Cognitive disorder detection method and device based on energy capture internet-of-things system

Abstract

The invention discloses a cognitive impairment detection method and device based on an energy capture internet of things system, and the method comprises the steps: deploying an RFID and a solar panel, collecting the signal intensity and energy collection intensity in real time, generating the moving original data of a to-be-detected target in a room, carrying out the track segmentation of the original data, and carrying out the single-mode feature extraction of a track segment. A speed feature image is generated, a conditional variation auto-encoder based on a Transform architecture is constructed, a decoder of the conditional variation auto-encoder is utilized to generate a forged image to realize data expansion, and the encoder of the conditional variation auto-encoder is utilized to extract deep submerged space features of the image. And finally, inputting the extracted high-dimensional feature vector into a classifier, and outputting a classification result, thereby realizing judgment of the cognitive state. According to the method, the problems of data scarcity and insufficient long-time-sequence behavior capture are solved through generative data enhancement, and high-precision non-sensitive monitoring is realized.

Description

Technical Field

[0001] This application belongs to the field of behavior recognition technology, and in particular relates to a cognitive impairment detection method and device based on an energy-capturing IoT system. Background Technology

[0002] Home safety and health monitoring for the elderly is a crucial issue in public health. In this context, abnormal behaviors (such as wandering, repetitive pacing, and signs of impending falls) are often early warning signs of deteriorating health. Traditional monitoring methods primarily rely on wearable devices (such as wristbands) or visual monitoring (such as cameras). However, wearable devices require frequent charging and wearing, making intervention too costly for elderly individuals with memory loss or resistance to wearing them, and device maintenance (such as battery replacement) increases the burden of care. While visual monitoring can capture behavior, it presents serious privacy issues and is significantly affected by lighting and obstructions.

[0003] Furthermore, despite the progress made in deep learning-based behavior recognition methods, significant challenges remain in practical applications, the most critical being the scarcity of high-quality labeled data (Small Sample Problem). Collecting behavioral data from patients with cognitive impairment involves privacy protection, ethical approvals, and long-term monitoring, making it extremely difficult to obtain large-scale, diverse real-world training samples. Traditional discriminative models (such as CNNs or standard RNNs) often struggle to learn robust feature representations when faced with small sample sizes, easily exhibiting overfitting and resulting in severely insufficient generalization ability in real-world scenarios. Moreover, existing data augmentation methods (such as simple geometric transformations) are inadequate to simulate the complex spatiotemporal dynamics of human behavior, failing to fundamentally address the data scarcity problem. Summary of the Invention

[0004] The purpose of this application is to provide a cognitive impairment detection method and device based on an energy-capture IoT system, in order to solve the problems of high maintenance costs, serious privacy violations, and scarce training samples in the prior art for dedicated monitoring equipment.

[0005] To achieve the above objectives, the technical solution of this application is as follows:

[0006] A cognitive impairment detection method based on an energy-capturing IoT system includes:

[0007] In a cognitive impairment detection environment, RFID tags and solar panels are deployed, and several collection points are set up. The received signal strength and photocurrent intensity when a person stands at each collection point are collected, generating a first database that includes the location of the collection point and the collected received signal strength and photocurrent intensity.

[0008] The received signal intensity and photocurrent intensity of the sampled object during daily activities in a cognitive impairment detection scenario are collected and compared with the first database to obtain the original trajectory data of the sampled object's movement.

[0009] The original trajectory data of the movement is segmented, and then the velocity changes of the segmented trajectory segments are converted into color gradient features to obtain velocity feature images. Corresponding labels are set to obtain the original training samples.

[0010] The conditional variational autoencoder was trained using the original training samples.

[0011] Synthetic data is generated by using the decoder in the trained conditional variational autoencoder and then merged with the original training samples to obtain the expanded training dataset.

[0012] A cognitive impairment detection model is constructed using the encoder and classifier in a conditional variational autoencoder. The cognitive impairment detection model is then trained using an expanded training dataset to obtain a well-trained cognitive impairment detection model.

[0013] The velocity feature image of the target to be detected is obtained and input into the trained cognitive impairment detection model to obtain the recognition result.

[0014] Preferably, the received signal intensity and photocurrent intensity of the sampled object during daily activities in the cognitive impairment detection scenario are compared with a first database to obtain the original trajectory data of the sampled object's movement, including:

[0015] Calculate the Euclidean distance between the sampling point and the acquisition point, considering the received signal strength and photocurrent intensity.

[0016] The location of the sampling point with the smallest Euclidean distance is used as the estimated location of the sampling point;

[0017] Connect the estimated locations of all sampling points sequentially in chronological order to generate the original trajectory data of the sampled object's movement.

[0018] Preferably, the trajectory segmentation of the original mobile trajectory data includes:

[0019] When the time interval between two consecutive locations in the trajectory exceeds the threshold, it is determined that the sampled object has been stationary or has not moved continuously during this period. The trajectory is then segmented between these two locations, with the earlier location defined as the end point of the previous trajectory segment and the later location defined as the start point of the next trajectory segment.

[0020] Preferably, the step of training the conditional variational autoencoder using the original training samples includes:

[0021] The velocity feature image is segmented and positionally encoded to generate an image word sequence; the labels of the velocity feature image are embedded to generate a conditional vector.

[0022] The latent vector posterior probability distribution is obtained by inputting the image word sequence and conditional vector into the encoder.

[0023] The latent vector is obtained by sampling from the posterior probability distribution of the latent vector;

[0024] The latent vector and conditional vector are input into the decoder to generate a reconstructed velocity feature map;

[0025] Calculate the loss and update the parameters of the conditional variational autoencoder through backpropagation.

[0026] Preferably, the step of using the decoder in a trained conditional variational autoencoder to generate synthetic data includes:

[0027] A target cognitive state label is defined, and a latent vector is randomly sampled from a standard normal distribution. The target cognitive state label and the randomly sampled latent vector are input into the decoder to reconstruct the synthetic velocity feature map.

[0028] Preferably, the loss includes reconstruction loss and KL divergence loss. The reconstruction loss measures the difference between the reconstructed velocity feature map output by the decoder and the original input velocity feature map; the KL divergence loss measures the difference between the approximate posterior distribution inferred by the encoder and the preset prior distribution.

[0029] This application also proposes a cognitive impairment detection device based on an energy-capturing IoT system, including a processor and a memory storing a number of computer instructions, which, when executed by the processor, implement the steps of the above-described cognitive impairment detection method based on an energy-capturing IoT system.

[0030] This application proposes a cognitive impairment detection method and device based on an energy-capturing IoT system. It generates a velocity feature map using RFID spatial positioning signals and solar energy fluctuation characteristics, employs a conditional variational autoencoder (CVAE) for data augmentation and deep feature extraction, and finally combines it with a multilayer perceptron (MLP) to achieve high-precision behavior analysis. By constructing a Transformer-based CVAE, the powerful self-attention mechanism of the Transformer is utilized to capture long-range global dependencies in the velocity feature image, while the generative capabilities of the CVAE are leveraged to learn the latent probability distribution of the data. This not only enables the generation of large amounts of high-fidelity synthetic data "on demand" to expand the training set and solve the small sample size problem, but also allows for the extraction of more discriminative deep semantic features through the encoder. Combined with the MLP, this significantly improves the accuracy and robustness of cognitive impairment detection. Attached Figure Description

[0031] Figure 1 This is a flowchart of the cognitive impairment detection method based on an energy-capturing IoT system proposed in this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0033] This application expands the functionality of a radio frequency and solar energy system originally used for indoor power supply based on the principle of "signal multiplexing." Specifically, the system is built on an energy-capturing IoT system composed of radio frequency tags (RFID) and solar panels. Under normal conditions, these devices serve as energy receiving and conversion units to maintain the operation of IoT nodes; however, in this application, they also serve as touchpoints for environmental sensing. By monitoring in real time the multipath attenuation of radio frequency signals during spatial transmission and the changes in photocurrent generated by indoor lights on solar panels, the disturbance of the environmental energy field by human movement is converted into behavioral characteristic data, thereby reconstructing the activity trajectory of the target object without the need for wearing any devices.

[0034] One embodiment of this application, such as Figure 1 As shown, a cognitive impairment detection method based on an energy-capturing IoT system is provided, including:

[0035] Step S1: Deploy RFID tags and solar panels in the cognitive impairment detection environment, and set up several collection points. Collect the received signal strength and photocurrent intensity when a person stands at each collection point, and generate a first database including the location of the collection point and the collected received signal strength and photocurrent intensity.

[0036] The environment used in this embodiment for cognitive impairment detection is typically a prepared room where multiple radio frequency tags (RFID) and solar panels are deployed. These can be mounted on the walls of the room, and the specific deployment location is not limited. The RFID tags and solar panels are connected to a data acquisition device to facilitate the acquisition of the received signal strength from the RFID tags and the photocurrent intensity from the solar panels.

[0037] Then, N collection points are determined in the room. When someone stands at a collection point, the received signal strength and photocurrent intensity of each RFID tag and solar panel are collected to generate the first database.

[0038] The first database includes the location of each acquisition point and the received signal strength and photocurrent intensity acquired:

[0039]

[0040]

[0041] in, This is the first database, containing data from all collection points. For the first Data vectors from each collection point The number of data collection points; It means that the person is standing at the first The sum of the received signal strength of the m-th RFID tag at each collection point. It means that the person is standing at the first The photocurrent intensity collected by the q-th solar panel at each collection point is M, and the number of RFID tags is Q. and They represent the first The x and y coordinates of each data collection point.

[0042] Step S2: Collect the received signal intensity and photocurrent intensity of the sampled object when it performs daily activities in the cognitive impairment detection scenario, and compare them with the first database to obtain the original trajectory data of the sampled object's movement.

[0043] In this embodiment, the sampling subjects are arranged to perform their daily activities in a room according to their own habits. Then, the received signal strength and photocurrent intensity are simultaneously collected via RFID and a solar panel. The collected real-activity data first undergoes outlier removal, that is, values ​​outside the normal range and zero values ​​are removed. Then, a dynamic sampling strategy is adopted, triggering sampling when data changes, resulting in sampled data:

[0044]

[0045]

[0046] In the formula, For the collection of results, For the first A vector of sampling points The total number of sampling points. The sequence number of the sampled object. This is the cognitive state label of the object. The timestamp of the sampling point This represents the received signal strength of the m-th RFID tag corresponding to sampling point n. This represents the photocurrent intensity collected by the q-th solar panel corresponding to sampling point n.

[0047] In this embodiment, multiple sampling subjects are arranged to perform their daily activities in a room according to their own habits, thereby collecting data and obtaining multiple sets of sampling data. By comparing the sampling data with the first database, the location of the sampling subject can be obtained, thus obtaining the original trajectory data of the sampling subject's movement.

[0048] Specifically, for any given sampling object, the received signal intensity and photocurrent intensity of the sampling object during daily activities in a cognitive impairment detection scenario are collected and compared with the first database to obtain the original trajectory data of the sampling object's movement. The process is as follows:

[0049] Step 2.1: Calculate the Euclidean distance between the sampling point and the acquisition point for the received signal strength and photocurrent intensity.

[0050] In this embodiment, the Euclidean distance formula is used to calculate the Euclidean distance between the sampling point and the acquisition point, which represents the received signal intensity and photocurrent intensity. The calculation formula is as follows:

[0051]

[0052] The first term is the sum of squares of the differences in RFID signal strength, and the second term is the sum of squares of the differences in the photocurrent intensity of the solar panel. This represents the received signal strength of the m-th RFID tag corresponding to sampling point n. This represents the photocurrent intensity collected by the q-th solar panel corresponding to sampling point n. This represents the received signal strength of the m-th RFID tag corresponding to collection point i. This represents the photocurrent intensity collected by the q-th solar panel corresponding to collection point i.

[0053] Step 2.2: Use the location of the sampling point with the smallest Euclidean distance as the estimated location of the sampling point.

[0054] Compare all the calculated distance values Find the minimum value among them. The minimum value corresponds to the first The physical coordinates of each collection point are recorded in the first database. This is determined as the current sampling point. exist Estimated location at a given time.

[0055] Step 2.3: Connect the estimated positions of all sampling points in chronological order to generate the original trajectory data of the sampled object's movement.

[0056] According to timestamp By sequentially connecting the coordinates of all the calculated sampling points in the order of their order, the original trajectory data of the continuous movement of the sampling object within the room is generated.

[0057] Step S3: Perform trajectory segmentation on the original trajectory data. Then, for the segmented trajectory segments, convert the velocity changes of the trajectory segments into color gradient features to obtain velocity feature images. Set the corresponding labels to obtain the original training samples.

[0058] In this embodiment, the original trajectory data is first processed to eliminate outliers. This involves detecting the movement speed between consecutive points and identifying outliers that exceed a preset threshold. Data points that are considered abnormal are removed; data points exceeding a preset threshold are removed by detecting the spatial distance between consecutive locations. Data points that are not in the data set are also considered abnormal and removed.

[0059] Then, an adjustable still time threshold is set using a smart segmentation method based on still time. Traverse the trajectory data, and when the time interval between two consecutive locations in the trajectory exceeds a certain threshold,... When the sampled object is determined to have either been stationary or engaged in discontinuous motion during this period, trajectory segmentation is performed between these two locations: the earlier location is defined as the end point of the previous trajectory segment, and the later location is defined as the starting point of the next trajectory segment. In this way, a long-time continuous trajectory is divided into multiple independent active trajectory segments.

[0060] Next, velocity features are extracted from the segmented trajectory segments. Based on a pre-defined velocity-color mapping relationship, the velocity changes of the trajectory segments are converted into color gradient features, generating velocity feature images. The velocity feature image corresponding to each trajectory segment is then processed. Associate with corresponding cognitive state tags {healthy, MCI, dementia}, to obtain training samples Among them, "healthy" indicates a healthy state, meaning that the subject's cognitive function is normal and there are no obvious signs of cognitive decline; "MCI" indicates mild cognitive impairment, meaning that the subject has a slight decline in cognitive ability, but has not yet reached the level of dementia and is in a transitional stage between health and dementia; "dementia" indicates a dementia state (such as Alzheimer's disease), meaning that the subject has severe cognitive impairment that affects daily living abilities.

[0061] It should be noted that each original moving trajectory is divided into multiple trajectory segments, and then a label is assigned to each trajectory segment to obtain a training sample. Multiple trajectory segments result in multiple training samples. The training samples generated from all the original moving trajectories constitute the final training sample set.

[0062] Step S4: Train the conditional variational autoencoder using the original training samples.

[0063] This embodiment uses a Conditional Variational Autoencoder (CVAE), consisting of an encoder and a decoder, both employing a multi-layer Transformer structure for feature extraction and data augmentation. Each layer of the encoder contains two core sub-layers: a Multi-Head Self-Attention (MHSA) layer to capture global dependencies in the velocity feature map, and a Forward Propagation Network (FFN) layer to perform non-linear feature transformation through fully connected layers and ReLU activation. Each sub-layer is followed by Layer Normalization and Dropout layers, and feature information is preserved through residual connections, ultimately mapping the input image and conditional labels into one-dimensional latent vectors. The decoder uses a similar structure to reconstruct the image based on the latent vector and conditional label.

[0064] Specifically, the training process of the conditional variational autoencoder is as follows:

[0065] Step 4.1: Segment and position encoding the velocity feature image to generate an image word sequence; embed the labels of the velocity feature image to generate a conditional vector.

[0066] Because the Transformer model cannot directly obtain the sequence order information, the input data first needs to be preprocessed. This includes the input velocity feature map. (dimension is) ) divided into A number of fixed-size patches, each with dimensions of [missing information]. Each tile is flattened and mapped using a trainable linear projection layer. Dimensional tokens form a token sequence. ,in This is the total dimension of the model. Adding sine position coding to the sequence By injecting spatial location information of the image tiles, a position-encoded image word sequence is obtained. .

[0067] velocity feature map Corresponding cognitive state labels (For example, "MCI") is converted through an embedding layer. dimensional condition vector .

[0068] Step 4.2: Input the image word sequence and conditional vector into the encoder to obtain the latent vector posterior probability distribution.

[0069] The input consists of an image word sequence and a conditional vector, and the input is based on a Transformer. In the encoder (composed of trainable parameters) (Definition), inferring the latent vector The approximate posterior normal distribution is given by the mean vector output by the encoder. Sum of variance vectors Common parameterization.

[0070] Specifically, the condition vector With image word sequence The sequences are concatenated and a [CLS] aggregation term is added to form the unified input sequence for the encoder. The input sequence will be unified. A multi-head self-attention module is incorporated to dynamically map the query (Q) and a set of key-value pairs (KV) to a weighted, aggregated output. This mechanism utilizes the dimension of a single attention head. To scale and The dot product, after being normalized by Softmax, is... Weighting is applied to prevent gradient vanishing:

[0071]

[0072] To effectively capture information from different feature subspaces, multi-head self-attention pairs are used. , , conduct Different linear projections (i.e.) Each head (or "head") performs scaled dot product attention in parallel, allowing the model to simultaneously focus on information from different feature subspaces. Finally, the outputs of each head are concatenated and fused through a linear transformation.

[0073]

[0074]

[0075] in All of these are learnable projection parameter matrices. It's about the number of heads to focus on. Representing the The output of each attention head.

[0076] The output processed by the attention module is transformed step-by-step using a fully connected feedforward neural network:

[0077]

[0078] in It is the input to the feedforward neural network module; These are the weights of the feedforward network; This is the bias vector.

[0079] Each coding layer contains the two sub-layers mentioned above (multi-head self-attention and feedforward neural network), and each sub-layer is followed by residual connections and layer normalization.

[0080] go through After stacking and computing the layer encoders, the final output vector corresponding to the [CLS] lexical is extracted. It is then fed into two independent fully connected "head" networks to predict the mean of the posterior distribution, respectively. and logarithmic variance :

[0081]

[0082] .

[0083] Step 4.3: Sample the latent vector from the posterior probability distribution of the latent vector.

[0084] To enable backpropagation of the gradient, a "reparameterization technique" is used to sample latent vectors from the distribution. :

[0085]

[0086] in From the standard normal distribution Random noise in the sampled data.

[0087] Step 4.4: Input the latent vector and conditional vector into the decoder to generate the reconstructed velocity feature map.

[0088] latent vector and condition vector As input, the input is based on Transformer. Decoder (consisting of trainable parameters) (Definition), this generates a reconstructed velocity feature map. .

[0089] Specifically, cognitive state labels Transformed into a condition vector through the embedding layer At the same time, the latent vectors obtained from sampling The dimensions are adjusted using a linear mapping. The two are concatenated to form the source domain context of the decoder. Using a set of learnable query vectors. Its length is , dimension This serves as the initial input, representing the output unit to be generated. In each layer of the decoder, First, a multi-head self-attention module is used to capture the dependencies within the generated sequence. The computation process is similar to that of the encoder, but here Q, K, and V all come from... The output of itself or the previous layer. The query (Q) comes from the generated sequence after self-attention. The key (K) and value (V) come from the source domain context. This allows for the generation of sequences actively from latent variables. and conditions Extract the desired features.

[0090] The output of cross-attention is passed through a fully connected feedforward neural network for nonlinear transformation:

[0091]

[0092] Each decoding layer contains the three sub-layers mentioned above (self-attention, cross-attention, and feedforward neural network), and each sub-layer is followed by a residual connection and layer normalization. The sequence passes through... After the layer decoder, the final output feature sequence is obtained. It is then fed into a linear projection layer. Map back to the feature map space and reshape it to the dimensionality of the target image. :

[0093]

[0094] This represents the reconstruction speed characteristic map.

[0095] Step 4.5: Calculate the loss and update the parameters of the conditional variational autoencoder through backpropagation.

[0096] The training objective of the CVAE model is to maximize the lower bound of evidence; therefore, a combined loss function is defined. Losses from reconstruction and KL divergence loss It consists of two parts, and simultaneously optimizes encoder parameters through the backpropagation algorithm. and decoder parameters .

[0097] Reconstruction loss measures the "reconstruction velocity feature map" output by the decoder. Compared with the original input "velocity feature map" The difference between them. Assuming the data follows a Gaussian distribution, this term corresponds to the negative log-likelihood, and is calculated as the mean squared error between the two:

[0098]

[0099] in, Represents the square of the L2 norm; It is the total number of elements in the image; and The original image and the reconstructed image are respectively in the 1st... The pixel value at each position.

[0100] The KL divergence term serves as a regularization constraint, used to measure the approximate posterior distribution inferred by the encoder. Compared with the pre-set prior distribution The differences between them. This invention assumes that the prior distribution is a standard normal distribution. The posterior distribution is a diagonal Gaussian distribution. In this case, the KL divergence has a closed-form analytical solution:

[0101]

[0102] in, It is a latent variable The dimension; and These are the first and second mean vectors and variance vectors output by the encoder, respectively. Each component.

[0103] During training, the Adam optimizer is used based on the total loss function. Calculate the gradient and update all trainable weights of the encoder and decoder until the model converges.

[0104] Step S5: Use the decoder in the trained conditional variational autoencoder to generate synthetic data and merge it with the original training samples to obtain the expanded training dataset.

[0105] To address the issue of insufficient model generalization ability due to the scarcity of samples in the original training set, the trained decoder is utilized. Generate synthetic data.

[0106] Specifically, a target cognitive state label is set. And randomly sample a latent vector from the standard normal distribution. .Will and The input decoder combines the randomness of latent variables with the constraints of labels through a cross-attention mechanism to reconstruct a novel synthetic velocity feature map. Repeat the above process. This generates a diverse batch of synthetic data.

[0107] These synthetic data are combined with the original training samples to construct an expanded training dataset. .

[0108] Step S6: Construct a cognitive impairment detection model using the encoder and classifier in the conditional variational autoencoder, and train the cognitive impairment detection model using the expanded training dataset to obtain a trained cognitive impairment detection model.

[0109] In this embodiment, the augmented dataset (containing both original and fake velocity feature maps) is input into the trained CVAE encoder to extract its output one-dimensional latent vector. As high-level features. Subsequently, the latent vectors... The input is fed into an MLP classifier, which performs dimensionality reduction and nonlinear transformation on the feature through several fully connected layers, and finally outputs the probability distribution through a Softmax function. During training, the Adam optimizer is used to update parameters, and the model performance is optimized by combining the cross-entropy loss function.

[0110] Specifically, using the trained encoder As a feature extractor, it maps the high-dimensional image space to the low-dimensional semantic latent space. This will augment the dataset. Each velocity feature map in Input encoder. Instead of random sampling, directly extract the posterior mean vector predicted by the encoder. As a stable feature representation of the image This results in a feature dataset consisting of feature vectors and their corresponding labels. .

[0111] This embodiment employs a multilayer perceptron (MLP) classifier to process the extracted feature vectors. It consists of an input layer, several hidden layers, and an output layer. Feature vectors are processed by linear weighting and activation functions in the hidden layers to extract discriminative features. The output of each hidden layer is The output of the last layer After processing with the Softmax function, predicted probability distributions belonging to each cognitive state category are generated. .

[0112] Training uses the cross-entropy loss function to measure the predicted probability distribution. With real labels The differences between (One-hot encoding) are addressed during training by employing the Adam optimizer and dynamically adjusting the learning rate using momentum estimation to iteratively update the weight parameters of the MLP, thereby minimizing the classification loss. .

[0113] Step S7: Obtain the velocity feature image of the target to be detected, input it into the trained cognitive impairment detection model, and obtain the recognition result.

[0114] After training, the cognitive impairment detection model can output the detection result of the target object's cognitive impairment based on the input RFID and solar panel velocity feature maps. The process of acquiring the velocity feature image of the target object is described in steps S1-S3 and will not be repeated here.

[0115] It's worth noting that to address the pain points of battery maintenance and endurance in traditional IoT devices, green IoT systems based on energy harvesting are gradually becoming more common in smart homes. In existing indoor infrastructure, radio frequency identification (RFID) and indoor photovoltaic (solar panels) are typically deployed as power sources: RFID readers emit radio frequency energy to activate passive tags, while indoor lights illuminate solar panels to generate photocurrent that powers low-power nodes. The core purpose of this architecture is to achieve "battery-free" and "self-sustaining" operation of devices.

[0116] This application's technical solution explores the "sensing potential" of the energy harvesting system, transforming it from a single "power supply end" into a "detection end." Its physical basis lies in the fact that the human body is an excellent reflector of radio frequency signals and a light blocker. When an elderly person moves indoors, they dynamically block or disturb the propagation path of RFID radio frequency signals, while simultaneously casting shadows on the sunlight projected onto the solar panel, causing minute but regular fluctuations in the received signal strength (RSSI) and photocurrent. This application utilizes this "integration of energy harvesting and sensing" mechanism to directly reuse existing indoor green energy supply facilities to infer human behavior trajectories without adding additional dedicated sensors (such as cameras or infrared probes). This design not only inherits the maintenance-free and low-carbon environmental advantages of energy harvesting systems but also achieves extremely low-cost, concealed monitoring.

[0117] To verify the effectiveness of this application, experimental data is used as follows: The experimental setting was an apartment space with an area of ​​102.7 square meters, in which 78 energy capture sensor nodes were deployed, and a total of 155 subjects were recruited to participate in the experimental data collection. Table 1 shows the accuracy comparison between this application and other recognition methods (such as multilayer perceptron MLP). It can be seen that the cognitive impairment recognition accuracy of this invention is the best.

[0118] Table 1

[0119]

[0120] Another embodiment of this application provides a cognitive impairment detection device based on an energy-capturing IoT system, including a processor and a memory storing a plurality of computer instructions, which, when executed by the processor, implement the steps of the above method.

[0121] Specific limitations regarding the cognitive impairment detection device based on the energy-capture IoT system can be found in the limitations of the cognitive impairment detection method based on the energy-capture IoT system mentioned above, and will not be repeated here. The aforementioned cognitive impairment detection device based on the energy-capture IoT system can be implemented entirely or partially through software, hardware, or a combination thereof. It can be embedded in the processor of a computer device in hardware form or independent of the processor, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0122] The memory and processor are electrically connected directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines. The memory stores a computer program that can run on the processor, which implements the method in the embodiments of the present invention by running the computer program stored in the memory.

[0123] The memory may be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory stores the program, and the processor executes the program upon receiving an execution instruction.

[0124] The processor may be an integrated circuit chip with data processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor.

[0125] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for detecting cognitive impairment based on an energy-harvesting IoT system, characterized in that, The cognitive impairment detection method based on an energy-capturing IoT system includes: In a cognitive impairment detection environment, RFID tags and solar panels are deployed, and several collection points are set up. The received signal strength and photocurrent intensity when a person stands at each collection point are collected, generating a first database that includes the location of the collection point and the collected received signal strength and photocurrent intensity. The received signal intensity and photocurrent intensity of the sampled object during daily activities in a cognitive impairment detection scenario are collected and compared with the first database to obtain the original trajectory data of the sampled object's movement. The original trajectory data of the movement is segmented, and then the velocity changes of the segmented trajectory segments are converted into color gradient features to obtain velocity feature images. Corresponding labels are set to obtain the original training samples. The conditional variational autoencoder was trained using the original training samples. Synthetic data is generated by using the decoder in the trained conditional variational autoencoder and then merged with the original training samples to obtain the expanded training dataset. A cognitive impairment detection model is constructed using the encoder and classifier in a conditional variational autoencoder. The cognitive impairment detection model is then trained using an expanded training dataset to obtain a well-trained cognitive impairment detection model. The velocity feature image of the target to be detected is obtained and input into the trained cognitive impairment detection model to obtain the recognition result.

2. The cognitive impairment detection method based on an energy-harvesting IoT system according to claim 1, characterized in that, The received signal intensity and photocurrent intensity of the sampled object during daily activities in the cognitive impairment detection scenario are collected and compared with the first database to obtain the original trajectory data of the sampled object's movement, including: Calculate the Euclidean distance between the sampling point and the acquisition point, considering the received signal strength and photocurrent intensity. The location of the sampling point with the smallest Euclidean distance is used as the estimated location of the sampling point; Connect the estimated locations of all sampling points sequentially in chronological order to generate the original trajectory data of the sampled object's movement.

3. The cognitive impairment detection method based on an energy-harvesting IoT system according to claim 1, characterized in that, The process of segmenting the original mobile trajectory data includes: When the time interval between two consecutive locations in the trajectory exceeds the threshold, it is determined that the sampled object has been stationary or has not moved continuously during this period. The trajectory is then segmented between these two locations, with the earlier location defined as the end point of the previous trajectory segment and the later location defined as the start point of the next trajectory segment.

4. The cognitive impairment detection method based on an energy-harvesting IoT system according to claim 1, characterized in that, The training of the conditional variational autoencoder using the original training samples includes: The velocity feature image is segmented and positionally encoded to generate an image word sequence; the labels of the velocity feature image are embedded to generate a conditional vector. The latent vector posterior probability distribution is obtained by inputting the image word sequence and conditional vector into the encoder. The latent vector is obtained by sampling from the posterior probability distribution of the latent vector; The latent vector and conditional vector are input into the decoder to generate a reconstructed velocity feature map; Calculate the loss and update the parameters of the conditional variational autoencoder through backpropagation.

5. The cognitive impairment detection method based on an energy-harvesting IoT system according to claim 4, characterized in that, The method of generating synthetic data using the decoder in a trained conditional variational autoencoder includes: A target cognitive state label is defined, and a latent vector is randomly sampled from a standard normal distribution. The target cognitive state label and the randomly sampled latent vector are input into the decoder to reconstruct the synthetic velocity feature map.

6. The cognitive impairment detection method based on an energy-harvesting IoT system according to claim 4, characterized in that, The losses include reconstruction loss and KL divergence loss.

7. The cognitive impairment detection method based on an energy-harvesting IoT system according to claim 6, characterized in that, The reconstruction loss is used to measure the difference between the reconstructed velocity feature map output by the decoder and the original input velocity feature map; the KL divergence loss is used to measure the difference between the approximate posterior distribution inferred by the encoder and the preset prior distribution.

8. A cognitive impairment detection device based on an energy-harvesting IoT system, comprising a processor and a memory storing a plurality of computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 7.

Images