User physiological parameter prediction method and device based on multivariable time series

Abstract

The invention discloses a user physiological parameter prediction method and device based on a multivariable time sequence. The user physiological parameter prediction method and device are used for solving the problem that an existing time convolution network does not pay attention to interaction and physical relevance between different signals in the same time step. The method comprises the steps of transposing a multivariable time sequence corresponding to user physiological parameters within a set duration from a current moment to a variable dimension, and mapping and outputting sequence feature information of each variable; determining a final fusion result corresponding to a global feature and a local feature between each piece of sequence feature information through an encoder module; and determining a user physiological parameter prediction result according to the final fusion result. According to the method, the time sequence of each variable can be independently processed, the global dependency relationship and the subtle change in a short time are captured, and automatic extraction and fusion of global and local features among the variables are realized.

Description

Technical Field

[0001] The present application relates to the field of deep learning technology, and in particular to a method and device for predicting user physiological parameters based on multivariate time series. Background Art

[0002] With the rapid development of telemedicine technology, it has become possible to monitor the time series waveform data of physiological parameters. In particular, for continuous glucose monitoring, the time series waveform data not only includes changes in glucose levels, but also involves multiple channel information of the sensor, such as blank current, working current, reference current and temperature signal. There are complex physical correlations between these channel information, which work together to measure and interpret glucose levels, and are the key to achieving in-depth analysis of physiological signs, sensor signal pattern recognition and time series prediction.

[0003] Although the temporal convolutional network (TCN) can better capture the temporal dependency in time series than traditional recurrent neural networks (RNN) and convolutional neural networks (CNN), it mainly focuses on the temporal dependency of the time series itself, and does not pay attention to the interaction and physical correlation between different signals in the same time step. Summary of the invention

[0004] The embodiments of the present application provide a method and device for predicting user physiological parameters based on multivariate time series to solve the above technical problems.

[0005] On the one hand, an embodiment of the present application provides a method for predicting a user's physiological parameters based on a multivariate time series, comprising:

[0006] Transpose the multivariate time series corresponding to the user's physiological parameters within a set time from the current moment to the variable dimension, and map and output the sequence feature information of each variable;

[0007] The final fusion result of the global features and the local features between each sequence feature information is determined by an encoder module, where the encoder module includes multiple encoder layers, each of which includes a self-attention mechanism and multiple layers of temporal convolutional layers;

[0008] The user's physiological parameter prediction results are determined based on the final fusion results.

[0009] In one implementation of the present application, the final fusion result corresponding to the global feature and the local feature between each sequence feature information is determined by the encoder module, including:

[0010] For the first encoder layer, the global features between each sequence feature information are extracted through the self-attention mechanism, and the local features between each sequence feature information are extracted through multi-layer temporal convolution layers. The global features and local features are fused to obtain the fusion result;

[0011] Executing a data processing step, the data processing step includes: using the fusion result as an input of a next layer encoder, and the next layer encoder outputs a corresponding fusion result;

[0012] The data processing steps are repeated until the last layer of encoder outputs the final fusion result.

[0013] In one implementation of the present application, the global features between each sequence feature information are extracted through a self-attention mechanism, including:

[0014] Through the multi-head attention mechanism, the feature information of each sequence is projected to generate the QKV matrix;

[0015] The QKV matrix is ​​dimensionally correlated and weighted by scaling the dot product attention to obtain the global features between each sequence feature information.

[0016] In one implementation of the present application, local features between each sequence feature information are extracted through a multi-layer temporal convolution layer, including:

[0017] For the data set composed of each sequence feature information, local features are extracted through each layer of temporal convolution layer in turn, and the last layer of temporal convolution layer outputs the hidden state data set;

[0018] Transpose the hidden state dataset to obtain the local features between each sequence feature information.

[0019] In one implementation of the present application, each encoder layer further includes a feedforward network layer, and correspondingly, the global features and the local features are fused to obtain a fusion result, including:

[0020] Add the global features and local features to obtain the concatenated feature vector;

[0021] Perform nonlinear transformation and feature extraction operations on the concatenated feature vectors to generate a fusion result.

[0022] In one implementation of the present application, the multivariate time series corresponding to the user's physiological parameters within a set time from the current moment is transposed to the variable dimension, and the sequence feature information of each variable is mapped and output, including:

[0023] Obtain the multivariate time series corresponding to the user's physiological parameters within a set time from the current moment;

[0024] Transpose the multivariate time series to obtain the time series of each variable;

[0025] The time series of each variable is mapped to a high-dimensional representation space through a multi-layer perceptron to obtain the sequence feature information of each variable.

[0026] In one implementation of the present application, determining a prediction result of a user's physiological parameters according to the final fusion result includes:

[0027] The final fusion result is mapped to a high-dimensional representation space, and the preliminary prediction result corresponding to the multivariate time series is output through a multi-layer perceptron;

[0028] Transpose the preliminary prediction results to obtain the user parameter prediction results.

[0029] In one implementation of the present application, before mapping the final fusion result to the high-dimensional representation space, the method further includes:

[0030] Perform residual connection operation on the sequence feature information of each variable of the final fusion result and the initial input of the encoder module;

[0031] Perform layer normalization on the residual connection results to calculate the mean and variance of the residual connection results;

[0032] According to the mean and variance, the characteristic dimension of each variable is normalized and the data processing results are output;

[0033] Correspondingly, the final fusion result is mapped to a high-dimensional representation space, including:

[0034] Map the data processing results to a high-dimensional representation space.

[0035] On the other hand, an embodiment of the present application further provides a user physiological parameter prediction device based on a multivariate time series, the device comprising:

[0036] at least one processor;

[0037] and, a memory communicatively coupled to the at least one processor;

[0038] The memory stores instructions that can be executed by at least one processor, and the instructions are executed by at least one processor so that the at least one processor can execute the user physiological parameter prediction method based on multivariate time series as described above.

[0039] On the other hand, an embodiment of the present application further provides a non-volatile computer storage medium storing computer executable instructions. When the computer executable instructions are executed, a method for predicting user physiological parameters based on a multivariate time series as described above is implemented.

[0040] The present application provides a method and device for predicting user physiological parameters based on multivariate time series, which at least has the following beneficial effects:

[0041] By transposing the multivariate time series to the variable dimension and determining the sequence feature information corresponding to each variable, it is easier to more effectively capture the global and local features between the sequence feature information of different variables. The global and local features are concatenated and fused through the self-attention mechanism and the temporal convolution layer of the encoder module, and the overall and local information in each sequence feature information is fully utilized to obtain a comprehensive final fusion result. The user physiological parameter prediction result is then determined based on the final fusion result, thereby improving the accuracy and robustness of the user physiological parameter prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0043] Figure 1 A flowchart of a method for predicting user physiological parameters based on multivariate time series provided in an embodiment of the present application;

[0044] Figure 2 A schematic diagram of the model results provided in the embodiments of the present application;

[0045] Figure 3 A schematic diagram of a model workflow provided for an embodiment of the present application;

[0046] Figure 4 A schematic diagram of the dilated convolution process provided in an embodiment of the present application;

[0047] Figure 5 A schematic diagram of the internal structure of a user physiological parameter prediction device based on a multivariate time series provided in an embodiment of the present application. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0049] The technical solutions provided by various embodiments of the present application are described in detail below in conjunction with the accompanying drawings.

[0050] Figure 1 A flowchart of a method for predicting user physiological parameters based on multivariate time series provided in an embodiment of the present application.

[0051] The analysis method involved in the embodiments of the present application can be implemented by a terminal device or a server, and the present application does not impose any special restrictions on this. For the convenience of understanding and description, the following embodiments are described in detail by taking a server as an example.

[0052] It should be noted that the server may be a single device or a system consisting of multiple devices, that is, a distributed server, and this application does not make any specific limitation on this.

[0053] like Figure 1 As shown, an embodiment of the present application provides a method for predicting user physiological parameters based on a multivariate time series, comprising:

[0054] 101. Transpose the multivariate time series corresponding to the user's physiological parameters within a set time from the current moment to the variable dimension, and map and output the sequence feature information of each variable.

[0055] Figure 2 This is a schematic diagram of the model results provided in the embodiments of this application. Figure 2 As shown, the present application includes an embedding module, an encoder module and an output projection module, wherein the encoder module is the core part, which combines a multi-head attention mechanism with dilated convolution.

[0056] The multivariate time series is input into the embedding module and embedded through a multi-layer perceptron (MLP) to map the time series of each variable into a high-dimensional representation space.

[0057] The encoder module is the core part of the network and consists of multiple encoder layers. Each layer includes a multi-head attention mechanism, a dilated convolution module, a residual connection and layer normalization. The multi-head attention mechanism is used to capture the global correlation between different variables. The dilated convolution module is used to apply the dilated convolution operation on the variable dimension, expand the receptive field between variables, and capture the complex dependencies between variables. The residual connection is used to splice the fusion results of global features and local features with the original input multivariate time series to ensure the effective reflux of gradients and the stability of information transfer. Layer normalization is used to ensure the stability of the model and prevent the gradient from disappearing or exploding.

[0058] The output projection module is used to project the final output of the encoder module to the prediction space to obtain the final user physiological parameter prediction result.

[0059] Specifically, in one embodiment of the present application, the multivariate time series corresponding to the user's physiological parameters within a set time from the current moment is transposed to the variable dimension, and the sequence feature information of each variable is mapped and output, including:

[0060] Obtain the multivariate time series corresponding to the user's physiological parameters within a set time from the current moment;

[0061] Transpose the multivariate time series to obtain the time series of each variable;

[0062] The time series of each variable is mapped to a high-dimensional representation space through a multi-layer perceptron to obtain the sequence feature information of each variable.

[0063] In one embodiment, in order to solve the problem that the time convolution network (TCN) in the prior art does not focus on the interaction and physical correlation between different signals in the same time step, the present application first obtains the multivariate time series data corresponding to the user's physiological parameters within the set time from the continuous glucose monitoring system or other physiological parameter monitoring equipment, thereby providing the original data basis for subsequent feature extraction and prediction. It should be noted that the set time in the embodiment of the present application can be set according to actual conditions, for example: the past 24 hours. These data should include the original signals of all relevant channels, for example: glucose level, blank current, working current, reference current and temperature signal, etc.

[0064] Then, the data X∈R in the obtained multivariate time series T×N Transpose the data from the original T×N ("time step × variable") form to N×T ("variable × time step") form. In this way, each variable corresponds to a time series, which is convenient for subsequent feature extraction by variable and independent processing of its corresponding time series. The purpose of transposition is to reorganize the data so that it can be processed by variables instead of time steps, making it easier to capture the associations and features between different variables.

[0065] The multilayer perceptron is a feedforward neural network composed of multiple fully connected layers. In this application, the multilayer perceptron inputs the time series of each variable, maps it to a D-dimensional token representation space, and outputs the feature vector of the variable in the high-dimensional representation space, that is, the sequence feature information, to form the initial token representation H0∈R N×D This can transform the original time series data into more expressive feature vectors, which can better capture the complex patterns and characteristics in the time series.

[0066] 102. Determine the final fusion result corresponding to the global features and the local features between each sequence feature information through the encoder module.

[0067] It should be noted that the encoder module in the embodiment of the present application includes multiple encoder layers, each encoder layer includes a self-attention mechanism and multiple layers of temporal convolution layers.

[0068] Specifically, in one embodiment of the present application, the final fusion result corresponding to the global feature and the local feature between each sequence feature information is determined by the encoder module, including:

[0069] For the first encoder layer, the global features between each sequence feature information are extracted through the self-attention mechanism, and the local features between each sequence feature information are extracted through multi-layer temporal convolution layers. The global features and local features are fused to obtain the fusion result;

[0070] Executing a data processing step, the data processing step includes: using the fusion result as an input of a next layer encoder, and the next layer encoder outputs a corresponding fusion result;

[0071] The data processing steps are repeated until the last layer of encoder outputs the final fusion result.

[0072] It should be noted that the multi-layer encoder processing module consists of multiple improved encoder layers, each of which contains a multi-head attention mechanism and multi-layer temporal convolution layers.

[0073] In one embodiment, the input of the first encoder layer (i.e., the sequence feature information of each variable) is processed using a self-attention mechanism. The self-attention mechanism can capture the global dependencies, i.e., the global features between different variables, by calculating the correlation between different sequence feature information. At the same time, a multi-layer temporal convolutional layer is used to process the input of the first encoder layer. The temporal convolutional layer can learn local patterns in the sequence, such as short-term fluctuations and trend changes. The multi-layer structure helps to extract deeper local features. Then, the global features and local features are fused to obtain the output of the first encoder layer (i.e., the fusion result).

[0074] After the first encoder layer is processed, the fusion result output by the first encoder layer is used as the input of the next encoder layer. The next encoder layer also uses the self-attention mechanism and multi-layer temporal convolution layer to extract global features and local features and fuse them. This data processing step is repeated until the preset number of encoder layers is reached. After processing by multiple layers of encoders, the fusion result output by the last encoder layer is the final fusion result corresponding to the global features and local features between all sequence feature information.

[0075] It can be understood that this application combines the self-attention mechanism with multi-layer temporal convolutional layers, so that the model can simultaneously capture long-term dependencies and local patterns in time series, providing strong technical support for the fields of telemedicine and physiological parameter monitoring.

[0076] Furthermore, the global features between each sequence feature information are extracted through the self-attention mechanism, including:

[0077] Through the multi-head attention mechanism, the feature information of each sequence is projected to generate the QKV matrix;

[0078] The QKV matrix is ​​dimensionally correlated and weighted by scaling the dot product attention to obtain the global features between each sequence feature information.

[0079] In one embodiment, each sequence feature information is projected through a multi-head attention mechanism to generate a query matrix Q, a key matrix K, and a value matrix V, that is, a QKV matrix. The dimensions of the Q, K, and V matrices are usually set to [sequence length, feature dimension], where the sequence length is the number of time steps and the feature dimension is the dimension of the high-dimensional representation space.

[0080] Use scaled dot product attention to calculate the dimensional correlation of multidimensional sequences. Specifically, the dot product of Q and K is calculated to obtain the attention score matrix, which reflects the correlation between different sequence feature information. At the same time, in order to avoid the gradient disappearance or explosion caused by the dot product result being too large, the attention score matrix is ​​scaled, usually by dividing by the square root of the feature dimension. Then the Softmax function is applied to normalize the scaled attention score matrix to obtain the attention weight matrix, which is used to represent the degree of attention of each sequence feature information to other feature information, that is, to obtain the corresponding correlation coefficient.

[0081] The V matrix is ​​then weighted and summed using the attention weight matrix to obtain a global feature representation of each sequence feature information, which integrates the correlation between all sequence feature information and can capture long-term dependencies in the time series. The multi-head attention mechanism repeats the above steps multiple times, using a different linear projection layer each time to generate the Q, K, and V matrices and perform independent attention calculations. The global feature representations obtained from each attention calculation are then concatenated and fused through another fully connected layer to obtain the final global feature, which is the weighted result representation of the generated attention mechanism.

[0082] It can be understood that the present application effectively extracts the global features between each sequence feature information through the multi-head attention mechanism, which can not only capture the long-term dependencies in the time series, but also handle the interactions between different variables. At the same time, the introduction of the multi-head attention mechanism also enhances the expressiveness and robustness of the model, enabling it to better adapt to complex and changeable physiological parameter time series data.

[0083] Figure 3 This is a schematic diagram of the model workflow provided in the embodiment of this application. Figure 3As shown, first, the present application transposes the multivariate time series and projects it into a QKV matrix, calculates the dot product of Q and K, and scales it to prevent the dot product result from being too large, so that the corresponding correlation coefficient is obtained. Secondly, the V matrix is ​​multiplied by the correlation coefficient for weighted summation to obtain a global feature representation, which is then activated. At the same time, the dilated convolution operation is applied to the variable dimension through the multi-layer time convolution layer to extract local features. Afterwards, the global variables obtained by the multi-head attention mechanism and the local features obtained by the multi-layer time convolution layer are fused through the fully connected layer in the feedforward network layer to obtain a fusion result. Then, the fusion result is connected to the original input multivariate time series to ensure the effective reflux of the gradient and the stability of information transfer, and finally output the prediction results of the user's physiological parameters.

[0084] Furthermore, the local features between each sequence feature information are extracted through multiple layers of temporal convolutional layers, including:

[0085] For the data set composed of each sequence feature information, local features are extracted through each layer of temporal convolution layer in turn, and the last layer of temporal convolution layer outputs the hidden state data set;

[0086] Transpose the hidden state dataset to obtain the local features between each sequence feature information.

[0087] In one embodiment, in each encoder layer, the embedding passes through a multi-layer temporal convolutional layer TCN, and a data set composed of multiple sequence feature information is used as the input of the multi-layer temporal convolutional layer.

[0088] The input data set is first processed by the first temporal convolution layer. The first layer uses the basic convolution kernel to perform a one-dimensional convolution operation on the variable dimension, preliminarily captures the local dependencies between adjacent variables, and outputs the hidden state of the first layer.

[0089] Starting from the second layer, each temporal convolution layer uses a dilated convolution operation. Dilated convolution inserts an interval, i.e., a dilation rate, into the convolution kernel, allowing the convolution kernel to cover more distant variables, thereby expanding the receptive field. Moreover, as the number of layers increases, the dilation rate gradually increases, allowing the high layers to capture a wider range of local features.

[0090] Each layer of temporal convolution continues to perform convolution operations on the variable dimension based on the output of the previous layer, capturing and extracting the local dependencies between sequence feature information layer by layer. This step-by-step process enables each layer to generate a new hidden state, thus reflecting a higher level of feature abstraction.

[0091] After that, after being processed by multiple layers of temporal convolutional layers, the final hidden state dataset is output through the last layer of temporal convolutional layers, which contains relatively rich local features extracted after multiple layers of convolutional operations. In addition, in order to more intuitively analyze the local features between each sequence feature information, the hidden state dataset output by the last layer of temporal convolutional layers needs to be transposed. The transposed data converts the hidden states originally arranged in chronological order into a form arranged by variable dimensions, so that the interaction features and local dependencies between different variables can be directly observed.

[0092] By introducing dilated convolution, the Transformer encoder can operate on the variable dimension, capture the convolution correlation between multiple channels or multiple dimensions, and especially reveal the deep dependency between variables at the same time step.

[0093] Figure 4 Schematic diagram of the dilated convolution process provided in the embodiment of the present application. Figure 4 As shown in the figure, the dilated convolution includes multiple layers of temporal convolution layers. Since the multivariate time series is transposed to the variable dimension, in each encoder layer, the dilated convolution module no longer expands along the time dimension, but performs convolution operations along the variable dimension. Through this operation, the model can expand the receptive field of the dependency relationship between variables without increasing the depth of the convolution layer, thereby improving the efficiency of the model. Dilated convolution can capture deeper interaction features between variables and plays a key role in multivariate time series prediction.

[0094] Furthermore, each layer of the encoder in the embodiment of the present application further includes a feedforward network layer, and correspondingly, the global features and the local features are fused to obtain a fusion result, including:

[0095] Add the global features and local features to obtain the concatenated feature vector;

[0096] Perform nonlinear transformation and feature extraction operations on the concatenated feature vectors to generate a fusion result.

[0097] In one embodiment, the global feature vector extracted by the self-attention mechanism and the local feature vector extracted by the multi-layer temporal convolution layer are added to obtain a concatenated feature vector. Then, the concatenated feature vector is input into the feed-forward network layer. The feed-forward network layer is usually composed of one or more fully connected layers (also called dense layers), each of which may be followed by a nonlinear activation function (such as ReLU, tanh, etc.).

[0098] In the feedforward network layer, the concatenated feature vectors will undergo a series of nonlinear transformations and feature extraction operations, so that more complex nonlinear relationships between features can be learned, thereby generating more expressive fusion features. The feature vector processed by the feedforward network layer is the fusion result, and the fusion result has undergone nonlinear transformation and feature extraction, so it has stronger expressiveness and generalization capabilities. The introduction of the feedforward network layer increases the nonlinear expression ability of the model, allowing the model to learn more complex feature relationships.

[0099] 103. Determine the prediction result of the user's physiological parameters according to the final fusion result.

[0100] Specifically, in one embodiment of the present application, determining the prediction result of the user's physiological parameters according to the final fusion result includes:

[0101] The final fusion result is mapped to a high-dimensional representation space, and the preliminary prediction result corresponding to the multivariate time series is output through a multi-layer perceptron;

[0102] Transpose the preliminary prediction results to obtain the user parameter prediction results.

[0103] In one embodiment, first, the final fusion result is projected or mapped to a high-dimensional representation space, which can more comprehensively capture and represent the complex features and patterns of time series data. Then, the mapped high-dimensional representation is further processed and predicted using a multi-layer perceptron MLP, and the complex relationship between the data is learned and fitted through multi-layer nonlinear transformations. In this embodiment, the input of the multi-layer perceptron MLP is the mapped high-dimensional representation, and the output is the preliminary prediction result corresponding to the multivariate time series, which contains the prediction information of the user's physiological parameters based on the fusion features.

[0104] At this time, since the preliminary prediction results output by the multi-layer perceptron MLP are presented in the variable dimension, in order to obtain a form that actually corresponds to the user's physiological parameters, it is necessary to transpose the preliminary prediction results, that is, project them back to the prediction space of the original time series. After the transposition operation, the prediction results that actually correspond to the user's physiological parameters are obtained, and the final prediction results are returned as the output of the entire prediction process, which can be directly used for subsequent analysis, decision-making or display to the user, realizing accurate prediction of the user's physiological parameters.

[0105] In one embodiment of the present application, before mapping the final fusion result to the high-dimensional representation space, the method further includes:

[0106] Perform residual connection operation on the sequence feature information of each variable of the final fusion result and the initial input of the encoder module;

[0107] Perform layer normalization on the residual connection results to calculate the mean and variance of the residual connection results;

[0108] According to the mean and variance, the characteristic dimension of each variable is normalized and the data processing results are output;

[0109] Correspondingly, the final fusion result is mapped to a high-dimensional representation space, including:

[0110] Map the data processing results to a high-dimensional representation space.

[0111] In one embodiment, before mapping the final fusion result to a high-dimensional representation space, a key step is performed, that is, a residual connection operation is performed on the final fusion result and the sequence feature information of each variable initially input by the encoder. This is to combine the original input information with the fusion information after complex transformation, thereby retaining the original features of the input data and introducing deep features after model processing.

[0112] Next, in order to maintain the stability of model training and prevent the gradient from disappearing or exploding, the residual connection results are subjected to layer normalization (LayerNorm). LayerNorm is a normalization technique applied between neural network layers. It calculates the mean and variance of the activation values ​​of all neurons in the current layer, and then normalizes the activation value of each neuron to make its distribution more stable.

[0113] After the layer normalization operation, a more stable and reasonably distributed feature representation is obtained, that is, the data processing result, which not only retains the original input information and the fused high-level features, but also ensures the inter-layer balance and consistency of information through normalization. Finally, the data processing results after layer normalization are mapped to a high-dimensional representation space, so that low-dimensional features can be mapped to a higher-dimensional space to better capture the complex relationship between data.

[0114] It can be understood that the present application not only enhances the fusion capability of the model through residual connection, layer normalization and final mapping operations, but also ensures the consistency of information and balance between layers, thereby effectively mapping the final fusion result to a high-dimensional representation space, providing a more stable and rich feature representation for subsequent model processing.

[0115] In one embodiment, the overall forward propagation of the encoder is:

[0116] Given historical observations X = {x1, x2, …, x T}, where T represents the number of time steps, i.e., the embedding dimension, and N is the number of variables. Based on these historical data, the observation values ​​of the next S time steps are predicted, i.e., {x T+1 ,…,x T+S}. X[t,:] is used to represent the observed values ​​of all variables recorded at time step t. These observations not only reflect the data of multiple channels at each time point, but also contain rich cross-channel dependencies and time series change information.

[0117] In multivariate time series prediction, the receptive field can be expanded by performing convolution operations and feature extraction in parallel on the time step, especially in the time series convolutional network layer TCN, the adjustment of the receptive field is particularly important. The spatial receptive field is used to represent the different variable information that the model can perceive at time step t, which directly determines the feature dimension that the model can capture.

[0118] The input multivariate time series is X∈R T×N , where T is the number of time steps and N is the number of variables. Let the length of the prediction be S, which is the length of the future time steps. First, transpose the input multivariate time series so that its dimension changes from X∈R T×N becomes X∈R N×T , that is, the sequence characteristic information of each variable.

[0119] Through a multi-layer perceptron MLP, the time series of each variable is mapped to a D-dimensional token representation space, and X is mapped to form the initial embedding representation H0, where H0∈R N×D The purpose of this operation is to convert the time series into tokens with feature representation for subsequent processing.

[0120] These embedded tokens are then processed through multiple encoder layers, each of which contains an attention mechanism and multiple layers of temporal convolutional layers.

[0121] Attention mechanism: l-1 (i.e. the first - 1 layer) to generate the query (Q) matrix, key (K) matrix and value (V) matrix. Then a multi-head attention mechanism is used to capture the correlation between variables, and the output is H att ∈R N×D , which is the new sequence representation after being processed by the attention mechanism.

[0122] In each encoder layer, the embedding passes through multiple layers of temporal convolutional layers. Each temporal convolutional layer captures the relationship between the variable dimensions layer by layer. The process of TCN is to gradually apply convolution operations in each layer to generate the hidden state output of the mth TCN layer After completing all the time convolution layers, transpose the results back to get H attnS ∈R N×D , H attnS It reflects the interaction characteristics between different variables.

[0123] Fusion of H through activation and feed-forward networksatt With H attnS , which can effectively extract global and local features among multiple variables, and finally maintain information consistency and inter-layer balance through residual connection and layer normalization.

[0124] The output of the last encoder layer is H L , project it back to the prediction space of the original time series through the multi-layer perceptron MLP, and output the prediction result Finally, the prediction results are transposed and become And return the final prediction result.

[0125] The above is an embodiment of the method proposed in this application. Based on the same inventive concept, the embodiment of this application also provides a user physiological parameter prediction device based on multivariate time series, whose structure is as follows: Figure 5 shown.

[0126] Figure 5 The internal structure diagram of a user physiological parameter prediction device based on multivariate time series provided in an embodiment of the present application. Figure 5 As shown, the device includes:

[0127] at least one processor;

[0128] and, a memory communicatively coupled to the at least one processor;

[0129] The memory stores instructions that can be executed by at least one processor, and the instructions are executed by at least one processor to enable the at least one processor to:

[0130] Transpose the multivariate time series corresponding to the user's physiological parameters within a set time from the current moment to the variable dimension, and map and output the sequence feature information of each variable;

[0131] The final fusion result of the global features and the local features between each sequence feature information is determined by an encoder module, where the encoder module includes multiple encoder layers, each of which includes a self-attention mechanism and multiple layers of temporal convolutional layers;

[0132] The user's physiological parameter prediction results are determined based on the final fusion results.

[0133] The present application also provides a non-volatile computer storage medium storing computer executable instructions. When the computer executable instructions are executed, they can:

[0134] Transpose the multivariate time series corresponding to the user's physiological parameters within a set time from the current moment to the variable dimension, and map and output the sequence feature information of each variable;

[0135] The final fusion result of the global features and the local features between each sequence feature information is determined by an encoder module, where the encoder module includes multiple encoder layers, each of which includes a self-attention mechanism and multiple layers of temporal convolutional layers;

[0136] The user's physiological parameter prediction results are determined based on the final fusion results.

[0137] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0138] The above describes specific embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying 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.

[0139] The devices and media provided in the embodiments of the present application correspond one-to-one to the methods. Therefore, the devices and media also have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0140] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0141] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0142] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0143] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0144] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0145] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0146] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0147] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0148] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for predicting user physiological parameters based on multivariate time series, characterized in that: The method comprises: Transpose the multivariate time series corresponding to the user's physiological parameters within a set time from the current moment to the variable dimension, and map and output the sequence feature information of each variable; Determine the final fusion result corresponding to the global features and the local features between each sequence feature information through an encoder module, wherein the encoder module includes a plurality of encoder layers, each encoder layer includes a self-attention mechanism and a multi-layer temporal convolution layer; The user physiological parameter prediction result is determined according to the final fusion result.

2. The method according to claim 1, characterized in that The final fusion result of the global features and local features between each sequence feature information is determined by the encoder module, including: For the first encoder layer, the global features between each sequence feature information are extracted through the self-attention mechanism, the local features between each sequence feature information are extracted through the multi-layer temporal convolution layer, and the global features and the local features are fused to obtain the fusion result; Executing a data processing step, the data processing step comprising: using the fusion result as an input of a next layer encoder, and the next layer encoder outputting a corresponding fusion result; The data processing steps are repeated until the last layer of encoder outputs the final fusion result.

3. The method according to claim 2, characterized in that The global features between each sequence feature information are extracted through the self-attention mechanism, including: Through the multi-head attention mechanism, each sequence feature information is projected to generate a QKV matrix; The QKV matrix is ​​subjected to dimensional correlation calculation and weighted calculation by scaling dot product attention to obtain the global features between each sequence feature information.

4. The method according to claim 2, characterized in that: The local features between each sequence feature information are extracted through multiple layers of temporal convolutional layers, including: For the data set composed of the feature information of each sequence, local feature extraction is performed in each time convolution layer in turn, and the last time convolution layer outputs the hidden state data set; The hidden state data set is transposed to obtain local features between each sequence feature information.

5. The method according to claim 2, characterized in that: Each layer of the encoder further includes a feedforward network layer, and correspondingly, the global features and the local features are fused to obtain a fusion result, including: Adding the global feature and the local feature to obtain a concatenated feature vector; Nonlinear transformation and feature extraction operations are performed on the concatenated feature vectors to generate a fusion result.

6. The method according to claim 1, characterized in that Transpose the multivariate time series corresponding to the user's physiological parameters within the set time from the current moment to the variable dimension, and map and output the sequence feature information of each variable, including: Obtain the multivariate time series corresponding to the user's physiological parameters within a set time from the current moment; Transposing the multivariate time series to obtain a time series of each variable; The time series of each variable is mapped to a high-dimensional representation space through a multi-layer perceptron to obtain the sequence feature information of each variable.

7. The method according to claim 6, characterized in that Determining the prediction result of the user's physiological parameter according to the final fusion result includes: Mapping the final fusion result to the high-dimensional representation space, and outputting the preliminary prediction result corresponding to the multivariate time series through the multi-layer perceptron; The preliminary prediction result is transposed to obtain the user parameter prediction result.

8. The method according to claim 7, characterized in that Before mapping the final fusion result to the high-dimensional representation space, the method further includes: Performing a residual connection operation on the final fusion result and the sequence feature information of each variable initially input to the encoder module; Perform layer normalization on the residual connection results to calculate the mean and variance of the residual connection results; Normalizing the feature dimension of each variable according to the mean and the variance, and outputting a data processing result; Correspondingly, mapping the final fusion result to the high-dimensional representation space includes: The data processing result is mapped to the high-dimensional representation space.

9. A user physiological parameter prediction device based on multivariate time series, characterized in that: The device comprises: at least one processor; and, a memory communicatively coupled to the at least one processor; Wherein, the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the user physiological parameter prediction method based on multivariate time series as described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer executable instructions, characterized in that: When the computer executable instructions are executed, the user physiological parameter prediction method based on multivariate time series as described in any one of claims 1 to 8 is implemented.

Images