force image tensor after phase consistency and edge verification; S3.1: Extract the synthesized acceleration, wearable end angular velocity instantaneous amplitude, speed change curve, and arm end linear velocity time series from the ball hitting window data containing speed information to generate a ball hitting dynamics time series; Further, the synthesized acceleration, wearable end angular velocity instantaneous amplitude, speed change curve, and arm end linear velocity time series are sequentially extracted from the ball hitting window data containing speed information according to a unified time axis, and integrity check is performed, samples with missing time indexes are deleted and single-point gaps are filled in the channel by linear interpolation, to ensure that each time index is defined, and arranged in a fixed order as a synthesized acceleration channel, a wearable end angular velocity instantaneous amplitude channel, a speed change curve channel, and an arm end linear velocity time series channel, to output a ball hitting dynamics time series.

[0038] S3.2: Resample and amplitude normalize the ball hitting dynamics time series, calculate the pairwise distance matrix of time points, and obtain four recursive graphs by converting distance to gray scale; Further, the ball hitting dynamics time series is resampled on a unified time axis to obtain a consistent sampling frequency, and amplitude normalization is performed on each channel to map the minimum value to zero and the maximum value to one. The pairwise distance matrix of time points is calculated on each channel, the distance measure adopts the absolute value of point value difference, and the absolute value is mapped to gray scale in a linear proportion, to obtain a gray matrix consistent with the time length, as four recursive graphs.

[0039] S3.3: The resampled and amplitude-normalized ball hitting dynamics time series is overlapped using a Hanning window, and the time-varying energy distribution is calculated to generate a short-time frequency spectrum graph; Further, the resampled and amplitude-normalized ball hitting dynamics time series is overlapped and framed using a Hanning window on each channel, the frequency energy distribution of each frame is calculated and arranged in time sequence to form a two-dimensional energy time-frequency plane, which is stretched in the frequency direction and the time direction respectively to keep consistent with the size of the recursive graph, and the two-dimensional energy time-frequency planes of the four channels are combined and output in a fixed channel order to generate a short-time frequency spectrum graph.

[0040] S3.4: Channel splice four recursive graphs and short-time frequency spectrum graphs to generate a preliminary force image tensor; Further, the four recursive graphs and the short-time spectrograms are spliced in a fixed order to form a preliminary force image tensor with a multi-channel structure. The splicing order is: synthesized acceleration recursive graph, wearing end angular velocity instantaneous amplitude recursive graph, speed change curve recursive graph, arm end linear velocity time series recursive graph, synthesized acceleration short-time spectrogram, wearing end angular velocity instantaneous amplitude short-time spectrogram, speed change curve short-time spectrogram, and arm end linear velocity time series short-time spectrogram.

[0041] S3.5: Normalizing the preliminary force image tensor to generate an unverified force image tensor; Further, the preliminary force image tensor is subjected to amplitude normalization in the channel dimension to linearly map the pixel intensity of each channel to the interval of zero to one, maintain the amplitude scale consistency between channels, and preserve the channel order unchanged, thereby outputting the unverified force image tensor.

[0042] S3.6: Based on the unverified force image tensor, multi-scale and multi-directional phase consistency is measured, and a phase consistency graph is calculated. Further, multi-scale and multi-directional phase consistency is measured on each channel based on the unverified force image tensor, a set of scale and direction bandpass filter kernels are used to calculate local phase consistency responses in each channel and aggregate results in scale and direction, thereby obtaining a phase consistency graph consistent with the channel size, and outputting a set of phase consistency graphs in the order of channels.

[0043] S3.7: According to the phase consistency graph, low consistency regions are screened out, and after performing edge connectivity statistics based on connected domains, pixel filling and amplitude stretching are performed to generate a force image tensor. Further, according to the phase consistency graph, low consistency regions are screened out on each channel and a binary mask is generated, isolated small domains are removed using edge connectivity statistics based on connected domains, pixel filling is performed on the mask coverage area using a bilinear method of adjacent valid pixels, and after completing pixel filling in all channels, the unverified force image tensor is subjected to amplitude stretching again to restore the channel contrast, thereby outputting the force image tensor.

[0044] S4: The force image tensor is input into a deep learning network combining convolution and attention for feature extraction, the speed change curve is fitted to the air resistance and effective mass parameters, and the momentum conservation relationship is used to constrain in the differentiable physical layer, thereby obtaining a physically corrected badminton hitting force result and generating a detection report. S4.1: The force image tensor is subjected to uniform size and amplitude stretching, and the speed change curve is aligned to the time axis of the force image tensor, thereby generating a uniform force image tensor and speed change curve after amplitude stretching; Further, the force image tensor is bilinearly interpolated in the spatial dimension to unify the height and width to a fixed size while keeping the channel order unchanged. The force image tensor is amplitude stretched in the channel dimension to map the amplitude of each channel pixel to a unified range. The speed change curve is read and aligned according to the unified time axis corresponding to the force image tensor, and the amplitude stretching is performed on the speed change curve to eliminate the dimensional difference, thereby obtaining the unified force image tensor and the speed change curve.

[0045] S4.2: input the unified force image tensor into the deep learning network combining convolution and attention, and output a deep feature map; Further, the unified force image tensor is input into the deep learning network combining convolution and attention, local edges and textures are extracted through shallow convolution blocks, multi-scale structural features are extracted through middle convolution blocks, channel attention and spatial attention are added at the end of each convolution block to strengthen the texture area related to impact, and finally stable high-level representations are obtained through deep convolution blocks, thereby outputting a deep feature map.

[0046] It should be noted that the construction process of the deep learning network combining convolution and attention includes setting a unified size force image tensor as input at the input end, and sequentially stacking shallow convolution blocks, middle convolution blocks and deep convolution blocks; wherein the shallow convolution block is composed of two convolution layers and one nonlinear activation layer, and is used to extract local edges and fine-grained texture features in the force image tensor; the middle convolution block is composed of multi-scale convolution kernels and stride convolution layers, and is used to capture multi-scale spatial structural features caused by changes in speed and posture during the hitting process; the deep convolution block is composed of multiple stacked convolution layers and residual connections, and is used to aggregate global semantic features at different scales while keeping the gradient stable, so that the output feature map can comprehensively reflect the overall spatial distribution of the hitting force, and the channel attention and spatial attention are added at the end of each convolution block, respectively, so that the deep learning network combining convolution and attention can assign feature weights to highlight the regions related to the hitting force.

[0047] The training process of the deep learning network combining convolution and attention is to input the sample force image tensor containing the real force label into the network, use the mean square error function as the loss function, obtain the predicted force result through forward propagation, update the convolution layer and attention weight parameters based on the gradient descent algorithm in the backward propagation, and repeat the training until the loss function converges, thereby obtaining the deep learning network combining convolution and attention which can stably extract force features and accurately predict the hitting force.

[0048] S4.3: perform global average pooling on the deep feature map to generate a global representation vector; Further, global average pooling is performed on the deep feature map in the spatial dimension, and the spatial information of each channel is compressed into a single statistic, which is spliced into a vector arranged by channel to obtain a global feature vector.

[0049] S4.4: The global feature vector is input into the regression head, and the preliminary badminton hitting force result is output. Further, the global feature vector is input into the regression head. The regression head adopts a two-layer fully connected structure. The first layer of fully connected structure completes nonlinear transformation to improve the representation ability, and the second layer of fully connected structure completes numerical mapping to output the preliminary badminton hitting force result of a single hitting. The preliminary badminton hitting force result and the corresponding time index are recorded for physical constraint.

[0050] S4.5: A fixed time range of the early deceleration segment after impact is intercepted from the unified speed change curve, and a joint fitting is performed combined with the minimum residual criterion to generate the air resistance and effective mass parameters. Further, the hitting impact starting time index and the unified time axis are read from the unified speed change curve. After the hitting impact starting point, the first-order difference is calculated along the unified time axis. When the first-order difference transitions from positive to non-positive, and the first-order difference near the peak value remains non-positive within a short-time sliding window, the time index of the maximum speed is marked as the speed peak time index. The starting point is the speed peak time index, and the ending point is the time index at which any of the following conditions occurs for the first time. Condition one: the absolute value of the first-order difference is lower than the median statistic of the absolute value of the first-order difference in the speed peak vicinity short-time sliding window within the short-time sliding window. Condition two: the sign of the first-order difference appears positive and negative mixed within the short-time sliding window, which breaks the monotone decrease. The earliest time index that meets the condition is taken as the ending time index. The speed sequence between the speed peak and the ending time index is taken as the early deceleration segment and the corresponding unified time axis index is reserved to complete the determination of the early deceleration segment.

[0051] The first-order difference of the unified speed change curve is estimated to obtain the instantaneous deceleration in the early deceleration segment, and the median of the short sliding window is used to smooth the difference noise, so as to obtain the corresponding sample of the deceleration time sequence and the speed time sequence. The joint expression of air resistance and effective mass is established, which includes two parts. Part one is the air dynamic deceleration consistency expression, which aims to minimize the quadratic speed term relationship residual between the deceleration time sequence and the speed time sequence, so as to identify the ratio of the air resistance parameter and the effective mass parameter. Part two is the momentum consistency expression, which aims to minimize the difference between the speed change and the initial impact momentum change in the early deceleration segment, so as to identify the effective mass parameter. The residual of the air dynamic deceleration consistency expression and the residual of the momentum consistency expression are combined into the total residual according to the fixed weight, and the joint solution is carried out by using the minimum residual criterion. In the solving process, the parameter space is constrained by monotonicity and non-negativity to ensure physical interpretability, and the air resistance parameter and the effective mass parameter are output, which completes the generation of the air resistance and the effective mass parameter.

[0052] The length of the short sliding window is fixed after registration according to the sampling frequency and the statistical length of the hitting window, which is used to ensure that the determination of the speed peak and the early deceleration segment does not depend on specific numerical threshold and has reproducibility. The initial impact momentum change used in the momentum consistency expression is composed of the speed difference near the speed peak and the effective mass parameter. The identifiable separation between the air resistance parameter and the effective mass parameter is realized by minimizing the total residual in the joint solution.

[0053] S4.6: input the preliminary badminton hitting force result, the air resistance and the effective mass parameter, and the unified speed change curve into the differentiable physical layer, and calculate the physical residual according to the momentum conservation relationship; Further, the preliminary badminton hitting force result, the air resistance and the effective mass parameter, and the unified speed change curve are simultaneously input into the differentiable physical layer. The differentiable physical layer establishes the mechanical constraint between the preliminary badminton hitting force result and the speed change curve according to the momentum conservation relationship, calculates the physical residual, and the physical residual is used to measure the consistency degree of the preliminary badminton hitting force result and the momentum conservation relationship.

[0054] It should be noted that the micro-physical layer is a computational layer that embeds physical constraint relationships in a deep learning network. The core is to add physical laws in a computable form to the deep learning network structure while maintaining the gradient transferable. The micro-physical layer is based on the momentum conservation relationship, taking the preliminary badminton hitting strength result, air resistance parameter, effective mass parameter and unified speed change curve as input. The physical residual between the theoretical momentum change and the predicted momentum change is calculated, and the derivative of the calculation process to the network parameters is kept continuous, so that the physical constraint can participate in the back propagation training. In this way, the micro-physical layer considers both data fitting error and physical consistency in the optimization process of the deep learning network, so that the output badminton hitting strength result not only conforms to the observation data trend, but also meets the aerodynamics and momentum conservation principle.

[0055] It should be noted that the core of the physical law is the law of conservation of momentum, that is, in the process of hitting the ball, the total momentum change of the system after the badminton and racket contact is equal to the integral of external force, and the external force mainly includes air resistance and gravity component. In the calculation, the preliminary badminton hitting strength result is regarded as the initial momentum output at the moment of hitting the ball, the air resistance parameter and the effective mass parameter are substituted into the air dynamics deceleration expression, and the theoretical momentum decay curve is calculated by time integration of the speed change curve. The momentum conservation relationship is used to ensure that the theoretical momentum decay curve is consistent with the momentum change process predicted by the deep learning network, thereby forming a physical residual.

[0056] S4.7: using the physical residual to physically correct the preliminary badminton hitting strength result to generate a physically corrected badminton hitting strength result; Further, the physical residual is read and a physical correction is performed on the preliminary badminton hitting strength result. The physical correction adjusts the numerical value of the preliminary badminton hitting strength result by weighting the physical residual, so that the adjusted value is more consistent with the momentum conservation relationship. The physically corrected badminton hitting strength result is output, and the corresponding time index is recorded to ensure consistency with the unified speed change curve.

[0057] S4.8: combining the physically corrected badminton hitting strength result, the corresponding timestamp, the speed change curve summary index, and the air resistance and effective mass parameters into a detection report; Further, the physically corrected badminton hitting strength result is paired with the corresponding timestamp, the summary index of the unified speed change curve in the continuous time range is calculated, and the air resistance and effective mass parameters are arranged together into a structured item. The structured items are combined in time sequence into a detection report, which includes the physically corrected badminton hitting strength result, the corresponding timestamp, the speed change curve summary index, and the air resistance and effective mass parameters.

[0058] To sum up, the application realizes visual representation and deep feature extraction of multi-source motion signals by converting the ball hitting window data containing speed information into a recurrence plot and a short-time spectrum and splicing them into a force image tensor, and improves the accuracy of the ball hitting force detection; meanwhile, the speed change curve is combined to fit the air resistance and effective mass parameters, and the network output is constrained by the momentum conservation relationship in the differentiable physical layer, thereby realizing the physical consistency and stability of the force result.

[0059] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A badminton flight strength detection method based on deep learning, characterized by: The application relates to a method for detecting badminton hitting force, and belongs to the technical field of sports science. Ball-end three-axis acceleration and wearing-end six-degree-of-freedom motion data are collected, time alignment is performed through Bluetooth time marking and impact peak cross-correlation, and multi-source synchronous data is generated; The hitting impact starting point is detected in the multi-source synchronous data, and a hitting window is cropped, the ball-end three-axis acceleration is integrated, the speed change curve is obtained in combination with the zero-speed constraint, the arm end linear velocity is calculated according to the wearing-end six-degree-of-freedom motion data, and the hitting window data containing the speed information is generated; The hitting window data containing the speed information is converted into four recursive graphs and a short-time frequency spectrum graph, and is spliced into an unverified force image tensor, the force image tensor is obtained after phase consistency and edge verification, The force image tensor is input into a deep learning network combining convolution and attention for feature extraction, the speed change curve is fitted to air resistance and effective mass parameters, and the momentum conservation relationship is used for constraint in the differentiable physical layer, and the badminton hitting force result is obtained after physical correction, and a detection report is generated.

2. The badminton flight strength detection method based on deep learning according to claim 1, characterized in that: The ball-end three-axis acceleration and wearing-end six-degree-of-freedom motion data are collected, time alignment is performed through Bluetooth time marking and impact peak cross-correlation, and multi-source synchronous data is generated, and the specific steps are as follows, Ball-end three-axis acceleration and wearing-end six-degree-of-freedom motion data are collected, and sampling time stamps and Bluetooth receiving time stamps are recorded to generate double-time-marked data sequences; The double-time-marked data sequences are roughly aligned according to the Bluetooth receiving time stamps to generate preliminary synchronous data; The impact peak position is monitored from the ball-end three-axis acceleration, and the swing peak position is monitored from the wearing-end six-degree-of-freedom motion data, and the cross-correlation function between the impact peak and the swing peak is calculated; The time offset corresponding to the maximum value of the cross-correlation function between the impact peak and the swing peak is used as the fine alignment parameter; The preliminary synchronous data is time-shifted and corrected using the fine alignment parameter to generate time-aligned multi-source synchronous data.

3. The badminton flight strength detection method based on deep learning according to claim 2, characterized in that: The hitting impact starting point is detected in the multi-source synchronous data, and a hitting window is cropped, the ball-end three-axis acceleration is integrated, the speed change curve is obtained in combination with the zero-speed constraint, the arm end linear velocity is calculated according to the wearing-end six-degree-of-freedom motion data, and the hitting window data containing the speed information is generated, and the specific steps are as follows, The ball-end three-axis acceleration and wearing-end six-degree-of-freedom motion data are extracted from the multi-source synchronous data; The synthesized acceleration is calculated according to the ball-end three-axis acceleration, and is subjected to sliding median filtering to obtain a robust threshold; The time point when the synthesized acceleration exceeds the robust threshold for the first time is taken as the hitting impact starting point; The multi-source synchronous data is divided into a preparation segment and an early flight segment forward and backward respectively with the time point of the hitting impact starting point as the center to generate a hitting window, and the hitting window data and the hitting window time range are obtained.

4. The badminton flight strength detection method based on deep learning according to claim 3, characterized in that: The ball-end three-axis acceleration is integrated, the speed change curve is obtained in combination with the zero-speed constraint, and the arm end linear velocity is calculated according to the wearing-end six-degree-of-freedom motion data to generate the hitting window data containing the speed information, and the specific steps are as follows, The ball-end three-axis acceleration is subjected to bias correction and numerical integration in the hitting window, and the zero-speed constraint is performed in combination with the zero-speed baseline of the speed average of the preparation segment to calculate the constrained synthesized acceleration and generate the speed change curve; The arm end linear velocity is calculated according to the wearing-end six-degree-of-freedom motion data, and is aligned with the speed change curve sample by sample to generate the aligned arm end linear velocity time sequence; The ball hitting window data, the speed change curve and the hand arm end linear velocity time sequence are uniformly packaged in the ball hitting window time range to generate the ball hitting window data containing speed information.

5. The badminton flight strength detection method based on deep learning according to claim 4, characterized in that: The ball hitting window data containing speed information is converted into four recursive graphs and a short-time frequency spectrum graph, and the specific steps are as follows, The synthesized acceleration, the wearing end angular velocity instantaneous amplitude, the speed change curve and the hand arm end linear velocity time sequence are extracted from the ball hitting window data containing speed information to generate the ball hitting dynamics time sequence. The ball hitting dynamics time sequence is resampled and amplitude normalized, the distance matrix of time points is calculated, and four recursive graphs are obtained through the distance to gray conversion.

6. The badminton flight strength detection method based on deep learning according to claim 5, characterized in that: The resampled and amplitude normalized ball hitting dynamics time sequence is overlapped using the Hanning window, and the time-varying energy distribution is calculated to generate a short-time frequency spectrum graph. The four recursive graphs and the short-time frequency spectrum graph are channel spliced to generate a preliminary force image tensor. After phase consistency and edge checking, the force image tensor is obtained, and the specific steps are as follows, 7. The badminton flight strength detection method based on deep learning according to claim 6, characterized in that: Based on the unverified force image tensor, multi-scale and multi-directional phase consistency is measured, and a phase consistency graph is calculated. According to the phase consistency graph, low consistency areas are screened out, edge connectivity statistics based on connected domains are performed, pixel filling and amplitude stretching are performed to generate a force image tensor. The specific steps of the feature extraction of the force image tensor input into the deep learning network combined with convolution and attention are as follows, 8. The badminton flight strength detection method based on deep learning according to claim 7, characterized in that: The force image tensor is uniformly sized and amplitude stretched, and the speed change curve is aligned to the time axis of the force image tensor. After amplitude stretching, a uniform force image tensor and speed change curve are generated. The uniform force image tensor is input into the deep learning network combined with convolution and attention, and a deep feature map is output. The global average of the deep feature map is aggregated to generate a global representation vector. The global representation vector is input into the regression head to output a preliminary badminton hitting force result. The specific steps of combining the speed change curve to fit the air resistance and effective mass parameters, and constraining them in the differentiable physical layer based on the momentum conservation relationship to obtain the physically modified badminton hitting force result and generate a detection report are as follows, 9. The badminton flight strength detection method based on deep learning according to claim 8, characterized in that: The fixed time range of the early deceleration segment after impact is intercepted from the uniform speed change curve, and the air resistance and effective mass parameters are jointly fitted based on the minimum residual criterion to generate the air resistance and effective mass parameters. The preliminary badminton hitting force result, the air resistance and effective mass parameters, and the uniform speed change curve are input into the differentiable physical layer, and the physical residual is calculated based on the momentum conservation relationship. The preliminary badminton hitting force result is physically modified using the physical residual to generate a physically modified badminton hitting force result. The physically modified badminton hitting force result, the corresponding timestamp, the speed change curve summary index and the air resistance and effective mass parameters are combined into a detection report. The deep learning network combined with convolution and attention is composed of shallow convolution blocks, middle convolution blocks and deep convolution blocks, and channel attention and spatial attention are added at the tail of the convolution blocks.

10. The badminton flight strength detection method based on deep learning according to claim 8, characterized in that: ​ The regression head is composed of two full connection structures, the first full connection structure is used for nonlinear mapping, and the second full connection structure is used for numerical regression.

Images