Old people living behavior analysis method for smart pension

Abstract

The invention relates to the technical field of data model analysis, in particular to an old people living behavior analysis method for intelligent old people care. Comprising the following steps: acquiring multi-modal data of living behaviors of old people to obtain synchronized multi-modal data; preprocessing the synchronized multi-modal data, and performing feature extraction and optimization on the preprocessed multi-modal data to obtain comprehensive feature data; acquiring historical living behavior data of old people to obtain historical comprehensive feature data; and based on the historical comprehensive feature data, constructing a hierarchical time sequence dynamic neural network model to obtain a trained and verified hierarchical time sequence dynamic neural network model, and processing the comprehensive feature data through the trained and verified hierarchical time sequence dynamic neural network model to obtain an old people living behavior analysis result. The technical problems that the multi-modal data feature extraction of the living behavior is not accurate enough and the sensitivity of an analysis model is low in the aged living behavior analysis are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of data model analysis, and in particular to a method for analyzing the daily living behavior of the elderly for smart elderly care. Background Art

[0002] With the intensification of global aging, health management and safety monitoring of the elderly have become issues of widespread concern in society. Many elderly people live alone at home and face health risks such as falls, sudden illnesses, and insufficient daily activities. Therefore, how to monitor the living behavior of the elderly in real time, detect abnormal activities in a timely manner, and ensure the safety of the elderly has become a technical problem that needs to be solved urgently. At present, traditional health monitoring of the elderly mainly relies on wearable devices or monitoring systems based on a single sensor. These technical means can usually only collect data of a single modality (such as heart rate or acceleration), and there are certain limitations in data accuracy, real-time and adaptability, making it difficult to fully and accurately identify the behavior patterns and activity status of the elderly.

[0003] However, the existing methods for analyzing the living behavior of the elderly have the following technical problems: the extraction of multimodal data features of the living behavior of the elderly is not accurate enough and the analysis model has low sensitivity. Summary of the invention

[0004] The present invention provides a method for analyzing the living behavior of the elderly for smart elderly care, so as to solve the technical problems of inaccurate extraction of multimodal data features of the living behavior and low sensitivity of the analysis model in the existing analysis of the living behavior of the elderly.

[0005] The present invention provides a method for analyzing the daily living behavior of the elderly for smart elderly care, which specifically includes the following technical solutions:

[0006] A method for analyzing the living behavior of the elderly for smart elderly care includes the following steps:

[0007] S1. Acquire multimodal data of the living behavior of the elderly, align the multimodal data to a unified time axis, and obtain synchronized multimodal data; preprocess the synchronized multimodal data to obtain preprocessed multimodal data; and extract and optimize features of the preprocessed multimodal data to obtain comprehensive feature data;

[0008] S2. Obtain historical daily living behavior data of the elderly, and obtain historical comprehensive feature data through the processing of S1 for the historical daily living behavior data of the elderly; construct a hierarchical temporal dynamic neural network model based on the historical comprehensive feature data, and train and verify the hierarchical temporal dynamic neural network model; process the comprehensive feature data based on the trained and verified hierarchical temporal dynamic neural network model to obtain the analysis results of the daily living behavior of the elderly.

[0009] Preferably, the S1 specifically includes:

[0010] A hierarchical adaptive nested interactive optimization and weighted feedback feature extraction algorithm is introduced to extract behavioral features from the preprocessed multimodal data sources.

[0011] Preferably, the S1 specifically includes:

[0012] In the process of implementing the hierarchical adaptive nested interaction optimization and weighted feedback feature extraction algorithm, preliminary feature extraction is performed on the preprocessed multimodal data to obtain preliminary modal feature data; based on the preliminary modal feature data, a nested interaction matrix between modalities is established to obtain an initial feature data set;

[0013] Any element in the nested interaction matrix is ​​defined as follows:

[0014]

[0015] Among them, H a，b is the nested interaction eigenvalue of mode a and mode b; d a d b is the characteristic dimension of mode a and mode b; is the kth eigenvalue of mode a; is the sth eigenvalue of mode b; Represents the mutual influence between the eigenvalues ​​of mode a and mode b; is a nonlinear factor.

[0016] Preferably, the S1 specifically includes:

[0017] In the process of implementing the hierarchical adaptive nested interactive optimization and weighted feedback feature extraction algorithm, the initial feature data set is subjected to cross-feedback and hierarchical weighted optimization, and a cross-feedback weight matrix is ​​introduced to weight the initial feature data set to generate hierarchically optimized feature data F( l ):

[0018]

[0019] Among them, F(l) is the optimized feature data of the lth layer; W( l ) is the cross-feedback weight matrix of the lth layer; F init is the initial feature data set; α is the dynamic balance factor; F init，i is the i-th feature in the initial feature dataset; is the cross-feedback weight of the i-th feature in the initial feature dataset of the l-th layer; dd is the feature dimension of the initial feature dataset.

[0020] Preferably, the S1 specifically includes:

[0021] In the process of implementing the hierarchical adaptive nested interactive optimization and weighted feedback feature extraction algorithm, a hierarchical feature cumulative feedback matrix is ​​constructed based on the optimized feature data; based on the cumulative feedback matrix, a dimensionality reduction weight matrix is ​​introduced to calculate the variance contribution of the feature to obtain the final comprehensive feature data. The specific formula is:

[0022]

[0023] Among them, F final is the final comprehensive feature data; D (l) is the dimension reduction weight matrix; L (i) is the cumulative feedback matrix; K is the total number of layers.

[0024] Preferably, the S2 specifically includes:

[0025] The historical comprehensive feature data are divided into a training set and a validation set, and data expansion is performed on the historical comprehensive feature data used as the training set to obtain the expanded historical comprehensive feature data.

[0026] Preferably, the S2 specifically includes:

[0027] The hierarchical time series dynamic neural network model includes an input layer, a dynamic time series layer, an information fusion layer and a decision output layer; the expanded historical comprehensive feature data is used as the input of the input layer, and the output of the input layer is obtained after weighted processing, and the output of the input layer is used as the input of the dynamic time series layer.

[0028] Preferably, the S2 specifically includes:

[0029] In the hierarchical time series dynamic neural network model, the dynamic time series layer combines the long short-term memory neural network and the gated recurrent unit neural network to capture the dynamic characteristics of time series data; in the information fusion layer, a dynamic weight update mechanism is introduced. The update formula is as follows:

[0030]

[0031] in, is the dynamic weight at time step t; is the dynamic weight at time step t-1; η is the basic learning rate; is the error feedback term; It is in The prediction error at the moment; is the time exponential decay function; is the time decay rate; is the normalization coefficient; is the sum of the absolute values ​​of the prediction errors; It is The prediction error at the moment; is the oscillation adjustment coefficient; sin(θ·t) is the periodic variation function; θ is the frequency factor.

[0032] Preferably, the S2 specifically includes:

[0033] After the dynamic weights are updated, the hierarchical temporal dynamic neural network model enters the decision output layer, and the output of the dynamic temporal layer is synthesized using a deep fully connected network. The trained hierarchical temporal dynamic neural network model is then verified using the historical comprehensive feature data as a verification set to obtain a trained and verified hierarchical temporal dynamic neural network model.

[0034] The beneficial effects of the technical solution of the present invention are:

[0035] 1. A hierarchical adaptive nested interaction optimization and weighted feedback feature extraction algorithm is introduced. By establishing a multimodal nested interaction matrix, dynamic interaction and weighted optimization of modal data are realized. The algorithm extracts features from the relationship between different modal data, and captures the deep correlation between features through hierarchical weight adjustment and nonlinear amplification, effectively improving the feature discrimination ability. This optimization process makes the extracted features more representative, robust and information-rich, which can effectively cope with the complex daily behavior patterns of the elderly and provide high-quality input for the hierarchical temporal dynamic neural network model.

[0036] 2. In the hierarchical temporal dynamic neural network model, the dynamic weight update mechanism is used to achieve rapid adaptive learning of the model to historical comprehensive feature data. The dynamic weight update mechanism dynamically adjusts the weights through multiple factors such as time exponential decay and periodic oscillation, so that the model can be quickly updated according to real-time error feedback and adapt to the periodicity and variability of the elderly's behavior, thereby improving the generalization and real-time performance of the model. It not only enhances the model's sensitivity to short-term behavioral changes in the elderly, but also can adapt to long-term changes in behavioral patterns of the elderly, making detection of abnormal behaviors of the elderly more sensitive. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flow chart of a method for analyzing the daily living behavior of the elderly for smart elderly care described in the present invention. DETAILED DESCRIPTION

[0038] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0040] The following is a detailed description of a method for analyzing the daily living behavior of the elderly for smart elderly care provided by the present invention in conjunction with the accompanying drawings.

[0041] See attached Figure 1 , which shows a flow chart of a method for analyzing the daily living behavior of the elderly for smart elderly care provided by an embodiment of the present invention, the method comprising the following steps:

[0042] S1. Acquire multimodal data of the living behavior of the elderly, align the multimodal data to a unified time axis, and obtain synchronized multimodal data; preprocess the synchronized multimodal data to obtain preprocessed multimodal data; and extract and optimize features of the preprocessed multimodal data to obtain comprehensive feature data;

[0043] The multimodal data of the elderly's daily life behavior is obtained through sensors installed in the elderly's living space, including real-time sensor data such as video, acceleration, heart rate, temperature, humidity, etc.; in order to ensure the synchronization of multimodal data, the multimodal data is aligned to a unified time axis using a timestamp alignment method to obtain synchronized multimodal data. Furthermore, the synchronized multimodal data is preprocessed to obtain preprocessed multimodal data; the preprocessing processes such as denoising, outlier processing, and normalization all use existing technologies well known in the art and will not be elaborated here.

[0044] Furthermore, a hierarchical adaptive nested interaction optimization and weighted feedback feature extraction algorithm is introduced to extract high-quality behavioral features from the preprocessed multimodal data source. The hierarchical adaptive nested interaction optimization and weighted feedback feature extraction algorithm is based on the multimodal nested interaction matrix, optimizes the behavioral features layer by layer through dynamic weighting and adaptive feedback, and finally generates comprehensive feature data. The specific implementation process is as follows:

[0045] First, the existing feature engineering technology is used to perform preliminary feature extraction on the preprocessed multimodal data to obtain preliminary modal feature data, which is recorded as Where a∈{1, 2, …, M} represents the modal number, d a Represents the characteristic dimension of mode a. The nested interaction matrix between modes is established through preliminary modal feature data, laying the foundation for the optimization of the final behavioral characteristics.

[0046] Furthermore, nested interaction features between modalities are constructed based on the preliminary modal feature data. The interaction between modalities is established through the nested interaction matrix H to capture the association information between modalities. Any element in the nested interaction matrix is ​​defined as follows:

[0047]

[0048] Among them, H a，b is the nested interaction eigenvalue of mode a and mode b, which is used to capture the complex relationship between different modal data; d a d b It is the characteristic dimension of mode a and mode b, which is used to standardize the characteristic scale of modal data; is the kth eigenvalue of mode a; is the sth eigenvalue of mode b; Represents the mutual influence between the eigenvalues ​​of mode a and mode b; It is a nonlinear factor used to amplify the nonlinear correlation between high-dimensional data features and suppress the influence of noise data. By defining and constructing the nested interaction matrix, the interaction between multiple modalities is completed, and the set of nested interaction matrices H is obtained, which is used as the initial feature data set F init .

[0049] Furthermore, the initial feature data set is subjected to cross-feedback and hierarchical weighted optimization processing; in order to optimize the contribution of features at different levels, the cross-feedback weight matrix W is introduced (l) , W (l) is the cross-feedback weight matrix of layer l; it is designed to adjust the weights of different features at different levels. Any element in the cross-feedback weight matrix is ​​defined by the following formula:

[0050]

[0051] in, is the cross-feedback weight of the i-th feature (element) of the initial feature data set in the l-th layer; F init，i is the i-th feature in the initial feature dataset; F init，j is the jth feature in the initial feature dataset; is the local mean of the i-th feature in the l-th layer, reflecting the overall distribution trend of the feature in this layer; σ l is the standard control parameter in the lth layer, used to adjust the weight distribution within the layer; βl is the nonlinear power control parameter in the lth layer, which is used to adjust the amplification effect of feature differences in weights; d is the feature dimension of the initial feature data set.

[0052] Then, using the cross-feedback weight matrix W (l) For the initial feature dataset F init Weighted to generate hierarchically optimized feature data F (l) :

[0053]

[0054] Among them, F (l) is the optimized feature data of the lth layer, which is used to generate the dynamically weighted optimized feature data; α is the dynamic balance factor, which is used to adjust the nonlinear contribution of the optimization level.

[0055] Furthermore, in order to further accumulate feedback on the optimization results at different levels and make the hierarchical relationship between features more stable, a hierarchical feature cumulative feedback matrix L is constructed. (l) :

[0056]

[0057] Among them, L (l) is the cumulative feedback matrix of the lth layer, which is used for layer-by-layer feedback accumulation to optimize the balance of the feature vector; It is Feature data after layer optimization; It is Feature data after layer optimization; It is Layer cross-feedback weight matrix; γ is the feedback attenuation factor, which controls the decrease of cumulative feedback between layers.

[0058] Furthermore, in order to further reduce feature redundancy and reduce computational overhead, the dimension reduction weight matrix D is introduced (l) Provides a basis for adaptive dimensionality reduction. For each level of the cumulative feedback matrix L (l) , calculate the variance contribution of the feature, and define the dimension reduction weight matrix D by the following formula (l) Any element in

[0059]

[0060] in, is the first dimension reduction weight matrix in the lth layer Line Column elements represent the weights of the l-th layer features after dimensionality reduction; is the first cumulative feedback matrix in the lth layer Line Column elements; is the first cumulative feedback matrix in the lth layer Line Column elements;

[0061] After the dimensionality reduction weight matrix calculation is completed at all levels, the final comprehensive feature data is generated by the following formula:

[0062]

[0063] Among them, F final It is the final comprehensive feature data, which integrates the results of multi-layer feature optimization and feedback; D (l) is the dimension reduction weight matrix of the lth layer; K is the total number of layers, which is used for normalization.

[0064] Through the above-mentioned nested hierarchical optimization, cross-feedback and nonlinear dimensionality reduction process, the precise extraction and optimization of multimodal data features are structurally achieved, generating stable and information-rich comprehensive feature data with high distinctiveness and representativeness, which is suitable for applications in complex scenarios such as the analysis of the daily living behavior of the elderly.

[0065] S2. Obtain historical daily living behavior data of the elderly, and obtain historical comprehensive feature data through the processing of S1 on the historical daily living behavior data; construct a hierarchical temporal dynamic neural network model based on the historical comprehensive feature data, and train and verify the hierarchical temporal dynamic neural network model; process the comprehensive feature data based on the trained and verified hierarchical temporal dynamic neural network model to obtain the analysis results of the daily living behavior of the elderly.

[0066] First, the historical daily living behavior data of the elderly are obtained from the existing database, and the historical daily living behavior data of the elderly are processed through the process of step S1 to obtain historical comprehensive feature data. Based on the historical comprehensive feature data, a hierarchical temporal dynamic neural network model, i.e., an intelligent behavior analysis model, is constructed. The construction process is as follows:

[0067] The historical comprehensive feature data are divided into a training set and a validation set. The existing data enhancement technology is used to expand the historical comprehensive feature data used as the training set to obtain the expanded historical comprehensive feature data, thereby ensuring the adaptability of the hierarchical temporal dynamic neural network model to different scenarios.

[0068] The expanded historical comprehensive feature data is used as the input of the input layer, and the output of the input layer is obtained after weighted processing. The output of the input layer is used as the input of the dynamic time series layer. The dynamic time series layer combines the existing long short-term memory and gated recurrent unit neural network structures to capture the dynamic characteristics of time series data; the long short-term memory neural network can maintain the long-term dependence of information; the gated recurrent unit neural network flexibly controls the flow of information through the gating mechanism; as the long short-term memory and gated recurrent unit states are generated, they enter the information fusion layer, and a dynamic weight update mechanism is introduced in the information fusion layer. The dynamic weight update mechanism ensures that the hierarchical time series dynamic neural network model can adaptively learn the historical comprehensive feature data, and the dynamic weights The update formula is as follows:

[0069]

[0070] in, is the dynamic weight at time step t, which represents the weight coefficient used to calculate the model output at the current time step, and it is adjusted with the time step; is the dynamic weight at time step t-1, which is used for the output of the previous time step and is the basis of the current dynamic weight, determining the cumulative effect of the dynamic weight update; η is the basic learning rate, which determines the intensity of the influence of the error feedback term on the dynamic weight update; is the error feedback term; It is in The prediction error at a certain moment is defined as the difference between the predicted value and the true value. It describes the difference between the model prediction and the true label and is the core basis for dynamic weight update. is a time exponential decay function, which is used to decay the prediction error of past time steps, ensuring that recent errors have a greater impact on the dynamic weight update; It is the time decay rate, which is used to control the influence of the prediction error of the past time step on the current dynamic weight. The larger its value is, the smaller the influence of the historical error far away from the current time step on the current dynamic weight update; It is the normalization coefficient, which prevents the dynamic weight from fluctuating too much due to large errors. By normalizing the absolute value of the prediction error, the dynamic weight update is made smoother and more stable. It is the sum of the absolute values ​​of the prediction errors, which is used to adjust the impact of the errors on the dynamic weights and avoid extreme updates caused by excessive errors. Its function is to accumulate the absolute values ​​of all prediction errors to balance the update speed of the dynamic weights. It is The prediction error at the moment; is the oscillation adjustment coefficient, which is used to control the intensity of periodic changes; sin(θ·t) is a periodic change function. By introducing periodic oscillations, the dynamic weight update has a certain periodic adjustment and can adapt to periodic behavior characteristics more sensitively; θ is the frequency factor, which is used to control the periodicity of the sine function and ensure that the oscillation frequency matches the periodic characteristics of the behavior data.

[0071] After the dynamic weights are updated, the hierarchical temporal dynamic neural network model enters the decision output layer. The output of the dynamic temporal layer is synthesized using a deep fully connected network, and the final analysis results of the elderly's living behavior are generated through the deep fully connected layer. The trained hierarchical temporal dynamic neural network model is then verified using the existing verification algorithm using the historical comprehensive feature data as the verification set to obtain a trained and verified hierarchical temporal dynamic neural network model.

[0072] The comprehensive feature data were processed through a trained and verified hierarchical temporal dynamic neural network model to obtain the analysis results of the living behavior of the elderly, including: classification of behavior types, including stillness, walking, sitting, standing up and lying down; abnormal behaviors, including falls, long-term stillness, rapid movements, etc.; behavior frequency and duration analysis, including the number of steps walked per day, the duration of each sitting and lying down, etc.; daily activity patterns, including daily activity peak and trough time periods, etc.; sleep behavior, including the number of turning over at night, the frequency of getting up at night, the number of sleep interruptions, etc.

[0073] In summary, a method for analyzing the living behavior of the elderly for smart elderly care has been completed.

[0074] The order of the embodiments of the invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0075] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0076] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A method for analyzing the daily living behavior of the elderly for smart elderly care, characterized in that: The following steps are involved: S1. Obtain multimodal data of the living behavior of the elderly, align the multimodal data to a unified time axis, and obtain synchronized multimodal data; Preprocessing the synchronized multimodal data to obtain preprocessed multimodal data; and extracting and optimizing features of the preprocessed multimodal data to obtain comprehensive feature data; S2. Obtain historical daily living behavior data of the elderly, and obtain historical comprehensive feature data through the processing of S1 on the historical daily living behavior data; construct a hierarchical temporal dynamic neural network model based on the historical comprehensive feature data, and train and verify the hierarchical temporal dynamic neural network model; process the comprehensive feature data based on the trained and verified hierarchical temporal dynamic neural network model to obtain the analysis results of the daily living behavior of the elderly.

2. The method for analyzing the daily living behavior of the elderly for smart elderly care according to claim 1 is characterized in that: The S1 specifically includes: A hierarchical adaptive nested interactive optimization and weighted feedback feature extraction algorithm is introduced to extract behavioral features from the preprocessed multimodal data sources.

3. The method for analyzing the daily living behavior of the elderly for smart elderly care according to claim 2 is characterized in that: The S1 specifically includes: In the process of implementing the hierarchical adaptive nested interaction optimization and weighted feedback feature extraction algorithm, preliminary feature extraction is performed on the preprocessed multimodal data to obtain preliminary modal feature data; based on the preliminary modal feature data, a nested interaction matrix between modalities is established to obtain an initial feature data set; Any element in the nested interaction matrix is ​​defined as follows: Among them, H a，b is the nested interaction eigenvalue of mode a and mode b; d a ,d b is the characteristic dimension of mode a and mode b; is the kth eigenvalue of mode a; is the sth eigenvalue of mode b; Represents the mutual influence between the eigenvalues ​​of mode a and mode b; is a nonlinear factor.

4. The method for analyzing the daily living behavior of the elderly for smart elderly care according to claim 3 is characterized in that: The S1 specifically includes: In the process of implementing the hierarchical adaptive nested interactive optimization and weighted feedback feature extraction algorithm, the initial feature data set is subjected to cross-feedback and hierarchical weighted optimization, and a cross-feedback weight matrix is ​​introduced to weight the initial feature data set to generate hierarchically optimized feature data F( l ): Among them, F (l) is the optimized feature data of the lth layer; W (l) is the cross-feedback weight matrix of the lth layer; F init is the initial feature data set; α is the dynamic balance factor; F init，i is the i-th feature in the initial feature dataset; is the cross-feedback weight of the i-th feature in the initial feature dataset of the l-th layer; d is the feature dimension of the initial feature dataset.

5. The method for analyzing the daily living behavior of the elderly for smart elderly care according to claim 4 is characterized in that: The S1 specifically includes: In the process of implementing the hierarchical adaptive nested interactive optimization and weighted feedback feature extraction algorithm, a hierarchical feature cumulative feedback matrix is ​​constructed based on the optimized feature data; based on the cumulative feedback matrix, a dimensionality reduction weight matrix is ​​introduced to calculate the variance contribution of the feature to obtain the final comprehensive feature data. The specific formula is: Among them, F final is the final comprehensive feature data; D (l) is the dimension reduction weight matrix; L (l) is the cumulative feedback matrix; K is the total number of layers.

6. The method for analyzing the daily living behavior of the elderly for smart elderly care according to claim 1 is characterized in that: The S2 specifically includes: The historical comprehensive feature data are divided into a training set and a validation set, and data expansion is performed on the historical comprehensive feature data used as the training set to obtain the expanded historical comprehensive feature data.

7. The method for analyzing the daily living behavior of the elderly for smart elderly care according to claim 1 is characterized in that: The S2 specifically includes: The hierarchical time series dynamic neural network model includes an input layer, a dynamic time series layer, an information fusion layer and a decision output layer; the expanded historical comprehensive feature data is used as the input of the input layer, and the output of the input layer is obtained after weighted processing, and the output of the input layer is used as the input of the dynamic time series layer.

8. The method for analyzing the daily living behavior of the elderly for smart elderly care according to claim 7 is characterized in that: The S2 specifically includes: In the hierarchical time series dynamic neural network model, the dynamic time series layer combines the long short-term memory neural network and the gated recurrent unit neural network to capture the dynamic characteristics of time series data; in the information fusion layer, a dynamic weight update mechanism is introduced. The update formula is as follows: in, is the dynamic weight at time step t; is the dynamic weight at time step t-1; η is the basic learning rate; is the error feedback term; It is in The prediction error at the moment; is the time exponential decay function; is the time decay rate; is the normalization coefficient; is the sum of the absolute values ​​of the prediction errors; It is The prediction error at the moment; is the oscillation adjustment coefficient; sin(θ·t) is the periodic variation function; θ is the frequency factor.

9. The method for analyzing the daily living behavior of the elderly for smart elderly care according to claim 8 is characterized in that: The S2 specifically includes: After the dynamic weights are updated, the hierarchical temporal dynamic neural network model enters the decision output layer, and the output of the dynamic temporal layer is synthesized using a deep fully connected network. The trained hierarchical temporal dynamic neural network model is then verified using the historical comprehensive feature data as a verification set to obtain a trained and verified hierarchical temporal dynamic neural network model.

Images