Bed tail non-contact type periodic sleep limb dyskinesia and sleep body position combined monitoring device

Abstract

The invention discloses a bed tail non-contact type periodic sleep limb dyskinesia and sleep position combined monitoring device. The device is implemented as follows: collecting multi-view continuous echo signals of leg actions and sleep positions of a human body; signal processing is carried out on the echo signals, and fragment segmentation is carried out through a dynamic adjustment window algorithm; time-distance-frequency multi-spectral-domain joint analysis is carried out on the multi-view-angle echo signal segments, time-frequency diagrams and time-distance diagrams under different view angles are generated, then the time-frequency diagrams and the time-distance diagrams are input into a sleep body movement identification model based on a multi-view-angle multi-task architecture, and sleep leg movement type and sleep posture type judgment results are output; judging whether the judgment result of the sleep leg movement meets the characteristics of periodic limb movement or not, and if yes, marking as a periodic leg movement sequence; and whether sleep periodic limb dyskinesia exists or not is judged. According to the invention, joint identification and monitoring of periodic sleep limb dyskinesia and sleep body positions are realized, and a flexible and convenient monitoring means is provided for home intelligent sleep monitoring.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical sleep monitoring, and in particular relates to a non-contact periodic sleep limb movement disorder and sleep posture combined monitoring device at the foot of the bed. Background Art

[0002] Periodic Limb Movement Disorder (PLMD) is a sleep disorder that occurs during sleep. The main symptom is the repeated occurrence of periodic and involuntary limb movements during sleep, usually manifested as extension of the big toe and dorsiflexion of the ankle joint, and sometimes involving the knee and hip joints. Periodic sleep limb movement disorder has multiple harms. In terms of sleep, it will disrupt the continuity and stability of sleep, resulting in a lack of deep sleep and causing chronic fatigue. In daily life, it causes daytime sleepiness, affects work and study, triggers emotional problems, and leads to a decline in cognitive function. In terms of health, it increases the risk of cardiovascular diseases, weakens the immune system function, and disrupts the balance of the endocrine system. Sleep posture is an important piece of information in sleep monitoring and is of great significance for aspects such as the health assessment of subjects. In terms of health assessment, it can assist in diagnosing diseases such as sleep apnea syndrome. Different postures affect the condition, and it can also evaluate the cardiovascular burden; it can warn of health problems caused by long-term bad postures. In terms of improving sleep quality, it can select a suitable bedding according to the posture preference to optimize the environment, adjust habits, and adjusting the posture for those with sleep disorders may help sleep. In terms of safety guarantee, it can prevent accidents for special populations such as the elderly and pregnant women, and also provide nursing support for rehabilitation patients to ensure the correct posture and promote rehabilitation.

[0003] In the field of clinical medicine, periodic sleep limb movement disorder and sleep posture are usually monitored by polysomnography (PSG). This monitoring method has the following disadvantages: (1) High operation difficulty. Polysomnography is a professional medical device that includes various types of sensitive sensors and electrodes. The attachment methods of different types of sensors are very different, so professional sleep medicine technicians are required to operate and implement it; (2) High test cost. The PSG device itself is expensive, and only a few sleep centers have it. Moreover, it requires professional medical technicians to operate, resulting in a high test cost for patients; (3) Low measurement comfort. When using PSG for sleep measurement, a large number of electrodes or sensors need to be worn or attached, which is extremely likely to cause discomfort to the subjects, thus affecting their sleep quality and test results; (4) Heavy workload for result interpretation. The multiple physiological signals output by PSG now require professional medical technicians to interpret them according to relevant medical guidelines, with a heavy workload and low execution efficiency. The above problems directly prevent the PSG device from being widely promoted and applied, especially unable to meet the needs of long-term home sleep monitoring.

[0004] Radar sensors have been proven to be able to accurately perceive the motion information of human targets, with the advantages of non-contact, non-intrusive, all-day and all-weather. They have been widely used in applications such as human target positioning, human posture recognition, and gesture recognition. Artificial intelligence technology can automatically classify results and identify targets based on radar echo data, thereby reducing the difficulty and workload of data interpretation and mining more in-depth hidden information from radar echo information.

[0005] In summary, placing a radar sensor at the foot of the bed facing the human body and supplemented with artificial intelligence can achieve non-contact monitoring and identification of periodic sleep limb movement disorders and sleep postures. Summary of the Invention

[0006] The purpose of the present invention is to provide a non-contact combined monitoring device for periodic sleep limb movement disorders and sleep postures at the foot of the bed in view of the problems existing in the above-mentioned prior art.

[0007] The technical solution for achieving the purpose of the present invention is: a non-contact combined monitoring device for periodic sleep limb movement disorders and sleep postures at the foot of the bed, the device includes a signal acquisition module, a signal processing module, a feature extraction module, a classification module, a sleep periodic limb movement sequence judgment module, and a sleep periodic limb movement disorder judgment module;

[0008] The signal acquisition module is used to non-contact and continuously collect echo signals of human leg movements and sleeping postures from different perspectives through a number of distributed FMCW radars;

[0009] The signal processing module is used to process the radar echo signals from different perspectives and identify and segment adaptive leg movement segments through a dynamic window adjustment algorithm;

[0010] The feature extraction module is used to perform time-frequency analysis and time-distance analysis on the echo signal segments to generate time-frequency diagrams and time-distance diagrams of corresponding perspectives;

[0011] The classification module is used to input the time-frequency diagram and time-distance diagram into a deep network model to simultaneously determine the leg movement type and sleeping posture type of the corresponding segment; the deep network model is a leg movement-sleeping posture multi-task classification model based on multi-perspective and multi-spectrum domain information fusion;

[0012] The sleep periodic limb movement sequence judgment module is used to judge whether the leg movement output meets the characteristics of periodic limb movement. If it meets, mark the periodic leg movement sequence;

[0013] The sleep periodic limb movement disorder judgment module is used to judge whether there is a sleep periodic limb movement disorder.

[0014] Further, the signal acquisition module non - contactly and continuously acquires the echo signals of human leg movements and sleeping postures from different perspectives through a number of distributed FMCW radars, specifically including:

[0015] Step 2 - 1, data acquisition: The subject lies flat on the bed initially. A number of FMCW radar sensors are distributed and installed at the end of the bed, and the antennas are all oriented towards the human legs; different FMCW radar sensors acquire radar leg echo signals from different perspectives.

[0016] The echo signal at each perspective is expressed as: ), where r is the original radar signal, and represents the time sequence number of slow - time sampling, represents the total number of slow - time sampling times, represents the serial number of the range gate, and N represents the total number of range gates; and represent the sampling moments of slow - time and fast - time respectively;

[0017] Step 2 - 2, moving target detection. The collected leg echo signals R at each perspective are rearranged after fast - and - slow - time sampling to obtain the echo matrix, where X represents the number of fast - time sampling points and Y represents the number of slow - time sampling points;

[0018] Step 2 - 3, use a loop function to traverse the leg echo signal R. In each loop, the data at a fixed time interval is subtracted from the first - column data to remove background clutter interference.

[0019] Step 2 - 4, filter the leg echo signal R after removing background clutter interference to remove the tail - shadow clutter generated by moving target detection, and obtain the pre - processed leg echo signal.

[0020] Further, the signal processing module processes the radar echo signals from different perspectives and identifies and segments the adaptive leg - movement segments through a dynamic window adjustment algorithm, specifically including:

[0021] Step 3 - 1, select the leg echo signals of a fixed range gate for demodulation to obtain the phase signal ;

[0022] Step 3 - 2, motion information extraction: Perform third - order difference processing on the phase signal to obtain a further enhanced leg motion signal ;

[0023] Step 3 - 3, sliding window detection: Use a 2 sliding window to traverse the enhanced leg motion signal , calculate the variance within each window, and the current window Window average variance of the past N non-leg movement events NLM and determine the current window according to the threshold condition shown in the following formula whether the signal is a leg movement event LM, and then execute Step 3-4;

[0024]

[0025] If the signal of the current window satisfies the above formula, it indicates a leg movement event LM; if it does not satisfy the above formula, the window average variance is updated ;

[0026] wherein, is an empirical threshold, is the segment sampling length;

[0027] Step 3-4, dynamic window adjustment: If the threshold condition is satisfied in Step 3-3, further detect the peak points and valley points in the window. If there is a peak point between two adjacent valley points, it indicates a complete leg movement event LM, and save the time segment of the t seconds before or after the window where the peak point is located; if not, it indicates that the leg movement event LM is incomplete, move the dynamic window backward by t seconds, and continue to detect the peak points and valley points to ensure obtaining a complete time segment of the leg movement event LM:

[0028]

[0029] wherein, , respectively represent the nth and (n + 1)th valley points, represents the mth peak point;

[0030] If the threshold condition is not satisfied in Step 3-3, save the time segment in the window as a non-leg movement segment NLM, thereby obtaining a complete leg movement segment LM and a non-leg movement segment NLM.

[0031] Furthermore, the feature extraction module performs time-frequency analysis and time-distance analysis on the echo signal segment to generate a time-frequency diagram and a time-distance diagram of the corresponding perspective, specifically including:

[0032] Step 4-1, obtain the leg echo signal segment, perform time-frequency analysis on the complex signal of the fixed range gate through short-time Fourier transform, and obtain a complete time-frequency diagram reflecting the positive and negative frequency changes;

[0033] Step 4-2, obtain the leg echo signal segment, and obtain a high-resolution time-distance diagram through an optimized algorithm for the time-distance diagram of adaptive adjacent signal enhancement.

[0034] Furthermore, Step 4-1 specifically includes:

[0035] Step 4-1-1: Obtain an LM segment of t seconds, containing n sampling points, through a continuous monitoring algorithm for sleep limb activities based on energy envelope; take the leg movement signal lm[n] at a fixed distance gate within this time period.

[0036] Step 4-1-2: Set the window length, overlap length, and the number of points for STFT calculation, and perform short-time Fourier transform, expressed as:

[0037]

[0038] where is a rectangular window function aligned at the position and sliding according to the window length and overlap length, is the number of points for STFT calculation, is the leg movement signal at time is the window length, is the time-frequency matrix, and ([[]] [[]]) represents the coordinate points of the time-frequency matrix; is the time-frequency matrix, and ([[]] [[]]) represents the coordinate points of the time-frequency matrix; represents the coordinate points of the time-frequency matrix;

[0039] Step 4-1-3: Convert the obtained time-frequency matrix into a two-dimensional heat map as one of the subsequent model inputs.

[0040] Furthermore, Step 4-2 specifically includes:

[0041] Step 4-2-1: Form a time-distance matrix from the obtained leg movement signal segments, convert it into a two-dimensional heat map, i.e., a time-distance map, and at the same time limit the matrix range within several distance gates, ensuring that it covers the distance range of leg movements when a person lies flat during sleep;

[0042] Step 4-2-2: Smooth the signal in Step 4-2-1;

[0043] Step 4-2-3: Trajectory extraction: If signals are detected at two adjacent distance gates and the signal intensity detected at any other distance gate is lower than the preset threshold, i.e., no signal is detected, it is determined that the signals detected at the first two distance gates originate from the same leg movement; if signals are detected at non-adjacent distance gates, it is judged that these signals may come from movements of multiple parts. At this time, select the distance gate with the strongest signal as the estimated target position;

[0044] Step 4-2-4: Binarization processing: Introduce morphological dilation and erosion processing to further optimize the time-distance map;

[0045] Step 4-2-5: The obtained time-frequency matrix Convert it into a two-dimensional heat map as one of the subsequent model inputs.

[0046] Further, the classification module inputs the time-frequency map and the time-distance map into a deep network model to synchronously determine the leg movement type and the sleeping posture type of the corresponding segment, specifically including:

[0047] Step 5-1: Input the time-frequency map and the time-distance map into the deep network model respectively to obtain the initial leg movement / sleeping posture characteristics and their initial predictions.

[0048] Step 5-2: Multimodal distillation, including the calculation and combination of the affinity matrix, feature fusion, and feature reconstruction, to achieve feature interaction for the two tasks of leg movement classification and sleeping posture classification, specifically including:

[0049] Step 5-2-1: Perform pixel-level affinity matrix calculation, specifically including: the first step is dimensionality reduction and activation, and the second step is to restore the number of channels and normalize, so as to obtain the affinity matrix for the leg movement classification task and the affinity matrix for the sleeping posture classification task ;

[0050] Step 5-2-2: Introduce learnable weight parameters and , and perform weighted combination on the affinity matrix of each task, expressed as:

[0051]

[0052]

[0053] In the formula, , are respectively the affinity matrix for the leg movement classification task and the affinity matrix for the sleeping posture classification task after weighted combination;

[0054] Step 5-2-3: Combine the initial features and the combined affinity matrix obtained in Step 5-2-2, that is, perform a pixel-by-pixel multiplication operation on the weighted affinity matrix and the original task features to obtain the refined leg movement features and the sleeping posture features :

[0055]

[0056]

[0057] Step 5-2-4: Transmit the refined features to their respective feature reconstruction modules to obtain the reconstructed feature maps; the specific feature reconstruction module is:

[0058] The feature reconstruction module is an upsampling module designed to gradually restore the spatial resolution of the feature map through deconvolution operations. During the convolution operation, 1-pixel padding is applied to the input. After upsampling, ordinary convolution is used to eliminate the interpolation artifacts caused by deconvolution. A non-linear transformation is introduced through the ReLU activation function to suppress irrelevant information and alleviate the problem of gradient vanishing.

[0059] The feature reconstruction module performs two deconvolution-convolution operations. The first deconvolution layer doubles the spatial size of the input feature map and halves the number of channels, followed by a convolution operation to keep the number of channels unchanged. The second deconvolution layer doubles the feature map again and restores the number of channels to the initial state, followed by a convolution operation to keep the number of channels unchanged.

[0060] Step 5-3: Input the reconstructed feature map into the classifier to achieve the final classification and recognition, and output the types of leg movements and the types of sleeping postures ;

[0061] Among them, the types of leg movements include two types of periodic leg movements, PLM1 and PLM2, double-leg twitching, crossed legs, flexed and extended legs, turning over, and non-leg movement NLM; the types of sleeping postures include Supine, Side, Prone, and non-sleeping posture LM.

[0062] Furthermore, the deep network model described in step 5-1 is specifically:

[0063] The deep learning model includes a shared feature extraction layer, two specific task classification layers, and a dynamic weight averaging layer; the two specific task classification layers include a leg movement classification task layer and a sleeping posture classification task layer.

[0064] Among them, the structure in the shared feature extraction layer is:

[0065] (1) Use ResNet18 for pre-training. The initial layer is a 7×7 convolutional layer with a stride of 2, followed by a 3×3 max pooling layer with a stride of 2. Then, it passes through four residual groups in sequence. Each residual group contains two basic residual blocks, and each residual block contains two 3×3 convolutional layers. After the last residual group, a global average pooling layer is used to convert the feature map into a fixed-length feature vector.

[0066] (2) Remove the final fully connected layer in the pre-trained ResNet18 and only retain its convolutional and pooling layers to extract features from the input multi-view time-frequency map and time-distance map respectively.

[0067] (3) Add an SE block to each perspective of the time-frequency map and the time-distance map respectively to perform channel attention calibration on the extracted feature maps, and then weighted sum the three-channel input signals of the time-frequency map and the time-distance map to obtain the time-frequency map and time-distance map features after multi-perspective fusion;

[0068] (4) The time-frequency map and time-distance map after multi-perspective fusion are calibrated through an SE block respectively, and then the two-channel features are concatenated to obtain the multi-perspective time-frequency and time-distance fusion features;

[0069] (5) Pass the multi-perspective time-frequency and time-distance fusion features through a fully connected layer to output the classification result as the initial training result;

[0070] (6) Through the bottleneck layer, including two 1×1 convolutional layers, the number of channels is sequentially reduced from 1024 to 512 and 128 to further compress the feature information;

[0071] (7) Finally, flatten the feature map through global average pooling;

[0072] In the leg movement classification task layer, first use a fully connected layer with a ReLU activation function to reduce the 128-dimensional feature to 64 dimensions, and then output 6-class prediction results through another fully connected layer;

[0073] In the sleep posture classification task layer, adopt the same method as the leg movement classification task layer to output 3-class prediction results.

[0074] In the dynamic weight averaging layer, the loss of each task is weighted and adjusted according to the dynamic weighted adjustment strategy, and the weight of the current task is calculated based on the previous round of classification loss to balance the training of the two tasks; the calculation formula of the dynamic weight is:

[0075]

[0076] Among them,

[0077]

[0078] In the formula, leg movement classification and sleep posture classification correspond to Take 1 and 2, represent the weight of task is the loss of task at the step, is the training speed calculated at the step, the larger the value, the faster the loss drops. At this time, the weight is made smaller to reduce the loss proportion of this task; T is the temperature constant. When T = 1 there is no effect of weight adjustment. The larger T is, the more stable the weight adjustment is. After obtaining the dynamic weight, the overall loss L can be calculated: There is no effect of weight adjustment. The larger T is, the more stable the weight adjustment is. After obtaining the dynamic weight, the overall loss L can be calculated:

[0079]

[0080] Finally, according to the weighted network parameters, the leg movement category and the sleeping posture category are respectively output in the leg movement classification layer and the sleeping posture classification layer; among them, when the signal processing module determines that it is in the non-leg movement segment NLM, the leg movement prediction result is directly output as NLM, and the sleeping posture prediction does not pass through the multi-task learning network and is directly output through the sleeping posture classifier.

[0081] Furthermore, the sleep periodic limb movement sequence judgment module judges whether the leg movement output satisfies the characteristics of periodic limb movement. If it is satisfied, the periodic leg movement sequence is marked, specifically including:

[0082] Step 6-1, judge the leg movement output Whether it is 0 or 1. If so, mark the data segment during this period as periodic leg movement PLM;

[0083] Step 6-2, when the coherence of the periodic leg movement PLM meets the preset conditions, it is considered as a group of periodic leg movement sequences PLMS; at the same time, save the sleeping posture output .

[0084] Furthermore, the sleep periodic limb movement disorder judgment module judges whether there is a sleep periodic limb movement disorder, specifically including:

[0085] Step 7-1, count the number of occurrences of the periodic leg movement sequences PLMS during the whole night's sleep, and divide the number by the total sleep time to obtain the periodic limb movement index PLMI;

[0086] Step 7-2, judge whether there is a sleep periodic limb movement disorder. For children, the periodic limb movement index PLMI>5 / h, and for adults, the periodic limb movement index PLMI>15 / h, which indicates the existence of a sleep periodic limb movement disorder, otherwise it indicates the non-existence of a sleep periodic limb movement disorder.

[0087] Compared with the prior art, the remarkable advantages of the present invention are:

[0088] (1) The non-contact periodic sleep limb movement disorder and sleep posture combined monitoring method proposed by the present invention uses a non-contact FMCW radar to collect leg movement and sleeping posture data, avoiding the interference of wearing devices to the subjects, improving the comfort and convenience of monitoring, and can perform long-term monitoring without disturbing the user.

[0089] (2) The radar signal processing flow proposed by the present invention includes phase demodulation and short-time Fourier transform, which can perform high-resolution time-frequency and time-distance analysis on radar echo signals, and can extract characteristic information of leg movement and sleeping posture. This flow can greatly improve the accuracy of the leg movement and sleeping posture acquisition database, ensure the effectiveness of model input, and improve the accuracy of leg movement classification and sleeping posture classification.

[0090] (3) The present invention adopts a multi-task learning model, which can share the feature representations of the leg movement classification and sleeping posture classification tasks, and improve the performance of the classification tasks through joint training, significantly improving the training efficiency and generalization ability.

[0091] (4) Multi-view data acquisition is carried out through multiple FMCW radar sensors installed at the end of the bed. Through the fusion of multi-view multi-spectrum domain signals, the detection blind area can be effectively reduced, the signal quality can be enhanced, and no additional multi-sensor devices or complex hardware systems are required, reducing the hardware cost and system complexity. At the same time, the convenience and robustness of data acquisition are improved.

[0092] The present invention will be further described in detail below with reference to the accompanying drawings. Description of the Drawings

[0093] Figure 1 It is a schematic diagram of the implementation of the non-contact periodic sleep limb movement disorder and sleep posture joint monitoring device at the end of the bed in an embodiment.

[0094] Figure 2 It is a test scenario diagram in an embodiment.

[0095] Figure 3 It is a time-frequency diagram of different leg movements and sleeping postures in an embodiment, where Figure 3 (a) to (f) in it are the time-frequency diagrams corresponding to PLM1, PLM2, twitch, crossed legs, flexed and extended legs, and turning over respectively.

[0096] Figure 4 It is a time-distance diagram of different leg movements and sleeping postures in an embodiment, where Figure 4 (a) to (f) in it are the time-distance diagrams corresponding to PLM1, PLM2, twitch, crossed legs, flexed and extended legs, and turning over respectively.

[0097] Figure 5 It is a model block diagram in an embodiment. Detailed Embodiment

[0098] In order to make the purpose, technical solution and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present application and are not used to limit the present application.

[0099] It should be noted that if there are directional indications (such as up, down, left, right, front, back...) involved in the embodiments of the present invention, the directional indications are only used to explain the relative positional relationship, movement conditions, etc. between components in a certain specific posture (as shown in the drawings). If the specific posture changes, the directional indications will also change accordingly.

[0100] In one embodiment, a combined non-contact periodic sleep limb movement disorder and sleep posture monitoring device at the foot of the bed is provided. Figure 1 The device includes: a signal acquisition module, a signal processing module, a feature extraction module, a classification module, a sleep periodic limb movement sequence judgment module, and a sleep periodic limb movement disorder judgment module;

[0101] The signal acquisition module is used to non-contact and continuously acquire echo signals of the human leg movements and sleep postures from different perspectives through a number of distributed FMCW radars;

[0102] The signal processing module is used to process the radar echo signals from different perspectives and identify and segment the adaptive leg movement segments through a dynamic window adjustment algorithm;

[0103] The feature extraction module is used to perform time-frequency analysis and time-distance analysis on the echo signal segments to generate time-frequency diagrams and time-distance diagrams of corresponding perspectives;

[0104] The classification module is used to input the time-frequency diagrams and time-distance diagrams into a deep network model to simultaneously determine the leg movement types and sleep posture types of corresponding segments; the deep network model is a leg movement-sleep posture multi-task classification model based on multi-perspective multi-spectrum domain information fusion;

[0105] The sleep periodic limb movement sequence judgment module is used to judge whether the leg movement output meets the characteristics of periodic limb movement. If it meets, mark the periodic leg movement sequence;

[0106] The sleep periodic limb movement disorder judgment module is used to judge whether there is a sleep periodic limb movement disorder.

[0107] Furthermore, in one of the embodiments, the signal acquisition module non-contact and continuously acquires echo signals of the human leg movements and sleep postures from different perspectives through a number of distributed FMCW radars, specifically including:

[0108] Step 2-1, data acquisition: The subject initially lies flat on the bed. A number of FMCW radar sensors are distributed and installed at the foot of the bed, and the antennas are all oriented towards the human legs; preferably, the distance between the radar sensor and the measured part of the human body is about 30 centimeters; the radar leg echo signals are acquired by different FMCW radar sensors from different perspectives;

[0109] Exemplarily here, the test scenario is as Figure 2 shown. Continuous data of periodic leg movements, common human leg movements, and three basic sleeping postures are collected. The specific descriptions of each leg movement and each sleeping posture are shown in Table 5.1 and Table 5.2 respectively.

[0110] Table 5.1 Description of Leg Movements

[0111] Leg movement Detailed description Periodic leg movement 1 Brief, stereotyped extension of the big toe and dorsiflexion of the foot Periodic leg movement 2 Brief, stereotyped extension of the big toe and dorsiflexion of the foot, along with flexion of the knee and hip, similar to kicking. Twitch Slight up-and-down twitch of the calf, which is a rapid and involuntary slight contraction or tremor of the leg muscles. Crossed legs The legs are overlapped, and the process is that one leg is placed on top of the other. Flexion and extension of the legs The movement of the leg from bent to straight, or from straight back to bent. Turn over Lateral rotation of the body, accompanied by movement and adjustment of the legs.

[0112] Table 5.2 Description of Sleep Posture States

[0113] Sleeping position state Detailed description Supine position Lie flat on the back on the bed, with the back touching the bed surface, the legs and arms placed naturally on both sides of the body, and the soles of the feet facing the radar. Lateral position Lie on the side on the bed, with the head, shoulders and hips in an inclined line, the legs can be bent or straight, and the soles of the feet are rotated 90 degrees to face the radar. Prone position Lie flat on the stomach on the bed, with the chest and abdomen touching the bed surface, the head turned to one side, the legs and arms can be naturally extended or slightly bent, and the soles of the feet no longer block the radar.

[0114] The echo signal in each perspective is expressed as: ), where r is the original radar signal, represents the time sequence number of slow-time sampling, denotes the total number of slow-time sampling times, represents the serial number of the range gate, and N represents the total number of range gates; and respectively represent the sampling moments of slow time and fast time;

[0115] Step 2-2, moving target detection. The leg echo signal R collected from each perspective is rearranged after fast and slow time sampling to obtain the echo matrix, where X represents the number of fast-time sampling points and Y represents the number of slow-time sampling points;

[0116] Step 2-3, using a loop function to traverse the leg echo signal R. In each loop, the data at a fixed time interval is subtracted from the first column data to remove background clutter interference;

[0117] Step 2-4, filtering the leg echo signal R after removing background clutter interference to remove the tail shadow clutter generated by moving target detection, and obtaining the preprocessed leg echo signal. Here, preferably, median filtering is used.

[0118] Furthermore, in one embodiment, the signal processing module processes the radar echo signals from different perspectives and identifies and segments the adaptive leg movement segments through a dynamic window adjustment algorithm, specifically including:

[0119] Step 3-1, selecting the leg echo signal of a fixed range gate for demodulation to obtain a phase signal ;

[0120] Step 3-2, motion information extraction: performing a third-order difference processing on the phase signal to obtain a further enhanced leg motion signal ;

[0121] Step 3-3, sliding window detection: Use a sliding window of 2 to traverse the enhanced leg movement signal , calculate the variance within each window , and the average variance of the windows of the past N non-leg movement events NLM of the current window . Then, determine whether the signal of the current window is a leg movement event LM according to the threshold condition shown in the following formula, and then execute Step 3-4; If the signal of the current window

[0122]

[0123] satisfies the above formula, it indicates a leg movement event LM; if it does not satisfy the above formula, the window average variance is updated; ;

[0124] where, is the empirical threshold, is the segment sampling length;

[0125] Here, according to the AASM periodic leg movement judgment standard, the cycle length between leg movement events LM is 5 to 90 s, so it is reasonable to set the segment length to 5 s, and the number of sampling points of the segment ;

[0126] Step 3-4, dynamic window adjustment: If the threshold condition is satisfied in Step 3-3, further detect the peak points and valley points within the window. If there are peak points between two adjacent valley points, it indicates a complete leg movement event LM, and save the time segment of t seconds before or after the window where the peak point is located; if not, it indicates that the leg movement event LM is incomplete, and the dynamic window moves backward by t seconds, and continue to detect the peak points and valley points to ensure obtaining a complete time segment of the leg movement event LM:

[0127]

[0128] where, , respectively represent the nth and (n + 1)th valley points, represents the mth peak point;

[0129] If the threshold condition is not satisfied in Step 3-3, the time segment within the window is saved as a non-leg movement segment NLM, thereby obtaining a complete leg movement segment LM and non-leg movement segment NLM. The obtained leg movement segments are subsequently input into a multi-task learning model for simultaneous classification of leg movement and sleep posture, and the non-leg movement segments directly output the sleep posture classification through a classifier.

[0130] Here, preferably, the t seconds is 5 seconds.

[0131] Further, in one embodiment, the feature extraction module performs time-frequency analysis and time-distance analysis on the echo signal segment to generate a time-frequency map and a time-distance map of the corresponding perspective, specifically including:

[0132] Step 4-1, obtain the leg echo signal segment, perform time-frequency analysis on the fixed-range gate complex signal through short-time Fourier transform, and obtain a complete time-frequency map reflecting the positive and negative frequency changes;

[0133] Step 4-2, obtain the leg echo signal segment, and obtain a high-resolution time-distance map through the time-distance map optimization algorithm of adaptive adjacent signal enhancement.

[0134] Further, in one embodiment, step 4-1 specifically includes:

[0135] Step 4-1-1, obtain the LM segment of t seconds through the continuous monitoring algorithm of sleep limb movement based on energy envelope, which contains n sampling points (preferably, t takes 2 and n is 250); take the fixed-range gate leg movement signal lm[n] in this time period;

[0136] Step 4-1-2, set the window length, overlap length, and the number of points for STFT calculation, and perform short-time Fourier transform, which is expressed as:

[0137]

[0138] Among them, is the rectangular window function aligned at the position to and slides according to the window length and overlap length, is the number of points for STFT calculation, is the leg movement signal at the moment, is the time-frequency matrix;

[0139] Step 4-1-3, convert the obtained time-frequency matrix into a two-dimensional heat map , which is used as one of the subsequent model inputs. Since the leg movement and sleeping posture information are coupled in the time-frequency map and it is difficult to directly observe the characteristics of different sleeping postures, the time-frequency maps of different sleeping postures are not shown. Figure 3 The time-frequency maps of different leg movements are shown.

[0140] Further, in one embodiment, step 4-2 specifically includes:

[0141] Step 4-2-1: Form a time-distance matrix from the obtained leg movement signal segments, convert it into a two-dimensional heat map (instantaneous distance map), and at the same time limit the matrix range within several distance gates, ensuring that the distance range of a person's leg movement when lying flat in a sleeping state is covered.

[0142] Preferably here, the matrix range is limited to 1 to 3 distance gates, that is, 0 to 1.32 meters.

[0143] Step 4-2-2: Perform signal smoothing on the signal in Step 4-2-1.

[0144] Preferably here, first perform two-dimensional median filtering through a sliding window of size 3×3 to suppress isolated noise points, and then remove noise through Gaussian filtering.

[0145] Step 4-2-3: Trajectory extraction: If signals are detected in two adjacent distance gates and the signal intensity detected in any other distance gate is lower than the preset threshold (i.e., no signal is detected), it is determined that the signals detected in the first two distance gates originate from the same leg movement; if signals are detected in non-adjacent distance gates, it is judged that these signals may come from the movements of multiple parts. At this time, select the distance gate with the strongest signal as the estimate of the target position.

[0146] Step 4-2-4: Binary processing: Introduce morphological dilation and erosion processing to further optimize the instantaneous distance map.

[0147] Step 4-2-5: Convert the obtained time-frequency matrix into a two-dimensional heat map as one of the subsequent model inputs. Since the leg movement and sleep posture information are coupled in the instantaneous distance map and it is difficult to directly observe the characteristics of different sleep postures, the instantaneous distance maps of different sleep postures are not shown. Figure 4 The instantaneous distance maps of different leg movements are shown.

[0148] Furthermore, in one embodiment, the classification module inputs the time-frequency map and the instantaneous distance map into a deep network model to simultaneously determine the leg movement type and sleep posture type of the corresponding segment, specifically including:

[0149] Step 5-1: Input the time-frequency map and the instantaneous distance map into the deep network model respectively to obtain the initial leg movement / sleep posture characteristics and their initial predictions.

[0150] Step 5-2: Multimodal distillation, including the calculation and combination of the affinity matrix, feature fusion, and feature reconstruction, to achieve feature interaction for the two tasks of leg movement classification and sleep posture classification, specifically including:

[0151] Step 5-2-1: Perform pixel-level affinity matrix calculation, specifically including: The first step is dimensionality reduction and activation, and the second step is to restore the number of channels and normalize, so as to obtain the affinity matrix for the leg movement classification task. Affinity matrix for sleep position classification task ;

[0152] Here, the feature extractor in ResNet18 downsamples the input multiple times, and finally outputs a feature map with a size of 2×2 and 128 channels, that is, with a size of [128, 2, 2]. Preferably, the calculation of the affinity matrix includes two steps. The first step is dimensionality reduction and activation. A 1×1 convolutional kernel is used to reduce the input channels to 8, and the ReLU activation function is applied. The second step is to restore the number of channels and normalize. The number of channels is restored to 128 through another 1×1 convolutional kernel, and the Sigmoid activation function is applied to limit the output between [0, 1].

[0153] Step 5-2-2, introducing learnable weight parameters and , performing a weighted combination of the affinity matrices for each task, expressed as:

[0154]

[0155]

[0156] In the formula, 、 are the affinity matrices of the leg movement classification task and the sleep position classification task after weighted combination respectively;

[0157] Step 5-2-3, combining the initial features and the combined affinity matrix obtained in step 5-2-2, that is, performing a pixel-by-pixel multiplication operation on the weighted affinity matrix and the original task features to obtain refined leg movement features and sleep position features :

[0158]

[0159]

[0160] Step 5-2-4, passing the refined features to their respective feature reconstruction modules to obtain the reconstructed feature maps; the feature reconstruction module is specifically:

[0161] The feature reconstruction module is an upsampling module, which aims to gradually restore the spatial resolution of the feature map through deconvolution operations (preferably, the deconvolution uses a 4×4 convolutional kernel with a stride of 2), and pads the input by 1 pixel during the convolutional operation. After upsampling, ordinary convolution is used to eliminate the interpolation artifacts brought by deconvolution, and the ReLU activation function is introduced for non-linear transformation to suppress irrelevant information and alleviate the problem of gradient disappearance;

[0162] The feature reconstruction module performs two deconvolution-convolution operations. The first deconvolution layer first doubles the spatial size of the input feature map and reduces the number of channels by half (from 128 to 64), and then performs a convolution operation to keep the number of channels (64) unchanged. The second deconvolution layer doubles the feature map again, restores the number of channels to the initial state, and then performs a convolution operation to keep the number of channels unchanged.

[0163] Step 5-3: Input the reconstructed feature map into the classifier to achieve the final classification and recognition, and output the types of leg movements and the types of sleeping postures ;

[0164] Among them, the types of leg movements include two types of periodic leg movements (PLM1, PLM2), double-leg twitching, crossed legs, extended and flexed legs, turning over (NPLM1, NPLM2, NPLM3, NPLM4), and non-leg movement NLM, which are respectively labeled as 0, 1, 2, 3, 4, 5, 6; the types of sleeping postures include Supine, Side, and Prone, which are respectively labeled as 0, 1, 2.

[0165] Furthermore, in one embodiment, combined with Figure 5 , the deep network model described in step 5-1 is specifically:

[0166] The deep learning model includes a shared feature extraction layer, two specific task classification layers, and a dynamic weight averaging layer; the two specific task classification layers include a leg movement classification task layer and a sleep posture classification task layer;

[0167] Among them, the structure in the shared feature extraction layer is:

[0168] (1) Use ResNet18 for pre-training. The initial layer is a 7×7 convolutional layer with a stride of 2, followed by a 3×3 max pooling layer with a stride of 2; then sequentially pass through four residual groups, each residual group contains two basic residual blocks, and each residual block contains two 3×3 convolutional layers; after the last residual group, use a global average pooling layer to convert the feature map into a fixed-length feature vector;

[0169] (2) Remove the final fully connected layer in the pre-trained ResNet18, and only retain its convolutional and pooling layers to extract features from the input multi-view time-frequency map and time-distance map respectively;

[0170] (3) Add an SE block for each view of the time-frequency map and the time-distance map to perform channel attention calibration on the extracted feature map, and then weight and add the three-way input signals of the time-frequency map and the time-distance map to obtain the features of the multi-view fused time-frequency map and time-distance map;

[0171] (4) The time-frequency map and time-distance map after multi-perspective fusion are respectively calibrated by an SE block, and then the two-way features are concatenated to obtain multi-perspective time-frequency and time-distance fusion features;

[0172] (5) The multi-perspective time-frequency and time-distance fusion features are output through a fully connected layer to obtain the classification result as the initial training result;

[0173] (6) Through the bottleneck layer, including two 1×1 convolutional layers, the number of channels is sequentially reduced from 1024 to 512 and 128 to further compress the feature information;

[0174] (7) Finally, the feature map is flattened through global average pooling;

[0175] In the leg movement classification task layer, first, the 128-dimensional features are reduced to 64 dimensions by a fully connected layer with a ReLU activation function, and then 6-class prediction results are output through another fully connected layer;

[0176] In the sleep posture classification task layer, 3-class prediction results are output in the same way as in the leg movement classification task layer.

[0177] In the dynamic weight averaging layer, the loss of each task is weighted according to the dynamic weighted adjustment strategy, and the weight of the current task is calculated based on the previous round of classification loss to balance the training of the two tasks; the calculation formula for the dynamic weight is:

[0178]

[0179] Among them,

[0180]

[0181] In the formula, leg movement classification and sleep posture classification respectively correspond to Take 1 and 2, represent the weight of task , is the loss of task at the th step, is the training speed calculated at the th step. The larger the value, the faster the loss drops. At this time, the weight is made smaller to reduce the loss ratio of this task; T is the temperature constant. When T = 1 there is no effect of weight adjustment. The larger T is, the more stable the weight adjustment is. After obtaining the dynamic weight, the overall loss L can be calculated:

[0182]

[0183] Finally, the leg movement category and the sleep posture category are respectively output at the leg movement classification layer and the sleep posture classification layer according to the weighted network parameters; among them, when the signal processing module determines that it is in the non-leg movement segment NLM, the leg movement prediction result is directly output as NLM, and the sleep posture prediction does not pass through the multi-task learning network and is directly output through the sleep posture classifier.

[0184] Further, in one embodiment, the sleep periodic limb movement sequence judgment module judges whether the leg movement output satisfies the characteristics of periodic limb movement. If it is satisfied, the periodic leg movement sequence is marked, and specifically includes:

[0185] Step 6-1, judge the leg movement output Whether it is 0 or 1. If so, mark the data segment during this period as periodic leg movement PLM;

[0186] Step 6-2, when the coherence of the periodic leg movement PLM meets the preset conditions, it is considered as a group of periodic leg movement sequences PLMS; at the same time, save the sleep posture output .

[0187] Preferably here, the preset condition is that the periodic leg movement PLM appears coherently more than 4 times, that is, the interval between adjacent periodic leg movements PLM is between 5 and 90 s.

[0188] Further, in one embodiment, the sleep periodic limb movement disorder judgment module judges whether there is a sleep periodic limb movement disorder, and specifically includes:

[0189] Step 7-1, count the number of occurrences of the periodic leg movement sequence PLMS during the whole night's sleep, and divide the number by the total sleep time to obtain the periodic limb movement index PLMI;

[0190] Step 7-2, judge whether there is a sleep periodic limb movement disorder. For children, the periodic limb movement index PLMI>5 / h, and for adults, the periodic limb movement index PLMI>15 / h, which means there is a sleep periodic limb movement disorder, otherwise it means there is no sleep periodic limb movement disorder.

[0191] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A non-contact periodic limb movement disorder and sleep position combined monitoring device at the end of the bed, characterized in that: The device includes a signal acquisition module, a signal processing module, a feature extraction module, a classification module, a sleep periodic limb movement sequence judgment module and a sleep periodic limb movement disorder judgment module; The signal acquisition module is used to non-contactly and continuously acquire echo signals of human leg movements and sleeping postures at different viewing angles through a plurality of distributedly placed FMCW radars; The signal processing module is used to process radar echo signals at different viewing angles and to identify and segment adaptive leg movement segments through a dynamic window adjustment algorithm; The feature extraction module is used to perform time-frequency analysis and time-distance analysis on the echo signal segments to generate a time-frequency diagram and a time-distance diagram of the corresponding viewing angle; The classification module is used to input the time-frequency graph and the time-distance graph into the deep network model to simultaneously determine the leg movement type and the sleeping posture type of the corresponding segment; the deep network model is a leg movement-sleeping posture multi-task classification model based on multi-view and multi-spectral domain information fusion; The sleep periodic limb movement sequence judgment module is used to judge whether the leg movement output meets the characteristics of periodic limb movement, and if so, mark the periodic leg movement sequence; The sleep periodic limb movement disorder judgment module is used to judge whether there is sleep periodic limb movement disorder.

2. The non-contact periodic limb movement disorder and sleep position combined monitoring device at the end of the bed according to claim 1, characterized in that: The signal acquisition module collects the echo signals of the human leg movements and sleeping postures at different viewing angles in a non-contact and continuous manner through a number of distributedly placed FMCW radars, specifically including: Step 2-1, data collection: The subject initially lies flat on the bed, and several FMCW radar sensors are distributed and installed at the end of the bed, with the antennas facing the legs of the subject; different FMCW radar sensors collect radar leg echo signals at different viewing angles; The echo signal at each viewing angle is expressed as: ), r is the original radar signal, where Represents the time sequence number of slow time sampling, Indicates the total amount of slow sampling time, Represents the serial number of the range gate, N represents the total number of range gates; and represent the sampling moments of slow and fast time respectively; Step 2-2, moving target detection, the collected leg echo signal R of each view angle is sampled at fast and slow times and then rearranged to obtain The echo matrix of , where X represents the number of sampling points in the fast time, and Y represents the number of sampling points in the slow time; Step 2-3, using a loop function to traverse the leg echo signal R, in each cycle, the data at a fixed time interval is subtracted from the first column of data to remove background clutter interference; Step 2-4, filtering the leg echo signal R after removing the background clutter interference, removing the tail clutter generated by the moving target detection, and obtaining the preprocessed leg echo signal.

3. The non-contact periodic limb movement disorder and sleep position combined monitoring device at the end of the bed according to claim 1, characterized in that: The signal processing module processes radar echo signals at different viewing angles and recognizes and segments adaptive leg motion segments through a dynamic window adjustment algorithm, specifically including: Step 3-1, select the leg echo signal of the fixed range gate for demodulation to obtain the phase signal ; Step 3-2, motion information extraction: perform third-order difference processing on the phase signal to obtain a further enhanced leg motion signal ; Step 3-3, sliding window detection: use 2 The sliding window traverses the enhanced leg movement signal , calculate the variance within each window , and the current window The window average variance of the past N non-leg movement events NLM , and judge the current window according to the threshold condition shown in the following formula Whether the signal is a leg movement event LM, then execute steps 3-4; ； If the current window If the signal satisfies the above formula, it indicates a leg movement event LM; if it does not satisfy the above formula, the window average variance is updated ; in, is the empirical threshold, is the segment sampling length; Step 3-4, dynamic window adjustment: If the threshold condition is met in step 3-3, the peak points and valley points in the window are further detected. If there are two adjacent valleys containing a peak point, it indicates that it is a complete leg movement event LM, and the time segment t seconds before or after the window where the peak point is located is saved; if it does not exist, it indicates that the leg movement event LM is incomplete, and the dynamic window is moved back t seconds, and the peak points and valley points are continuously detected to ensure that a complete leg movement event LM time segment is obtained: ； in, , Respectively represent the nth and n+1th valley points, represents the mth peak point; If the threshold condition is not met in step 3-3, the time segment in the window is saved as a non-leg movement segment NLM, thereby obtaining a complete leg movement segment LM and a non-leg movement segment NLM.

4. The non-contact periodic limb movement disorder and sleep position combined monitoring device at the end of the bed according to claim 1, characterized in that: The feature extraction module performs time-frequency analysis and time-distance analysis on the echo signal segments to generate a time-frequency diagram and a time-distance diagram of the corresponding viewing angle, specifically including: Step 4-1, obtaining the leg echo signal fragment, performing time-frequency analysis on the fixed range gate complex signal by short-time Fourier transform, and obtaining a complete time-frequency diagram reflecting the positive and negative frequency changes; Step 4-2, obtain the leg echo signal fragments, and obtain a high-resolution time-distance map through the time-distance map optimization algorithm of adaptive neighboring signal enhancement.

5. The non-contact periodic limb movement disorder and sleep position combined monitoring device at the end of the bed according to claim 4, characterized in that: Step 4-1 specifically includes: Step 4-1-1, obtain a LM segment of t seconds, including n sampling points, through the continuous monitoring algorithm of sleep limb activity based on energy envelope; obtain the fixed-range gate leg movement signal lm[n] of this time period; Step 4-1-2, set the window length, overlap length and the number of points for STFT calculation, and perform short-time Fourier transform, expressed as: ； in, is Position Alignment The rectangular window function slides according to the window length and overlap length. is the number of points at which the STFT is calculated, for The leg movement signal at each moment, is the window length, is the time-frequency matrix, ( ) represents the coordinate point of the time-frequency matrix; Step 4-1-3, the obtained time-frequency matrix Convert to a 2D heatmap , as one of the subsequent model inputs.

6. The non-contact periodic limb movement disorder and sleep position combined monitoring device at the end of the bed according to claim 4, characterized in that: Step 4-2 specifically includes: Step 4-2-1, the acquired leg movement signal segments are formed into a time-distance matrix, and converted into a two-dimensional heat map, i.e., a time distance map, and the matrix range is limited to a number of distance gates to ensure that the distance range of the leg movement of a person lying flat in sleep is covered; Step 4-2-2, performing signal smoothing processing on the signal of step 4-2-1; Step 4-2-3, trajectory extraction: If two adjacent range gates detect signals and the strength of the signal detected by any other range gate is lower than the preset threshold, that is, no signal is detected, it is determined that the signals detected by the first two range gates originate from the same leg movement; if non-adjacent range gates all detect signals, it is determined that these signals may come from the movement of multiple parts, and the range gate with the strongest signal is selected as the estimate of the target position; Step 4-2-4, binarization processing: introduce morphological expansion and corrosion processing to further optimize the time-distance map; Step 4-2-5, the obtained time-frequency matrix Converted into a two-dimensional heat map as one of the subsequent model inputs.

7. The non-contact periodic limb movement disorder and sleep position combined monitoring device at the end of the bed according to claim 1, characterized in that: The classification module inputs the time-frequency graph and the time-distance graph into the deep network model to simultaneously determine the leg movement type and sleeping posture type of the corresponding segment, specifically including: Step 5-1, input the time-frequency graph and the time-distance graph into the deep network model respectively to obtain the initial leg movement / sleeping position features and their initial predictions; Step 5-2, multimodal distillation, includes the calculation and combination of affinity matrices, feature fusion, and feature reconstruction to achieve feature interaction between the two tasks of leg movement classification and sleeping position classification, including: Step 5-2-1, calculate the pixel-level affinity matrix, which includes: the first step is dimensionality reduction and activation, the second step is to restore the number of channels and normalize, so as to obtain the affinity matrix of the leg movement classification task and the affinity matrix for the sleep position classification task ; Step 5-2-2, introduce learnable weight parameters and , the affinity matrix of each task is weighted and combined, expressed as: ； ； In the formula, , They are the affinity matrix of the leg movement classification task and the affinity matrix of the sleeping position classification task after weighted combination; Step 5-2-3, combine the initial features and the combined affinity matrix obtained in step 5-2-2, that is, perform pixel-by-pixel multiplication of the weighted affinity matrix and the original task features to obtain the refined leg movement features and sleeping posture characteristics : ； ； Step 5-2-4, passing the refined features to respective feature reconstruction modules to obtain reconstructed feature maps; the feature reconstruction modules are specifically: The feature reconstruction module is an upsampling module that aims to gradually restore the spatial resolution of the feature map through deconvolution operations, and pad the input with 1 pixel during the convolution operation. After upsampling, ordinary convolution is used to eliminate the interpolation artifacts caused by deconvolution. The activation function ReLU is used to introduce nonlinear transformations to suppress irrelevant information and alleviate the problem of gradient disappearance. The feature reconstruction module performs two deconvolution-convolution operations. The first deconvolution layer doubles the spatial size of the input feature map and reduces the number of channels by half, and then performs a convolution operation to keep the number of channels unchanged. The second deconvolution layer doubles the feature map again, restores the number of channels to the initial state, and then performs a convolution operation to keep the number of channels unchanged. Step 5-3: Input the reconstructed feature map into the classifier to achieve the final classification and output the leg movement type. and sleeping position ; Among them, leg movement types Includes two types of periodic leg movements PLM1 and PLM2, leg twitching, leg crossing, leg flexion and extension, turning over, and non-leg movement NLM; Sleeping position types Including Supine, Side, Prone and non-sleeping positions LM.

8. The non-contact periodic limb movement disorder and sleep position combined monitoring device at the foot of the bed according to claim 3 or 7, characterized in that: The deep network model described in step 5-1 is specifically: The deep learning model includes a shared feature extraction layer, two specific task classification layers and a dynamic weight averaging layer; the two specific task classification layers include a leg movement classification task layer and a sleeping position classification task layer; Among them, the structure in the shared feature extraction layer is: (1) Pre-training is performed using ResNet18. The initial layer is a 7×7 convolutional layer with a stride of 2, followed by a 3×3 maximum pooling layer with a stride of 2. Then, four residual groups are passed in sequence. Each residual group contains two basic residual blocks, and each residual block contains two 3×3 convolutional layers. After the last residual group, a global average pooling layer is used to convert the feature map into a fixed-length feature vector. (2) Remove the final fully connected layer in the pre-trained ResNet18 and only retain its convolution and pooling layers to extract features from the input multi-view frequency-time map and time-distance map respectively; (3) Add an SE block to each view of the time-frequency map and the time-distance map to perform channel attention calibration on the extracted feature map, and then perform weighted addition on the three input signals of the time-frequency map and the time-distance map to obtain the time-frequency map and time-distance map features after multi-view fusion; (4) The time-frequency graph and time-distance graph after multi-view fusion are calibrated by an SE block respectively, and then the two features are spliced ​​to obtain the multi-view time-frequency and time-distance fusion features; (5) The multi-view time-frequency-time-distance fusion features are output through a fully connected layer as the classification result as the initial training result; (6) Through the bottleneck layer, including two 1×1 convolutional layers, the number of channels is reduced from 1024 to 512 and 128 respectively, further compressing the feature information; (7) Finally, the feature map is flattened through global average pooling; In the leg movement classification task layer, the 128-dimensional features are first reduced to 64 dimensions using a fully connected layer with a ReLU activation function, and then another fully connected layer outputs 6 categories of prediction results; In the sleep position classification task layer, the same method as the leg movement classification task layer is used to output three types of prediction results; In the dynamic weight averaging layer, the loss of each task is adjusted according to the dynamic weight adjustment strategy, and the weight of the current task is calculated based on the previous round of classification loss to balance the training of the two tasks; the calculation formula of the dynamic weight is: ； in, ； In the formula, the leg movement classification and sleeping posture classification correspond to Take 1 and 2, Representative tasks The weight of The task is The loss of step time, It is The training speed calculated at step 1. The larger the value, the faster the loss decreases. At this time, the weight becomes smaller, thereby reducing the loss ratio of the task. T is the temperature constant. When T = 1 There is no effect of weight adjustment. The larger T is, the more stable the weight adjustment is. After obtaining the dynamic weight, the overall loss L can be calculated: ； Finally, according to the weighted network parameters, the leg movement category and sleeping position category are output in the leg movement classification layer and the sleeping position classification layer respectively; among them, when the signal processing module determines that it is in the non-leg movement segment NLM, the leg movement prediction result is directly output as NLM, and the sleeping position prediction does not go through the multi-task learning network, but is directly output through the sleeping position classifier.

9. The non-contact periodic limb movement disorder and sleep position combined monitoring device at the foot of the bed according to claim 1, characterized in that: The sleep periodic limb movement sequence judgment module judges whether the leg movement output meets the characteristics of periodic limb movement, and if so, marks the periodic leg movement sequence, specifically including: Step 6-1, determine the leg movement output Is it 0 or 1? If so, the data segment during this period is marked as periodic leg movement PLM; Step 6-2: When the continuity of the periodic leg movement PLM meets the preset conditions, it is considered to be a set of periodic leg movement sequences PLMS; at the same time, the sleeping posture output is saved. .

10. The non-contact periodic limb movement disorder and sleep position combined monitoring device at the end of the bed according to claim 9, characterized in that: The sleep periodic limb movement disorder judgment module judges whether there is sleep periodic limb movement disorder, specifically including: Step 7-1, counting the number of periodic leg movement sequences PLMS occurring during the whole night of sleep, and dividing the number by the total sleep time to obtain the periodic limb movement index PLMI; Step 7-2, determine whether periodic limb movement disorder of sleep exists. If the periodic limb movement index PLMI of children is >5 / h and the periodic limb movement index PLMI of adults is >15 / h, it indicates the presence of periodic limb movement disorder of sleep; otherwise, it indicates the absence of periodic limb movement disorder of sleep.

Images