A method and system for identifying mattress users based on pressure array sensors

Abstract

This invention discloses a method and system for identifying mattress users based on a pressure array sensor, belonging to the field of smart home and health monitoring technology. To address the problem of deep integration of weight prediction and user identification, this invention places a pressure array sensor in the center of the mattress to collect pressure data and weight data during the user's sleep process. This data is then filtered, including row and column filtering, to obtain a dataset of processed pressure data during the user's sleep process. A deep learning-based user weight prediction and user identification model and its loss function are constructed. The deep learning-based user weight prediction and user identification model is trained using the dataset to obtain a trained model for user identification. This invention achieves a dual improvement in user identification accuracy and robustness, ensuring the reliability of the core identification function.

Description

Technical Field

[0001] This invention belongs to the field of smart home and health monitoring technology, specifically relating to a method and system for identifying mattress users based on a pressure array sensor. Background Technology

[0002] With the rapid development of smart home and health monitoring technologies, smart devices that interact with the human body, such as smart mattresses, sleep monitoring systems, and smart scales, are widely used in multi-user shared environments such as homes and elderly care facilities. One of the core requirements for these devices is to achieve accurate identification of different users in order to provide personalized services for each user (such as customized sleep interventions and exclusive health data statistics), while ensuring the privacy and security of user data. However, current user identification technologies based on biometrics or physical characteristics still suffer from low accuracy and poor robustness in multi-user shared environments, making it difficult to meet practical application needs.

[0003] Currently, user identification solutions for devices such as smart mattresses and sleep monitoring systems mainly rely on two types of technical approaches: one is identification methods based on traditional biometrics, such as facial recognition and fingerprint recognition; the other is physical feature identification methods based on pressure sensing data, such as collecting the user's pressure distribution profile (such as body contact area, pressure peak position, etc.) through pressure array sensors and using it as the characteristic basis for distinguishing different users.

[0004] While traditional biometric identification technologies such as facial recognition and fingerprint recognition are highly mature in the field of identity verification, they have significant limitations in scenarios such as smart mattresses and sleep monitoring. On the one hand, these technologies have stringent requirements for the collection environment. For example, facial recognition requires the user to be in specific lighting conditions and have an unobstructed face. However, users often have their faces covered or their postures are not fixed when they are sleeping or resting, which can lead to interruption or failure of recognition. On the other hand, some biometric collection processes (such as fingerprint pressing and facial scanning) may disturb the user, which violates the design concept of "seamless interaction" of smart devices and significantly reduces the user experience.

[0005] The second type of technology, physical feature recognition based on pressure array sensors, has become the mainstream choice for smart mattresses and sleep monitoring systems due to its advantages such as non-invasiveness and seamless integration with daily use scenarios (e.g., data collection can be completed while the user is lying on the mattress). Existing technologies typically extract geometric features (such as pressure centroid coordinates and the size of the circumscribed rectangle of the pressure region), texture features (such as the gray-level co-occurrence matrix of pressure values), or statistical features (such as the pressure mean and variance) from the pressure distribution image to construct a user feature model, and then use pattern matching algorithms (such as support vector machines and K-nearest neighbors) to achieve user identification. However, this type of method has significant drawbacks: when different users have similar body shapes and weights, the differences in their pressure distribution contours are small, resulting in insufficient feature discrimination and a high likelihood of misidentification.

[0006] Weight, as one of the core physiological characteristics of users, is characterized by high stability and significant individual differences (even users with similar body types may have different weights). Combining weight information with pressure distribution characteristics can effectively improve the distinguishability of user characteristics. In existing technologies, weight measurement methods based on pressure array sensors are relatively mature. The principle is to collect and sum the pressure values ​​of all sensors in the pressure array to obtain the user's weight (ignoring the sensor's own weight and environmental interference). However, current technical solutions generally treat "user identification" and "weight prediction" as two independent functional modules, initiating the weight measurement process only when user weight data is needed, without integrating weight prediction results as auxiliary information into the user identification process. This "functional separation" design results in the underutilization of the data source value of the pressure array sensor. User identification still relies on a single pressure distribution characteristic, failing to use weight information to compensate for the insufficient distinguishability of pressure distribution characteristics, thus making it difficult to solve the problem of misidentification of users with similar body types. Summary of the Invention

[0007] The problem this invention aims to solve is to achieve deep integration of weight prediction and user identification, and proposes a method and system for identifying mattress users based on a pressure array sensor.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A method for identifying mattress users based on a pressure array sensor includes the following steps:

[0010] S1. Place the pressure array sensor in the middle of the mattress to collect pressure data and user weight data during sleep;

[0011] S2. Filter the stress data during the user's sleep process collected in step S1, including row filtering and column filtering, to obtain the processed stress data during the user's sleep process;

[0012] S3. Construct a dataset based on the processed user stress data and user weight data during sleep;

[0013] S4. Construct a deep learning-based user weight prediction and user identification model and its loss function;

[0014] S5. Using the dataset obtained in step S3, train a deep learning-based user weight prediction and user identification model to obtain a trained deep learning-based user weight prediction and user identification model for user identification.

[0015] Furthermore, in step S1, the pressure array sensor is a flexible thin-film pressure array sensor, and the pressure data collected during the user's sleep process is two-dimensional array data, with each two-dimensional array data having m rows and n columns.

[0016] Furthermore, the specific implementation method of step S2 includes the following steps:

[0017] S2.1. The user's sleep stress data collected in step S1 is processed using a differential filtering strategy, expressed as follows:

[0018] (1)

[0019] (2)

[0020] (3)

[0021] in, The local threshold for the i-th row. The pressure value in the i-th row and j-th column is... Let be the local threshold coefficient for the i-th row;

[0022] S2.2. The user's sleep stress data collected in step S1 is processed using a differential filtering strategy, expressed as follows:

[0023] (4)

[0024] (5)

[0025] in, is the local threshold for the j-th column.

[0026] Furthermore, the specific implementation method of step S4 includes the following steps:

[0027] S4.1. Construct a deep learning-based user weight prediction and user identification model, including a user feature extractor, a weight feature extractor, a feature concatenation unit, a user classifier, and a weight predictor; the user feature extractor and the weight feature extractor are respectively connected to the feature concatenation unit, the feature concatenation unit is connected to the user classifier to output classification results, and the weight feature extractor is connected to the weight predictor to output prediction results.

[0028] S4.2. Construct a loss function for a deep learning-based user weight prediction and user identification model, which consists of user classification loss and weight prediction loss.

[0029] User classification loss Cross-entropy loss is used to measure the difference between the user classifier's predictions and the actual user labels. The expression is:

[0030] (6)

[0031] in, Category tags representing real users, This represents the probability distribution of user categories predicted by the model, and N represents the batch size of the input model data.

[0032] Weight Predicted Loss Mean squared error is used to measure the deviation between predicted and actual weight values. The expression is:

[0033] (7)

[0034] in, This indicates the user's actual weight. M represents the user's predicted weight, and M represents the batch size input to the model.

[0035] The loss function of the deep learning-based user weight prediction and user identification model is obtained. The expression is:

[0036] (8)

[0037] in, and These are the weighting coefficients for user classification loss and weight prediction loss, respectively.

[0038] Furthermore, the specific implementation method of step S4.1 includes the following steps:

[0039] S4.1.1. The first, second, third, and fourth convolutional units and the fully connected layer in the user feature extractor are connected sequentially to obtain the expression for the output of the user feature extractor:

[0040]

[0041] in, The output of the user feature extractor For the fully connected layer parameters of the user feature extractor, This is the output of the fourth convolutional unit. For the bias term of the user feature extractor;

[0042] S4.1.2. The first convolutional unit, the second convolutional unit, and the fully connected layer in the weight feature extractor are connected sequentially to obtain the output of the weight feature extractor. The expression is:

[0043]

[0044] in, It is a non-linear activation function. The parameters of the fully connected layer for the weight feature extractor are... This is the output of the second convolutional unit. This is the bias term for the weight feature extractor;

[0045] S4.1.3. The specific operation of the feature splicer is as follows:

[0046]

[0047] in, This is the output of the feature splicer;

[0048] S4.1.4. The specific operation of the user classifier is as follows, and the expression for linear layer regression is:

[0049]

[0050] in, For the output of the user classifier, For the parameters of the linear regression layer in the user classifier, For the bias term of the user classifier;

[0051] S4.1.5. The specific operation of the weight predictor is as follows, and the expression after linear layer regression is:

[0052]

[0053] in, The output of the weight predictor These are the parameters of the linear regression layer in the weight predictor. This is the bias term for the weight predictor.

[0054] Furthermore, the trained deep learning-based user weight prediction and user identification model obtained in step S5 is used for user identification. User identification is performed when the user is in a stable state. The determination of the user's stable state is based on a sliding window. If the difference between any two adjacent frames within the sliding window is less than a threshold, the sliding window is considered to be in a stable state at that time. A frame is selected within the stable sliding window as the input for user identification. The formula for calculating the difference between adjacent frames is:

[0055]

[0056] in, It represents the difference between adjacent frames. The data is a two-dimensional array at time t. This is the two-dimensional array data at time t+1. Let be the pressure value in the i-th row and j-th column at time t.

[0057] A user identification system based on a pressure array sensor includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed, it implements the steps of a method for identifying a user using a mattress based on a pressure array sensor as described above.

[0058] The beneficial effects of this invention are:

[0059] This invention discloses a method for identifying mattress users based on a pressure array sensor. By introducing weight prediction results as an auxiliary feature into the user identification process, it effectively supplements the information gap of a single identification dimension, significantly enhancing the distinguishability between different user features. This solves the problem of easy confusion between users with similar physical characteristics and improves the recognition algorithm's resistance to environmental interference, such as slight changes in posture and minor sensor errors. Ultimately, it achieves a dual improvement in user identification accuracy and robustness, significantly reducing misidentification and ensuring the reliability of core functions. Comparing the feature distribution of the joint user identification and weight prediction task training with that of the single user identification task training, it can be seen that the model trained by the joint task, i.e., introducing weight prediction auxiliary features, has clear cluster boundaries for user features in the feature space and significantly higher inter-class distinguishability. In contrast, the user features of the single-task training model without introducing weight prediction auxiliary features show partial category overlap. This further confirms the effectiveness of the technical solution of this invention in improving the core performance of user identification, ultimately achieving a dual improvement in user identification accuracy and robustness, and ensuring the reliability of core identification functions. Attached Figure Description

[0060] Figure 1 A flowchart of a method for identifying mattress users based on a pressure array sensor, as described in this invention;

[0061] Figure 2 The images are comparisons of data before and after filtering according to the present invention, where (a) is the image before flat filtering, (b) is the image after flat filtering, (c) is the image before side filtering, and (d) is the image after side filtering.

[0062] Figure 3 This is a schematic diagram of the pressure array sensor of the present invention arranged on a mattress;

[0063] Figure 4 This is a structural diagram of a deep learning-based user weight prediction and user identification model according to the present invention.

[0064] Figure 5 This is a comparison chart of the effects of single-task training and multi-task joint training of the present invention, where (a) is single-task training and (b) is multi-task training.

[0065] Figure 6 The diagram shows the changes in accuracy and loss during joint training of the model of this invention, where (a) is the accuracy change diagram and (b) is the loss change diagram.

[0066] Figure 7 A graph showing user characteristics in a reduced-dimensional format. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.

[0068] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.

[0069] To further understand the invention's content, features, and effects, the following specific embodiments are provided, along with accompanying drawings. Figure 1 - Appendix Figure 7Detailed explanation is as follows:

[0070] Example 1:

[0071] A method for identifying mattress users based on a pressure array sensor includes the following steps:

[0072] S1. Place the pressure array sensor in the middle of the mattress to collect pressure data and user weight data during sleep;

[0073] Furthermore, in step S1, the pressure array sensor is a flexible thin-film pressure array sensor, and the pressure data collected during the user's sleep process is two-dimensional array data, with each two-dimensional array data having m rows and n columns.

[0074] Furthermore, the dataset used in this embodiment includes sleep data from 12 users, including pressure array sensor data collected during sleep and corresponding user weight information. The pressure sensor data is in the format of a 40×26 two-dimensional array, which can completely reflect the force distribution of the user on the bed surface. User weight data are shown in Table 1:

[0075] Table 1

[0076]

[0077] Furthermore, the sensor is laid on a mattress, such as Figure 3 As shown, the pressure sensor is placed as close to the center of the mattress as possible. When a user lies on the surface of the pressure sensor, the raw signal is easily mixed with environmental noise. To ensure accurate extraction of effective pressure information, data from areas with weaker pressure, such as the legs and hands, is also retained.

[0078] S2. Filter the stress data during the user's sleep process collected in step S1, including row filtering and column filtering, to obtain the processed stress data during the user's sleep process;

[0079] Furthermore, the specific implementation method of step S2 includes the following steps:

[0080] S2.1. The user's sleep stress data collected in step S1 is processed using a differential filtering strategy, expressed as follows:

[0081] (1)

[0082] (2)

[0083] (3)

[0084] in, The local threshold for the i-th row. The pressure value is in the i-th row and j-th column. Let be the local threshold coefficient for the i-th row;

[0085] S2.2. The user's sleep stress data collected in step S1 is processed using a differential filtering strategy, expressed as follows:

[0086] (4)

[0087] (5)

[0088] in, is the local threshold for the j-th column.

[0089] Furthermore, A differentiated strategy is adopted in the row filtering: for the upper body (such as the torso), the threshold parameter is set relatively large due to the higher overall pressure value; while for the lower body (such as the legs), the threshold parameter is appropriately reduced to balance noise suppression and weak signal preservation, as shown in Formula 2, where the upper body... =1, lower half of the body =0.75. After processing by formula (3), the updated pressure data is obtained. In practical applications, when the intensity of environmental noise signals in the same row of data is close to the amplitude of torso pressure, simple row filtering is difficult to achieve effective separation. To address this problem, a secondary filter is further introduced in the column direction. Column filtering calculates the local threshold of each column using formula (4) and updates the data using formula (5), thereby achieving effective elimination of residual weak noise.

[0090] Through the aforementioned row and column bidirectional filtering process, noise can be suppressed in both the horizontal and vertical directions, improving the purity and effectiveness of the pressure signal. The filtered data effectively preserves signals from weak pressure areas such as legs and hands, while significantly reducing environmental noise interference, thus achieving a balance between filtering effectiveness and information preservation. Figure 2 As shown, Figure 2(a) and (c) show the data before filtering, while (b) and (d) show the data after filtering. A comparison clearly shows that the filtered data effectively preserves information from weak pressure areas such as the legs and hands. This is mainly due to the row and column filtering methods used in this study. These methods focus on the current horizontal or vertical data characteristics, significantly reducing over-filtering of weak pressure areas through directional processing, ensuring that subtle pressure signals are not weakened, thus achieving a balance between filtering effect and information preservation. Further observation of the data in the figure below shows that the combined application of row and column filtering has a significant noise suppression effect: through dual processing in both horizontal and vertical directions, abnormal fluctuations in the data are effectively eliminated, greatly improving the purity and effectiveness of the pressure signal.

[0091] S3. Construct a dataset based on the processed user stress data and user weight data during sleep;

[0092] S4. Construct a deep learning-based user weight prediction and user identification model and its loss function;

[0093] Furthermore, the specific implementation method of step S4 includes the following steps:

[0094] S4.1. Construct a deep learning-based user weight prediction and user identification model, including a user feature extractor, a weight feature extractor, a feature concatenation unit, a user classifier, and a weight predictor; the user feature extractor and the weight feature extractor are respectively connected to the feature concatenation unit, the feature concatenation unit is connected to the user classifier to output classification results, and the weight feature extractor is connected to the weight predictor to output prediction results; as shown... Figure 4 As shown;

[0095] Furthermore, the specific implementation method of step S4.1 includes the following steps:

[0096] S4.1.1. The first, second, third, and fourth convolutional units and the fully connected layer in the user feature extractor are connected sequentially to obtain the expression for the output of the user feature extractor:

[0097]

[0098] in, The output of the user feature extractor For the fully connected layer parameters of the user feature extractor, This is the output of the fourth convolutional unit. For the bias term of the user feature extractor;

[0099] Furthermore, the specific structure is detailed as follows:

[0100] First convolutional unit:

[0101] Convolutional layer: Input channel Output channel kernel size stride=1, p=1.

[0102] Batch normalization layer: Normalizes the output of convolution.

[0103] Activation layer: Apply ReLU to any element value d in the batch normalization layer, which is defined as follows: .

[0104] Second convolutional unit:

[0105] Convolutional layer: Input channel Output channel kernel size stride=1, p=1.

[0106] Batch normalization layer: Normalizes the output of convolution.

[0107] Activation layer: Apply ReLU to any element value d in the batch normalization layer, which is defined as follows: .

[0108] Pooling layer: Max pooling (kernel=2, stride=2), halves the width and height.

[0109] Third convolutional unit:

[0110] Convolutional layer: Input channel Output channel kernel size stride=1, p=1.

[0111] Batch normalization layer: Normalizes the output of convolution.

[0112] Activation layer: Apply ReLU to any element value d in the batch normalization layer, which is defined as follows: .

[0113] Fourth convolutional unit:

[0114] Convolutional layer: Input channel Output channel 256, kernel size stride=1, p=1.

[0115] Batch normalization layer: Normalizes the output of convolution.

[0116] Activation layer: Apply ReLU to any element value d in the batch normalization layer, which is defined as follows: .

[0117] Pooling layer: Max pooling (kernel=2, stride=2), halves the width and height.

[0118] Fully connected layer:

[0119] First, flatten the features: ,in It is the output of the fourth convolutional unit.

[0120] Then it goes through a fully connected layer: ;

[0121] S4.1.2. The first convolutional unit, the second convolutional unit, and the fully connected layer in the weight feature extractor are connected sequentially to obtain the output of the weight feature extractor. The expression is:

[0122]

[0123] in, It is a non-linear activation function. The parameters of the fully connected layer for the weight feature extractor are... This is the output of the second convolutional unit. This is the bias term for the weight feature extractor;

[0124] Furthermore, the detailed structure of the weight feature extractor is as follows:

[0125] First convolutional unit:

[0126] Convolutional layer: Input channel Output channel kernel size stride=1, p=1.

[0127] Activation layer: For any element in the convolutional layer Applying ReLU, which is defined as Batch normalization layer: Normalizes the output of convolution.

[0128] Pooling layer: Max pooling (kernel=2, stride=2), halves the width and height.

[0129] Second convolutional unit:

[0130] Convolutional layer: Input channel Output channel kernel size stride=1, p=1.

[0131] Activation layer: Apply ReLU to any element value d in the convolutional layer, which is defined as follows: .

[0132] Batch normalization layer: Normalizes the output of convolution.

[0133] Pooling layer: Max pooling (kernel=2, stride=2), halves the width and height.

[0134] Fully connected layer:

[0135] First layer: ,in This is the output of the pooling part from the previous layer.

[0136] Second layer: .

[0137] S4.1.3. The specific operation of the feature splicer is as follows:

[0138]

[0139] in, This is the output of the feature splicer; ;

[0140] S4.1.4. The specific operation of the user classifier is as follows, and the expression for linear layer regression is:

[0141]

[0142] in, For the output of the user classifier, For the parameters of the linear regression layer in the user classifier, For the bias term of the user classifier;

[0143] S4.1.5. The specific operation of the weight predictor is as follows, and the expression after linear layer regression is:

[0144]

[0145] in, The output of the weight predictor These are the parameters of the linear regression layer in the weight predictor. This is the bias term for the weight predictor.

[0146] S4.2. Construct a loss function for a deep learning-based user weight prediction and user identification model, which consists of user classification loss and weight prediction loss.

[0147] User classification loss Cross-entropy loss is used to measure the difference between the user classifier's predictions and the actual user labels. The expression is:

[0148] (6)

[0149] in, Category tags representing real users, This represents the probability distribution of user categories predicted by the model, and N represents the batch size of the input model data.

[0150] Weight Predicted Loss Mean squared error is used to measure the deviation between predicted and actual weight values. The expression is:

[0151] (7)

[0152] in, This indicates the user's actual weight. M represents the user's predicted weight, and M represents the batch size input to the model.

[0153] The loss function of the deep learning-based user weight prediction and user identification model is obtained. The expression is:

[0154] (8)

[0155] in, and These are the weighting coefficients for user classification loss and weight prediction loss, respectively.

[0156] Furthermore, as shown in the model structure diagram, the user features output by the user feature extractor serve as input to both the user classifier and the weight predictor. Specifically, in the weight predictor, the user features and weight features are concatenated as input. Therefore, the loss function for the user classification task is used to update the parameters of the user classifier and the user feature extractor; the loss function for the weight prediction task is used to update the parameters of the weight predictor, the weight feature extractor, and the user feature extractor, thereby achieving joint optimization of user classification and weight prediction. The joint loss function is used to balance the importance of the classification task and the weight prediction task. Its value can be set according to different application scenarios and experimental results. For example, in scenarios where it is necessary to enhance the model's ability to distinguish different user features, the value can be appropriately increased. The value of is chosen to enhance the discriminative power between samples of different categories; however, in scenarios where it is necessary to enhance the consistency of features across different states of the same user, the value can be appropriately increased. The value of is determined to ensure the robustness of the model on samples of the same class.

[0157] Furthermore, before training begins, a calibration dataset is provided, and candidate weight combinations are performed based on this calibration dataset. , The search is performed using methods such as grid search and Bayesian optimization. Subsequently, a comprehensive evaluation index is calculated on the validation set, and the weight combination corresponding to the optimal index is used as the final weight. The comprehensive evaluation index can simultaneously consider the performance of the user classification task and the weight prediction task. The performance of the classification task is represented by classification accuracy (Acc), and the performance of the weight prediction task is represented by mean absolute error (MAE), and the weight is further evaluated using a preset importance coefficient. The two are weighted and combined. When the comprehensive index When the minimum value is obtained, ( , The weight combination corresponding to ) is taken as the optimal solution and is used in a fixed manner throughout the training process.

[0158] S5. Using the dataset obtained in step S3, train a deep learning-based user weight prediction and user identification model to obtain a trained deep learning-based user weight prediction and user identification model for user identification.

[0159] Furthermore, the trained deep learning-based user weight prediction and user identification model obtained in step S5 is used for user identification. User identification is performed when the user is in a stable state. The determination of the user's stable state is based on a sliding window. If the difference between any two adjacent frames within the sliding window is less than a threshold, the sliding window is considered to be in a stable state at that time. A frame is selected within the stable sliding window as the input for user identification. The formula for calculating the difference between adjacent frames is:

[0160]

[0161] in, It represents the difference between adjacent frames. The data is a two-dimensional array at time t. This is the two-dimensional array data at time t+1. Let be the pressure value in the i-th row and j-th column at time t.

[0162] Furthermore, when continuously acquiring frame sequences , of which each frame , t=1, ...,T. User identification and weight prediction are only performed when the user is in a stable state (i.e., the inter-frame fluctuations are small). Stability is determined using a sliding window: if the difference between any two adjacent frames within the window (measured by a calculable scalar) is less than a threshold, then the window is considered to be in a stable state at the corresponding time point; a frame is selected within the stable window as the identification / prediction input.

[0163] This embodiment describes a method for identifying mattress users based on a pressure array sensor. By introducing weight prediction results as an auxiliary feature into the user identification process, it effectively supplements the information gap of a single identification dimension, significantly enhancing the distinguishability between different user characteristics. This solves the problem of easy confusion between users with similar physical characteristics and improves the identification algorithm's resistance to environmental interference (such as slight changes in posture or minor sensor errors). Ultimately, it achieves a dual improvement in user identification accuracy and robustness, significantly reducing misidentification and ensuring the reliability of core functions. Figure 5 As shown, comparing the feature distributions of the "user identification-weight prediction" joint task training and the single user identification task training reveals that the model trained with the joint task (i.e., incorporating weight prediction auxiliary features) exhibits clearer cluster boundaries for user features in the feature space, with significantly higher inter-class discriminative power. In contrast, the single-task training model (without incorporating weight prediction auxiliary features) shows partial category overlap in user features. This further confirms the effectiveness of this technical solution in improving the core performance of user identification, ultimately achieving a dual improvement in user identification accuracy and robustness, ensuring the reliability of the core identification function. For typical home / multi-user sharing scenarios such as smart mattresses and sleep monitoring systems, this invention can accurately distinguish different users, avoiding the functional failure problem of "single-person exclusive" technical solutions in shared environments. For example, in the home smart mattress scenario, it can automatically identify the current user's identity and match their personalized sleep monitoring parameters (such as pressure sensitivity thresholds and sleep report push accounts), enabling shared devices to have a "personalized" adaptability, expanding the application scope of the technology.

[0164] from Figure 6 As can be seen, this embodiment uses a dataset containing 11 samples for verification experiments, and presents the dynamic changes of the loss function during training and the evolution trend of the accuracy of the user recognition task on the test set in the form of curves (the results are shown in the figure below). The experimental results show that as the number of training rounds increases, the model's test accuracy shows a steady upward trend, reaching a peak of 95% in the 8th training round; at the same time, the loss function value continues to decrease with the advancement of training rounds, and the two trends are consistent, confirming the effectiveness and convergence of the model training.

[0165] To visually represent the distribution characteristics of the high-dimensional features output by the model, this embodiment uses Principal Component Analysis (PCA) to reduce its dimensionality to a two-dimensional feature space, and then displays the results through visualization. Figure 7 As shown. From Figure 7 It can be clearly observed that the two-dimensional features corresponding to different users exhibit significant clustering discrimination in space, with clear boundaries between feature points of each category and no feature overlap or confusion. This indicates that the high-dimensional features extracted by the model have good user discrimination ability and can effectively represent the unique attributes of different users.

[0166] Example 2:

[0167] A user identification system based on a pressure array sensor includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed, it implements the steps of a method for identifying a user using a mattress based on a pressure array sensor as described in Embodiment 1.

[0168] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0169] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for identifying users of a mattress based on a pressure array sensor, characterized in that, Includes the following steps: S1. Place the pressure array sensor in the middle of the mattress to collect pressure data and user weight data during sleep; In step S1, the pressure array sensor is a flexible thin-film pressure array sensor. The pressure data collected during the user's sleep process is two-dimensional array data, with each two-dimensional array data having m rows and n columns. S2. Filter the stress data during the user's sleep process collected in step S1, including row filtering and column filtering, to obtain the processed stress data during the user's sleep process; The specific implementation method of step S2 includes the following steps: S2.

1. The user's sleep stress data collected in step S1 is processed using a differential filtering strategy, expressed as follows: (1) (2) (3) in, The local threshold for the i-th row. The pressure value in the i-th row and j-th column is... Let be the local threshold coefficient for the i-th row; S2.

2. The user's sleep stress data collected in step S1 is processed using a differential filtering strategy, expressed as follows: (4) (5) in, Let be the local threshold for the j-th column; S3. Construct a dataset based on the processed user stress data and user weight data during sleep; S4. Construct a deep learning-based user weight prediction and user identification model and its loss function; A deep learning-based user weight prediction and user identification model is constructed, including a user feature extractor, a weight feature extractor, a feature concatenation unit, a user classifier, and a weight predictor. The user feature extractor and the weight feature extractor are respectively connected to the feature concatenation unit. The feature concatenation unit is connected to the user classifier to output the classification result, and the weight feature extractor is connected to the weight predictor to output the prediction result. S5. Using the dataset obtained in step S3, train a deep learning-based user weight prediction and user identification model to obtain a trained deep learning-based user weight prediction and user identification model for user identification. In step S5, a trained deep learning-based user weight prediction and user identification model is used for user identification. User identification is performed when the user is in a stable state. The determination of a stable user state is achieved through a sliding window. If the difference between any two adjacent frames within the sliding window is less than a threshold, the sliding window is considered to be in a stable state at that moment. A frame is selected within the stable sliding window as the input for user identification. The formula for calculating the difference between adjacent frames is: in, It represents the difference between adjacent frames. The data is a two-dimensional array at time t. This is the two-dimensional array data at time t+1. Let be the pressure value in the i-th row and j-th column at time t.

2. The method for identifying a mattress user based on a pressure array sensor according to claim 1, characterized in that, The specific implementation method of step S4 includes the following steps: S4.

1. Construct a deep learning-based model for user weight prediction and user identification; S4.

2. Construct a loss function for a deep learning-based user weight prediction and user identification model, which consists of user classification loss and weight prediction loss. User classification loss Cross-entropy loss is used to measure the difference between the user classifier's predictions and the actual user labels. The expression is: (6) in, Category tags representing real users, This represents the probability distribution of user categories predicted by the model, and N represents the batch size of the input model data. Weight Predicted Loss Mean squared error is used to measure the deviation between predicted and actual weight values. The expression is: (7) in, This indicates the user's actual weight. M represents the user's predicted weight, and M represents the batch size input to the model. The loss function of the deep learning-based user weight prediction and user identification model is obtained. The expression is: (8) in, and These are the weighting coefficients for user classification loss and weight prediction loss, respectively.

3. The method for identifying a mattress user based on a pressure array sensor according to claim 2, characterized in that, The specific implementation method of step S4.1 includes the following steps: S4.1.

1. The first, second, third, and fourth convolutional units and the fully connected layer in the user feature extractor are connected sequentially to obtain the expression for the output of the user feature extractor: in, The output of the user feature extractor For the fully connected layer parameters of the user feature extractor, This is the output of the fourth convolutional unit. For the bias term of the user feature extractor; S4.1.

2. The first convolutional unit, the second convolutional unit, and the fully connected layer in the weight feature extractor are connected sequentially to obtain the output of the weight feature extractor. The expression is: in, It is a non-linear activation function. The parameters of the fully connected layer for the weight feature extractor are... This is the output of the second convolutional unit. This is the bias term for the weight feature extractor; S4.1.

3. The specific operation of the feature splicer is as follows: in, This is the output of the feature splicer; S4.1.

4. The specific operation of the user classifier is as follows, and the expression for linear layer regression is: in, For the output of the user classifier, For the parameters of the linear regression layer in the user classifier, For the bias term of the user classifier; S4.1.

5. The specific operation of the weight predictor is as follows, and the expression after linear layer regression is: in, The output of the weight predictor These are the parameters of the linear regression layer in the weight predictor. This is the bias term for the weight predictor.

4. A user identification system based on a pressure array sensor, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed, implements the steps of a method for identifying a user of a mattress based on a pressure array sensor as described in any one of claims 1-3.

Images