Event-triggered double-memory time series modeling method for smart wearable devices

Abstract

The application provides an event-triggered double-memory time sequence modeling method and system for a smart wearable device. The method and system accurately capture key moments through an event-driven mechanism, balance short-term response and long-term pattern maintenance by using a double-memory structure, significantly improve the identification accuracy of abnormal events under the premise of strictly limiting the end-side power consumption and storage budget, and guarantee user privacy and security. Since the application adopts an end-side original event anchor point memory and an adaptive gating fusion mechanism, the "forgetting" and "false alarm" problems of low-power devices in processing complex time sequences are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology for smart wearable devices, specifically to a time series modeling scheme that runs on a resource-constrained, low-power edge processor. The core function is to construct an anomaly monitoring and data analysis system model based on an event-triggered mechanism and a dual-memory recurrent neural network (LSTM / GRU). Background Technology

[0002] Smart wearable devices can continuously collect multi-dimensional physiological signals such as heart rate, blood oxygen, skin conductance, and cadence for health monitoring and exercise analysis. However, smart wearable devices often encounter the following problems:

[0003] 1) The continuously collected data are affected by sensor sampling and are often accompanied by a lot of noise. At the same time, the data distribution is significantly drifted due to the wearing environment and individual differences.

[0004] 2) When capturing low-frequency critical physiological abnormalities and sudden movement events, extremely high real-time performance is required;

[0005] 3) The device itself has extremely limited computing resources, storage space, and battery life, so it is not advisable to upload to the cloud frequently.

[0006] Existing edge-side solutions are mostly based on standard LSTM or attention mechanism models, which struggle to simultaneously achieve both keen capture of sudden events and effective memorization of long-term historical patterns at extremely low power consumption. Traditional replay-based memory mechanisms or solutions relying on cloud collaboration not only increase communication and storage overhead but also introduce potential privacy risks. Therefore, there is an urgent need for a general modeling method that can operate independently on the edge, is sensitive to critical events, and efficiently utilizes limited computing resources. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this invention provides an event-triggered dual-memory time series modeling method and system for smart wearable devices. This invention aims to provide an event-triggered dual-memory time series modeling method and system specifically designed for low-power smart wearable devices. This method accurately captures key moments through an event-driven mechanism, utilizes a dual-memory structure to balance short-term response and long-term pattern retention, and significantly improves the accuracy of abnormal event identification while strictly limiting edge power consumption and storage budget, while ensuring user privacy and security.

[0008] The present invention adopts the following technical solution, and the specific steps are as follows:

[0009] a) Lightweight backbone processing;

[0010] The collected time series input sequence is denoted as xt. The time series input sequence xt is input into the recurrent neural network backbone model deployed on the edge device. The recurrent neural network backbone model is a lightweight LSTM or GRU, which is used to update the hidden state ht and the initial output yt at the current time in real time. The recurrent neural network backbone model constitutes a short time memory path to retain the temporal dependencies of the most recent sampling times.

[0011] b) Adaptive event detection;

[0012] The change intensity index Et is calculated based on the current input xt, the predicted residual et, and the hidden state ht. The change intensity index Et measures the degree of anomaly compared to the nearest time window. Preferably, the change intensity index Et is calculated in the following form:

[0013]

[0014] Among them, R t V is the sliding window residual statistic, representing the difference between the current sampled value and the predicted value of the backbone model; t C is the variance abrupt change statistic for the sliding window, representing the degree of variance change between the current window and the previous window; t For cumulative and CUSUM statistics; U t Output uncertainty score for the model; a1, a2, a3 and a4 are all non-negative weight coefficients, and satisfy a1+a2+a3+a4=1;

[0015] Based on the set of change intensity indices {E( t-N+1 ), …, E t Dynamically update the trigger threshold θ t The trigger threshold is preferably a quantile threshold or a median absolute deviation threshold, specifically:

[0016]

[0017] Where median(E) is the median of the set of change intensity indicators, MAD(E) is the absolute deviation of the median, and λ is the threshold adjustment coefficient; when Et≥θ t When the current time t is determined to be a critical event time, the event memory construction and retrieval enhancement process is triggered; when Et < θ t At that time, only the recursive update of the backbone model is performed;

[0018] c) Anchor point construction and restricted storage;

[0019] When the current moment is determined to be a critical event moment, linear projection compression is performed on the current hidden state ht to obtain a low-dimensional projection vector pt. The specific steps are as follows:

[0020]

[0021] Among them, W p Let b be the projection matrix. p p is the bias term. t The dimension is d p Then, statistical features and frequency domain features are extracted from the preset window before and after the current time to form an auxiliary feature vector ft; the statistical features include at least one of mean, variance, extreme value or slope, and the frequency domain features include at least one of main frequency energy, frequency band energy ratio or spectral centroid.

[0022] The low-dimensional projection vector pt, the auxiliary feature vector ft, and the corresponding timestamp τt are concatenated or encapsulated according to a predetermined format to form a composite anchor point A. t ,Right now:

[0023]

[0024] Among them, u t Count the historical utilization rate of the anchor point; the composite anchor point A t The event memory EM is written to the event memory library. The event memory library EM is implemented using a circular buffer, which includes a head pointer and a tail pointer. It also uses a bucket index structure based on quantized projection coding to record the index bucket to which each anchor point belongs, so as to support subsequent fast retrieval.

[0025] d) Dynamic budget management;

[0026] To accommodate limited storage resources on the endpoint, the event memory is set to a fixed capacity budget C, where C represents the maximum number of anchors that the event memory can accommodate. When the event memory EM reaches its capacity limit C and there are new composite anchors to be written, the budget management strategy is activated.

[0027] The budget management strategy includes calculating an elimination score Si for each anchor point within the database, wherein the elimination score Si is preferably defined as:

[0028]

[0029] Among them, Sim i Age represents the similarity between the i-th anchor point and its nearest neighbor anchor points, used to characterize the degree of redundancy; i This represents the normalized result of the anchor dwell time; Use iThis represents the normalized result of the frequency of the anchor point being hit during the historical retrieval process; λ1, λ2, and λ3 are weight coefficients. The system prioritizes eliminating the anchor point with the largest Si, thereby deleting anchor points with high redundancy, obsolescence, and low utilization, ensuring that the event memory always retains key event patterns with strong representativeness and high diversity.

[0030] e) Fusion of contextual retrieval and gating;

[0031] During the model inference phase, a query vector qt is constructed based on the current hidden state ht, preferably:

[0032]

[0033] Using the query vector qt, a similarity retrieval is performed in the event memory EM. First, candidate anchor points are filtered within the index bucket corresponding to qt and its adjacent index buckets. Then, the candidate anchor points are sorted according to the cosine similarity. The cosine similarity calculation formula is as follows:

[0034]

[0035] The top-k historical anchor points with the highest similarity are selected as the search results; the highest similarity refers to the anchor point with the largest cosine similarity value with the query vector.

[0036] Subsequently, a weighted aggregation is performed on the Top-k historical anchors to obtain the historical context vector rt, preferably using Softmax normalized weights:

[0037]

[0038]

[0039] Where T is the temperature coefficient. Indicates that from anchor point A i The feature representation obtained by mapping the low-dimensional projection vector and auxiliary feature vector in the model;

[0040] The current hidden state ht and the historical context vector rt are then input into the fusion gating module to calculate the gating coefficient g. t :

[0041]

[0042] Where [ht; rt] represents the splicing operation, and Wg and bg are the parameters of the fusion gating module, which generate the enhanced hidden state based on the gating coefficients. :

[0043]

[0044] Here, ⊙ represents element-wise multiplication, ultimately based on the enhanced hidden state. Generate corrected output ;

[0045] f) Multi-objective joint training;

[0046] During the training phase, a joint loss function L is constructed to jointly optimize the backbone model, event detection parameters, projection parameters, and gating fusion parameters. The joint loss function L includes the task prediction loss Li. pred Anchor point consistency regularization term L anchor And memory diversity regularization term L mem Specifically:

[0047]

[0048] Where μ and ν are non-negative hyperparameters;

[0049] The task prediction loss L pred The error between the model output and the label is constrained, preferably by cross-entropy loss or mean squared error loss; the anchor consistency regularization term L... anchor Hidden state after constraint enhancement The distance between the retrieved historical context vector rt and the target vector rt in the feature space is preferably defined as:

[0050]

[0051] The memory diversity regularization term L mem To suppress excessive similarity between anchors within the event memory, it is preferably defined as:

[0052]

[0053] Where M is the number of anchors currently participating in training, and m is the threshold of the upper limit of similarity tolerance. When the cosine similarity between two anchors is higher than the threshold m, a penalty is imposed on the anchor to reduce redundant memory and improve pattern coverage.

[0054] g) Low-power optimization implementation;

[0055] The method runs on the edge processor of the smart wearable device, which includes an MCU, DSP or low-power SoC. In order to reduce system power consumption, the backbone model, projection module and gating fusion module are implemented using fixed-point arithmetic or 8-bit quantization operators.

[0056] The event detection module and the retrieval module are activated when a triggering condition is met, wherein the triggering condition is Et ≥ θ. tDuring non-critical periods, i.e. when Et < θt is satisfied for multiple consecutive sampling periods, the system only executes the backbone model recursion and updates the event detection statistics in a low duty cycle manner, while pausing anchor point writing and historical retrieval operations to reduce average computing load and power consumption.

[0057] h) Privacy protection and collaborative expansion;

[0058] In edge-cloud collaboration scenarios, edge devices only upload anchor point summaries or statistical features that have undergone low-dimensional projection compression and anonymization, without uploading the original physiological signals; the cloud performs aggregation training or parameter fine-tuning based on the anchor point summaries uploaded by multiple edge devices, and sends the updated model parameters to the edge devices.

[0059] Furthermore, differential privacy noise can be superimposed on the anchor point summary to be synchronized on the edge or cloud side, or a secure aggregation protocol can be used during multi-device joint training to prevent local features of a single user from being inverted into the original data, thereby further enhancing data privacy protection capabilities.

[0060] The calculation method for the change intensity index Et includes at least one of sliding window residual analysis, variance mutation detection, cumulative sum CUSUM algorithm, and uncertainty scoring.

[0061] The dimension dp of the low-dimensional projection vector pt is 64 to 128, the number of retrievals Top-k is 3 to 5, and the capacity budget C of the event memory is 256 to 2048 anchor units.

[0062] The similarity metric uses a cosine similarity algorithm, and the fusion gating coefficient gt is calculated using a sigmoid activation function.

[0063] The data structure of the composite anchor At includes a low-dimensional projection vector of the hidden state, statistical or frequency domain feature descriptors of the nearest time window, timestamp information, and anchor utilization statistics.

[0064] The memory diversity regularization term punishes pairs of anchors in the event memory bank whose similarity exceeds a threshold, thereby suppressing redundant anchors and improving the representativeness of coverage for different event patterns.

[0065] The edge-cloud collaboration mechanism includes the following steps: edge devices periodically upload low-dimensional anchor point summaries; the cloud aggregates and trains the anchor point summaries uploaded by multiple devices and updates the shared model parameters; the edge devices receive and load the shared model parameters for subsequent local inference and event detection.

[0066] A system for implementing the method includes:

[0067] The sensor acquisition module is used to acquire raw physiological or motion signals;

[0068] The preprocessing module is used to filter, denoise, segment, and standardize the raw physiological or motion signals.

[0069] The recurrent neural network backbone module is used to extract time series features and output the hidden state ht and the initial output yt.

[0070] The event detection module is used to identify key event moments in real time based on the change intensity index Et and the adaptive threshold θt.

[0071] The event memory module is used to build, store, and manage composite anchors;

[0072] The retrieval module is used to retrieve similar historical anchors from the event memory based on the current query vector;

[0073] The fusion gating module is used to fuse the current hidden state with the historical context vector and generate an enhanced hidden state;

[0074] The budget control module is used to manage the event memory's capacity budget, anchor point eviction, and compression strategies.

[0075] The security and privacy module is used to perform anonymization, encryption, or privacy protection processing on the data to be synchronized;

[0076] The edge-cloud collaboration module is used to upload anchor point summaries, receive global model updates, and synchronize parameters when needed.

[0077] The modules described above work together to complete time-series modeling, anomaly detection, and memory-enhanced reasoning under resource-constrained conditions on the smart wearable device side, following the steps outlined in the method.

[0078] The budget controller is responsible for real-time monitoring of the insertion rate, anchor utilization, storage capacity usage, and system energy consumption of the memory bank. When the capacity limit is exceeded, it automatically triggers anchor eviction and compression strategies.

[0079] A computer-readable storage medium having stored thereon computer-executable instructions that, when loaded and executed by a processor, cause the processor to perform the steps of the method as described.

[0080] The backbone model is implemented using 8-bit quantization or fixed-point operators during deployment, and event detection and retrieval operations are only executed when the trigger condition Et≥θt is met, in order to reduce the overall power consumption of the system.

[0081] Optimal parameter settings:

[0082] The hidden layer dimension of the recurrent backbone network is set to 64 to 128; the low-dimensional projection dimension is controlled between 64 and 128; the event memory capacity budget C is set to 256 to 2048 anchor points; the retrieval Top-k value is set to 3 to 5; the Sigmoid activation function is selected for fusion gating; and cosine similarity is used as the similarity measure.

[0083] The beneficial effects of this invention are: by employing the native event anchor memory and adaptive gating fusion mechanism on the edge side, this invention effectively solves the problems of "forgetting" and "false alarms" in low-power devices when processing complex time series. Its core advantage lies in:

[0084] 1. High robustness: It can keenly capture sudden anomalies and structural data changes, significantly improving the recognition accuracy.

[0085] 2. Resource efficiency: Through strict capacity budget management and low-dimensional data compression, the demand for on-chip storage and communication bandwidth is significantly reduced.

[0086] 3. Privacy compliance: Only anonymized feature summaries are stored and transmitted, ensuring the security of user data from the architectural level.

[0087] 4. Low latency and low power consumption: The event-triggered computing strategy and quantized execution scheme ensure the system's real-time response capability under extremely low power consumption. Attached Figure Description

[0088] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0089] Figure 2 This is a flowchart of the method of the present invention;

[0090] Figure 3 This is a schematic diagram of the event memory structure and retrieval of the present invention;

[0091] Figure 4 This is a schematic diagram of the fusion gated computation of the present invention;

[0092] Figure 5 This is a schematic diagram of the low-power execution and resource budgeting of the present invention.

[0093] The components are as follows: 100 - Sensor acquisition module; 110 - Preprocessing module; 120 - Cyclic backbone (LSTM / GRU); 130 - Event detector; 140 - Event memory (EM); 150 - Retrieval unit; 160 - Fusion gating module; 170 - Budget controller; 180 - Security and privacy module; 190 - End-to-cloud collaboration interface; 210 - Anchor entry; 211 - Low-dimensional projection vector; 212 - Near-window statistics / frequency domain features; 213 - Timestamp; 220 - Index structure; 230 - Top-k search results; 240 - Weighted aggregation; 250 - Current hidden state ht; 251 - Aggregation anchor; 260 - Gated computing unit; 261 - Gated output gt; 270 - Enhanced state h't; 300 - Processor / MCU; 310 - Storage resources; 320 - Duty cycle control; 330 - Capacity budget C; 340 - Power consumption statistics; 350 - Inference latency. Detailed Implementation

[0094] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0095] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0096] 5.1 Examples of Low-Power Devices

[0097] This embodiment operates on a smart wearable device equipped with an MCU, DSP, or low-power SoC, with on-chip available memory ranging from 0.5MB to 16MB. The system adaptively adjusts the sampling frequency according to the sensor type, with the PPG sensor sampling frequency ranging from 25Hz to 128Hz and the accelerometer sampling frequency ranging from 25Hz to 250Hz.

[0098] The first step is model deployment: a single-layer or dual-layer LSTM / GRU is used as the backbone, the hidden layer dimension is set to 64-128, and 8-bit quantization or fixed-point arithmetic is fully applied to adapt to the hardware accelerator.

[0099] The second step is event triggering: Calculate the sliding window residual statistic Rt and the cumulative sum statistic Ct for the continuously sampled signal, and calculate the dynamic threshold θt based on the set of change intensity indicators for the most recent N=64 time points. When Et ≥ θt, the current time point is determined to be a critical event moment.

[0100] The third step is anchor point construction: the current hidden state ht is compressed into a 64-dimensional or 128-dimensional low-dimensional vector pt using the projection matrix Wp; at the same time, the mean, variance, peak-to-peak value and main frequency energy ratio are extracted from the nearest neighbor window composed of 8 sampling points before and after, forming an auxiliary feature vector ft; then combined with the timestamp τt to form a composite anchor point At, and written into the event memory.

[0101] The fourth step is memory management: The event memory uses a circular buffer structure to achieve fixed-capacity storage. When a new event triggers a storage request and the capacity is full, the system calculates an elimination score based on the "least recently used" and "similarity redundancy" statistics, and deletes the old anchor point with the highest score.

[0102] The fifth step is to enhance reasoning: using the query vector qt to perform cosine similarity retrieval on the event memory, select the top-3 historical anchors and use Softmax weights to aggregate and generate a historical context vector rt, and then calculate the fusion gating coefficient gt through affine transformation and Sigmoid activation function to achieve dynamic fusion of the current state and historical memory.

[0103] 5.2 Training and Deployment Process

[0104] 1. Warm-up phase: Only the recurrent neural network backbone module is enabled, and 3 to 5 rounds of basic training are performed using training samples to establish the ability to extract temporal features.

[0105] 2. Joint Training Phase: Enable the event detection module, event memory module, and fusion gating module, and perform 5 to 10 rounds of joint training according to the joint loss function L = Lpred +μ·Lanchor +ν·Lmem; in this phase, perform grid search or segmented adjustment on μ and ν to balance prediction accuracy and memory diversity.

[0106] 3. Deployment phase: Complete model quantization and compression, and solidify the projection matrix, gating parameters, and retrieval temperature coefficient; in actual operation, anchor point writing, historical retrieval, and fusion enhancement paths are only activated when an event is triggered, and low-power standby or low duty cycle operation is maintained under normal conditions.

[0107] In a series of simulation experiments, synthetic heart rate sequence data were tested based on a 64-dimensional hidden-state LSTM backbone. The noise level σ of the synthetic heart rate sequence was 0.1. After adopting a CUSUM-based triggering strategy, the average computational cost was reduced by approximately 30%. After introducing Top-3 anchor point retrieval, the F1 score of the anomaly detection task improved from 0.85 for the standard LSTM to 0.92, and the false alarm rate decreased from 15% to 5%.

[0108] 5.3 Applicable Scenarios

[0109] This invention is applicable to wearable health and exercise monitoring scenarios such as abnormal heart rate warning, respiratory pattern analysis, blood oxygen saturation monitoring, gait recognition, and sleep quality assessment. Without departing from the concept of this invention, the backbone network can be replaced with a lightweight Transformer or a one-dimensional convolutional network, and the similarity metric can be replaced with inner product or Euclidean distance calculation depending on the characteristics of the hardware instruction set.

Claims

1. An event-triggered dual-memory time series modeling method for smart wearable devices, characterized in that... Includes the following steps: a) Lightweight backbone processing; The collected time series input sequence is denoted as xt. The time series input sequence xt is input into the recurrent neural network backbone model deployed on the edge device. The recurrent neural network backbone model is a lightweight LSTM or GRU, which is used to update the hidden state ht and the initial output yt at the current time in real time. The recurrent neural network backbone model constitutes a short time memory path to retain the temporal dependencies of the most recent sampling times. b) Adaptive event detection; The change intensity index Et is calculated based on the current input xt, the predicted residual et, and the hidden state ht. The change intensity index Et measures the degree of anomaly compared to the nearest time window. Preferably, the change intensity index Et is calculated in the following form: ； Among them, R t V is the sliding window residual statistic, representing the difference between the current sampled value and the predicted value of the backbone model; t C is the variance abrupt change statistic for the sliding window, representing the degree of variance change between the current window and the previous window; t For cumulative and CUSUM statistics; U t Output uncertainty score for the model; a1, a2, a3 and a4 are all non-negative weight coefficients, and satisfy a1+a2+a3+a4=1; Based on the set of change intensity indices {E( t-N+1 ), …, E t Dynamically update the trigger threshold θ t The trigger threshold is preferably a quantile threshold or a median absolute deviation threshold, specifically: ； Where median(E) is the median of the set of change intensity indicators, MAD(E) is the absolute deviation of the median, and λ is the threshold adjustment coefficient; when Et≥θ t When the current time t is determined to be a critical event time, the event memory construction and retrieval enhancement process is triggered; when Et < θ t At that time, only the recursive update of the backbone model is performed; c) Anchor point construction and restricted storage; When the current moment is determined to be a critical event moment, linear projection compression is performed on the current hidden state ht to obtain a low-dimensional projection vector pt. The specific steps are as follows: ； Among them, W p Let b be the projection matrix. p p is the bias term. t The dimension is d p Then, statistical features and frequency domain features are extracted from the preset window before and after the current time to form an auxiliary feature vector ft; the statistical features include at least one of mean, variance, extreme value or slope, and the frequency domain features include at least one of main frequency energy, frequency band energy ratio or spectral centroid. The low-dimensional projection vector pt, the auxiliary feature vector ft, and the corresponding timestamp τt are concatenated or encapsulated according to a predetermined format to form a composite anchor point A. t ,Right now: ； Among them, u t Count the historical utilization rate of the anchor point; the composite anchor point A t The event memory EM is written to the event memory EM. The event memory EM is implemented using a circular buffer, which includes a head pointer and a tail pointer. It also uses a bucket index structure based on quantized projection coding to record the index bucket to which each anchor point belongs, so as to support subsequent fast retrieval. d) Dynamic budget management; To accommodate limited storage resources on the endpoint, the event memory is set to a fixed capacity budget C, where C represents the maximum number of anchors that the event memory can accommodate. When the event memory EM reaches its capacity limit C and there are new composite anchors to be written, the budget management strategy is activated. The budget management strategy includes calculating an elimination score Si for each anchor point within the database, wherein the elimination score Si is preferably defined as: ； Among them, Sim i Age represents the similarity between the i-th anchor point and its nearest neighbor anchor points, used to characterize the degree of redundancy; i This represents the normalized result of the anchor dwell time; Use i This represents the normalized result of the frequency of the anchor point being hit during the historical retrieval process; λ1, λ2, and λ3 are weight coefficients. The system prioritizes eliminating the anchor point with the largest Si, thereby deleting anchor points with high redundancy, obsolescence, and low utilization, ensuring that the event memory always retains key event patterns with strong representativeness and high diversity. e) Fusion of contextual retrieval and gating; During the model inference phase, a query vector qt is constructed based on the current hidden state ht, preferably: ； Using the query vector qt, a similarity retrieval is performed in the event memory EM. First, candidate anchor points are filtered within the index bucket corresponding to qt and its adjacent index buckets. Then, the candidate anchor points are sorted according to the cosine similarity. The cosine similarity calculation formula is as follows: ； The top-k historical anchor points with the highest similarity are selected as the search results; the highest similarity refers to the anchor point with the largest cosine similarity value to the query vector. Subsequently, a weighted aggregation is performed on the Top-k historical anchors to obtain the historical context vector rt, preferably using Softmax normalized weights: ； ； Where T is the temperature coefficient. Indicates that from anchor point A i The feature representation obtained by mapping the low-dimensional projection vector and auxiliary feature vector in the model; The current hidden state ht and the historical context vector rt are then input into the fusion gating module to calculate the gating coefficient g. t : ； Where [ht; rt] represents the splicing operation, and Wg and bg are the parameters of the fusion gating module, which generate the enhanced hidden state based on the gating coefficients. : ； Here, ⊙ represents element-wise multiplication, ultimately based on the enhanced hidden state. Generate corrected output ; f) Multi-objective joint training; During the training phase, a joint loss function L is constructed to jointly optimize the backbone model, event detection parameters, projection parameters, and gating fusion parameters. The joint loss function L includes the task prediction loss Li. pred Anchor point consistency regularization term L anchor And memory diversity regularization term L mem Specifically: ； Where μ and ν are non-negative hyperparameters; The task prediction loss L pred The error between the model output and the label is constrained, preferably by cross-entropy loss or mean squared error loss; the anchor consistency regularization term L... anchor Hidden state after constraint enhancement The distance between the retrieved historical context vector rt and the target vector rt in the feature space is preferably defined as: ； The memory diversity regularization term L mem To suppress excessive similarity between anchors within the event memory, it is preferably defined as: ； Where M is the number of anchors currently participating in training, and m is the threshold of the upper limit of similarity tolerance. When the cosine similarity between two anchors is higher than the threshold m, a penalty is imposed on the anchor to reduce redundant memory and improve pattern coverage. g) Low-power optimization implementation; The method runs on the edge processor of the smart wearable device, which includes an MCU, DSP or low-power SoC. In order to reduce system power consumption, the backbone model, projection module and gating fusion module are implemented using fixed-point arithmetic or 8-bit quantization operators. The backbone model is implemented using 8-bit quantization or fixed-point operators during deployment, and the event detection and retrieval operations are only executed when the trigger condition Et≥θt is met, so as to reduce the overall power consumption of the system. The event detection module and the retrieval module are activated when a triggering condition is met, wherein the triggering condition is Et ≥ θ. t During non-critical periods, i.e. when Et < θt is satisfied for multiple consecutive sampling periods, the system only executes the backbone model recursion and updates the event detection statistics in a low duty cycle manner, while pausing anchor point writing and historical retrieval operations to reduce average computing load and power consumption. h) Privacy protection and collaborative expansion; In edge-cloud collaboration scenarios, edge devices only upload anchor point summaries or statistical features that have undergone low-dimensional projection compression and anonymization, without uploading the original physiological signals; the cloud performs aggregation training or parameter fine-tuning based on the anchor point summaries uploaded by multiple edge devices, and sends the updated model parameters to the edge devices.

2. The event-triggered dual-memory time series modeling method for smart wearable devices according to claim 1, characterized in that: Differential privacy noise can be superimposed on the anchor point summary to be synchronized on the edge or cloud side, or a secure aggregation protocol can be used during joint training of multiple devices to prevent local features of a single user from being reverse-engineered into the original data, thereby further enhancing data privacy protection capabilities.

3. The event-triggered dual-memory time series modeling method for smart wearable devices according to claim 1, characterized in that: The calculation method for the change intensity index Et includes at least one of sliding window residual analysis, variance mutation detection, cumulative sum CUSUM algorithm, and uncertainty scoring.

4. The event-triggered dual-memory time series modeling method for smart wearable devices according to claim 1, characterized in that: The dimension dp of the low-dimensional projection vector pt is 64 to 128, the number of retrievals Top-k is 3 to 5, and the capacity budget C of the event memory is 256 to 2048 anchor units.

5. The event-triggered dual-memory time series modeling method for smart wearable devices according to claim 1, characterized in that: The similarity metric uses a cosine similarity algorithm, and the fusion gating coefficient gt is calculated using a sigmoid activation function.

6. The event-triggered dual-memory time series modeling method for smart wearable devices according to claim 1, characterized in that: The data structure of the composite anchor At includes a low-dimensional projection vector of the hidden state, statistical or frequency domain feature descriptors of the nearest time window, timestamp information, and anchor utilization statistics.

7. The event-triggered dual-memory time series modeling method for smart wearable devices according to claim 1, characterized in that: The memory diversity regularization term punishes pairs of anchors in the event memory bank whose similarity exceeds a threshold, thereby suppressing redundant anchors and improving the representativeness of coverage for different event patterns.

8. The event-triggered dual-memory time series modeling method for smart wearable devices according to claim 1, characterized in that: The edge-cloud collaboration mechanism includes the following steps: edge devices periodically upload low-dimensional anchor point summaries; the cloud aggregates and trains the anchor point summaries uploaded by multiple devices and updates the shared model parameters; the edge devices receive and load the shared model parameters for subsequent local inference and event detection.

9. A system according to any one of claims 1-8, characterized in that: The system includes: The sensor acquisition module is used to acquire raw physiological or motion signals; The preprocessing module is used to filter, denoise, segment, and standardize the raw physiological or motion signals. The recurrent neural network backbone module is used to extract time series features and output the hidden state ht and the initial output yt. The event detection module is used to identify key event moments in real time based on the change intensity index Et and the adaptive threshold θt. The event memory module is used to build, store, and manage composite anchors; The retrieval module is used to retrieve similar historical anchors from the event memory based on the current query vector; The fusion gating module is used to fuse the current hidden state with the historical context vector and generate an enhanced hidden state; The budget control module is used to manage the event memory's capacity budget, anchor point eviction, and compression strategies. The security and privacy module is used to perform anonymization, encryption, or privacy protection processing on the data to be synchronized; The edge-cloud collaboration module is used to complete anchor point summary uploading, global model update reception, and parameter synchronization when needed; The above modules work together to complete time series modeling, anomaly detection, and memory-enhanced reasoning under resource-constrained conditions on the smart wearable device side, following the described method steps. The budget controller is responsible for real-time monitoring of the insertion rate, anchor utilization, storage capacity usage, and system energy consumption of the memory bank. When the capacity limit is exceeded, it automatically triggers anchor eviction and compression strategies.

10. A computer-readable storage medium having stored thereon computer-executable instructions, which, when loaded and executed by a processor, cause the processor to perform the method of any one of claims 1-8.

Images