Smart bracelet multi-modal data fusion prediction method adopting deep learning

Abstract

The invention discloses a smart bracelet multi-modal data fusion prediction method adopting deep learning, and belongs to the technical field of wireless communication, and the method comprises the steps: collecting multi-modal data in real time; evaluating the confidence of the positioning source through a CNN-LSTM model, and dynamically distributing a fusion weight; when the GPS fails, a triple decision-making mechanism is triggered to optimize UWB positioning, and gyroscope short-time compensation is synchronously started; fusing the multi-modal data based on a double-branch Transform model, and predicting a behavior state prediction result in a second preset time period; querying a user path discretization segmented link quality knowledge graph, and determining a pre-switching time sequence in combination with the real-time signal strength; during link switching, a hierarchical fragmentation and sliding window caching strategy is adopted to transmit data; according to the invention, precise positioning, active link switching and reliable data transmission are realized, and the positioning precision and communication stability of the smart bracelet in a complex scene are improved.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically a method for multimodal data fusion and prediction of smart bracelets using deep learning. Background Technology

[0002] With the widespread adoption of smart wearable devices, smart bracelets are increasingly used in crowded and complex environments such as amusement parks. Their positioning accuracy and communication stability directly impact user experience and the provision of related services. However, existing smart bracelet positioning methods often rely on a single positioning source. GPS signals are prone to failure in scenarios with building obstruction or interference from dense crowds. UWB and Bluetooth positioning sources also suffer from insufficient accuracy or weak anti-interference capabilities when used alone. Furthermore, link switching often employs a passive triggering mechanism, typically initiating switching only after signal quality deteriorates, easily leading to communication lag and interruptions. Additionally, the lack of adaptation strategies for link switching scenarios during data transmission increases the risk of data loss. Moreover, traditional data fusion and state prediction methods struggle to effectively mine the spatiotemporal correlation features of multimodal data, resulting in low prediction accuracy and poor timeliness, failing to provide reliable support for decisions such as link switching. Therefore, there is an urgent need for a smart bracelet data processing method that can integrate multi-source data, accurately predict states, proactively switch links, and ensure reliable data transmission. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a deep learning-based multimodal data fusion and prediction method for smart bracelets. This method involves real-time acquisition of multimodal data; evaluation of location source confidence using a CNN-LSTM model and dynamic allocation of fusion weights; triggering a triple-decision mechanism to optimize UWB positioning when GPS fails, simultaneously initiating short-term gyroscope compensation; fusing multimodal data based on a dual-branch Transformer model to predict behavioral state prediction results for a second preset time period; querying a discretized segmented link quality knowledge graph of the user path and determining the pre-switching timing based on real-time signal strength; and employing a hierarchical segmentation and sliding window caching strategy for data transmission during link switching. This invention achieves accurate positioning, proactive link switching, and reliable data transmission, improving the positioning accuracy and communication stability of smart bracelets in complex scenarios.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A multimodal data fusion and prediction method for smart bracelets employing deep learning includes:

[0006] The wristband collects multimodal data in real time; the multimodal data includes GPS signals, UWB positioning signals deployed in the park, Bluetooth beacon signals, gyroscope data, acceleration data, motion data, and historical communication link data.

[0007] A signal quality assessment model based on CNN-LSTM is constructed. The inputs are GPS signal, UWB positioning signal deployed in the park and Bluetooth beacon signal. The output is the confidence score of each positioning source. The fusion weight of each positioning source is dynamically allocated based on the confidence score to generate the positioning source fusion result.

[0008] When the signal quality assessment indicates that GPS has entered a failure area, a triple decision mechanism is triggered to obtain the UWB positioning result after weight adjustment, start short-term gyroscope compensation, and generate the compensation result for the first preset time period.

[0009] Based on the location source fusion results and compensation results, the motion data collected by the wristband and historical communication link data are fused, and multimodal state prediction is performed through a dual-branch Transformer model to output the behavior state prediction results for the second preset time period; the first preset time period is shorter than the second preset time period;

[0010] Based on the behavioral state prediction results, the pre-constructed user path discretized segmented link quality knowledge graph is queried, and the pre-handover timing sequence is determined in combination with the real-time signal strength to generate a link handover instruction frame.

[0011] Based on the generated link switching instruction frame, during the triggered link switching process, a hierarchical fragmentation and sliding window caching strategy is adopted to transmit multimodal data.

[0012] Specifically, the signal quality assessment model is a hybrid model containing parallel one-dimensional convolutional neural network channels and long short-term memory network layers; the number of parallel one-dimensional convolutional neural network channels corresponds to the type of positioning source, and is used to extract local spatiotemporal features of GPS signal sequences, UWB positioning signal sequences and Bluetooth beacon signal sequences respectively; the long short-term memory network layer is used to fuse the local spatiotemporal features extracted by each channel and model the temporal dependency of signal quality.

[0013] The construction process of the CNN-LSTM-based signal quality assessment model includes: training the hybrid model using a historical multimodal dataset labeled with the true confidence scores of each location source; inputting real-time acquired GPS signal sequences, UWB positioning signal sequences, and Bluetooth beacon signal sequences into the trained hybrid model; and outputting a confidence score for each location source at the current moment from the hybrid model. The true confidence scores are calculated based on the error between the current positioning result and the reference trajectory. The confidence scores are one-dimensional vectors, with each element corresponding to the credibility of a location source.

[0014] Specifically, when the signal quality assessment indicates that GPS has entered a failure zone, a triple decision-making mechanism is triggered, including:

[0015] The GPS confidence score output by the signal quality assessment model is monitored in real time. When the GPS confidence score is detected to be lower than the first preset threshold, a timer is started. If the GPS confidence score continues to be lower than the first preset threshold within the preset time, the GPS signal is determined to have entered a failure zone. Based on the determination result of the failure zone, the triple decision mechanism is triggered.

[0016] The first step of the triple decision-making mechanism is to perform historical trajectory feature matching based on the Transformer-XL model, specifically: The UWB positioning results generated from the UWB positioning signal sequence within a time period are input into the pre-trained Transformer-XL model. The Transformer-XL model matches the motion pattern segment with the highest similarity from the historical trajectory database and outputs the trajectory continuity correction coefficient.

[0017] The second decision of the triple decision mechanism is to identify signal mutation patterns. Specifically, it involves analyzing the instantaneous change rate of the arrival time difference of each base station signal in the current UWB positioning signal sequence. If the instantaneous change rate exceeds the second preset threshold, it is determined that there is a signal mutation caused by multipath or occlusion, and a signal stability discount coefficient is generated.

[0018] The third decision in the triple decision-making mechanism is to adjust the weights according to the occlusion level. Specifically, the comprehensive occlusion level of the current environment is evaluated based on the attenuation model of the received signal strength indication value in the Bluetooth beacon signal sequence, and mapped to the basic weight value.

[0019] Specifically, the calculation process of the UWB positioning result after weight adjustment includes:

[0020] Based on the trajectory continuity correction coefficient, signal stability discount coefficient, and weight base value output by the triple decision-making mechanism, the dynamic adjustment weight of the UWB positioning signal is calculated.

[0021] The dynamically adjusted weights are weighted together with the UWB confidence score currently output by the signal quality assessment model to generate fusion weights applicable to the UWB positioning results at the current time.

[0022] Using the aforementioned fusion weights and combining them with the Bluetooth positioning results, the weighted UWB positioning result for the current moment is generated according to the weighted fusion method.

[0023] Specifically, the calculation process for the dynamic adjustment weight of the UWB positioning signal is as follows:

[0024] Preset reference weights for UWB positioning sources under normal conditions As a basis for adjustment;

[0025] Calculate the trajectory continuity correction coefficient complementary values ;

[0026] The complementary value Multiplying this by the preset empirical adjustment coefficient k yields the adjustment amount used to adjust the baseline weights due to trajectory discontinuity risk. ;

[0027] The adjustment amount With the benchmark weight Adding them together yields the initial weights after continuous adjustment. ;

[0028] The weights after the initial continuity adjustment With the signal stability discount factor Multiply to obtain the stability-corrected weights. ;

[0029] The stability-corrected weights With the aforementioned weight base value Multiply to generate dynamically adjusted weights for the UWB positioning signal used in data fusion at the current moment. .

[0030] Specifically, the short-time compensation for activating the gyroscope includes:

[0031] When the triple decision-making mechanism is triggered, a short-term gyroscope compensation process is initiated synchronously. The short-term gyroscope compensation adopts a physically embedded Transformer model. The input of the physically embedded Transformer model is preprocessed gyroscope data and acceleration data within a time window prior to the triggering moment, and the output is a sequence of trajectory prediction values ​​for each moment within a first preset time period starting from the triggering moment. The length of the time window is less than the first preset time period.

[0032] The physically embedded Transformer model embeds the basic kinematic equations of the inertial navigation system into the encoder as a physical constraint layer to ensure that the predicted trajectory conforms to physical laws in the feature space.

[0033] Specifically, the step of performing multimodal state prediction using a two-branch Transformer model and outputting the behavioral state prediction result for a second preset time period includes:

[0034] The dual-branch Transformer model consists of a spatial relation branch and a temporal dependency branch connected in parallel. The spatial relation branch receives preprocessed multimodal data and learns the intrinsic correlation features between positioning information, motion state, and communication link state at the same time point through a self-attention mechanism, outputting a first high-level feature vector. The temporal dependency branch receives the same multimodal data and learns the evolution rules and trends of each modality data along the time dimension through a causal self-attention mechanism, outputting a second high-level feature vector.

[0035] The first and second high-level feature vectors output by the spatial relationship branch and the temporal dependency branch are concatenated, and then fused and dimensionality reduced through a fully connected layer to finally output the behavioral state prediction result for each moment in the second preset time period; the behavioral state prediction result includes at least the user's location, movement speed, movement mode and expected communication quality level; the movement mode includes at least walking, running and standing still.

[0036] Specifically, the method for constructing the user path discretized segmented link quality knowledge graph includes:

[0037] The map of the target amusement park is pre-divided into multiple logically connected areas based on its physical layout and functional attributes, with each area associated with at least one wireless communication access point.

[0038] Obtain historical datasets; the historical datasets contain continuous spatiotemporal trajectory data of multiple users within the park and communication link data synchronously collected at each recording point of the trajectory;

[0039] For each continuous spatiotemporal trajectory data of a user in the historical dataset, the continuous spatiotemporal trajectory data is mapped to a sequence of regional nodes according to the divided regions. The transfer relationships between all adjacent regional nodes are extracted from the regional node sequence, and each directed transfer relationship from one regional node to the next regional node is defined as a user path.

[0040] For each extracted user path, the communication link data collected when all historical users traversed the corresponding user path are statistically analyzed to form a link quality profile of the user path; the statistical quantities included in the link quality profile include at least the average signal strength, signal strength variance, and connection bandwidth stability index.

[0041] Based on the extracted user paths, the communication behavior of users moving along the user paths is analyzed, wireless access point handover events occurring in the user paths are identified, the total number of historical handover events and the number of successful completions on each user path are counted, and the historical handover success rate of each user path is calculated.

[0042] Using the divided regions as nodes, the extracted user paths as directed edges, and the link quality profile and historical handover success rate as edge attributes, a discretized segmented link quality knowledge graph of user paths is constructed.

[0043] Specifically, the step of querying a pre-constructed user path discretized segmented link quality knowledge graph based on the behavior state prediction results, and determining the pre-handover timing in conjunction with real-time signal strength, includes:

[0044] Analyze the behavioral state prediction results within the second preset time period and extract the user's location and movement speed;

[0045] The coordinates of each user's location are mapped to the nearest spatial node in the user path discretized segmented link quality knowledge graph, forming a predicted node access sequence.

[0046] On the user path discretized segmented link quality knowledge graph, query the link quality profile attributes associated with each edge traversed by the predicted node access sequence, and generate a communication link profile sequence corresponding to the predicted node access sequence; the link quality profile attributes include historical average signal strength, signal stability and link handover success rate.

[0047] Obtain the real-time signal strength and bit error rate measured by the wristband at the current moment, as the current link status observation value;

[0048] The current link state observations and the communication link profile sequence are aligned and fused along the time dimension to form the input feature vector.

[0049] The input feature vector is fed into a pre-trained temporal prediction model; the temporal prediction model is a gated recurrent unit network used to predict the predicted quality degradation point when the signal quality on the current communication link degrades to the switching threshold.

[0050] Based on the predicted quality degradation point, a preset protective delay is shifted forward, and the final calculated time is determined as the pre-switching timing for performing active link switching. The protective delay is used to cover the signaling interaction and execution time of the link switching command.

[0051] Specifically, the step of transmitting multimodal data using a hierarchical fragmentation and sliding window caching strategy during the triggered link switching process, based on the generated link switching instruction frame, includes:

[0052] Upon receiving the link switching instruction frame and initiating the switching process, the multimodal data waiting to be transmitted is prioritized based on the timeliness of data generation and the criticality of the application, generating three priority queues of data to be transmitted: critical, important, and ordinary.

[0053] Based on the expected link quality of the target access point carried in the link switching instruction frame and the assessment of the network conditions before switching, the data transmission unit size applicable to the current and expected link conditions is dynamically calculated, and the data in each priority queue is divided into data fragments of fixed length according to the data transmission unit size.

[0054] Data transmission is initiated before the switching execution time indicated by the link switching instruction frame. Multimodal data is retrieved from the critical priority data queue, encapsulated into data fragments according to the determined data transmission unit size, and sent to the server through the currently maintained connection link.

[0055] On both the wristband and the server, a pair of state-synchronized sliding window buffers are instantiated for each data stream. Before sending each data fragment, the sending end places it in the local sliding window buffer and marks it as sent but not acknowledged. After successfully receiving and verifying the data fragment, the receiving end places it in the corresponding sliding window buffer and marks it as received, and sends back an acknowledgment message containing the sequence number of the received data fragment to the sending end.

[0056] After receiving the acknowledgment from the receiver, the sender compares the status of the local sliding window buffer with the acknowledgment information, identifies the data fragments that have been marked as sent but not acknowledged, but have not been acknowledged as received, determines the corresponding data fragments as lost or erroneous, and initiates retransmission.

[0057] During the period when the network connection is temporarily interrupted due to the link switching action, the sending end stops sending any data fragments to the network, but continues to append newly generated multimodal data to the corresponding data queue to be transmitted according to its priority.

[0058] Once the network connection with the new target access point is re-established and a stable transmission state is reached, transmission recovery operations are performed sequentially.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] 1. This invention proposes a multimodal data fusion and prediction method for smart bracelets using deep learning. By collecting multimodal positioning and motion data in real time, and combining a CNN-LSTM model to dynamically evaluate the confidence of each positioning source and assign fusion weights, a triple decision-making mechanism is triggered to optimize UWB positioning results when GPS fails. Simultaneously, a gyroscope with embedded physical constraints is activated for short-term compensation. This effectively solves the problems of positioning interruption or insufficient accuracy caused by occlusion and multipath interference in complex amusement park environments, achieving accurate and continuous positioning across all scenarios. Furthermore, a dual-branch Transformer model deeply fuses multimodal data to accurately predict user location, movement speed, and other behavioral states within a second preset time period, providing a forward-looking decision-making basis for link switching and breaking through the limitations of traditional short-term, single-dimensional prediction.

[0061] 2. This invention proposes a deep learning-based multimodal data fusion and prediction method for smart bracelets. Relying on a pre-constructed user path discretized segmented link quality knowledge graph, and combining real-time signal strength with a time-series prediction model, it determines the pre-switching timing, achieving proactive prediction and precise planning of link switching, thus avoiding the communication lag and interruption risks caused by traditional passive switching. Furthermore, the hierarchical fragmentation and sliding window caching strategy employed during link switching prioritizes data transmission and dynamically adapts the transmission unit size. Through bidirectional caching and precise retransmission mechanisms, it ensures the integrity and reliability of multimodal data during network fluctuations or interruptions, improving the communication stability and data transmission efficiency of smart bracelets in complex scenarios, and comprehensively optimizing the user experience. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the multimodal data fusion and prediction method for smart bracelets using deep learning, as described in this invention.

[0063] Figure 2 This is a flowchart illustrating the principle of the deep learning-based multimodal data fusion prediction method for smart bracelets in this invention. Detailed Implementation

[0064] Example 1:

[0065] Please see Figure 1 and Figure 2 The present invention provides an embodiment of a multimodal data fusion prediction method for smart bracelets using deep learning, the method comprising S1 to S6, including the following steps:

[0066] S1: Real-time acquisition of multimodal data from the wristband; the multimodal data includes GPS signals, UWB positioning signals deployed in the park, Bluetooth beacon signals, gyroscope data, acceleration data, motion data, and historical communication link data;

[0067] Furthermore, the real-time acquisition of multimodal data from the wristband includes:

[0068] The GPS module integrated into the wristband collects raw GPS observations at the first sampling frequency, including pseudorange, carrier phase and satellite ephemeris information, and generates a GPS signal sequence.

[0069] The wristband integrates a UWB receiver module to receive time difference of arrival signals from at least three UWB positioning base stations deployed in the park, and analyzes the relative distance information between the wristband and each base station to generate a UWB positioning signal sequence.

[0070] The wristband uses an integrated Bluetooth scanning module to periodically scan the broadcast signals of multiple pre-deployed Bluetooth beacons within the park, records the received signal strength indication value and the corresponding beacon ID, and generates a Bluetooth beacon signal sequence.

[0071] The wristband integrates an inertial measurement unit to simultaneously collect angular velocity data from a three-axis gyroscope and relative force data from a three-axis accelerometer. The angular velocity data and relative force data are then subjected to zero bias correction and scale factor compensation to generate pre-processed gyroscope data and acceleration data.

[0072] Using the built-in motion coprocessor of the wristband, based on the pre-processed acceleration data, after attitude calculation and step frequency detection, the output includes motion data including step count, step length estimation, and motion mode;

[0073] The wristband's communication module records the association status, received signal strength, bit error rate, and switching events with various communication access points within the park at different historical moments, constructing a time-stamped historical communication link data sequence.

[0074] S2: Construct a signal quality assessment model based on CNN-LSTM, input GPS signal, UWB positioning signal deployed in the park and Bluetooth beacon signal, output confidence score of each positioning source, dynamically allocate fusion weight of each positioning source based on confidence score, and generate positioning source fusion result;

[0075] Furthermore, based on the confidence score, the fusion weights of each location source are dynamically allocated to generate the location source fusion result, including:

[0076] (1) Normalize the confidence score vector output by the signal quality assessment model and convert each score value into a weight coefficient so that the sum of all weight coefficients is 1;

[0077] (2) According to the weighting coefficients, the real-time positioning results from GPS, UWB and Bluetooth beacons are weighted and summed. The GPS positioning result is obtained by solving the GPS signal sequence, the UWB positioning result is obtained by solving the UWB positioning signal sequence through the trilateration method, and the Bluetooth positioning result is obtained by fingerprint matching or weighted centroid method from the Bluetooth beacon signal sequence. The fingerprint matching or weighted centroid method is the prior art in this field and is not an inventive solution of this application. It will not be described in detail here.

[0078] (3) Output the weighted summed coordinate values ​​as the current location source fusion result.

[0079] It is important to emphasize that traditional location source evaluation methods often rely solely on simple indicators such as signal strength at a single moment, neglecting the temporal correlation of signals and the characteristic differences among different location sources. This leads to one-sided evaluation results, and the weight allocation cannot adapt to the dynamic changes of signals. In contrast, the hybrid model constructed in this method achieves comprehensive and in-depth analysis of location signals through the synergistic effect of parallel one-dimensional convolutional neural network channels and long short-term memory network layers. The number of parallel one-dimensional convolutional neural network channels strictly corresponds to the type of location source, enabling targeted extraction of local spatiotemporal features of GPS signal sequences, UWB location signal sequences, and Bluetooth beacon signal sequences. For example, it captures the temporal fluctuation characteristics of GPS signals, the multi-base station collaborative transmission characteristics of UWB location signals, and the attenuation characteristics of Bluetooth beacon signals. The unique signal attributes of each location source can be accurately captured, avoiding the problem of insufficient adaptation of a single convolutional structure to different signal characteristics. The Long Short-Term Memory (LSTM) network layer further addresses the temporal dependence of signal quality. Signal quality is not isolated; the current signal state is often closely related to the signal change trend in the previous time period. The LSTM layer effectively models this temporal correlation, fusing the local spatiotemporal features extracted from each channel with historical temporal features to form a comprehensive understanding of signal quality. Crucially, the model training process utilizes historical multimodal datasets labeled with true confidence levels. These true confidence levels are calculated based on the error between the localization result and the high-precision reference trajectory at that moment, ensuring the accuracy of the model's evaluation after training and enabling it to output a precise one-dimensional vector score reflecting the current confidence level of each localization source.

[0080] S3: When the signal quality assessment indicates that GPS has entered a failure area, a triple decision mechanism is triggered to obtain the UWB positioning result after weight adjustment, start short-term compensation of the gyroscope, and generate the compensation result for the first preset time period.

[0081] It is important to emphasize that the GPS failure determination mechanism has extremely high accuracy. By monitoring the GPS confidence score in real time and starting a timer, the GPS is only determined to have entered a failure zone when the built-in confidence score is continuously lower than the first preset threshold for a preset time. This effectively avoids misjudgments caused by instantaneous signal fluctuations and provides a reliable premise for triggering the subsequent triple decision mechanism. The design of the triple decision mechanism further optimizes the UWB positioning results in all aspects from three core dimensions: trajectory continuity, signal stability, and environmental adaptability, forming a progressive and complementary optimization system. The first layer of decision-making involves matching historical trajectory features using the Transformer-XL model. Current and recent UWB positioning results are input into the pre-trained model, which then matches the most similar motion pattern segments from the historical trajectory database and outputs a trajectory continuity correction coefficient. This design precisely addresses the potential trajectory breakage problem in UWB positioning after GPS failure. Traditional positioning methods rely solely on current UWB data and cannot utilize the correlation of historical trajectories, easily leading to jumps in positioning results. The Transformer-XL model, however, can uncover the intrinsic connection between historical motion patterns and the current motion state, ensuring the continuity of the UWB positioning trajectory through the trajectory continuity correction coefficient, making the positioning results more consistent with the user's actual movement patterns. The second decision-making step analyzes the instantaneous change rate of the time difference of arrival of each base station signal in the UWB positioning signal to identify signal abrupt changes caused by multipath propagation or obstruction, generating a signal stability discount coefficient. In the complex environment of amusement parks, users may frequently encounter building obstruction, crowd interference, etc., during their movement, causing instantaneous changes in the UWB positioning signal. Traditional methods cannot effectively distinguish between normal signal changes and abrupt signals, easily including abnormal signals in the positioning calculation, leading to a decrease in positioning accuracy. This method, however, captures the characteristics of signal abrupt changes and generates a signal stability discount coefficient, which can effectively reduce the interference of abnormal signals on the positioning results and ensure the stability of the positioning results. The third decision-making step estimates the comprehensive obstruction level based on the attenuation model of Bluetooth beacon signals and maps it to the weight base value. The degree of obstruction in different areas directly affects the propagation quality of UWB positioning signals. The attenuation of Bluetooth beacon signals can intuitively reflect the obstruction level of the current environment. Based on this, the weight base value of UWB positioning can be adjusted to make the weight allocation more in line with the actual environmental conditions and avoid positioning deviations caused by environmental obstruction. These three decision-making mechanisms do not exist in isolation, but rather work synergistically, comprehensively covering the problems that UWB positioning may face after GPS failure from three dimensions: trajectory, signal, and environment. The comprehensiveness and synergy of their optimization cannot be achieved by traditional single-dimensional optimization methods.

[0082] It is also important to emphasize that, based on this, the activated short-term gyroscope compensation strategy further ensures the continuity of positioning. The physically embedded Transformer model used breaks through the limitations of the traditional pure data-driven compensation model. It embeds the basic kinematic equations of the inertial navigation system into the encoder as a physical constraint layer, ensuring that the predicted trajectory strictly conforms to physical laws. Traditional gyroscope compensation methods often rely solely on statistical data, which can easily lead to prediction results that do not conform to the actual motion logic, such as sudden changes in instantaneous velocity or no logical jumps in direction. The existence of the physical constraint layer can force the predicted trajectory to follow the principle of inertial motion, making the predicted trajectory highly consistent with the user's actual motion state. The pre-processed gyroscope and acceleration data within the time window preceding the model input trigger moment can accurately capture the user's motion characteristics and output a sequence of trajectory prediction values ​​within the first preset time period. During the transition phase when GPS fails and UWB positioning is being adjusted, it seamlessly fills the gaps in positioning data, ensuring uninterrupted positioning services. The design of a time window length shorter than the first preset time period ensures both the timeliness of the input data and the ability to make effective predictions for a future period. This synergy between short-term compensation and the triple decision-making mechanism ensures that the positioning results in GPS-failed areas are stable, accurate, and continuous, breaking through the positioning bottleneck in extreme environments. Its positioning guarantee capability is irreplaceable by any single compensation or optimization method.

[0083] S4: Based on the location source fusion result and compensation result, the motion data collected by the wristband and historical communication link data are fused, and multimodal state prediction is performed through a dual-branch Transformer model to output the behavior state prediction result for the second preset time period; the first preset time period is shorter than the second preset time period;

[0084] Furthermore, based on the location source fusion results and compensation results, the motion data collected by the wristband and historical communication link data are fused, including:

[0085] (1) Construct a time-aligned multimodal data tensor that is aligned in the time dimension and contains the following channels:

[0086] Positioning channel: When GPS is not lost, it is the sequence of generated positioning source fusion results; when GPS is lost and compensation is being performed, it is the sequence of generated compensation results.

[0087] Motion status channel: a sequence of motion data consisting of collected and generated step frequency, step length, and motion pattern;

[0088] Link status channel: This refers to the segments of historical communication link data sequences that are relevant to the current time, including signal strength and bit error rate, that are collected and constructed.

[0089] (2) Normalize and fill missing values ​​in the multimodal data tensor.

[0090] It's important to note that traditional state prediction methods often only target single-modal data or a single prediction objective, such as predicting only user location or signal strength, and the prediction time is relatively short, failing to provide sufficient support for decisions requiring advance planning, such as link switching. In contrast, our proposed method's dual-branch Transformer model, through the parallel design of spatial relationship branches and temporal dependency branches, achieves deep fusion and comprehensive feature extraction of multimodal data. The spatial relationship branch, through a self-attention mechanism, can uncover the inherent correlation features between location information, motion state, and communication link state at the same point in time. For example, the correlation between a user's specific location in a certain area and the signal strength of wireless access points in that area; the correlation between a user's motion pattern and positioning accuracy; and the correlation between the current communication link quality and the user's location. These cross-modal spatial correlation features cannot be captured by traditional single-modal analysis, allowing the model to more comprehensively understand the current multimodal data scenario and avoiding feature omissions caused by isolated analysis of different modalities. The time-dependent branch learns the evolution patterns and trends of each modality's data along the time dimension through a causal self-attention mechanism. This includes, for example, the changing trend of communication signal strength with user movement, the temporal fluctuations in user movement speed, and the temporal correlation of positioning results. The causal self-attention mechanism effectively captures the temporal dynamics of the data, avoiding prediction bias caused by ignoring the time dimension, allowing the model to accurately grasp the long-term changing patterns of the data. The first and second high-level feature vectors output from the two branches are concatenated and then fused and reduced through a fully connected layer, achieving a deep integration of spatial and temporal features. The generated feature vectors comprehensively and accurately reflect the core information of the multimodal data. The final output of the behavioral state for the second preset time period covers multiple key dimensions such as user location, movement speed, movement pattern, and expected communication quality level. The movement pattern is further subdivided into specific types such as walking, running, and stationary. The prediction duration far exceeds the first preset time period, providing ample planning time for subsequent link switching decisions. This multi-dimensional, long-term accurate prediction allows the system to anticipate changes in users' future movement trajectories, motion states, and communication needs. For example, it can predict that a user will move from area A to area B within a certain period, with walking as the movement pattern, and that the communication link quality in area B may degrade. Based on these predictions, the system can initiate link switching planning in advance, avoiding passively waiting for signal quality to deteriorate before triggering a switch. The comprehensiveness, accuracy, and long-term nature of its predictions are unmatched by traditional prediction methods. This accurate state prediction provides a crucial basis for determining the pre-switching timing based on knowledge graphs, making proactive link switching possible. It is the core support for the entire method to achieve seamless communication, and its irreplaceable predictive capability directly determines the timeliness and effectiveness of subsequent link switching.

[0091] S5: Based on the behavioral state prediction results, query the pre-built user path discretized segmented link quality knowledge graph, combine it with real-time signal strength to determine the pre-handover timing sequence, and generate link handover instruction frames;

[0092] Furthermore, the link switching instruction frame is a structured data frame that includes at least the following fields: target access point identifier, switching execution timestamp, preloaded frequency band and channel information, context summary information of the current session, and instruction validity credential signed and authenticated by the current access point. At a predetermined time before the pre-switching timeline arrives, the wristband generates and temporarily stores the link switching instruction frame. The target access point identifier is the optimal access point to which the user should switch in the pre-switching timeline based on the user path discretized segmented link quality knowledge graph, and the switching execution timestamp is the determined pre-switching timeline.

[0093] S6: Based on the generated link switching instruction frame, during the triggered link switching process, a hierarchical fragmentation and sliding window caching strategy is used to transmit multimodal data.

[0094] Furthermore, the method also includes a dynamic update step for the user path discretized segmented link quality knowledge graph: after each successful link switch, the actual communication link data monitored on the new link is compared with the predicted profile of the corresponding path in the user path discretized segmented link quality knowledge graph; if there is a continuous and significant difference between the actual data and the historical profile, the update process of the user path discretized segmented link quality knowledge graph is triggered, and the link quality profile attribute of the path is updated using a moving average to achieve the self-evolution of the graph. Here, the moving average is the prior art in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0095] The signal quality assessment model is a hybrid model containing parallel one-dimensional convolutional neural network channels and long short-term memory network layers. The number of parallel one-dimensional convolutional neural network channels corresponds to the type of positioning source and is used to extract local spatiotemporal features of GPS signal sequences, UWB positioning signal sequences, and Bluetooth beacon signal sequences, respectively. The long short-term memory network layer is used to fuse the local spatiotemporal features extracted from each channel and model the temporal dependency of signal quality.

[0096] Furthermore, the construction process of the one-dimensional convolutional neural network channels and long short-term memory network layers includes:

[0097] (1) Define independent input ports for the three types of positioning source data. Each port receives a fixed-length time sequence of the corresponding signal source. For the GPS signal sequence, the preprocessing includes cycle slip detection and gross error removal for pseudorange and carrier phase. For the UWB positioning signal sequence, the preprocessing includes consistency verification of arrival time difference data from different base stations and conversion into relative distance feature vectors. At the same time, the signal strength is standardized. For the Bluetooth beacon signal sequence, the preprocessing includes environmental attenuation compensation for the received signal strength indication value and encoding it into a spatial feature vector according to the beacon physical layout. All sequences are normalized before input so that their amplitude is within the numerical range of efficient learning of the neural network.

[0098] (2) Constructing parallel one-dimensional convolutional neural network feature extraction channels: For three preprocessed time-series signals, three structurally independent but functionally similar one-dimensional convolutional neural network channels are constructed. Each channel is composed of a continuous stack of multiple layers of one-dimensional convolutional modules. The specific operation of each layer of convolutional modules is to use a set of learnable convolutional kernels to perform sliding convolution calculation on the input features along the time dimension. Each convolutional kernel focuses on extracting specific patterns of the signal within a local time window, such as the sudden change trend of signal intensity, periodic fluctuations, or specific distortion shapes. After the convolution calculation, a nonlinear activation function, such as a modified linear unit, is immediately applied to introduce nonlinear transformation capability. Then, max pooling is performed, which gradually compresses the time dimension, retains the most significant feature response, and improves the robustness of the model to micro-time-series shifts. The convolutional layer depth and the number of filters for these three channels are configured independently according to the characteristics of each signal. For example, a deeper convolutional layer can be used for Bluetooth beacon signals with more obvious noise to perform more powerful denoising feature extraction. A flattening layer is connected to the end of each channel to convert the final output multidimensional feature map into a one-dimensional high-level feature vector. Thus, each positioning source signal is encoded into a high-dimensional representation containing its local spatiotemporal pattern.

[0099] (3) The high-level feature vectors output by the three parallel channels are concatenated in the feature dimension to form a joint feature vector that integrates multi-source signal information. This joint feature vector is then fed into the long short-term memory network layer for temporal dependency modeling. The long short-term memory network layer operates through a fine gating mechanism, including an input gate, a forget gate, and an output gate. The input gate controls the proportion of information from the current fused feature vector flowing into the cell state. The forget gate determines which historical information in the cell state at the previous moment needs to be retained or discarded. The output gate calculates the hidden state output at the current moment based on the current cell state. Through this gating mechanism, the network can adaptively memorize long-term historical patterns related to the current signal quality assessment, such as the slow attenuation trend of GPS signals before entering the obstruction area, while ignoring irrelevant short-term fluctuations. The stacking of multiple long short-term memory networks further extracts deeper temporal abstract features. The final hidden state of the last long short-term memory network contains a comprehensive contextual representation after integrating multi-source information and understanding its temporal evolution law.

[0100] (4) The comprehensive context representation output by the Long Short-Term Memory (LSTM) network layer is fed into a fully connected feedforward neural network layer. This fully connected feedforward neural network layer usually consists of two layers. The first layer further combines and abstracts the features through linear transformation and nonlinear activation functions. The second layer outputs a vector with the same dimension as the number of localization sources. Finally, a Softmax activation function is used to normalize the vector into a probability distribution, where the value of each element represents the confidence score of the corresponding localization source at the current moment. The sum of all scores is 1. During the model training phase, a labeled historical dataset is used. The true confidence label of the localization source at each moment in the historical dataset can be generated by back-calculating the localization error obtained by comparing with the high-precision reference trajectory. The backpropagation algorithm is used to optimize the model parameters. Minimizing the difference between the confidence score distribution predicted by the model and the true confidence label distribution typically involves using the cross-entropy loss function, along with a gradient descent optimizer to iteratively adjust all convolutional kernel parameters, long short-term memory network unit parameters, and fully connected layer weights in the model until the model can accurately assess the real-time reliability of each signal source from the input original multimodal localization signal sequence. During training, techniques such as dropout layers are introduced to prevent overfitting, and a validation set is used to monitor the model's generalization performance. The resulting hybrid model can process the original signal sequence end-to-end and output a dynamic confidence score with temporal consistency. The backpropagation algorithm, cross-entropy loss function, and gradient descent optimizer are existing technologies in this field and are not inventive solutions for this application; therefore, they will not be elaborated upon here.

[0101] The construction process of the CNN-LSTM-based signal quality assessment model includes:

[0102] S2.1: The hybrid model is trained using a historical multimodal dataset labeled with the true confidence scores of each location source; the true confidence scores are calculated based on the error between the current location result and the high-precision reference trajectory.

[0103] Furthermore, the specific steps of S2.1 include:

[0104] (1) Deploy a high-precision reference measurement device independent of the positioning system to be evaluated. The high-precision reference measurement device continuously outputs the true position of the carrier in space to form a high-precision reference trajectory. During the data acquisition stage, the various positioning sensors of the wristband are fixed to the reference measurement device on the same carrier, and a unified high-precision clock source is used to stamp the output data of all sensors and the reference measurement device with a synchronization timestamp to complete the spatiotemporal synchronization of multi-sensor data and generate a spatiotemporally aligned multi-source raw data stream.

[0105] (2) For the spatiotemporally aligned multi-source raw data stream, for the data of each positioning source, calculate the geometric position error between the output positioning result and the corresponding high-precision reference trajectory at each time point to obtain the error time series sequence of each positioning source. Then, based on the global statistical characteristics of the error time series sequence of each positioning source, convert the instantaneous error value of each positioning source at each time point into its corresponding true confidence score through a predefined monotonically decreasing mapping function from error value to confidence score. The true confidence score represents the reliability quantification truth value of the positioning source at that time. This conversion is completed for all positioning sources and all time points, that is, a multidimensional true confidence label sequence that is strictly aligned with the multi-source raw data stream is generated.

[0106] (3) With a fixed time window length, the generated spatiotemporally aligned multi-source raw data stream is slicing in time sequence. Each slice is used as a model input sample. At the same time, the label value of the true confidence label sequence at the center point of the corresponding time window is extracted as the supervision target of the input sample. All input-target sample pairs are divided into a training sample subset for model parameter learning, a verification sample subset for training process monitoring, and a test sample subset for final performance evaluation according to a preset ratio.

[0107] (4) Iterative training and optimization of the model: First, all learnable parameters of the hybrid model are initialized. Then, the iterative training loop is entered. In each iteration, a batch of samples is extracted from the training sample subset. The multi-source original data slices in the samples are input into the model. Forward propagation calculation is performed to obtain the confidence score distribution predicted by the model. The difference between the predicted distribution and the sample supervision target, i.e. the true confidence label, is calculated. The difference value is quantified by the loss function. Then, backpropagation calculation is performed to solve the gradient of the loss value with respect to each learnable parameter in the model. The gradient indicates the direction and magnitude of parameter adjustment. Finally, based on the solved gradient, all parameters of the model are updated through the gradient optimization algorithm to make the model prediction closer to the true confidence label. During the training process, the current model performance is evaluated periodically using the validation sample subset. The training strategy is adjusted according to the evaluation results to prevent overfitting. The training loop is terminated when the model performance meets the preset conditions. The forward propagation, backpropagation, and gradient optimization algorithms are existing technologies in this field and are not the inventive solutions of this application. They will not be described in detail here.

[0108] (5) After training is terminated, the optimized hybrid model is tested by forward inference using a subset of test samples that have not participated in the training and validation process. The consistency and accuracy of the model’s predicted confidence score and the true confidence label on the independent test sample subset are evaluated. When the test performance reaches the predetermined application standard, the model is confirmed to have completed training.

[0109] S2.2: Input the real-time collected GPS signal sequence, UWB positioning signal sequence and Bluetooth beacon signal sequence into the trained hybrid model, and output the confidence score of each positioning source at the current time by the hybrid model; the confidence score is a one-dimensional vector, and each element value corresponds to the credibility of a positioning source.

[0110] When signal quality assessment indicates that GPS has entered a failure zone, a triple decision-making mechanism is triggered, including:

[0111] The GPS confidence score output by the signal quality assessment model is monitored in real time. When the GPS confidence score is detected to be lower than the first preset threshold, a timer is started. If the GPS confidence score continues to be lower than the first preset threshold within the preset time, the GPS signal is determined to have entered a failure zone. Based on the determination result of the failure zone, the triple decision mechanism is triggered.

[0112] The first step of the triple decision-making mechanism is to perform historical trajectory feature matching based on the Transformer-XL model, specifically: The UWB positioning results generated from the UWB positioning signal sequence within a given time period are input into a pre-trained Transformer-XL model. The Transformer-XL model matches the motion pattern segment with the highest similarity from the historical trajectory database and outputs trajectory continuity correction coefficients. Time is expressed as a time range including the current period and the most recent period;

[0113] Furthermore, the specific process of the first-level decision-making includes:

[0114] (1) In the offline stage, collect the historical continuous trajectory data of the user in the target area, cut it into trajectory segments of equal length to form a model training set, and use the model training set to train the Transformer-XL model. The training objective is to predict the position of subsequent points based on the sequence of preceding points in the trajectory segments. Through this self-supervised learning, the Transformer-XL model can master the temporal pattern and common forms of trajectory patterns in the area. After training, the pre-trained Transformer-XL model is obtained.

[0115] (2) When the online system triggers the first decision, the coordinate points calculated from the UWB positioning signal sequence in the current and past period are obtained to form a real-time UWB positioning result sequence. The real-time UWB positioning result sequence is smoothed and filtered to suppress noise, and a pre-processed real-time UWB trajectory segment is obtained.

[0116] (3) Input the pre-processed real-time UWB trajectory fragment into the pre-trained Transformer-XL model, and use the encoding capability of the pre-trained Transformer-XL model to transform and condense the input coordinate sequence into a high-dimensional feature vector. This high-dimensional feature vector encodes the deep motion pattern and morphological features of the current trajectory fragment.

[0117] (4) Using the extracted high-dimensional feature vector, calculate the overall similarity between it and the historical trajectory pattern features learned by the Transformer-XL model during the training phase and embedded in its network parameters. The similarity value is converted into a scalar coefficient between 0 and 1 by a preset mapping function, which is the trajectory continuity correction coefficient. The higher the similarity, the larger the output trajectory continuity correction coefficient value, indicating that the current trajectory conforms to the historical common pattern to a higher degree and the continuity is better. The similarity calculation is the prior art in this field and is not an inventive solution of this application, so it will not be elaborated here.

[0118] The second decision of the triple decision mechanism is to identify signal mutation patterns. Specifically, it involves analyzing the instantaneous change rate of the arrival time difference of each base station signal in the current UWB positioning signal sequence. If the instantaneous change rate exceeds the second preset threshold, it is determined that there is a signal mutation caused by multipath or occlusion, and a signal stability discount coefficient is generated.

[0119] The third decision in the triple decision-making mechanism is to adjust the weights according to the occlusion level. Specifically, the comprehensive occlusion level of the current environment is evaluated based on the attenuation model of the received signal strength indication value in the Bluetooth beacon signal sequence, and mapped to the basic weight value.

[0120] Furthermore, based on the attenuation model of the received signal strength indication value in the Bluetooth beacon signal sequence, the overall obstruction level of the current environment is evaluated and mapped to a weighted base value, including:

[0121] (1) During the system deployment phase, the received signal strength of each Bluetooth beacon is measured in advance under an unobstructed environment. The relationship between the signal strength and propagation distance of each Bluetooth beacon under ideal propagation conditions is established, forming a reference propagation model for each Bluetooth beacon, and stored in the model database.

[0122] (2) When the system is running online, the wristband collects the signal strength values ​​from multiple nearby Bluetooth beacons in real time to form a real-time signal observation set. Based on the current location source fusion result or preliminary location information, the reference propagation model of the corresponding beacon is retrieved from the model database. Based on the reference propagation model and the current location information, the theoretical signal strength value of each beacon at the current location is calculated. Then, the signal strength value in the real-time signal observation set is compared with the calculated theoretical signal strength value to obtain the signal observation attenuation of each effective beacon. The observation attenuation is the difference between the actual observation value and the theoretical prediction value.

[0123] (3) Statistical analysis is performed on the observed attenuation of a set of effective beacons, and the overall distribution and consistency characteristics are analyzed. Based on the collective behavior pattern of this set of observed attenuation, the comprehensive obstruction level of the current environment on the propagation of wireless signals is inferred. Specifically, if the observed attenuation is very small, the obstruction level is inferred to be low; if the observed attenuation is generally significantly negative and the values ​​are similar, the obstruction level is inferred to be high; if the observed attenuation differs significantly, the obstruction level is inferred to be medium or heterogeneous. This inference process is completed by preset rules or models, and a quantitative comprehensive obstruction level is output.

[0124] (4) Input the inferred comprehensive occlusion level into a preset monotonically non-increasing mapping function. The monotonically non-increasing mapping function maps the comprehensive occlusion level to a value between 0 and 1. This value is the basic weight value used to adjust the fusion weight of UWB positioning signals. The higher the occlusion level, the smaller the basic weight value of the mapping output, so as to reflect the trust discount of UWB positioning results under adverse propagation environment.

[0125] The calculation process for the weighted UWB positioning result includes:

[0126] S3.1: Calculate the dynamic adjustment weight of the UWB positioning signal based on the trajectory continuity correction coefficient, signal stability discount coefficient, and weight base value output by the triple decision-making mechanism.

[0127] S3.2: The dynamically adjusted weights are weighted together with the UWB confidence score currently output by the signal quality assessment model to generate a fusion weight that is applied to the UWB positioning result at the current time.

[0128] S3.3: Using the aforementioned fusion weights and combining them with the Bluetooth positioning results, generate the adjusted weighted UWB positioning result for the current moment according to the weighted fusion method.

[0129] The calculation process for the dynamic adjustment weight of the UWB positioning signal is as follows:

[0130] S3.1.1: Preset reference weights for UWB positioning sources under normal conditions As a basis for adjustment;

[0131] S3.1.2: Calculate the trajectory continuity correction coefficient. complementary values ;

[0132] S3.1.3: The complementary value Multiplying this by the preset empirical adjustment coefficient k yields the adjustment amount required to adjust the baseline weights due to the risk of trajectory discontinuity. ;

[0133] S3.1.4: Adjust the amount With the benchmark weight Adding them together yields the initial weights after continuous adjustment. ;

[0134] S3.1.5: Adjust the weights after the initial continuity adjustment. With the signal stability discount factor Multiply by the weights to incorporate the instability discount caused by signal abrupt changes, and obtain the stability-corrected weights. ;

[0135] S3.1.6: Adjust the stability-corrected weights With the aforementioned weight base value Multiply to generate dynamically adjusted weights for the UWB positioning signal used in data fusion at the current moment. .

[0136] The short-time compensation for gyroscope activation includes:

[0137] When the triple decision-making mechanism is triggered, a short-term gyroscope compensation process is initiated synchronously. The short-term gyroscope compensation adopts a physically embedded Transformer model. The input of the physically embedded Transformer model is preprocessed gyroscope data and acceleration data within a time window prior to the triggering moment, and the output is a sequence of trajectory prediction values ​​for each moment within a first preset time period starting from the triggering moment. The length of the time window is less than the first preset time period.

[0138] Furthermore, the output process of the trajectory prediction value sequence includes:

[0139] (1) When the triple decision-making mechanism is triggered, the compensation process starts immediately. It extracts a fixed-time-length, pre-processed gyroscope data sequence and acceleration data sequence from a real-time inertial data ring buffer maintained by the wristband, with the trigger time as the reference point. This sequence is used as the historical input fragment of the model.

[0140] (2) The extracted historical input segments are processed by the recursive integration principle of inertial navigation: taking the reliable motion state of the previous moment as the initial condition, the attitude update, coordinate transformation, gravity elimination and integration operation are performed on the data segment step by step to generate a multi-dimensional state sequence containing the calculated position, calculated velocity, calculated attitude and sensor original values. This multi-dimensional state sequence transforms the original sensor readings into a physical feature sequence that conforms to the kinematic laws.

[0141] (3) The constructed physical feature sequence is input into the pre-trained physical embedded Transformer model. The physical embedded Transformer model incorporates kinematic constraints in its internal attention and feedforward mechanism, and can understand the motion pattern in the input sequence. The physical embedded Transformer model encodes and decodes the physical feature sequence and outputs a trajectory prediction sequence representing the motion of the carrier in the first preset time period from the triggering time. This trajectory prediction sequence is a position change sequence in the relative coordinate system calculated based on inertial data and physical laws.

[0142] (4) Obtain the reliable absolute position coordinates at the trigger time, use these reliable absolute position coordinates to perform coordinate transformation on the generated trajectory prediction sequence located in the relative coordinate system, transform it to the global coordinate system, and finally output the trajectory prediction value sequence in absolute coordinates that can be fused with other positioning sources in subsequent steps.

[0143] Furthermore, the construction process of the physically embedded Transformer model includes:

[0144] (1) Based on the encoder-decoder Transformer structure, the physical state variables processed by the model are defined. The physical state variables include, but are not limited to, position, velocity, attitude and raw readings of inertial sensors. These physical state variables together constitute the state vector of the model input and output, ensuring that the data flow has a clear kinematic meaning in the physical dimension. Transformer is the existing technology in this field and is not the inventive solution of this application. It will not be described in detail here.

[0145] (2) In the self-attention calculation layer of the encoder and decoder, a relative motion prior derived from the basic kinematic equation is introduced. Specifically, when calculating the attention weight between state vectors at any two time points, in addition to the standard data-based correlation calculation, an additional physical law-based correlation term is fused. This correlation term evaluates whether the change between the two states conforms to the expected motion trend constrained by kinematic variables such as velocity and acceleration. Thus, when the model understands the sequence, its attention mechanism naturally tends to physically coherent context segments. The kinematic equation is the prior art in this field and is not an inventive solution of this application. It will not be elaborated here.

[0146] (3) In the feedforward neural network module after the self-attention calculation layer, its function is structured so that its output is related to the rate of change of the physical state. Specifically, the output of the feedforward network is decomposed into the update amount of the state vector. This update amount should match the state change at the next moment obtained by discrete time integration in a physical sense. Through this design, the feedforward network implicitly follows the dynamic law described by the kinematic differential equation in the feature transformation process, ensuring the smoothness and physical rationality of the state evolution inside the model.

[0147] (4) The historical trajectory data generated by the high-precision motion capture system that strictly conforms to the physical laws is used as the training set. During training, the model takes the historical state sequence as input and the state sequence at the preset time as the supervision target. In the optimization process, the loss function not only includes the error term between the predicted position and the actual position, but also adds a physical consistency constraint term. This constraint term is used to punish the violation of the kinematic laws in the predicted sequence, such as sudden position change and discontinuous velocity. The model parameters are jointly optimized by backpropagation and gradient descent algorithm so that the trained model can achieve high-precision trajectory prediction and compliance with the underlying physical laws at the same time. Finally, a well-trained physical embedded Transformer model is obtained.

[0148] The physically embedded Transformer model embeds the basic kinematic equations of the inertial navigation system into the encoder as a physical constraint layer to ensure that the predicted trajectory conforms to physical laws in the feature space.

[0149] Furthermore, the trajectory prediction value sequence is fused and calibrated with the weighted UWB positioning results to generate a compensation result, specifically including:

[0150] (1) Obtain the sequence of trajectory prediction values ​​output by the physically embedded Transformer model within the first preset time period;

[0151] (2) Obtain the UWB positioning result sequence after weight adjustment generated within the same time period;

[0152] (3) Design an adaptive Kalman filter, using the trajectory prediction value as the prior information for state prediction, and the UWB positioning result after weight adjustment as the observation value; the observation noise covariance matrix of the adaptive Kalman filter is dynamically adjusted according to the UWB confidence score output by the signal quality assessment model and the signal stability discount coefficient. The Kalman filter is the prior art in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0153] (4) Run the adaptive Kalman filter to fuse the trajectory prediction value sequence with the UWB positioning result sequence after weight adjustment, and output a smooth and continuous trajectory sequence as the compensation result of the short-time compensation of the gyroscope.

[0154] The process of predicting multimodal states using a two-branch Transformer model and outputting behavioral state prediction results for a second preset time period includes:

[0155] S4.1: The dual-branch Transformer model consists of a spatial relation branch and a temporal dependency branch connected in parallel. The spatial relation branch receives preprocessed multimodal data and learns the intrinsic correlation features between positioning information, motion state, and communication link state at the same time point through a self-attention mechanism, outputting a first high-level feature vector. The temporal dependency branch receives the same multimodal data and learns the evolution rules and trends of each modality data along the time dimension through a causal self-attention mechanism, outputting a second high-level feature vector.

[0156] Furthermore, the specific steps of S4.1 include:

[0157] (1) Receive a multimodal data stream containing positioning results, motion data and link status, and precisely align all modal data in the time dimension to form a multimodal input tensor with fixed time length and synchronized modal features;

[0158] (2) Extract the instantaneous feature vector corresponding to each independent moment from the generated multimodal input tensor, add position encoding to each instantaneous feature vector, and input it into an encoder stack based on self-attention mechanism. The encoder allows features of different modes at the same time point, such as positioning coordinates, step frequency, and signal strength, to interact globally and calculate the correlation between them, thereby extracting the spatial correlation features that characterize the cross-modal integrated state at that moment, and outputting the first high-level feature vector. The encoder based on self-attention mechanism is the prior art in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0159] (3) The generated multimodal input tensor is segmented along the feature dimension to form independent time series of each modality. The independent time series of each modality is added with position encoding and then input into an encoder stack based on causal self-attention mechanism. The encoder based on causal self-attention mechanism restricts each time point to only focus on its own and historical information and prohibits access to information after the current time, thereby capturing the evolution law and long-term dependency relationship of each modality state along the time axis, extracting the temporal dependency features, and outputting the second high-level feature vector.

[0160] S4.2: The first and second high-level feature vectors output by the spatial relationship branch and the temporal dependency branch are concatenated, and then fused and dimensionality reduced through a fully connected layer to finally output the behavioral state prediction results for each moment in the second preset time period; the behavioral state prediction results include at least the user's location, movement speed, movement mode and expected communication quality level; the movement mode includes at least walking, running and standing still.

[0161] Furthermore, the first high-level feature vector and the second high-level feature vector are concatenated time-by-time along the feature dimension to form a joint feature sequence. Subsequently, the joint feature sequence is processed through a feedforward fusion network, which performs nonlinear transformation, information fusion and dimensionality reduction on the concatenated features, and finally outputs a high-order fusion feature that characterizes the basis of behavioral state prediction within the second preset time period.

[0162] Example 2:

[0163] The method for constructing the user path discretized segmented link quality knowledge graph in this embodiment includes:

[0164] S5.1: The map of the target amusement park is divided into multiple logically connected areas based on the physical layout and functional attributes, and each area is associated with at least one wireless communication access point.

[0165] S5.2: Obtain historical dataset; the historical dataset contains continuous spatiotemporal trajectory data of multiple users in the park and communication link data synchronously collected at each recording point of the trajectory;

[0166] S5.3: For each continuous spatiotemporal trajectory data of a user in the historical dataset, the continuous spatiotemporal trajectory data is mapped to a sequence of regional nodes according to the divided regions. The transfer relationships between all adjacent regional nodes are extracted from the regional node sequence, and each directed transfer relationship from one regional node to the next regional node is defined as a user path.

[0167] Furthermore, the specific steps in S5.3 include:

[0168] (1) For each user's continuous spatiotemporal trajectory data, based on the pre-divided set of regions with clear geometric boundaries, each original trajectory point with geographic coordinates in the user's continuous spatiotemporal trajectory data is mapped one by one to the specific region where its spatial location is located, thereby generating a preliminary mapping sequence of the same length as the original trajectory point sequence and composed of region identifiers.

[0169] (2) Perform time series analysis on the generated preliminary mapping sequence, merge the identical region identifiers that appear consecutively into a time-continuous access event, and then apply time-based filtering rules to clean the merged event; set a minimum effective dwell time threshold for all events. If the duration of any access event is lower than this minimum effective dwell time threshold, the event is removed from the sequence, and these short-term access records are merged into the boundary noise of adjacent dominant events or directly discarded, thereby obtaining a pure region access sequence consisting of effective access events and free from short-term oscillations and positioning drift noise;

[0170] (3) For any two adjacent different valid access events in the pure region access sequence, the region identifier of the previous event is used as the starting point and the region identifier of the next event is used as the ending point. A unique directed transition relationship from the starting region to the ending region is defined and recorded. This is a user path.

[0171] S5.4: For each extracted user path, statistical analysis is performed on the communication link data collected when all historical users traversed the corresponding user path to form a link quality profile of the user path; the statistical parameters included in the link quality profile include at least the average signal strength, signal strength variance, and connection bandwidth stability index.

[0172] Furthermore, the specific steps in S5.4 include:

[0173] (1) For each extracted user path, based on the precise start and end timestamps of entering the starting area and leaving the ending area recorded in the corresponding original trajectory segment, all communication measurement data points within the same time window are extracted from the historical communication link measurement data stream, so as to establish a communication data subsequence that strictly corresponds to each user's passage through the user path.

[0174] (2) For all the communication data subsequences of different access instances belonging to the same user path, data cleaning is performed to remove outliers caused by equipment or transient interference, and standardization is performed to eliminate systematic biases introduced by differences in user equipment hardware and network operators, so that the communication data from different access instances are comparable on a unified scale, and a normalized set of communication data subsequences is obtained.

[0175] (3) Define a unified logical segmentation model for each user path that is proportional to time. Divide all normalized communication data subsequences of the path into the same number of continuous stages in the time dimension according to the logical segmentation model. Then, for each logical stage, calculate the average signal strength, variance of signal strength, and bandwidth stability index of all passing instances in that stage. Combine these statistics of all stages to form the link quality profile of the user path.

[0176] (4) Store the link quality profile of each user path in the knowledge graph as the attribute of the corresponding path. When new passage instances and related communication data are generated, the existing profile is updated incrementally. The new data statistics are integrated with the historical profile data in the form of moving average or exponential weighted average to dynamically reflect the long-term changes in the path communication environment.

[0177] S5.5: Based on the extracted user path, analyze the communication behavior of the user when moving along the user path, identify the wireless access point handover events that occur in the user path, count the total number of historical handover events and the number of successful completions on each user path, and calculate the historical handover success rate of each user path.

[0178] Furthermore, the specific steps of S5.5 include:

[0179] (1) For communication data subsequences that are strictly time-aligned with each user path, the system sequentially scans each data reporting point. When it detects that the current service access point identifier is different from the previous reporting point identifier, an observation period is started. If the new service access point identifier remains stable and dominant in the subsequent observation period, a service point change is confirmed, and the start time, source access point identifier, and target access point identifier of the change are recorded and marked as a switching event candidate.

[0180] (2) For each marked switching event candidate, the system verifies the signaling logs and application layer performance data collected at the same time. If the signaling logs record the standard successful switching process in full, and the application layer performance data does not show service interruption or quality degradation exceeding the preset threshold within the time window corresponding to the event, then the switching event candidate is determined to be a successful switching event; otherwise, it is determined to be a failed switching attempt.

[0181] (3) For each user path, summarize all its historical passage instances, count the total number of times it is judged as a candidate for switching event in all instances, and take it as the total number of attempts to switch for that path. At the same time, count the total number of switching events judged as successful, and take it as the total number of successful switching for that path. Divide the total number of successful switching by the total number of attempts to switch to calculate the historical switching success rate of that path.

[0182] (4) The historical switching success rate of each user path is calculated and used as a key performance attribute of the path, and is integrated and stored in its corresponding link quality profile.

[0183] S5.6: Using the divided regions as nodes, the extracted user paths as directed edges, and the link quality profile and historical handover success rate as edge attributes, complete the construction of a discretized segmented link quality knowledge graph of user paths.

[0184] The process of querying a pre-constructed discretized segmented link quality knowledge graph of the user path based on the behavior state prediction results, and determining the pre-handover timing sequence in conjunction with real-time signal strength, includes:

[0185] A1: Analyze the behavioral state prediction results within the second preset time period and extract the user's location and movement speed;

[0186] A2: Map the coordinates of each user's location to the nearest spatial node in the user path discretized segmented link quality knowledge graph, forming a predicted node access sequence;

[0187] A3: Query the link quality profile attributes associated with each edge traversed by the predicted node access sequence along the user path discretized segmented link quality knowledge graph, and generate a communication link profile sequence corresponding to the predicted node access sequence; the link quality profile attributes include historical average signal strength, signal stability and link handover success rate.

[0188] A4: Obtain the real-time signal strength and bit error rate measured by the wristband at the current moment, as the current link status observation value;

[0189] A5: Align and fuse the current link state observations with the communication link profile sequence along the time dimension to form an input feature vector;

[0190] A6: Feed the input feature vector into a pre-trained time-series prediction model; the time-series prediction model is a gated recurrent unit network, used to predict the predicted quality degradation point when the signal quality on the current communication link degrades to the switching threshold. The time-series prediction model is existing technology in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0191] A7: Based on the obtained predicted quality degradation point, a preset protective delay is shifted forward, and the final calculated time is determined as the pre-switching timing for performing active link switching. The protective delay is used to cover the signaling interaction and execution time of the link switching command.

[0192] The step of transmitting multimodal data using a hierarchical fragmentation and sliding window caching strategy during the triggered link switching process, based on the generated link switching instruction frame, includes:

[0193] S6.1: When the link switching instruction frame is received and the switching process is started, the multimodal data waiting to be transmitted is classified into three priority levels—critical, important, and ordinary—based on the timeliness of data generation and the criticality of application.

[0194] The key priority data queue includes real-time positioning compensation results, real-time behavior state prediction results, and handover control signaling; the important priority data queue includes unprocessed raw positioning signals and preprocessed inertial data; and the ordinary priority data queue includes historical trajectory archive data and backups of raw sensor data.

[0195] S6.2: Based on the expected link quality of the target access point carried in the link switching instruction frame and the assessment of the network conditions before switching, dynamically calculate the data transmission unit size applicable to the current and expected link conditions, and divide the data in each priority queue into data fragments of fixed length according to the data transmission unit size;

[0196] Furthermore, the specific steps in S6.2 include:

[0197] (1) Before data sharding, the quality of the current service link and the target access point link is evaluated. The quality of the current link is evaluated based on the real-time measurement data at the moment before the handover. The quality of the target link is estimated based on the historical link quality profile and behavior state prediction results carried in the link handover instruction frame and obtained from the user path discretized segmented link quality knowledge graph. This forms the expectation of the link state after the handover, i.e. the expected quality of the target link.

[0198] (2) Input the current link quality and the expected quality of the target link into the preset decision model. The decision model outputs a set of initial transmission unit sizes applicable to data of different priorities based on the absolute quality level, relative superiority and inferiority relationship and expected stability of the two links. The core decision logic is: the worse or more unstable the expected quality of the target link, the smaller the calculated initial transmission unit size, so as to improve transmission reliability; the better and more stable the expected quality of the target link, the larger the calculated initial transmission unit size, so as to improve transmission efficiency.

[0199] (3) Based on the calculated initial transmission unit size applicable to each priority data, the corresponding data queue is divided. Starting from the critical priority data queue to be transmitted, the data stream is divided into fixed-length segments according to the determined unit size, and necessary control header information is added to each segment to generate independent data segment packets for transmission.

[0200] (4) After the data fragmentation starts transmission, continuously monitor its transmission performance feedback, including acknowledgment delay and retransmission rate. Based on this real-time feedback, dynamically adjust the transmission unit size of subsequent data fragments. If the feedback indicates that the link quality is better than expected, gradually increase the transmission unit size; if the feedback indicates that the link quality is worse than expected, decrease the transmission unit size to achieve adaptation to the actual link conditions.

[0201] S6.3: Before the switching execution time indicated by the link switching instruction frame, start high-priority data transmission, retrieve multimodal data from the critical-priority pending data queue, encapsulate it into data fragments according to the determined data transmission unit size, and send it to the server through the currently connected link;

[0202] S6.4: On the wristband and server, a pair of state-synchronized sliding window buffers are instantiated for each data stream. Before sending each data fragment, the sending end places it in the local sliding window buffer and marks it as sent but not acknowledged. After successfully receiving and verifying the data fragment, the receiving end places it in the corresponding sliding window buffer and marks it as received, and sends back an acknowledgment message containing the sequence number of the received data fragment to the sending end.

[0203] Furthermore, on both the wristband and server sides, a pair of state-synchronized sliding window caches are instantiated for each data stream, including:

[0204] (1) At the start of data transmission, based on a unified data stream identifier, a sliding window cache instance is created on both the wristband and the server for the data stream. Each instance contains a buffer space for storing data fragments, a bitmap for recording the status of each sequence number fragment, and control variables for recording the window start sequence number, current window size, and maximum window size. When created, the start sequence number and initial window parameters of the instances on both sides are set to the same value to complete the static initialization of the cache.

[0205] (2) After static initialization is completed, a data transmission connection is established. During this process, the sending end and the receiving end exchange and confirm the final sliding window parameters, including the starting sequence number, the initial window size and the maximum sequence number space, to ensure that the two parties reach a strict consensus on the initial state of the window before the logical transmission begins, so as to establish a synchronization benchmark for subsequent collaborative sliding.

[0206] (3) When the sending end sends data, it assigns a continuous sequence number to each data fragment and marks it as sent but not acknowledged. After the receiving end successfully receives the fragment, it marks it as received. The receiving end sends cumulative acknowledgment information to the sending end periodically or when conditions are triggered. The cumulative acknowledgment information indicates that the receiving end has successfully received all fragments with any specific sequence number Y. After receiving this acknowledgment, the sending end marks all fragments with sequence numbers less than or equal to Y as acknowledged and updates the local window start sequence number to Y+1, releasing the cache resources of the acknowledged fragments, thereby completing one forward sliding of the sending end window;

[0207] (4) The receiver continuously monitors the fragment status at the starting sequence number of its window. Once the fragment at the starting position is marked as received, the receiver immediately delivers the fragment data to the upper layer application and advances the starting sequence number of the local window by one position. This process is repeated to ensure that the data is delivered in order. At the same time, the receiver uses its window state bitmap to detect sequence number gaps and reports non-contiguous high sequence number fragments that have been successfully received to the sender through a selective acknowledgment mechanism to assist the sender in performing fast retransmission, thereby maintaining the continuous synchronization of the dual-end window state when dealing with packet loss and out-of-order delivery.

[0208] S6.5: After receiving the acknowledgment information from the receiver, the sender compares the status of the local sliding window buffer with the acknowledgment information, identifies the data fragments that have been marked as sent but not acknowledged but not received, determines the corresponding data as lost or erroneous, and initiates retransmission only for these fragments;

[0209] S6.6: During the period when the network connection is temporarily interrupted due to the execution of the link switching action, the sending end stops sending any data fragments to the network, but continues to append newly generated multimodal data to the corresponding data queue to be transmitted according to its priority.

[0210] S6.7: After the network connection with the new target access point is re-established and a stable transmission state is reached, perform the following transmission recovery operations in sequence:

[0211] The sending end first checks the local sliding window buffer, and retrieves and retransmits all high-priority data fragments with a status of sent but not acknowledged.

[0212] After the sending end records the last data fragment that has been successfully confirmed consecutively in its sliding window buffer, it retrieves data from the data queue corresponding to each priority level in the order of critical, important, and normal, and then fragments and transmits the data according to the determined transmission unit size.

[0213] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A multimodal data fusion and prediction method for smart bracelets using deep learning, characterized in that: include: The wristband collects multimodal data in real time; the multimodal data includes GPS signals, UWB positioning signals deployed in the park, Bluetooth beacon signals, gyroscope data, acceleration data, motion data, and historical communication link data. A signal quality assessment model based on CNN-LSTM is constructed. The inputs are GPS signal, UWB positioning signal deployed in the park and Bluetooth beacon signal. The output is the confidence score of each positioning source. The fusion weight of each positioning source is dynamically allocated based on the confidence score to generate the positioning source fusion result. When the signal quality assessment indicates that GPS has entered a failure area, a triple decision mechanism is triggered to obtain the UWB positioning result after weight adjustment, start short-term gyroscope compensation, and generate the compensation result for the first preset time period. Based on the location source fusion results and compensation results, the motion data collected by the wristband and historical communication link data are fused, and multimodal state prediction is performed through a dual-branch Transformer model to output the behavior state prediction results for the second preset time period. The first preset time period is shorter than the second preset time period; Based on the behavioral state prediction results, the pre-constructed user path discretized segmented link quality knowledge graph is queried, and the pre-handover timing sequence is determined in combination with the real-time signal strength to generate a link handover instruction frame. Based on the generated link switching instruction frame, during the triggered link switching process, a hierarchical fragmentation and sliding window caching strategy is adopted to transmit multimodal data.

2. The method for multimodal data fusion and prediction of smart bracelets using deep learning as described in claim 1, characterized in that, The signal quality assessment model is a hybrid model containing parallel one-dimensional convolutional neural network channels and long short-term memory network layers. The number of parallel one-dimensional convolutional neural network channels corresponds to the type of positioning source and is used to extract local spatiotemporal features of GPS signal sequences, UWB positioning signal sequences, and Bluetooth beacon signal sequences, respectively. The long short-term memory network layer is used to fuse the local spatiotemporal features extracted from each channel and model the temporal dependency of signal quality. The construction process of the CNN-LSTM-based signal quality assessment model includes: training the hybrid model using a historical multimodal dataset labeled with the true confidence scores of each location source; inputting real-time acquired GPS signal sequences, UWB positioning signal sequences, and Bluetooth beacon signal sequences into the trained hybrid model; and outputting a confidence score for each location source at the current moment from the hybrid model. The true confidence scores are calculated based on the error between the current positioning result and the reference trajectory. The confidence scores are one-dimensional vectors, with each element corresponding to the credibility of a location source.

3. The method for multimodal data fusion and prediction of smart bracelets using deep learning as described in claim 2, characterized in that, When signal quality assessment indicates that GPS has entered a failure zone, a triple decision-making mechanism is triggered, including: The GPS confidence score output by the signal quality assessment model is monitored in real time. When the GPS confidence score is detected to be lower than the first preset threshold, a timer is started. If the GPS confidence score continues to be lower than the first preset threshold within the preset time, the GPS signal is determined to have entered a failure zone. Based on the determination result of the failure zone, the triple decision mechanism is triggered. The first step of the triple decision-making mechanism is to perform historical trajectory feature matching based on the Transformer-XL model, specifically: The UWB positioning results generated from the UWB positioning signal sequence within a time period are input into the pre-trained Transformer-XL model. The Transformer-XL model matches the motion pattern segment with the highest similarity from the historical trajectory database and outputs the trajectory continuity correction coefficient. The second decision of the triple decision mechanism is to identify signal mutation patterns. Specifically, it involves analyzing the instantaneous change rate of the arrival time difference of each base station signal in the current UWB positioning signal sequence. If the instantaneous change rate exceeds the second preset threshold, it is determined that there is a signal mutation caused by multipath or occlusion, and a signal stability discount coefficient is generated. The third decision in the triple decision-making mechanism is to adjust the weights according to the occlusion level. Specifically, the comprehensive occlusion level of the current environment is evaluated based on the attenuation model of the received signal strength indication value in the Bluetooth beacon signal sequence, and mapped to the basic weight value.

4. The method for multimodal data fusion and prediction of smart bracelets using deep learning as described in claim 3, characterized in that, The calculation process for the weighted UWB positioning result includes: Based on the trajectory continuity correction coefficient, signal stability discount coefficient, and weight base value output by the triple decision-making mechanism, the dynamic adjustment weight of the UWB positioning signal is calculated. The dynamically adjusted weights are weighted together with the UWB confidence score currently output by the signal quality assessment model to generate fusion weights applicable to the UWB positioning results at the current time. Using the aforementioned fusion weights and combining them with the Bluetooth positioning results, the weighted UWB positioning result for the current moment is generated according to the weighted fusion method.

5. The method for multimodal data fusion and prediction of smart bracelets using deep learning as described in claim 4, characterized in that, The calculation process for the dynamic adjustment weight of the UWB positioning signal is as follows: Preset reference weights for UWB positioning sources under normal conditions As a basis for adjustment; Calculate the trajectory continuity correction coefficient complementary values ; The complementary value Multiplying this by the preset empirical adjustment coefficient k yields the adjustment amount used to adjust the baseline weights due to trajectory discontinuity risk. ; The adjustment amount With the benchmark weight Adding them together yields the initial weights after continuous adjustment. ; The weights after the initial continuity adjustment With the signal stability discount factor Multiply to obtain the stability-corrected weights. ; The stability-corrected weights With the aforementioned weight base value Multiply to generate dynamically adjusted weights for the UWB positioning signal used in data fusion at the current moment. .

6. The method for multimodal data fusion and prediction of smart bracelets using deep learning as described in claim 5, characterized in that, The short-time compensation for gyroscope activation includes: When the triple decision-making mechanism is triggered, a short-term gyroscope compensation process is initiated synchronously. The short-term gyroscope compensation adopts a physically embedded Transformer model. The input of the physically embedded Transformer model is preprocessed gyroscope data and acceleration data within a time window prior to the triggering moment, and the output is a sequence of trajectory prediction values ​​for each moment within a first preset time period starting from the triggering moment. The length of the time window is less than the first preset time period. The physically embedded Transformer model embeds the basic kinematic equations of the inertial navigation system into the encoder as a physical constraint layer.

7. The method for multimodal data fusion and prediction of smart bracelets using deep learning as described in claim 6, characterized in that, The process of predicting multimodal states using a two-branch Transformer model and outputting behavioral state prediction results for a second preset time period includes: The dual-branch Transformer model consists of a spatial relation branch and a temporal dependency branch connected in parallel. The spatial relation branch receives preprocessed multimodal data and learns the intrinsic correlation features between positioning information, motion state, and communication link state at the same time point through a self-attention mechanism, outputting a first high-level feature vector. The temporal dependency branch receives the same multimodal data and learns the evolution rules and trends of each modality data along the time dimension through a causal self-attention mechanism, outputting a second high-level feature vector. The first and second high-level feature vectors output by the spatial relationship branch and the temporal dependency branch are concatenated, and then fused and dimensionality reduced through a fully connected layer to finally output the behavioral state prediction result for each moment in the second preset time period; the behavioral state prediction result includes at least the user's location, movement speed, movement mode and expected communication quality level; the movement mode includes at least walking, running and standing still.

8. The method for multimodal data fusion and prediction of smart bracelets using deep learning as described in claim 7, characterized in that, The method for constructing the user path discretized segmented link quality knowledge graph includes: The map of the target amusement park is pre-divided into multiple logically connected areas based on its physical layout and functional attributes, with each area associated with at least one wireless communication access point. Obtain historical datasets; the historical datasets contain continuous spatiotemporal trajectory data of multiple users within the park and communication link data synchronously collected at each recording point of the trajectory; For each continuous spatiotemporal trajectory data of a user in the historical dataset, the continuous spatiotemporal trajectory data is mapped to a sequence of regional nodes according to the divided regions. The transfer relationships between all adjacent regional nodes are extracted from the regional node sequence, and each directed transfer relationship from one regional node to the next regional node is defined as a user path. For each extracted user path, the communication link data collected when all historical users traversed the corresponding user path are statistically analyzed to form a link quality profile of the user path; the statistical quantities included in the link quality profile include at least the average signal strength, signal strength variance, and connection bandwidth stability index. Based on the extracted user paths, the communication behavior of users moving along the user paths is analyzed, wireless access point handover events occurring in the user paths are identified, the total number of historical handover events and the number of successful completions on each user path are counted, and the historical handover success rate of each user path is calculated. Using the divided regions as nodes, the extracted user paths as directed edges, and the link quality profile and historical handover success rate as edge attributes, a discretized segmented link quality knowledge graph of user paths is constructed.

9. The method for multimodal data fusion and prediction of smart bracelets using deep learning as described in claim 8, characterized in that, The process of querying a pre-constructed discretized segmented link quality knowledge graph of the user path based on the behavior state prediction results, and determining the pre-handover timing sequence in conjunction with real-time signal strength, includes: Analyze the behavioral state prediction results within the second preset time period and extract the user's location and movement speed; The coordinates of each user's location are mapped to the nearest spatial node in the user path discretized segmented link quality knowledge graph, forming a predicted node access sequence. On the user path discretized segmented link quality knowledge graph, query the link quality profile attributes associated with each edge traversed by the predicted node access sequence, and generate a communication link profile sequence corresponding to the predicted node access sequence; the link quality profile attributes include historical average signal strength, signal stability and link handover success rate. Obtain the real-time signal strength and bit error rate measured by the wristband at the current moment, as the current link status observation value; The current link state observations and the communication link profile sequence are aligned and fused along the time dimension to form the input feature vector. The input feature vector is fed into a pre-trained temporal prediction model; the temporal prediction model is a gated recurrent unit network used to predict the predicted quality degradation point when the signal quality on the current communication link degrades to the switching threshold. Based on the predicted quality degradation point, a preset protective delay is shifted forward, and the final calculated time is determined as the pre-switching timing for performing active link switching. The protective delay is used to cover the signaling interaction and execution time of the link switching command.

10. The method for multimodal data fusion and prediction of smart bracelets using deep learning as described in claim 9, characterized in that, The step of transmitting multimodal data using a hierarchical fragmentation and sliding window caching strategy during the triggered link switching process, based on the generated link switching instruction frame, includes: Upon receiving the link switching instruction frame and initiating the switching process, the multimodal data waiting to be transmitted is prioritized based on the timeliness of data generation and the criticality of the application, generating three priority queues of data to be transmitted: critical, important, and ordinary. Based on the expected link quality of the target access point carried in the link switching instruction frame and the assessment of the network conditions before switching, the data transmission unit size applicable to the current and expected link conditions is dynamically calculated, and the data in each priority queue is divided into data fragments of fixed length according to the data transmission unit size. Data transmission is initiated before the switching execution time indicated by the link switching instruction frame. Multimodal data is retrieved from the critical priority data queue, encapsulated into data fragments according to the determined data transmission unit size, and sent to the server through the currently maintained connection link. On both the wristband and the server, a pair of state-synchronized sliding window buffers are instantiated for each data stream. Before sending each data fragment, the sending end places it in the local sliding window buffer and marks it as sent but not acknowledged. After successfully receiving and verifying the data fragment, the receiving end places it in the corresponding sliding window buffer and marks it as received, and sends back an acknowledgment message containing the sequence number of the received data fragment to the sending end. After receiving the acknowledgment from the receiver, the sender compares the status of the local sliding window buffer with the acknowledgment information, identifies the data fragments that have been marked as sent but not acknowledged, but have not been acknowledged as received, determines the corresponding data fragments as lost or erroneous, and initiates retransmission. During the period when the network connection is temporarily interrupted due to the link switching action, the sending end stops sending any data fragments to the network, but continues to append newly generated multimodal data to the corresponding data queue to be transmitted according to its priority. Once the network connection with the new target access point is re-established and a stable transmission state is reached, transmission recovery operations are performed sequentially.

Images