Earphone loss association analysis method and system in user movement scenario combined with deep learning model

Abstract

This invention provides a method and system for analyzing headphone loss in user mobility scenarios using a deep learning model. Relating to the field of deep learning technology, the method acquires wireless connection state transition sequences and scene snapshot transition sequences during user movement. It performs wearing state jump detection and signal strength attenuation segment extraction on the connection state sequences to generate a set of candidate time intervals for headphone loss. Based on these candidate intervals, it extracts synchronization subsequences from the scene snapshot sequences to construct a spatiotemporal correlation graph structure. A deep learning model is then invoked to encode node features and transfer edge features from this graph structure. Cross-type node information aggregation generates a feature vector explaining the loss event. Based on this, the loss time, triggering scene semantics, motion modality transition sequences, and environmental interaction features are deduced. A spatiotemporal tracing record is generated and matched to generate a retrieval operation instruction sequence, which is then pushed to the user terminal. This invention achieves multi-dimensional joint attribution and intelligent retrieval guidance for loss events.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a method and system for analyzing the correlation between headphone loss and user mobility scenarios using a deep learning model. Background Technology

[0002] With the widespread use of wireless earphones in daily life, earphone loss has become a frequent problem for users. Current technologies for finding lost earphones mainly rely on two approaches: Bluetooth signal strength indication and last known location (WHM) positioning. The Bluetooth signal strength indication method roughly determines whether the earphone is lost by detecting the degree of signal attenuation between the earphone and the terminal device. However, this method cannot distinguish whether the signal attenuation is caused by the user actively moving away from the device or the earphone actually being removed from wear, resulting in a high false positive rate. The WHM positioning method records the geographical coordinates of the earphone when it disconnects and marks the location on a map for users to trace back. However, this method completely ignores the dynamic changes in the user's environment before and after the loss, such as the impact of scene changes like moving from indoors to outdoors or from a stationary state to walking on the ground, leading to a lack of targeted guidance for the retrieval process. Furthermore, existing technologies have not yet developed a solution for jointly modeling and analyzing the temporal changes in connection status and the semantic changes in the scene during user movement. Nor have they developed a technical path for using deep learning models to aggregate cross-type node information from multi-source heterogeneous temporal data to deduce the loss triggering conditions and environmental interaction characteristics after loss. Therefore, existing solutions have significant shortcomings in terms of accurate determination of the loss time, semantic attribution of the loss scene, and intelligent generation of retrieval operation instructions. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for analyzing the correlation between headphone loss and user mobility scenarios using a deep learning model, the method comprising:

[0004] The system acquires a wireless connection state transition sequence and a scene snapshot transition sequence during the user's mobility state transition process. The wireless connection state transition sequence includes multiple connection state recording units arranged in time sequence. Each connection state recording unit includes a wear status marker and a signal strength marker. The scene snapshot transition sequence includes multiple scene snapshot units arranged in time sequence. Each scene snapshot unit includes a location semantic marker and a motion mode marker.

[0005] The wireless connection state transition sequence is processed by wearing state jump detection and signal strength attenuation segment extraction. Based on the detected jump events from the wearing state to the wearing unwearing mark and the continuous attenuation trend of the signal strength mark before and after the jump event, a set of candidate time intervals for headphone loss is generated.

[0006] Based on the set of candidate time intervals for headphone loss, a scene snapshot subsequence synchronized with each candidate time interval for headphone loss is extracted from the scene snapshot transition sequence. The scene snapshot subsequence is then time-aligned with the signal strength attenuation segment within the corresponding interval to construct a spatiotemporal association graph structure with connection state recording unit and scene snapshot unit as nodes and temporal adjacency edge and synchronization association edge as connection relationships.

[0007] The pre-built deep learning model is invoked to encode node features and pass edge features to the spatiotemporal graph structure, generating hidden state representations for each node. Cross-type node information is aggregated based on the hidden states of the connection state record unit node and the hidden states of the scene snapshot unit node to generate a missing event explanation feature vector.

[0008] Based on the lost event interpretation feature vector, the time of loss of the target earphone, the semantic mark of the lost triggering scene, the motion modal transition sequence before loss, and the set of environmental interaction features after loss are derived. A spatiotemporal source tracing record of earphone loss is generated. Based on the semantic mark of the lost triggering scene and the set of environmental interaction features after loss, an earphone retrieval operation instruction sequence is generated and pushed to the user terminal.

[0009] Furthermore, embodiments of the present invention also provide a headphone loss correlation analysis system in user mobility scenarios that combines a deep learning model, comprising:

[0010] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the above-described headphone loss correlation analysis method in a user mobility scenario incorporating a deep learning model by executing the machine-executable instructions.

[0011] Based on the above, a spatiotemporal relational graph structure is constructed with connection state recording units and scene snapshot units as dual nodes and temporal adjacency edges and synchronous association edges as composite connections. This structure unifies the temporal changes in wireless connection state and the synchronous changes in scene semantics during user movement into the same graph representation space. On this basis, a pre-built deep learning model is invoked to perform node feature encoding and edge feature transfer processing on the spatiotemporal relational graph structure. This enables the wearing state jump information and signal strength attenuation trend carried by the connection state recording unit nodes to achieve cross-type node information aggregation with the position semantic information and motion modality information carried by the scene snapshot unit nodes under the message passing mechanism of the graph neural network. The resulting loss event explanation feature vector simultaneously integrates the connection state evolution law in the temporal dimension and the scene semantic change law in the spatial dimension, and has the ability to represent the headphone loss event in a multi-dimensional joint attribution. Furthermore, based on the feature vector of the lost event, the semantic markers of the lost earphone time, the lost triggering scene, the motion modal transition sequence before loss, and the set of environmental interaction features after loss are derived. This not only accurately locates the time node of the loss, but also traces back the user's motion state transition path before loss and identifies the key scene semantic conditions that led to the loss. At the same time, it captures the interaction features of the environment in which the earphone is located after loss, thus providing a complete decision basis for generating a scene-adaptive earphone retrieval operation instruction sequence. This significantly improves the accuracy and executability of the retrieval operation in the earphone loss scenario. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the execution flow of the headphone loss correlation analysis method in a user mobile scenario combined with a deep learning model, provided in an embodiment of the present invention.

[0013] Figure 2 This is a logical diagram illustrating the headphone loss correlation analysis method in a user mobile scenario that combines a deep learning model, as provided in an embodiment of the present invention.

[0014] Figure 3 This is a schematic diagram of the loss diagnosis and spatiotemporal tracing interface provided in an embodiment of the present invention.

[0015] Figure 4 This is a schematic diagram of the retrieval guidance and multi-source collaborative interface provided in the embodiments of the present invention. Detailed Implementation

[0016] Figure 1 This is a flowchart illustrating a method for analyzing headphone loss in a user mobile scenario that incorporates a deep learning model, provided by an embodiment of the present invention. A detailed description follows.

[0017] Step S110: Obtain the wireless connection state transition sequence and scene snapshot transition sequence of the user during the mobile state transition process. The wireless connection state transition sequence includes multiple connection state recording units arranged in time sequence. Each connection state recording unit includes a wearing status marker and a signal strength marker. The scene snapshot transition sequence includes multiple scene snapshot units arranged in time sequence. Each scene snapshot unit includes a location semantic marker and a motion mode marker.

[0018] The smartphone operating system continuously acquires wireless connection status data with the target earphone through the Bluetooth protocol stack interface, with a sampling period denoted as T, in seconds. At the i-th sampling time ti, the Bluetooth protocol stack returns a connection status data structure, which includes a wearing status field and a signal strength field. The wearing status field is a Boolean variable, taking a value of 0 or 1, where 1 indicates the earphone is in a wearing-while-wearing state, and 0 indicates the earphone is in a wearing-out state. The signal strength field is an integer variable, taking a value representing the received signal strength indication, in decibels and milliwatts. The wearing status field value at the i-th sampling time is denoted as Pi, and the signal strength field value is denoted as Si, together forming the i-th connection status recording unit Ci. All connection status recording units are arranged in chronological order to form a wireless connection status transition sequence, denoted as sequence C, with a sequence length of N.

[0019] The smartphone simultaneously acquires scene snapshot data through location services and motion sensors. Location services return the current location semantic label of the device. This label is derived in real-time from GPS coordinates and wireless network fingerprints by a pre-built location classification model on the device, and the output is an encoded value from a pre-defined set of location semantic labels. Motion sensors, including accelerometers and gyroscopes, process the motion sensor data using a pre-built motion recognition algorithm on the device, outputting the current motion modality label. This motion modality label is an encoded value from a pre-defined set of motion modality labels. The location semantic label value at the i-th sampling time is denoted as Li, and the motion modality label value is denoted as Mi, together forming the i-th scene snapshot unit Ei. All scene snapshot units are arranged chronologically to form a scene snapshot transition sequence, denoted as sequence E, with a sequence length of N.

[0020] Sequence C and sequence E are time-aligned by the same sampling time ti and are stored together in a circular data buffer in the smartphone's memory. The size of the circular data buffer is set to accommodate all record units covering the preset backtracking time.

[0021] Step S120: Perform wearing state jump detection and signal strength attenuation segment extraction processing on the wireless connection state transition sequence. Based on the detected wearing state jump events from the wearing hold mark to the wearing unwear mark and the continuous attenuation trend of the signal strength marks before and after the jump events, generate a set of candidate time intervals for headphone loss.

[0022] Step S121: Load the connection state record units in the wireless connection state transition sequence into the circular processing window in timestamp order, and obtain the current connection state record unit and the immediately preceding connection state record unit from the circular processing window in each processing cycle.

[0023] The circular processing window is a first-in, first-out (FIFO) queue data structure of length 2. The head of the queue stores the connection status record unit of the current processing cycle, and the tail of the queue stores the connection status record unit of the previous processing cycle. Let the index of the current processing cycle be k, where k ranges from 2 to N. In the k-th processing cycle, the connection status record unit Ck with index k is retrieved from the circular processing window as the current unit, and the connection status record unit Ck-1 with index k-1 is retrieved as the previous unit.

[0024] Step S122: Compare the wearing status mark of the current connection status recording unit with the wearing status mark of the previous connection status recording unit. When the wearing status mark is detected to change from the wearing hold mark to the wearing disengage mark, record the timestamp of the current connection status recording unit as the time of the transition event and generate a transition event trigger signal.

[0025] Extract the wearing status flag field value Pk from Ck, and extract the wearing status flag field value Pk-1 from Ck-1. Perform a conditional judgment: if Pk-1 is 1 and Pk is 0, a wearing status change event is determined to have occurred, and the change event index is recorded as k. The time of the change event is taken as the sampling time tk corresponding to Ck. Generate a change event trigger signal. The trigger signal data structure contains two fields: an integer value of the change index and a floating-point value of the change time. Write the trigger signal to the change event queue.

[0026] Step S123: In response to the transition event trigger signal, extract from the wireless connection state transition sequence a forward signal segment centered on the time of the transition event, extending forward to the connection state recording unit where the signal strength mark first changes to the full signal mark, and a backward signal segment extending backward to the time when the signal strength mark recovers to the full signal mark and the signal strength mark of the subsequent three consecutive connection state recording units remains at the full signal mark.

[0027] The trigger signals are retrieved one by one from the transition event queue. For each trigger signal, its transition index is recorded as k. The forward signal segment extraction method is to backtrack unit by unit starting from index k-1 in the direction of decreasing index, extract the signal strength marker field value of each backtracked unit, and compare it with the predefined full signal strength marker value Smax. Smax is the maximum received signal strength indication value of the wireless connection. When backtracking to the first unit where the signal strength marker field value is equal to Smax, the index of that unit is recorded as a, and backtracking stops. The forward signal segment coverage index range is from a to k. The backward signal segment extraction method is to search unit by unit starting from index k+1 in the direction of increasing index, extract the signal strength marker field value of each search unit, and check the signal strength marker field values ​​of the next three consecutive units. When searching to the first unit where the signal strength marker field value is equal to Smax and the next three units where the signal strength marker field values ​​are all equal to Smax, the index of that unit is recorded as b, and the search stops. The backward signal segment coverage index range is from k+1 to b.

[0028] Step S124: The forward signal segment and the backward signal segment are spliced ​​together end to end to obtain a complete signal attenuation segment. The signal strength markers of each connection state recording unit in the complete signal attenuation segment are extracted and arranged in chronological order to form a signal attenuation marker chain. By comparing the signal attenuation marker chain with a variety of pre-stored typical loss attenuation templates, the best matching loss attenuation template corresponding to the signal attenuation marker chain is determined. Based on the attenuation duration characteristics of the best matching loss attenuation template, the forward offset step number and the backward offset step number of the interval are determined.

[0029] The forward and backward signal segments are concatenated end-to-end in index order. The resulting complete segment covers the index range from a to b, and the number of connection state record units contained within the segment is denoted as Lseg. Each connection state record unit within the complete segment is traversed, and the signal strength marker field value of each unit is extracted in ascending index order to form a signal attenuation marker chain, denoted as D, where the length of D is equal to Lseg.

[0030] A set of pre-stored typical loss attenuation templates is denoted as set T, containing U templates. Each template is a sequence of signal strength variation patterns of a preset length, and the u-th template is denoted as Tu. A dynamic time warping algorithm is used to calculate the similarity distance between the signal attenuation marker chain D and each template Tu. The dynamic time warping algorithm constructs a cumulative cost matrix with rows and columns of lengths D and Tu, respectively. Matrix elements are the minimum value among the Euclidean distance of the corresponding signal strength values ​​plus the minimum value of the three adjacent matrix elements to the left, top, and top left. The value of the bottom-right element of the cumulative cost matrix is ​​the warped distance between D and Tu, denoted as Du. The U templates are traversed, and the template corresponding to the minimum warped distance is selected as the best-matching loss attenuation template, denoted as Tm. Each template pre-stores the corresponding forward offset step number Fm and backward offset step number Bm, and these two parameter values ​​are read from the best-matching loss attenuation template Tm.

[0031] Step S125: Based on the time of occurrence of the jump event, shift the sampling interval of the forward offset step of the interval by several connection state recording units to obtain the start timestamp of the candidate time interval for headphone loss, shift the sampling interval of the backward offset step of the interval by several connection state recording units to obtain the end timestamp of the candidate time interval for headphone loss, combine the start timestamp and end timestamp of the interval to generate a candidate time interval for headphone loss, and associate and store the candidate time interval for headphone loss with the corresponding time of occurrence of the jump event.

[0032] The start timestamp of the candidate time interval for headphone loss is denoted as Tstart, and the end timestamp is denoted as Tend. Tstart = tk - Fm × T, Tend = tk + Bm × T. Create a data structure for the candidate time interval for headphone loss, containing a start timestamp field, an end timestamp field, and an associated transition timestamp field. Write Tstart, Tend, and tk into their respective fields, and append this data structure object to the candidate interval list.

[0033] Step S126: Traverse each transition event trigger signal detected in the wireless connection state transition sequence to generate multiple candidate earphone loss time intervals. Perform time axis overlap detection on the multiple candidate earphone loss time intervals. When the time ranges of any two candidate earphone loss time intervals overlap or the interval interval is less than the preset merging interval, merge the corresponding two candidate earphone loss time intervals into a continuous candidate earphone loss time interval. Arrange all the candidate earphone loss time intervals after merging in chronological order to generate a set of candidate earphone loss time intervals.

[0034] The candidate interval list contains multiple candidate time intervals for headphone loss generated by various transition events. The candidate interval list is sorted in ascending order by the interval's start timestamp. The preset merging interval is denoted as Tgap, in seconds. An empty set is created as the merged interval set. The sorted candidate interval list is traversed. For the current interval Icurr and the next interval Inext, the merging condition is checked: if the end timestamp of Icurr plus Tgap is greater than or equal to the start timestamp of Inext, the merging condition is met, and the end timestamp of Icurr is updated to the larger of the end timestamps of Icurr and Inext, and comparison with subsequent intervals continues. If the merging condition is not met, Icurr is added to the merged interval set, and Inext is set as the new Icurr, continuing the traversal. The merged interval set obtained after the traversal is completed is the headphone loss candidate time interval set, denoted as set Ω, which contains multiple non-overlapping time intervals.

[0035] Step S130: Based on the set of candidate time intervals for headphone loss, extract the scene snapshot subsequence synchronized with each candidate time interval for headphone loss from the scene snapshot transition sequence, and time-align the scene snapshot subsequence with the signal strength attenuation segment in the corresponding interval to construct a spatiotemporal association graph structure with connection state recording unit and scene snapshot unit as nodes and temporal adjacency edge and synchronization association edge as connection relationship.

[0036] Step S131: For each candidate time interval of headphone loss in the set of candidate time intervals of headphone loss, based on the start timestamp and end timestamp of the interval, select all scene snapshot units whose timestamps fall within the candidate time interval of the headphone loss from the scene snapshot transition sequence, and form a scene snapshot subsequence corresponding to the candidate time interval of the headphone loss.

[0037] Iterate through each interval in the set Ω. For the v-th interval Iv, its start timestamp is denoted as Tv1 and its end timestamp as Tv2. From the scene snapshot transition sequence E, select all scene snapshot units Ei whose sampling times ti satisfy Tv1 <= ti and ti <= Tv2. Arrange the selected scene snapshot units in ascending order of sampling times to form the scene snapshot subsequence corresponding to interval Iv, denoted as Ev.

[0038] Step S132: Extract connection state recording unit segments from the wireless connection state transition sequence that are completely consistent with the time range of the candidate time interval of the earphone loss, as signal strength attenuation segments. Each connection state recording unit in the signal strength attenuation segment and the scene snapshot unit in the scene snapshot subsequence are associated with each other by collecting timestamps.

[0039] From the wireless connection state transition sequence C, select all connection state recording units Ci for sampling times ti that satisfy Tv1 <= ti and ti <= Tv2. Arrange the selected connection state recording units in ascending order of sampling time to form the signal strength attenuation segment corresponding to interval Iv, denoted as Cv. Since sequences C and E are acquired simultaneously at the same sampling time, Cv and Ev have equal lengths and each unit corresponds one-to-one.

[0040] Step S133: Create a connection state node for each connection state recording unit in the signal strength attenuation segment, and map the wearing status mark and signal strength mark contained in the connection state recording unit to the initial feature vector of the connection state node.

[0041] For the j-th connection state record unit Cj in Cv, create a connection state node, denoted as uj. Extract the wearing status flag field value Pj and the signal strength flag field value Sj from Cj. The initial feature vector of the connection state node is a two-dimensional numerical vector, denoted as vector A. The first component of A is the floating-point value of Pj, and the second component of A is the floating-point value of Sj after minimum-to-maximum normalization. The normalization process divides Sj by a predefined maximum signal strength value.

[0042] Step S134: Create a scene snapshot node for each scene snapshot unit in the scene snapshot subsequence, and map the position semantic tags and motion modality tags contained in the scene snapshot unit to the initial feature vector of the scene snapshot node.

[0043] For the j-th scene snapshot unit Ej in Ev, create a scene snapshot node, denoted as vj. Extract the location semantic tag field value Lj and motion modality tag field value Mj from Ej. The initial feature vector of the scene snapshot node is a two-dimensional numerical vector, denoted as vector B. The first component of B is the integer value of Lj, which is one-hot encoded and then mapped through the embedding matrix to obtain the embedding vector. The second component of B is the integer value of Mj, which is one-hot encoded and then mapped through the embedding matrix to obtain the embedding vector. The two embedding vectors are concatenated to form the initial feature vector.

[0044] Step S135: Establish directed temporal adjacency edges between adjacent connected state nodes according to the order of timestamps, and establish directed temporal adjacency edges between adjacent scene snapshot nodes. Each temporal adjacency edge carries the time difference attribute between the two nodes.

[0045] For any two adjacent nodes uj and uj+1 in the connected state node sequence, establish a directed edge, with the direction of the edge pointing from uj to uj+1, and the edge type is marked as a temporal adjacency edge. The attribute field of this edge stores the difference between the two nodes at the corresponding sampling time, and the difference is equal to T. Similarly, establish a directed temporal adjacency edge for any two adjacent nodes vj and vj+1 in the scene snapshot node sequence.

[0046] Step S136: For connection state nodes and scene snapshot nodes with the same acquisition timestamp, establish an undirected synchronous association edge, wherein the undirected synchronous association edge carries a node type identifier and an edge type identifier.

[0047] For the connected state node uj and the scene snapshot node vj with the same index j, establish an undirected edge connecting the two nodes. The edge type is marked as a synchronous association edge. The attribute field of this edge stores the edge type identifier, which takes the value of the synchronous association type code, and stores the node type identifiers of the two ends.

[0048] Step S137: For connection status nodes and scene snapshot nodes whose acquisition timestamps differ by one time unit, if their corresponding wearing status marker is a wearing-off marker and their motion mode marker is a stationary mode marker, then establish an undirected cross-time synchronization association edge and mark the undirected cross-time synchronization association edge with a cross-time association marker.

[0049] For the connection state node uj at index j and the scene snapshot node vj+1 at index j+1, extract the wearing state flag Pj of the connection state record unit corresponding to uj, and extract the motion modality flag Mj+1 of the scene snapshot unit corresponding to vj+1. If Pj is 0 and Mj+1 is a preset static modality encoding value, then an undirected edge is established between uj and vj+1. The edge type is marked as a time-synchronous association edge, and the attribute field of the edge stores the time-synchronous association flag.

[0050] Step S138: Combine the connection state node, scene snapshot node, directed temporal adjacency edge, undirected synchronization association edge, and undirected cross-temporal synchronization association edge to form a single-interval spatiotemporal association subgraph corresponding to the candidate time interval of the lost headphone. Cascade all single-interval spatiotemporal association subgraphs on the time boundary through temporal adjacency edges between scene snapshot nodes that cross the interval to generate a spatiotemporal association graph structure that covers the entire user's mobile state transition process.

[0051] For each interval Iv, a corresponding single-interval spatiotemporal relational subgraph Gv is created. The node set of Gv contains all connected state nodes and scene snapshot nodes within the interval, and the edge set of Gv contains all directed temporal adjacency edges, undirected synchronous relational edges, and undirected cross-temporal synchronous relational edges within the interval. For adjacent intervals Iv and Iv+1, a directed temporal adjacency edge is established between the scene snapshot node at the end of Iv and the scene snapshot node at the beginning of Iv+1. All subgraphs are connected by temporal adjacency edges that cross intervals to form a complete spatiotemporal relational graph structure G. The total number of nodes in graph G is the sum of the number of nodes in all subgraphs, and the total number of edges is the sum of the number of edges within each subgraph and the number of edges that cross intervals.

[0052] Step S140: Call the pre-built deep learning model to encode node features and pass edge features to the spatiotemporal relation graph structure, generate the hidden state representation of each node, and aggregate cross-type node information based on the hidden state of the connection state record unit node and the hidden state of the scene snapshot unit node to generate the missing event explanation feature vector.

[0053] Step S141: Input the spatiotemporal relational graph structure into the pre-built deep learning model. The pre-built deep learning model includes a node embedding transformation layer, a multi-layer graph message passing layer, and a cross-type attention readout layer. Through the node embedding transformation layer, linear mapping and non-linear activation processing are applied to the initial feature vector of each connected state node to generate a connected state embedding vector. Similarly, linear mapping and non-linear activation processing are applied to the initial feature vector of each scene snapshot node to generate a scene snapshot embedding vector, thus obtaining an initial node embedding set.

[0054] The node embedding transformation layer maintains two trainable weight matrices, denoted as Wc and We, and two bias vectors, denoted as bc and be. For a connected state node uj, its initial feature vector is vector A, and the connected state embedding vector is calculated as vector Zc, Zc = σ(Wc × A + bc). For a scene snapshot node vj, its initial feature vector is vector B, and the scene snapshot embedding vector is calculated as vector Ze, Ze = σ(We × B + be). σ is the sigmoid function, expressed as σ(x) = 1 / (1 + exp(-x)). The embedding vectors of all nodes constitute the initial node embedding set.

[0055] Step S142: Input the initial node embedding set into the first layer of the graph message passing layer. For each directed temporal adjacency edge, calculate the temporal message vector using the source node embedding vector, the target node embedding vector, and the edge attribute features. Transmit the temporal message vector from the source node to the target node along the directed temporal adjacency edge. For each undirected synchronous association edge and undirected cross-temporal synchronous association edge, calculate the synchronization message vector using the embedding vectors of the two endpoints and the edge type identifier. Transmit the synchronization message vector bidirectionally along the edge to the two endpoints.

[0056] The graph message passing layer maintains a trainable weight matrix for message computation. For each directed temporally adjacent edge, the source node embedding vector is denoted as `src`, the target node embedding vector as `dst`, and the edge's temporal difference attribute as `dt`. The temporal message vector is calculated by concatenating `src`, `dst`, and `dt` into a long vector, multiplying it by the trainable weight matrix `Wt`, adding a bias vector `bt`, and finally passing it through a Sigmoid activation function. The calculated temporal message vector is propagated along the edge direction from the source node to the target node. For each undirected synchronous association edge or cross-temporal synchronous association edge, the embedding vectors of the two endpoints are denoted as `z1` and `z2`, and the edge type identifier is encoded as `et`. The synchronous message vector is calculated by concatenating `z1`, `z2`, and `et`, multiplying it by the trainable weight matrix `Ws`, adding a bias vector `bs`, and then activating it through a Sigmoid function. The synchronous message vector is propagated bidirectionally along the edge, meaning both endpoints receive the message.

[0057] Step S143: At each connection state node and each scene snapshot node, aggregate all received time-series message vectors and synchronization message vectors, and combine the aggregated message vectors with the node's current embedded vector through a gating fusion mechanism to update and generate the node hidden state output by the graph message passing layer.

[0058] At each connected state node and each scene snapshot node, all received temporal message vectors and synchronization message vectors are aggregated. A gating fusion mechanism is used to combine the aggregated message vector with the node's current embedding vector, updating and generating the node's hidden state output by the graph's message passing layer. For any node n in graph G, its current embedding vector is denoted as zc, and the set of message vectors received by this node through all incident edges is denoted as M. The aggregated message vector is the element-wise summation of all message vectors in M, denoted as magg. The gating fusion mechanism calculates an update gate vector g, g = σ(Wg × [magg; zc] + bg), where σ is the Sigmoid function, [;] indicates vector concatenation, Wg is the weight matrix, and bg is the bias vector. The node's updated hidden state vector new = g⊙zc + (1-g)⊙tanh(Wn × magg + bn), where ⊙ represents element-wise multiplication, tanh is the hyperbolic tangent activation function, Wn is the weight matrix, and bn is the bias vector.

[0059] Step S144: Use the hidden state of the node output by the previous graph message passing layer as the input of the next graph message passing layer, and repeatedly perform message calculation, message passing and node state update operations. After a preset number of graph message passing processes, the final hidden state representation of each node is obtained.

[0060] The total number of layers in the graph message passing layer is denoted as L, where L is a preset integer value. Layer 1 takes the initial node embedding output in step S141 as input and outputs the hidden state of the nodes in Layer 1. Layer l takes the hidden state of the nodes output in Layer l-1 as input and outputs the hidden state of the nodes in Layer l, where l ranges from 2 to L. The hidden state of the nodes output in Layer l is the final hidden state representation of each node.

[0061] Step S145: Extract all connected state nodes and scene snapshot nodes within the range of the single interval spatiotemporal correlation subgraph corresponding to each candidate time interval of headphone loss from the spatiotemporal correlation graph structure to form an interval node subset. Input the final hidden state representation of all connected state nodes in the interval node subset into the first attention branch of the cross-type attention readout layer to generate a connected state-side aggregated feature vector. At the same time, input the final hidden state representation of all scene snapshot nodes into the second attention branch of the cross-type attention readout layer to generate a scene snapshot-side aggregated feature vector.

[0062] The cross-type attention readout layer comprises two independent attention branches. The first attention branch maintains a trainable attention query vector, denoted as qc. For the final hidden state vector hj of each connected state node uj in the subset of nodes within interval Iv, the attention score αj is calculated as the dot product of qc and hj, multiplied by an exponentiation operation, and then divided by the sum of the exponentiation results of the dot products of all connected state nodes. The connected state-side aggregated feature vector Rc is equal to the sum of each hj multiplied by αj. The second attention branch maintains another trainable attention query vector qe. The same attention-weighted aggregation operation is performed on the scene snapshot nodes to obtain the scene snapshot-side aggregated feature vector Re.

[0063] Step S146: The aggregated feature vector on the connection state side and the aggregated feature vector on the scene snapshot side are concatenated and input into the fully connected fusion layer. Through the multi-layer nonlinear transformation of the fully connected fusion layer, the loss event explanation feature vector corresponding to the candidate time interval of the lost headphone is output. The loss event explanation feature vectors corresponding to the candidate time intervals of the lost headphone are summarized to form a set of loss event explanation feature vectors.

[0064] Concatenate vectors Rc and Re along their respective vector dimensions to obtain the concatenated vector Rcat. The fully connected fusion layer consists of two fully connected layers. The first layer maps Rcat to an intermediate dimension vector, using the sigmoid activation function. The second layer maps the intermediate dimension vector to the lost event explanation feature vector in the output dimension, also using the sigmoid activation function. The lost event explanation feature vector corresponding to interval Iv is denoted as Fv. By traversing all intervals in the set Ω, the Fv corresponding to each interval is summed to form the lost event explanation feature vector set.

[0065] Step S150: Based on the lost event interpretation feature vector, derive the target earphone loss time, the semantic marker of the loss triggering scene, the motion modal transition sequence before loss, and the set of environmental interaction features after loss, generate a spatiotemporal tracing record of earphone loss, and generate an earphone retrieval operation instruction sequence based on the semantic marker of the loss triggering scene and the set of environmental interaction features after loss, and push the earphone retrieval operation instruction sequence to the user terminal.

[0066] Step S151: Input the lost event interpretation feature vector into the lost time regressor. The lost time regressor is composed of stacked fully connected layers. By transforming layer by layer, it outputs a fine-tuned time offset relative to the start timestamp of the candidate time interval of the lost earphone. Then, it adds the start timestamp of the interval to the fine-tuned time offset to obtain the lost time of the target earphone.

[0067] The lost-time regressor consists of two fully connected layers. The first layer maps the lost event explanation feature vector Fv to an intermediate dimension vector, using the Sigmoid activation function. The second layer maps the intermediate dimension vector to a scalar output, which is the fine-tuning time offset ΔTfine. The target earphone lost time Tloss = Tstart + ΔTfine.

[0068] Step S152: Input the lost event explanation feature vector into the scene semantic classifier. The scene semantic classifier uses a fully connected mapping layer and a scene semantic classification output layer to map the lost event explanation feature vector to a preset scene semantic label set, and selects the scene semantic label whose output probability meets the preset probability threshold as the lost event trigger scene semantic label.

[0069] The fully connected mapping layer of the scene semantic classifier maps Fv to an intermediate dimension vector. The scene semantic classification output layer maps the intermediate dimension vector to the score value of each label in the preset scene semantic label set. The score value is converted into a probability distribution through a softmax function. The label with the highest probability value is selected from the probability distribution. If the probability value is greater than or equal to a preset probability threshold, the label is used as the missing scene semantic label Lloss. If it is less than the threshold, the label with the second highest probability value is selected for further evaluation.

[0070] Step S153: Input the lost event interpretation feature vector into the motion sequence decoder. The motion sequence decoder uses the lost event interpretation feature vector as the initial hidden state, generates motion mode labels in each decoding step, and uses the motion mode labels output in the decoding step as the input of the next decoding step, gradually generating multiple consecutive motion mode labels, which are arranged in the order of generation to form a motion mode transition sequence before loss.

[0071] The motion sequence decoder uses a gated loop unit as its decoding core. Fv is used as the initial hidden state input to the gated loop unit. In the first decoding step, the gated loop unit generates the first motion mode label M1 based on the initial hidden state and the start marker symbol, and simultaneously updates the hidden state. In the r-th decoding step, the gated loop unit generates the current motion mode label Mr based on the motion mode label Mr-1 output from the previous decoding step and the previous hidden state. Decoding continues for a preset number of steps K before stopping, resulting in a motion mode label sequence, which is the motion mode transition sequence before loss.

[0072] Step S154: Input the lost event interpretation feature vector into the environment interaction reconstruction decoder. The environment interaction reconstruction decoder initializes the environment state vector according to the lost event interpretation feature vector, and gradually predicts the combination of object category identifiers and interaction action identifiers in the user's surrounding environment after the loss, generating an environment interaction feature set containing multiple sets of object category identifier and interaction action identifier pairs.

[0073] The environmental interaction reconstruction decoder has the same structure as the motion sequence decoder, using a gated recurrent unit as the decoding core, but its output dimension is the joint dimension of object category and interaction action. Using Fv as the initial hidden state, it progressively decodes to generate pairs of object category identifiers Oq and interaction action identifiers Aq. Decoding continues for a preset number of steps Q before stopping. The resulting Q pairs constitute the environmental interaction feature set.

[0074] Step S155: Combine the target earphone loss time, the semantic marker of the loss triggering scene, the motion mode transition sequence before loss, and the set of environmental interaction features after loss into a spatiotemporal source tracing record of earphone loss.

[0075] The spatiotemporal tracing record of lost headphones is a key-value pair data structure, which includes four keys: loss time key, scene marker key, motion sequence key, and interaction set key, corresponding to Tloss, Lloss, motion modality label sequence, and environmental interaction feature set, respectively.

[0076] Step S156: Input the semantic tag of the lost trigger scene into the preset retrieval strategy knowledge base. The preset retrieval strategy knowledge base stores the mapping from scene semantic tags to retrieval operation instruction templates. Extract the corresponding retrieval operation instruction template by tag matching. The retrieval operation instruction template includes an audio prompt instruction template and a visual guidance instruction template.

[0077] The default retrieval strategy knowledge base uses a key-value storage structure, where the key is a scene semantic tag and the value is an instruction template object. An exact match query is performed using Lloss as the key, returning the corresponding instruction template object. This object contains an audio prompt instruction template field and a visual guidance instruction template field. The audio prompt instruction template includes the file path and playback parameters of the prompt audio, while the visual guidance instruction template includes the interface display layout and the prompt text template.

[0078] Step S157: Extract object category identifier sequences from the set of environmental interaction features after loss, determine the environmental object paths that the user may encounter after losing the headphones, and retrieve the corresponding spatial coordinates from the pre-stored environmental spatial layout map according to the environmental object paths to generate a sequence of object coordinates along the path. Using the sequence of object coordinates along the path as navigation intermediate points, and combining it with the current position coordinates of the user's terminal, call the path planning engine to generate navigation path instructions that start from the current position and pass through the navigation intermediate points in sequence. Combine the navigation path instructions with the audio prompt instruction template and the visual guidance instruction template to generate an initial retrieval operation instruction sequence.

[0079] The system extracts object category identifier sequences from the environmental interaction feature set in decoding order. For each object category identifier in the sequence, the system retrieves the object's spatial coordinates (including longitude and latitude) from a pre-stored environmental spatial layout map. The user terminal's current location coordinates are obtained from the Global Positioning System (GPS) module. Using the current location as the starting point and the sequence of object coordinates as intermediate waypoints, the path planning engine generates navigation path instructions using a shortest path algorithm. The navigation path instructions, audio prompt instruction templates, and visual guidance instruction templates are combined to form the initial retrieval operation instruction sequence.

[0080] Step S158: Add semantic tags of the target earphone loss time and loss triggering scenario to each instruction entry in the initial retrieval operation instruction sequence. This is used by the terminal to present loss time retrospective information and scenario prompt information. The initial retrieval operation instruction sequence with the added information is encapsulated into a message data packet and pushed to the user terminal through the wireless communication network.

[0081] Iterate through each instruction entry in the initial retrieval operation instruction sequence, adding a loss time field and a scene marker field to the entry structure, and writing the values ​​of Tloss and Lloss respectively. Serialize the modified instruction sequence into a message data packet, encoded using a Lightweight Data Interchange format. Send the message data packet to the application running on the user terminal via a cellular mobile communication network or wireless LAN. Upon receiving the message data packet, the user terminal parses it and presents a retrieval operation guidance interface.

[0082] Combination Figure 2The content demonstrates the complete processing flow from multi-source data acquisition to retrieval command push. The first part shows the process of acquiring the wireless connection state transition sequence and the scene snapshot transition sequence. The wireless connection state transition sequence is presented as a time-series waveform, with each sampling point containing a wearing state marker and a signal strength marker. The scene snapshot transition sequence is presented as a video frame sequence diagram, with each frame containing a location semantic marker and a motion modality marker. The second part demonstrates the generation process of the candidate time interval set for headphone loss. By detecting the jump event from the wearing state marker to the wearing-off marker, and the continuous attenuation trend of the signal strength marker before and after the jump event, the candidate time interval set is extracted. The third part demonstrates the construction process of the spatiotemporal correlation graph structure. Based on the candidate time intervals, corresponding sub-sequences are extracted from the scene snapshot transition sequence. The extracted scene snapshot sub-sequences are time-aligned with the signal strength attenuation segments within the corresponding intervals to construct a spatiotemporal correlation graph structure with connection state recording units and scene snapshot units as nodes, and temporal adjacency edges and synchronization correlation edges as connections. Part Four demonstrates the generation process of the lost event explanation feature vector. The spatiotemporal graph structure is input into a pre-built deep learning model. Hidden states of each node are generated through node feature encoding and edge feature propagation. Cross-type node information is aggregated based on the hidden states of the connection state record unit nodes and the hidden states of the scene snapshot unit nodes, outputting the lost event explanation feature vector. Part Five demonstrates the generation process of the headphone loss spatiotemporal tracing record. Based on the lost event explanation feature vector, the time of headphone loss, the semantic marker of the loss-triggered scene, the motion modal transition sequence before loss, and the set of environmental interaction features after loss are derived, forming a complete spatiotemporal tracing record. Part Six demonstrates the generation and push process of the headphone retrieval operation command sequence. Operation commands are generated based on the semantic marker of the loss-triggered scene and the set of environmental interaction features after loss, and pushed to the user terminal via a wireless communication network.

[0083] Step S210: After pushing the earphone retrieval operation command sequence to the user terminal, receive feedback behavior data returned by the user terminal during the retrieval operation in real time. The feedback behavior data includes terminal motion trajectory change data and user response operation records to the retrieval command.

[0084] After the user terminal displays the retrieval operation guidance interface, it continuously collects terminal motion data through the device's built-in accelerometer, gyroscope, and magnetometer. Simultaneously, it records the user's response to audio prompts and visual guidance through the user interaction event interface provided by the operating system. The terminal motion trajectory change data includes a time-series sequence of the terminal's latitude and longitude coordinates and a time-series sequence of terminal altitude changes, with the sampling period consistent with the sampling period T in step S110. The user's response operation record to the retrieval command includes the time difference between the user's first screen operation after the audio prompt command is triggered and the record of the user clicking on interface elements after the visual guidance command is displayed. The terminal packages and sends the feedback behavior data to the server according to a preset reporting period through the established long-term communication connection. The server writes the received data into a feedback data receiving queue in memory.

[0085] Step S220: Extract the user's actual movement direction sequence and user's actual movement speed sequence based on the terminal motion trajectory change data. At the same time, extract the user's reaction time to audio prompts and the user's interface touch status to visual guidance commands based on the response operation records.

[0086] The latitude and longitude coordinate sequence in the terminal motion trajectory change data is processed by point-by-point difference. The displacement vector between two adjacent sampling points is obtained by subtracting the latitude and longitude coordinates of the previous point from the latitude and longitude coordinates of the subsequent point. The direction angle of the displacement vector is calculated using the arctangent function, and the direction angle is taken as a clockwise angle relative to true north. The direction angles of all sampling intervals are arranged in chronological order to form the sequence of the user's actual movement direction. The magnitude of the displacement vector is divided by the sampling period T to obtain the average movement speed of that sampling interval. The average movement speeds of all sampling intervals are arranged in chronological order to form the sequence of the user's actual movement speed. The user's reaction time to the audio prompt command is taken as the absolute value of the time difference between the audio prompt command trigger time and the user's first screen interaction time. The user's interface touch status for the visual guidance command is taken as a binary value: if the user clicks the interface element specified by the visual guidance command, the value is 1; otherwise, the value is 0.

[0087] Step S230: Compare the user's actual movement direction sequence with the motion mode transition sequence before loss to calculate the direction deviation pattern, and compare the user's actual movement speed sequence with the pre-stored object-finding speed baseline to generate the movement efficiency deviation pattern.

[0088] The motion mode transition sequence before loss is generated in step S153. Each motion mode label in the sequence corresponds to a preset reference direction interval. The directional angles in the user's actual movement direction sequence are matched with the reference direction interval of the corresponding motion mode label. If a directional angle is outside the reference direction interval, it is recorded as a directional deviation. The number and magnitude of deviations are counted to form a directional deviation pattern. The pre-stored search speed baseline is the normal walking speed range obtained from the user's historical search behavior. Each speed value in the user's actual movement speed sequence is compared with the search speed baseline range. If the speed value is lower than the lower limit of the baseline range, it is recorded as low movement efficiency; if the speed value is higher than the upper limit of the baseline range, it is recorded as high movement efficiency. The ratio of low to high speeds constitutes a movement efficiency deviation pattern.

[0089] Step S240: Based on the direction deviation mode and the movement efficiency deviation mode, determine the degree of execution deviation of the current retrieval operation instruction sequence. When the degree of execution deviation meets the instruction adjustment trigger condition, generate an adjustment demand signal.

[0090] The degree of execution deviation is calculated as a weighted sum of two indicators: the ratio of the number of directional deviations to the total number of sampling intervals and the proportion of low movement efficiency. A threshold for triggering instruction adjustment is set. When the calculated result of the degree of execution deviation is greater than or equal to this threshold, it is determined that the degree of execution deviation of the current retrieval operation instruction sequence meets the instruction adjustment trigger condition, and an adjustment demand signal is generated. The adjustment demand signal includes the current deviation value and the trigger timestamp.

[0091] Step S250: In response to the adjustment demand signal, re-select unused environmental object path branches from the lost environmental interaction feature set, and use the re-selected environmental object path branches as alternative navigation intermediate points.

[0092] The set of environmental interaction features after loss is generated in step S154 and contains multiple pairs of object category identifiers and interaction action identifiers. The used environmental object path branches are the object category identifiers selected as navigation intermediate points in step S157. From the set of environmental interaction features after loss, object category identifiers that do not appear in the used list are filtered out, and the first preset number are taken as candidate navigation intermediate points according to their original arrangement order in the set.

[0093] Step S260: Update the path planning for the candidate navigation intermediate points based on the current user location, generate a corrected navigation path instruction, and adjust the timing of the audio prompt instruction and the display position of the visual guidance instruction based on the reaction time and interface touch status.

[0094] The current user location is determined by the latest reported latitude and longitude coordinates from the feedback behavior data. Using the current user location as the new starting point and alternative navigation midpoints as waypoints, the path planning engine is invoked to regenerate and correct the navigation path instructions. The timing of audio prompts is adjusted based on the reaction time value; if the reaction time exceeds a preset threshold, the prompt timing is advanced by measuring the difference between the reaction time and the threshold. The display position of visual guidance instructions is adjusted based on the interface access status; if the access status is 0, indicating the user has not noticed the visual guidance, the display position is moved from the default position to the highlighted area in the center of the interface.

[0095] Step S270: Integrate the navigation path correction instruction, the audio prompt instruction after adjusting the reminder timing, and the visual guidance instruction after adjusting the display position into an updated retrieval operation instruction sequence. Replace the original earphone retrieval operation instruction sequence with the updated retrieval operation instruction sequence, push it to the user terminal again, and continue to monitor subsequent feedback behavior data until the user terminal returns a retrieval completion signal.

[0096] The revised navigation path instructions, adjusted audio prompts, and adjusted visual guidance instructions are repackaged into an instruction sequence, with the same sequence format as in steps S157 and S158. The updated retrieval operation instruction sequence replaces the original instruction sequence in the server cache and is pushed to the user terminal again via the wireless communication channel. The feedback monitoring and instruction adjustment cycle of steps S210 to S260 continues until a retrieval completion signal is received from the user terminal, which includes a confirmation that the earphone has been retrieved.

[0097] Step S310: After generating the spatiotemporal source tracing record of the lost earphone, extract all object category identifiers from the set of environmental interaction features after the loss, and input the object category identifiers into the environmental object association network. The environmental object association network outputs a set of neighboring object category identifiers that have a co-occurrence relationship with each object category identifier by querying a pre-built environmental object co-occurrence knowledge graph.

[0098] The environmental object association network maintains a pre-constructed environmental object co-occurrence knowledge graph. This knowledge graph is stored in a graph data structure, where nodes are object category identifiers, edges represent spatial co-occurrence relationships between objects, and edge weights represent the probability of two objects co-occurring in the same spatial scene. For each object category identifier Oq in the lost environmental interaction feature set, all outgoing edges originating from Oq are retrieved from the knowledge graph. The object category identifiers pointed to by outgoing edges whose weights are greater than a preset co-occurrence weight threshold are selected to form the set of neighboring object category identifiers for Oq. The neighbor sets of all object category identifiers are then deduplicated and merged to obtain the total set of neighboring object category identifiers.

[0099] Step S320: Based on the set of neighboring object category identifiers and the set of interaction action identifiers in the set of lost environmental interaction features, construct an environmental object state transition probability graph with object category identifiers as nodes and co-occurrence relationships and interaction action relationships as edges. Transform the environmental object state transition probability graph into a weighted directed transition graph, where nodes are environmental objects, the direction of the directed edges represents the possible transition directions between objects, and the weight of the directed edges represents the probability of transition determined by the interaction action identifier.

[0100] Create a node set for the environment object state transition probability graph, containing the deduplicated union of all object category identifiers in the lost environment interaction feature set and all object category identifiers in the neighboring object category identifier set. For any two object nodes, if there is an interaction action identifier from the previous object to the next object in the lost environment interaction feature set, then establish a directed edge between the two nodes, with the edge pointing from the previous object to the next object, and the edge weight being the normalized value of the frequency of the interaction action identifier appearing in the lost environment interaction feature set. If there is no direct interaction action relationship but the two objects have a co-occurrence relationship in the environment object co-occurrence knowledge graph, then establish a directed edge, and the edge weight is the co-occurrence probability value.

[0101] Step S330: Starting from the object node that the earphone last contacted, perform a weight-based path expansion search on the weighted directed transition graph. The search depth is a preset number of expansion steps. Generate multiple possible environmental object transfer paths. Map the object nodes on each possible environmental object transfer path to the pre-stored environmental space layout graph. Obtain the physical space coordinates of each object node and connect them in the transfer order to obtain the object transfer trajectory projection.

[0102] The last known object node contacted by the headphones is identified by the object category identifier decoded last from the set of environmental interaction features after the loss. Starting from this node, a breadth-first search is performed on the weighted directed transition graph, prioritizing the expansion of outgoing edges with larger weights, with the upper limit of the search depth set to a preset number of expansion steps. Each of the multiple paths generated by the search is a sequence of object nodes. For each path, the physical spatial coordinates of each object node are retrieved from the pre-stored environmental spatial layout graph, and the coordinates are connected sequentially according to the path to form a polyline segment, thus obtaining the object transfer trajectory projection.

[0103] Step S340: Project each object transfer trajectory, estimate the time distribution of trajectory points based on motion inertia characteristics and user walking speed range, generate a spatiotemporal search area, and perform a union operation on all spatiotemporal search areas to obtain the spatiotemporal coverage range of the area where the headphones may be left behind.

[0104] The motion inertia feature is taken as the average motion direction of the user in the motion mode transition sequence before loss. For adjacent trajectory points on the projection of each object transfer trajectory, the upper and lower bounds of the arrival time are obtained by dividing the distance between the two points by the range of the user's walking speed. With each trajectory point as the center, a preset spatial search radius as the spatial range, and the upper and lower bounds of the arrival time as the time range, a spatiotemporal search volume is generated for each trajectory point. The union of the spatiotemporal search volumes of all trajectory points constitutes the spatiotemporal coverage range of the area where the earphone may be left behind.

[0105] Step S350: Convert the spatiotemporal coverage of the area where the earphone may be left into a set of geofence coordinates, and embed it into the visual guidance instruction template of the earphone retrieval operation instruction sequence, replacing the original sequence of object coordinates along the way, generating an enhanced visual guidance instruction, updating the enhanced visual guidance instruction to the user terminal, so that the user terminal highlights the area corresponding to the set of geofence coordinates when displaying the navigation map.

[0106] The spatial boundary of the spatiotemporal coverage area is projected onto a horizontal plane, and the outer envelope polygon of the projected area is extracted. The vertex coordinate sequence of the polygon constitutes the geofence coordinate set. The path object coordinate sequence field in the visual guidance instruction template in step S156 is replaced with the geofence coordinate set field to generate an enhanced visual guidance instruction. The enhanced visual guidance instruction is sent to the user terminal through the message push channel in step S158. After parsing the instruction, the user terminal renders the polygon area corresponding to the geofence coordinate set on the navigation map interface with a highlighted fill color.

[0107] Step S410: After generating the candidate time interval set for headphone loss, obtain a historical headphone loss event record library. The historical headphone loss event record library contains multiple historical loss records. Each historical loss record consists of a historical loss interval, a historical signal attenuation marker chain, and a historical loss scene semantic marker. For each candidate time interval in the candidate time interval set, perform a sequence similarity comparison between its corresponding signal attenuation marker chain and the historical signal attenuation marker chain of each historical loss record in the historical headphone loss event record library to obtain a signal attenuation similarity sequence. Simultaneously, use the position semantic marker with the highest frequency in the scene snapshot subsequence corresponding to the candidate time interval as the dominant scene marker of the current interval, and perform a scene consistency comparison with the historical loss scene semantic marker of the historical loss record to obtain a scene overlap sequence.

[0108] Sequence similarity comparison employs the same dynamic time warping algorithm as step S124, calculating the warped distance between the current signal attenuation marker chain and the historical signal attenuation marker chain. The reciprocal of this warped distance, after normalization, is used as the signal attenuation similarity value. The dominant scene marker for the current interval is the mode value of the frequency of semantic marker occurrences at each position in the scene snapshot subsequence. The scene overlap value is 1 if the dominant scene marker for the current interval is the same as the semantic marker of a historical lost scene, and 0 otherwise.

[0109] Step S420: The signal attenuation similarity sequence and the scene overlap sequence are weighted and fused item by item to obtain a comprehensive historical matching degree sequence. The maximum comprehensive historical matching degree value in the comprehensive historical matching degree sequence is selected as the historical matching degree score of the candidate time interval of the headphone loss. When the historical matching degree score of any candidate time interval of the headphone loss is lower than the preset historical matching threshold, the candidate time interval of the headphone loss is marked as a historical deviation interval.

[0110] The signal attenuation similarity value is multiplied by a first weighting coefficient and the scene overlap value is multiplied by a second weighting coefficient. The sum of the first and second weighting coefficients is 1. The historical matching score is the maximum value in the comprehensive historical matching score sequence. A preset historical matching threshold is a constant value. If the historical matching score is less than the threshold, a historical deviation marker is added to the corresponding candidate time interval for headphone loss.

[0111] Step S430: For the candidate time intervals of headphone loss marked as historical deviation intervals, extract the frequency of changes in the wearing status marker and the amplitude of the signal strength marker fluctuation within a specified time window before and after the time of the jump event. If the frequency of change and the amplitude of fluctuation both fall within the feature range of the false triggering mode, then the historical deviation interval is further marked as a false triggering interval; otherwise, it is retained as an undetermined interval.

[0112] The specified time window length is the preset duration before and after the occurrence of the jump event. The frequency of wear status marker changes is the total number of jumps in the wear status marker within the time window. The signal strength marker fluctuation amplitude is the difference between the maximum and minimum signal strength marker values ​​within the time window. The false trigger mode characteristic range includes the combination of a change frequency greater than or equal to a preset frequency threshold and a fluctuation amplitude less than or equal to a preset fluctuation threshold. If both conditions are met, the historical deviation marker is updated to a false trigger marker; otherwise, the historical deviation marker is retained and the interval status is set to pending.

[0113] Step S440: Remove all candidate time intervals marked as false trigger intervals from the candidate time interval set of headphone loss, while retaining intervals with historical matching scores not lower than the historical matching threshold and undetermined intervals, and generate a set of candidate time intervals of headphone loss that has been filtered by historical verification, as the updated set of candidate time intervals of headphone loss, which will be used to construct the spatiotemporal correlation graph structure and generate the feature vector for explaining the loss event.

[0114] Iterate through the candidate time interval set Ω for headphone loss, and remove elements marked as false trigger intervals from the set. Retain all intervals not marked as false triggers to form the updated candidate time interval set for headphone loss. Subsequent steps S130 to S150 use this updated set to replace the original set Ω.

[0115] Step S510: During the execution of the headphone retrieval operation instruction sequence by the user terminal, a continuous video image frame sequence of the user's surrounding environment is acquired by the user terminal's camera device and input into a pre-constructed headphone visual detection network. The pre-constructed headphone visual detection network includes a feature extraction backbone network and a detection output network. The feature extraction backbone network performs multi-layer convolution and downsampling processing on each frame of the environmental image to generate a multi-scale feature map set. The detection output network predicts the headphone target presence confidence and target bounding box coordinates position by position on the multi-scale feature map set.

[0116] The feature extraction backbone network employs a convolutional neural network architecture, containing alternating stacked convolutional layers and max-pooling layers. The convolutional layers use kernels of a pre-defined size for feature extraction, while the pooling layers perform spatial downsampling to expand the receptive field. The backbone network outputs multi-scale feature maps with three different spatial resolutions. The detection output network pre-defines multiple prior boxes with different aspect ratios at each spatial location within each scale feature map. For each prior box, it predicts the presence confidence of the headphone target and the target bounding box coordinate offset. The confidence output is mapped to the 0-1 range using a sigmoid function, and the bounding box coordinate offset is used to adjust the position and size of the prior box.

[0117] Step S520: When the confidence level of the detected headphone target in any frame exceeds the preset detection confidence condition, extract the pixel region of the headphone target from the target bounding box coordinates corresponding to that frame, and record the acquisition timestamp of that frame.

[0118] The preset detection confidence condition is that the confidence level is greater than or equal to the preset detection confidence threshold. From the detection results that meet the condition, the target bounding box coordinates with the highest confidence level are selected. The target bounding box coordinates are in the format of the row and column coordinates of the center pixel of the bounding box and the width and height of the bounding box. The pixel region within the bounding box is extracted as the pixel region of the headphone target. The acquisition timestamp of this frame is recorded.

[0119] Step S530: Based on the pixel region of the headphone target, using the camera intrinsic parameter matrix of the user terminal and the attitude angle data of the user terminal when the frame was acquired, the center coordinates of the pixel region are back-projected to the camera coordinate system, and then transformed to the world coordinate system to obtain the estimated three-dimensional coordinates of the headphone target in the world coordinate system.

[0120] The camera intrinsic parameter matrix contains focal length and principal point coordinate parameters, obtained from the user terminal operating system. The center coordinates of the pixel region are back-projected from the intrinsic parameter matrix to obtain normalized coordinates in the camera coordinate system. A rotation matrix from the camera coordinate system to the world coordinate system is constructed using attitude angle data. The normalized coordinates are then transformed using the rotation matrix and combined with the camera height parameter to calculate the estimated 3D coordinates in the world coordinate system.

[0121] Step S540: Combine the current terminal positioning coordinates provided by the user terminal's global positioning system module to perform position fusion correction on the estimated three-dimensional coordinates to obtain the geographic coordinates of the earphone target. Evaluate the distance difference between the geographic coordinates of the earphone target and the navigation destination coordinates in the current earphone retrieval operation command sequence. When the distance difference is greater than the preset difference tolerance distance, generate a destination update command.

[0122] The estimated 3D coordinates are converted into latitude and longitude coordinates, and then fused with the current terminal positioning coordinates provided by the Global Positioning System (GPS) module using a weighted average. The fusion weights are preset values. The fused latitude and longitude coordinates represent the geographic coordinates of the earphone target. The current navigation destination coordinates are extracted from the endpoint field of the navigation path command in the earphone retrieval operation command sequence. The spherical distance between the earphone target's geographic coordinates and the navigation destination coordinates is calculated. If the distance is greater than a preset difference tolerance distance, a destination update command is generated.

[0123] Step S550: Based on the destination update instruction, using the geographic coordinates of the earphone target as the new navigation destination, replan the walking path from the user's current location to the new navigation destination, generate a corrected navigation path, replace the navigation path instruction in the original earphone retrieval operation instruction sequence with the corrected navigation path, while retaining the original audio prompt instructions and visual guidance instructions, generate a corrected earphone retrieval operation instruction sequence, and push the corrected earphone retrieval operation instruction sequence to the user terminal in real time to update the navigation guidance on the user terminal interface.

[0124] The path planning engine is invoked to generate a corrected navigation path, starting from the user's current location and ending at the target geographic coordinates of the headphones. This corrected navigation path is then encapsulated as navigation path instructions. The corrected headphone retrieval operation instruction sequence includes the corrected navigation path instruction, the original audio prompt instruction, and the original visual guidance instruction. These instructions are pushed to the user terminal in real time via the message push channel in step S158. Upon receiving the corrected instructions, the user terminal updates the map navigation guidance, moving the navigation endpoint marker to the target geographic coordinates of the headphones.

[0125] For example, in step S610: when a headphone loss event is detected, a locally pre-built deep learning model is used to process the local spatiotemporal correlation graph structure to generate a local loss event explanatory feature vector, and a local headphone loss spatiotemporal tracing record sample is formed accordingly. Each user terminal uploads the local headphone loss spatiotemporal tracing record sample to the cloud aggregation server. The cloud aggregation server receives local headphone loss spatiotemporal tracing record samples from different user terminals, forming a distributed headphone loss sample library.

[0126] The headphone loss correlation analysis function unit is integrated into the user terminal application as a software module. Upon detecting a headphone loss event locally, this function unit generates a local loss event explanation feature vector and a local headphone loss spatiotemporal tracing record according to steps S110 to S150. The data structure of the local headphone loss spatiotemporal tracing record sample includes the local loss event explanation feature vector, the target headphone loss time, the semantic marker of the loss triggering scene, the motion mode transition sequence before loss, and the set of environmental interaction features after loss. When the wireless network is available, the user terminal uploads the local samples to the cloud aggregation server via an encrypted channel. The cloud aggregation server stores the received samples from each user terminal in a distributed headphone loss sample database.

[0127] Step S620: The cloud aggregation server extracts the local loss event interpretation feature vector corresponding to each sample from the distributed headphone loss sample library as input, and uses the target headphone loss time, loss trigger scene semantic mark, motion mode change sequence before loss and environmental interaction feature set after loss in the sample as supervision signals to train the cloud global deep learning model. The parameters of the cloud global deep learning model are updated through gradient backpropagation to obtain the updated cloud global parameters.

[0128] The structure of the cloud-based global deep learning model is completely consistent with the structure of the pre-built deep learning model in step S140. During training, the local loss event interpretation feature vectors of a batch of randomly selected samples from the sample library are input into the cloud-based global deep learning model. The model generates predicted loss times, scene labels, motion sequences, and interaction sets through forward propagation in steps S141 to S146. The multi-task joint loss between the predicted values ​​and the supervision signals is calculated. The loss for loss times is the mean squared error, the loss for scene labels is the cross-entropy loss, the loss for motion sequences is the sum of the stepwise cross-entropy losses, and the loss for interaction sets is the sum of the cross-entropy losses for object categories and interaction actions. The total loss is the weighted sum of the losses of each task. An adaptive moment estimation optimizer is used for backpropagation to update the parameters of the cloud-based global deep learning model.

[0129] Step S630: The updated cloud global parameters are sent to the headphone loss correlation analysis function unit of each user terminal. Each user terminal uses the updated cloud global parameters to replace the parameters of the locally pre-built deep learning model to obtain the updated local deep learning model.

[0130] The cloud aggregation server serializes the updated global parameters into a data file and distributes it to each user terminal via push notification or active terminal retrieval. After receiving the parameter file, the user terminal loads the weights and bias parameters of each layer in the file into the parameter variables of the corresponding layer of the local deep learning model, completing the parameter update of the local model.

[0131] Step S640: In subsequent use, the user terminal uses the updated local deep learning model to process the newly occurring candidate time interval of headphone loss, generates a supplementary loss event explanation feature vector, and uploads the supplementary loss event explanation feature vector and the associated supervision signal to the cloud aggregation server.

[0132] When the user terminal detects the headphone loss event again, the updated local deep learning model is used to generate a supplementary loss event explanation feature vector and a corresponding supplementary headphone loss spatiotemporal source record according to steps S110 to S150. The supplementary sample is then uploaded to the cloud aggregation server.

[0133] Step S650: The cloud aggregation server calculates the distribution center vector and distribution dispersion vector of the supplementary missing event explanation feature vector based on the collected supplementary missing event explanation feature vector, and determines whether the cloud global deep learning model has reached the convergence state based on the distribution dispersion vector.

[0134] The distribution center vector is the arithmetic mean of all supplementary missing event explanatory feature vectors in each dimension. The distribution dispersion vector is the standard deviation of all supplementary missing event explanatory feature vectors in each dimension. If all dimensions of the distribution dispersion vector are less than the preset convergence threshold, the cloud-based global deep learning model is considered to have converged; otherwise, it is considered not to have converged.

[0135] Step S660: If convergence is not achieved, the cloud aggregation server uses the supplementary missing event interpretation feature vector to incrementally train the cloud global deep learning model, adjusts the cloud global parameters, and sends the new cloud global parameters back to the user terminal until the cloud global deep learning model reaches convergence, thus obtaining a cross-user generalized global deep learning model, so that each user terminal can use the cross-user generalized global deep learning model to perform subsequent headphone loss correlation analysis.

[0136] Incremental training uses the cloud-based global parameters obtained from the previous training round as initial parameters, and performs a few iterations of fine-tuning training using supplementary samples according to the training process in step S620. After fine-tuning, new cloud-based global parameters are obtained and sent back to the user terminal. The loop from steps S640 to S660 is repeated until step S650 determines that the model has converged. The converged cloud-based global deep learning model is the cross-user generalized global deep learning model.

[0137] Step S710: While acquiring the wireless connection state transition sequence and the scene snapshot transition sequence, acquire the auxiliary motion sensor data stream recorded by the auxiliary wearable device worn by the target user within the same time interval. The auxiliary motion sensor data stream includes a three-axis accelerometer data sequence and a three-axis gyroscope data sequence. Perform step cycle detection and motion intensity analysis on the three-axis accelerometer data sequence. Extract the step ground contact time sequence using peak detection and calculate the step frequency feature based on the step ground contact interval. Simultaneously, calculate the motion intensity feature based on the acceleration amplitude envelope to generate a first motion sub-feature sequence. Perform angular velocity change rate analysis and attitude angle detection on the three-axis gyroscope data sequence. Extract the body turning event sequence using angular velocity zero-crossing detection and calculate the turning amplitude feature based on the cumulative angle between turning events. Simultaneously, extract the body sway feature based on the periodic change of the pitch angle channel to generate a second motion sub-feature sequence. By using a preset motion modality classification mapping, the motion sub-features in the first motion sub-feature sequence and the second motion sub-feature sequence are mapped to auxiliary device motion modality labels, generating an auxiliary device motion modality label sequence that is consistent with the motion modality label type in the scene snapshot unit.

[0138] The auxiliary wearable device maintains data synchronization with the smartphone via a Bluetooth communication link. The data stream from the auxiliary motion sensor has the same time range as the wireless connection state transition sequence and the scene snapshot transition sequence. Step ground contact detection is achieved by searching for local maxima exceeding a preset acceleration threshold on the combined acceleration modulus signal from the three-axis accelerometer. The step frequency feature is the reciprocal of the difference between the ground contact times of adjacent steps. The acceleration amplitude envelope is extracted by performing a Hilbert transform on the combined acceleration modulus signal, and the motion intensity feature is the mean of the amplitude envelope. Angular velocity zero-crossing detection is achieved by searching for points of sign change in the combined angular velocity modulus signal from the three-axis gyroscope. The steering amplitude feature is the absolute value of the cumulative angle between adjacent zero-crossing points. The periodic variation of the pitch angle channel is extracted as a period parameter through autocorrelation analysis, and the body sway feature is the standard deviation of the pitch angle variation within the period. The preset motion modality classification is mapped to a decision tree classifier from motion sub-features to motion modality labels.

[0139] Step S720: Compare the motion modality marker of each scene snapshot unit in the scene snapshot transition sequence with the motion modality label of the auxiliary device with the same timestamp. When the comparison result is that the motion modality marker and the auxiliary device motion modality label are inconsistent, record the timestamp as a conflict timestamp and collect all conflict timestamps to form a motion modality conflict marker list.

[0140] The motion modal tag Mi of the i-th scene snapshot unit is compared with the motion modal tag of the i-th auxiliary device for comparison. If the two values ​​are not equal, they are determined to be inconsistent, and the corresponding timestamp ti is added to the motion modal conflict tag list.

[0141] Step S730: For each conflict timestamp in the motion mode conflict marker list, extract the connection state recording unit corresponding to the conflict timestamp from the wireless connection state transition sequence, extract the wearing state marker of the connection state recording unit, and when the wearing state marker is equal to the wearing disengagement marker, replace the motion mode marker of the original scene snapshot unit with the auxiliary device motion mode tag of the conflict timestamp to obtain the preliminary corrected scene snapshot transition sequence.

[0142] Extract the wearing status flag field value from the connection status record unit corresponding to the conflict timestamp. If the value is 0, update the motion modality flag field value of the original scene snapshot unit to the value of the auxiliary device motion modality label. After the replacement operation is completed, a preliminary corrected scene snapshot transition sequence is obtained.

[0143] Step S740: Perform a temporal logic consistency check on the motion modality labels adjacent to the replacement positions in the preliminary corrected scene snapshot transition sequence. When the replaced motion modality label contradicts the preceding and following motion modality labels, adjust the replaced motion modality label according to the continuity trend of multiple consecutive motion modality labels to obtain the corrected scene snapshot transition sequence. Replace the original scene snapshot transition sequence with the corrected scene snapshot transition sequence, and re-execute the operation of constructing a spatiotemporal correlation graph structure based on the candidate time interval set of headphone loss and calling the pre-built deep learning model for processing to generate a corrected loss event explanation feature vector. Based on the corrected loss event explanation feature vector, re-derive the corrected pre-loss motion modality transition sequence.

[0144] The temporal logic consistency check determines whether the motion mode label, the preceding motion mode label, and the following motion mode label at the replacement position satisfy the preset reasonable motion mode transition rules. The reasonable transition rules define the allowed direct transition relationships between each motion mode. If any of the three do not satisfy any reasonable transition rule, it is determined to be a transition contradiction, and the label with the highest frequency among multiple consecutive motion mode labels is taken as the adjusted motion mode label. After correction, steps S130 to S153 are re-executed to generate the corrected lost event explanation feature vector and the corrected pre-loss motion mode transition sequence.

[0145] Step S750: Based on the corrected pre-loss motion mode transition sequence, adjust the navigation path start direction constraint parameters mapped by the pre-loss motion mode transition sequence, and remap the object category identifier sequence in the corrected post-loss environmental interaction feature set to the coordinate sequence of the objects along the path, generating a navigation path instruction that fuses the motion information of the auxiliary device. Combine the navigation path instruction that fuses the motion information of the auxiliary device with the original audio prompt instruction and the original visual guidance instruction to generate a multi-source motion fusion headphone retrieval operation instruction sequence, and push the multi-source motion fusion headphone retrieval operation instruction sequence to the user terminal.

[0146] The starting direction constraint parameter of the navigation path is taken as the center value of the reference direction interval corresponding to the first motion mode label in the corrected motion mode transition sequence before the loss. The generation method of the coordinate sequence of the objects traversed is the same as in step S157. The navigation path command that integrates the motion information of the auxiliary equipment adds a starting direction constraint field to the original navigation path command. The command push process is the same as in step S158.

[0147] Step S810: After generating the spatiotemporal tracing record of the lost earphone, extract all the object category identifiers that appear from the set of environmental interaction features after the loss, and obtain the object material type and the characteristic resonance frequency of the object material type corresponding to each object category identifier by searching the pre-built material acoustic property knowledge base. The material acoustic property knowledge base stores the correspondence between material type and characteristic resonance frequency with the object category identifier as the index.

[0148] The material acoustic property knowledge base is a key-value database. The key is the object category identifier, and the value is a structure containing the material type and characteristic resonant frequency. The material type is stored as a string encoding, and the characteristic resonant frequency is stored as a floating-point number in Hertz. For each object category identifier Oq in the lost environmental interaction feature set, an exact query is performed in the material acoustic property knowledge base using Oq as the key, returning the corresponding object material type Mq and characteristic resonant frequency Fq. All query results are then summarized into a list of material resonant frequencies.

[0149] Step S820: Based on the extracted object material type, further obtain the sound wave absorption coefficient and sound wave reflection coefficient of each object material type from the material acoustic property knowledge base, and at the same time obtain the estimated spatial volume parameters of the headphone loss scenario. Calculate the sound reverberation decay time constant of the headphone loss scenario based on the sound wave absorption coefficient, sound wave reflection coefficient and estimated spatial volume parameters.

[0150] In the material acoustic property knowledge base, each material type stores its characteristic resonant frequency, as well as its acoustic absorption coefficient α and acoustic reflection coefficient β. The acoustic absorption coefficient is a dimensionless ratio, ranging from 0 to 1. The acoustic reflection coefficient is equal to 1 minus the acoustic absorption coefficient. The estimated spatial volume parameter Vroom is obtained from the scene spatial parameter mapping table based on the missing trigger scene semantic tags, and the unit is cubic meters. The reverberation decay time constant T60 is calculated as T60 = 0.161 × V / Σ(Si × αi), where V is the estimated spatial volume parameter, Si is the surface area of ​​the i-th material, and αi is the acoustic absorption coefficient of that material. The surface area of ​​each material is estimated from the geometric dimensions of each object in the environmental spatial layout diagram.

[0151] Step S830: Input the characteristic resonance frequency, sound absorption coefficient, sound reflection coefficient and sound reverberation decay time constant into the audio excitation parameter optimization unit. The audio excitation parameter optimization unit uses the weighted curve of the user's ear's hearing sensitivity to a specific frequency band as the optimization benchmark, searches for the excitation frequency and excitation duration that maximize the perceived loudness, and outputs the optimal excitation frequency parameter and the optimal excitation duration parameter.

[0152] The audio excitation parameter optimization unit employs a grid search optimization algorithm. The search space comprises all frequency points within a preset frequency range near the characteristic resonant frequency, and all time points between preset minimum and maximum durations for the excitation frequency. For each combination of candidate excitation frequencies and durations, the perceived loudness value is calculated. The expression for the perceived loudness value L is L = P × W(f) × G, where P is the signal sound pressure level at the excitation frequency, W(f) is the weighting coefficient of that frequency on the hearing sensitivity weighted curve, and G is the room gain factor determined by the sound wave reflection coefficient and the sound reverberation decay time constant. The combination with the largest perceived loudness value is selected as the optimal excitation frequency parameter and the optimal excitation duration parameter.

[0153] Step S840: Based on the optimal excitation frequency parameter and the optimal excitation duration parameter, generate an audio excitation signal waveform of a specific frequency through digital signal synthesis. The audio excitation signal waveform is a narrowband pulse sequence with the optimal excitation frequency as the center frequency or a swept frequency signal whose frequency changes linearly with time.

[0154] Digital signal synthesis employs direct digital frequency synthesis (DFD) technology. The narrowband pulse sequence is generated by using the optimal excitation frequency as the center frequency, a preset narrowband width as the bandwidth, a Hanning window function as the pulse envelope, and the optimal excitation duration as the pulse repetition period. The sweep signal is generated by multiplying the optimal excitation frequency by a preset start ratio for the start frequency and the optimal excitation frequency by a preset end ratio for the stop frequency; the frequency changes linearly with time, and the sweep duration is the optimal excitation duration. The synthesized audio excitation signal waveform is stored in pulse code modulation (PCM) format, with a preset audio sampling rate and a preset bit depth.

[0155] Step S850: Determine the object category identifier that the earphone last contacted from the set of environmental interaction features after loss, mark the object category identifier as a neighboring object of the earphone, and associate the audio excitation signal waveform with the neighboring object of the earphone to generate an object-specific audio excitation signal for the neighboring object.

[0156] The object category identifier last contacted by the headphones is taken from the last decoded object category identifier (OQ) in the environmental interaction feature set after loss. The audio excitation signal waveform is associated with and stored with the OQ. A target object identifier field is added to the signal metadata, and the value of the OQ is written to generate an object-specific audio excitation signal.

[0157] Step S860: Extract audio prompt instructions from the headphone retrieval operation instruction sequence, replace the general prompt sound data in the original audio prompt instructions with the object-specific audio excitation signal, and retain the playback interval and repetition number settings in the original audio prompt instructions to generate customized audio prompt instructions.

[0158] An audio prompt instruction template is extracted from the headphone retrieval operation instruction sequence generated in step S157. The audio prompt instruction template includes a prompt sound data field, a playback interval field, and a repetition count field. The value of the prompt sound data field is replaced with the binary data of the object-specific audio excitation signal, while the original values ​​of the playback interval field and the repetition count field are retained, thus generating a customized audio prompt instruction.

[0159] Step S870: Add an audio playback space trigger condition to the customized audio prompt instruction. The audio playback space trigger condition is set to start looping the target audio excitation signal of the object when the user terminal enters a spatial area centered on the position coordinates of the object near the headphones and with a radius of a preset proximity distance threshold.

[0160] The position coordinates of nearby objects are obtained from the pre-stored environmental spatial layout map described in step S157 and recorded as spatial coordinates. A preset proximity distance threshold is recorded as the geofence radius. The audio playback spatial trigger condition is a geographic location trigger condition. The trigger logic is to calculate the spherical distance between the user terminal's current position coordinates and the spatial coordinates in real time; the condition is met when the spherical distance is less than or equal to the geofence radius. The trigger condition and trigger logic are written into the additional attribute field of the customized audio prompt instruction.

[0161] Step S880: Recombine the customized audio prompt instruction with the visual guidance instruction and navigation path instruction in the earphone retrieval operation instruction sequence to generate an acoustically enhanced earphone retrieval operation instruction sequence. Push the acoustically enhanced earphone retrieval operation instruction sequence to the user terminal via a wireless communication channel, simultaneously instructing the user terminal to activate the microphone to pick up ambient reflected sound signals while playing the object-specific audio excitation signal. After playing the audio excitation signal, the user terminal collects the ambient reflected sound signals through the microphone, extracts the delay time and signal strength attenuation of the ambient reflected sound signals, and transmits the delay time and signal strength attenuation back to assess the real-time relative distance change trend between the user terminal and the earphones, and accordingly fine-tunes the timing of subsequent audio playback.

[0162] The acoustically enhanced headphone retrieval command sequence consists of customized audio prompts, visual guidance commands, and navigation path commands, with the same sequence format as step S157. It is pushed to the user terminal via a wireless communication channel. A microphone acquisition control field is added to the command message, with a value of "activated." After parsing this field, the user terminal activates the microphone and begins recording ambient audio while playing the object-specific audio excitation signal. After recording, the delay time of the echo signal is extracted; the delay time is the difference between the transmission time of the audio excitation signal and the reception time of the reflected echo. The signal strength attenuation is the difference between the sound pressure level of the transmitted signal and the sound pressure level of the received signal. The delay time and attenuation are transmitted back to the server via a feedback data channel. The server estimates the real-time relative distance based on the delay time and assesses the degree of obstruction of the signal propagation path based on the attenuation, thereby fine-tuning the playback timing of subsequent audio prompts.

[0163] Step S910: After generating the spatiotemporal tracing record of the lost headphones, the cloud server extracts the target user's preset social trust relationship list. The social trust relationship list stores multiple trusted user terminal identifiers authorized by the target user. The cloud server sends a headphone loss assistance request message to the corresponding trusted user terminal based on the trusted user terminal identifier. The assistance request message carries the target headphone's Bluetooth broadcast unique identifier and a rough description of the area where the headphone is lost.

[0164] The social trust relationship list is obtained from the authorization relationship database of the user account system. The trusted user terminal identifier is the unique device identifier of the user terminal. The data structure of the headphone loss assistance request message includes a message type field, a Bluetooth broadcast unique identifier field for the target headphone, and a headphone loss coarse area description field. The Bluetooth broadcast unique identifier is taken from the media access control address carried by the target headphone in the Bluetooth broadcast frame, and the headphone loss coarse area description is taken from the semantic tag of the loss trigger scenario and the GPS coordinates of the midpoint of the headphone loss candidate time interval.

[0165] Step S920: The trusted user terminal that receives the headphone loss assistance request message automatically activates its Bluetooth signal scanning module after obtaining confirmation of the trusted user operation. It continuously listens for Bluetooth broadcast signals in the wireless environment according to the specified scanning time window and scanning duty cycle, and matches the data packets of the detected Bluetooth broadcast signals with the Bluetooth broadcast unique identifier of the target headphone.

[0166] Upon confirmation of the assistance request on the trusted user terminal, the terminal application calls the operating system's Bluetooth application programming interface to activate the Bluetooth signal scanning module. The scanning time window is a window that continues for the preset total scanning duration from the confirmation time. The scanning duty cycle is the ratio of the scanning start time to the scanning cycle length within each scanning cycle. During each scanning start period, the Bluetooth signal scanning module receives over-the-air Bluetooth broadcast signal data packets in full-channel scanning mode, extracts the Media Access Control Address field from each data packet, and performs a string match with the target headset's Bluetooth broadcast unique identifier.

[0167] Step S930: When any trusted user terminal matches the Bluetooth broadcast signal of the target earphone, extract the received signal strength value of the Bluetooth broadcast signal and the timestamp of the signal reception time, and simultaneously record the GPS coordinates and electronic compass orientation angle of the trusted user terminal at the time of signal reception, and combine them into a single assisted scanning record.

[0168] The received signal strength value is extracted from the scan result data structure returned by the Bluetooth protocol stack, and the unit is decibels and milliwatts. The timestamp of the signal reception time is the current time of the terminal system. The GPS coordinates are obtained from the terminal's location service module, including longitude and latitude. The terminal's electronic compass orientation angle is obtained from the terminal's magnetometer sensor, and the value is the angle relative to true north. The above four fields are combined into a single assisted scan record, and the record format is a key-value pair structure.

[0169] Step S940: Each trusted user terminal uploads its generated single assisted scan record to the cloud server via the wireless communication network. The cloud server aggregates the single assisted scan records from different trusted user terminals to generate a multi-source Bluetooth signal scan record set, which contains multiple records from different geographical locations and at different times.

[0170] After the scanning time window ends, the trusted user terminal packages all individual assisted scan records generated during this period into a record batch and uploads it to the cloud server via an encrypted channel. Upon receiving the record batches uploaded by each trusted user terminal, the cloud server merges all records into a multi-source Bluetooth signal scan record set, appending an uploading terminal identifier and a receiving timestamp to each record in the set.

[0171] Step S950: For each single assisted scan record in the multi-source Bluetooth signal scan record set, the cloud server uses the preset Bluetooth signal free space attenuation relationship to convert the received signal strength value into a preliminary distance estimate, and corrects the directional deviation of the preliminary distance estimate according to the orientation angle of the terminal electronic compass to obtain the corrected relative distance value.

[0172] The preset Bluetooth signal free space attenuation relationship is a logarithmic distance path loss model: PL = Ptx - Prx = 10 × n × log10(d) + PL0, where PL is the path loss, Ptx is the transmit power, Prx is the received signal strength, n is the path loss exponent, d is the distance, and PL0 is the reference distance path loss. The transmit power Ptx is taken as the rated transmit power of the target earphone. The distance d is derived from this relationship as 10^((Ptx - Prx - PL0) / (10 × n)). Directional deviation correction is performed by multiplying the initial distance estimate d by a directional correction factor α based on the orientation angle of the terminal's electronic compass, α = 1 + |cos(θ)| × β, where θ is the orientation angle of the electronic compass, β is the preset deviation coefficient, and the corrected relative distance value d' = d × α.

[0173] Step S960: Based on the GPS coordinates and corrected relative distance values ​​of the trusted user terminal corresponding to each single assisted scan record, the cloud server uses multi-point spherical intersection positioning processing. It constructs a spatial sphere with each trusted user terminal as the center and the corrected relative distance value as the radius, solves the center coordinates of the intersection area of ​​multiple spatial spheres, and obtains the multi-source collaborative positioning coordinates of the target earphone.

[0174] For each of the two different assisted scanning records, the intersection circle of the two spatial spheres is calculated. The equation of the intersection circle is |P-C1|=r1 and |P-C2|=r2, where C1 and C2 are the coordinates of the sphere centers, and r1 and r2 are the radii. For the third record, the two intersection points between the spatial sphere and the intersection circle are calculated, and the coordinates of the intersection points satisfy |P-C3|=r3. The intersection point closer to the spatial sphere of the fourth record is selected as the candidate positioning point. The final positioning point coordinates are the geometric center of all candidate points, with center longitude = Σ(longitude_i) / n and center latitude = Σ(latitude_i) / n. These center coordinates are the multi-source cooperative positioning coordinates of the target earphone.

[0175] Step S970: The cloud server calculates the location confidence radius of the multi-source collaborative positioning coordinates based on the spatial dispersion of the intersection area, packages the multi-source collaborative positioning coordinates and the location confidence radius into a social collaborative positioning result, and pushes it to the target user terminal of the user who lost the headphones.

[0176] The location confidence radius is taken as a preset multiple of the standard deviation of the Euclidean distance from each candidate location point to the multi-source collaborative positioning coordinates. The multi-source collaborative positioning coordinates and the location confidence radius are serialized into a data structure for the social collaborative positioning result. The social collaborative positioning result is sent to the target user's terminal via a cloud-to-terminal message push channel.

[0177] Step S980: The target user terminal compares the confidence of the received social collaborative positioning result with the locally generated lost location estimate. When the positioning confidence radius of the social collaborative positioning result is smaller than the uncertainty range of the local lost location estimate, the social collaborative positioning result is selected as the priority for finding the destination coordinates.

[0178] The local lost location estimation is based on the spatial positioning information in the spatiotemporal tracing record of the lost earphone generated in step S150, with the uncertainty range set to a preset local positioning error radius. The positioning confidence radius is compared with the local positioning error radius. If the positioning confidence radius is smaller than the local positioning error radius, the multi-source collaborative positioning coordinates in the social collaborative positioning results are set as the priority destination coordinates for retrieval.

[0179] Step S990: Based on the preferred destination coordinates, update the navigation path instructions in the headphone retrieval operation instruction sequence, set the preferred destination coordinates as the navigation endpoint, and retain the original audio prompt instructions and visual guidance instructions to generate a social collaboration-enhanced headphone retrieval operation instruction sequence. Push the social collaboration-enhanced headphone retrieval operation instruction sequence to the target user terminal, and display the preferred destination coordinates with a highlighted mark on the map interface of the target user terminal.

[0180] The route planning engine is invoked to generate a new navigation route instruction starting from the user's current location and ending at the primary destination coordinates. The social collaboration-enhanced headset retrieval operation instruction sequence consists of the new navigation route instruction, the original audio prompt instruction, and the original visual guidance instruction. The push process is the same as step S158. After receiving the instruction, the target user's terminal highlights the point on the map interface that prioritizes the retrieval of the destination coordinates.

[0181] Combination Figure 3The content showcases the specific presentation of this method on the user side, including two vertical mobile app interfaces. The first interface is the loss diagnosis and spatiotemporal tracing interface, with the title "Earphone Loss Analysis" displayed in the top navigation bar. The core status card is marked with red warning text "Suspected Loss," and details the time of occurrence "14:35:22," the location "Company Reception Area / Front Desk," and the speculated cause "Slipped off while seated." Historical verification results are marked with a green checkmark icon indicating "Accidental touch ruled out." The spatiotemporal trajectory playback is divided into two parts. The upper part is a map view, showing the user's location and the earphone's location. The earphone's location is precisely marked by a red broken earphone icon falling on the POI of the "front desk table," surrounded by a semi-transparent red heat circle indicating a high-probability area. A gray trajectory line connects the user's location and the earphone's location. The lower part is a timeline view, including a signal strength bar and a motion mode bar. The signal strength bar suddenly drops sharply and turns red at 14:35:22. The motion mode bar shows an icon flow from walking deceleration to sitting still. A vertical red dotted line simultaneously penetrates the "signal disconnection point" and the "sitting icon," visually demonstrating the correlation between the loss event and the change in motion mode. The second interface is the retrieval guidance and multi-source collaboration interface. The environmental interaction and acoustic suggestion cards display "retrieval suggestions" items, including an acoustic optimization suggestion of "optimized prompt sound for wooden desktops" and a path planning suggestion of "please focus on checking under the front desk table and floor gaps," the latter with a "high probability" orange label. The social collaboration status bar displays the collaboration information "3 friends are nearby assisting in the scan," while the visual detection floating window simulates a camera viewfinder preview, with a green recognition box framing a suspected headphone target. The bottom operation bar centers on a solid green main button, "Start AR Navigation Retrieval," with auxiliary function areas including a camera icon, a vibration icon, and a refresh icon. A tiny text message at the bottom reads, "Swept-in watch motion data, more accurate trajectory," indicating sensor fusion. The entire front-end interface presents the analysis results and retrieval guidance in a user-friendly manner through intuitive visual elements and interactive controls.

[0182] In an exemplary embodiment, a headphone loss correlation analysis system in a user mobile scenario incorporating a deep learning model is provided. This system can be a terminal, server, etc., and its internal structure includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements a headphone loss correlation analysis method in a user mobile scenario incorporating a deep learning model. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or it can be a button, trackball, or touchpad set on the shell of the headphone loss correlation analysis system in user mobile scenarios combined with deep learning models, or it can be an external keyboard, touchpad, or mouse, etc.

[0183] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for analyzing the correlation between headphone loss and user mobility scenarios using a deep learning model, characterized in that, The method includes: The system acquires a wireless connection state transition sequence and a scene snapshot transition sequence during the user's mobility state transition process. The wireless connection state transition sequence includes multiple connection state recording units arranged in time sequence. Each connection state recording unit includes a wear status marker and a signal strength marker. The scene snapshot transition sequence includes multiple scene snapshot units arranged in time sequence. Each scene snapshot unit includes a location semantic marker and a motion mode marker. The wireless connection state transition sequence is processed by wearing state jump detection and signal strength attenuation segment extraction. Based on the detected jump events from the wearing state to the wearing unwearing mark and the continuous attenuation trend of the signal strength mark before and after the jump event, a set of candidate time intervals for headphone loss is generated. Based on the set of candidate time intervals for headphone loss, a scene snapshot subsequence synchronized with each candidate time interval for headphone loss is extracted from the scene snapshot transition sequence. The scene snapshot subsequence is then time-aligned with the signal strength attenuation segment within the corresponding interval to construct a spatiotemporal association graph structure with connection state recording unit and scene snapshot unit as nodes and temporal adjacency edge and synchronization association edge as connection relationships. The pre-built deep learning model is invoked to encode node features and pass edge features to the spatiotemporal graph structure, generating hidden state representations for each node. Cross-type node information is aggregated based on the hidden states of the connection state record unit node and the hidden states of the scene snapshot unit node to generate a missing event explanation feature vector. Based on the lost event interpretation feature vector, the time of loss of the target earphone, the semantic mark of the lost triggering scene, the motion modal transition sequence before loss, and the set of environmental interaction features after loss are derived. A spatiotemporal source tracing record of earphone loss is generated. Based on the semantic mark of the lost triggering scene and the set of environmental interaction features after loss, an earphone retrieval operation instruction sequence is generated and pushed to the user terminal.

2. The headphone loss correlation analysis method in user mobility scenarios combined with a deep learning model as described in claim 1, characterized in that, The process of detecting wear state transitions and extracting signal strength attenuation segments in the wireless connection state transition sequence generates a set of candidate time intervals for headphone loss based on the detected transition events from the wear-hold mark to the wear-disengage mark and the continuous attenuation trend of the signal strength marks before and after the transition events. This set includes: The connection state record units in the wireless connection state transition sequence are loaded into the circular processing window in timestamp order. In each processing cycle, the current connection state record unit and the immediately preceding connection state record unit are obtained from the circular processing window. Compare the wearing status mark of the current connection status recording unit with the wearing status mark of the previous connection status recording unit. When the wearing status mark is detected to change from the wearing-keeping mark to the wearing-disconnect mark, the timestamp of the current connection status recording unit is recorded as the time of the transition event, and a transition event trigger signal is generated. In response to the transition event trigger signal, a forward signal segment is extracted from the wireless connection state transition sequence, centered on the time of the transition event and extending forward to the connection state recording unit where the signal strength mark first changes to the full signal mark, and a backward signal segment is extracted, extending backward to the time when the signal strength mark recovers to the full signal mark and the signal strength mark of the subsequent three consecutive connection state recording units remains at the full signal mark. The forward signal segment and the backward signal segment are spliced ​​together end to end to obtain a complete signal attenuation segment. The signal strength markers of each connection state recording unit in the complete signal attenuation segment are extracted and arranged in chronological order to form a signal attenuation marker chain. By comparing the signal attenuation marker chain with a variety of pre-stored typical loss attenuation templates, the best matching loss attenuation template corresponding to the signal attenuation marker chain is determined. Based on the attenuation duration characteristics of the best matching loss attenuation template, the number of forward offset steps and the number of backward offset steps of the interval are determined. Based on the time of occurrence of the jump event, the sampling interval of the forward offset step of the interval is shifted forward by several connection state recording units to obtain the interval start timestamp of the headphone loss candidate time interval. The sampling interval of the backward offset step of the interval is shifted backward by several connection state recording units to obtain the interval end timestamp of the headphone loss candidate time interval. The interval start timestamp and interval end timestamp are combined to generate a headphone loss candidate time interval, and the headphone loss candidate time interval is associated with and stored with the corresponding jump event time. Each transition event trigger signal detected in the wireless connection state transition sequence is traversed to generate multiple candidate time intervals for headphone loss. Time axis overlap detection is performed on the multiple candidate time intervals for headphone loss. When the time ranges of any two candidate time intervals for headphone loss overlap or the interval interval is less than the preset merging interval, the corresponding two candidate time intervals for headphone loss are merged into a continuous candidate time interval for headphone loss. All candidate time intervals for headphone loss after merging are arranged in chronological order to generate a set of candidate time intervals for headphone loss.

3. The method for analyzing headphone loss in user mobility scenarios using a deep learning model as described in claim 1, characterized in that, Based on the set of candidate time intervals for headphone loss, the method involves extracting a scene snapshot subsequence synchronized with each candidate time interval from the scene snapshot transition sequence, and temporally aligning and associating the scene snapshot subsequence with the signal strength attenuation segment within the corresponding interval. This constructs a spatiotemporal association graph structure with connection state recording units and scene snapshot units as nodes, and temporal adjacency edges and synchronization association edges as connections. For each candidate time interval of headphone loss in the set of candidate time intervals of headphone loss, based on the start timestamp and end timestamp of the interval, select all scene snapshot units whose timestamps fall within the candidate time interval of the headphone loss from the scene snapshot transition sequence, and form a scene snapshot subsequence corresponding to the candidate time interval of the headphone loss. From the wireless connection state transition sequence, extract connection state recording unit segments that are completely consistent with the time range of the candidate time interval of the earphone loss, as signal strength attenuation segments. Each connection state recording unit in the signal strength attenuation segment and the scene snapshot unit in the scene snapshot subsequence are associated with each other by collecting timestamps. A connection state node is created for each connection state recording unit in the signal strength attenuation segment, and the wearing status mark and signal strength mark contained in the connection state recording unit are mapped to the initial feature vector of the connection state node. Create a scene snapshot node for each scene snapshot unit in the scene snapshot subsequence, and map the position semantic tags and motion modality tags contained in the scene snapshot unit to the initial feature vector of the scene snapshot node; Based on the order of timestamps, directed temporal adjacency edges are established between adjacent connected state nodes, and directed temporal adjacency edges are established between adjacent scene snapshot nodes. Each temporal adjacency edge carries the time difference attribute between the two nodes. For connection status nodes and scene snapshot nodes with the same acquisition timestamp, an undirected synchronous association edge is established, and the undirected synchronous association edge carries a node type identifier and an edge type identifier. For connection status nodes and scene snapshot nodes whose acquisition timestamps differ by one time unit, if their corresponding wearing status marker is a wearing-off marker and their motion mode marker is a stationary mode marker, then an undirected cross-time synchronization association edge is established, and the undirected cross-time synchronization association edge is marked with a cross-time association marker. The connection state nodes, scene snapshot nodes, directed temporal adjacency edges, undirected synchronization association edges, and undirected cross-temporal synchronization association edges are combined to form a single-interval spatiotemporal association subgraph corresponding to the candidate time interval of the headphone loss. All single-interval spatiotemporal association subgraphs are cascaded on the time boundary through temporal adjacency edges between scene snapshot nodes that cross the interval, generating a spatiotemporal association graph structure that covers the entire user's movement state transition process.

4. The headphone loss correlation analysis method in user mobility scenarios combined with a deep learning model as described in claim 1, characterized in that, The process involves calling a pre-built deep learning model to encode node features and pass edge features to the spatiotemporal graph structure, generating hidden state representations for each node. Based on the hidden states of the connection state record unit nodes and the hidden states of the scene snapshot unit nodes, cross-type node information aggregation is performed to generate a missing event explanation feature vector, including: The spatiotemporal relation graph structure is input into the pre-built deep learning model, which includes a node embedding transformation layer, a multi-layer graph message passing layer, and a cross-type attention readout layer. Through the node embedding transformation layer, linear mapping and nonlinear activation processing are applied to the initial feature vector of each connected state node to generate a connected state embedding vector, and linear mapping and nonlinear activation processing are applied to the initial feature vector of each scene snapshot node to generate a scene snapshot embedding vector, thus obtaining the initial node embedding set. The initial node embedding set is input into the first graph message passing layer. For each directed temporal adjacency edge, a temporal message vector is calculated using the source node embedding vector, the target node embedding vector, and the edge attribute features. The temporal message vector is then passed from the source node to the target node along the directed temporal adjacency edge. For each undirected synchronous association edge and undirected cross-time synchronous association edge, the synchronization message vector is calculated using the embedding vectors of the two end nodes and the edge type identifier, and the synchronization message vector is bidirectionally transmitted to the two end nodes along the edge. At each connection state node and each scene snapshot node, all received time-series message vectors and synchronization message vectors are aggregated. The aggregated message vectors are combined with the node's current embedded vector through a gating fusion mechanism to update and generate the node hidden state output by the graph message passing layer. The hidden state of the nodes output by the previous graph message passing layer is used as the input of the next graph message passing layer. The message calculation, message passing and node state update operations are repeatedly executed. After a preset number of graph message passing processes, the final hidden state representation of each node is obtained. Extract all connection state nodes and scene snapshot nodes within the single interval spatiotemporal correlation subgraph corresponding to each candidate time interval of headphone loss from the spatiotemporal correlation graph structure to form an interval node subset. Input the final hidden state representation of all connection state nodes in the interval node subset into the first attention branch of the cross-type attention readout layer to generate a connection state-side aggregated feature vector. At the same time, input the final hidden state representation of all scene snapshot nodes into the second attention branch of the cross-type attention readout layer to generate a scene snapshot-side aggregated feature vector. The aggregated feature vectors from the connection state side and the aggregated feature vectors from the scene snapshot side are concatenated and input into the fully connected fusion layer. Through multi-layer nonlinear transformation of the fully connected fusion layer, the loss event explanation feature vector corresponding to the candidate time interval of the lost headphones is output. The loss event explanation feature vectors corresponding to the candidate time intervals of the lost headphones are summarized to form a set of loss event explanation feature vectors.

5. The headphone loss correlation analysis method in user mobility scenarios combined with a deep learning model as described in claim 1, characterized in that, The process involves deriving the target earphone's loss time, loss triggering scene semantic markers, pre-loss motion modal transition sequence, and post-loss environmental interaction feature set based on the loss event interpretation feature vector, generating a spatiotemporal tracing record of the earphone loss, and matching and generating an earphone retrieval operation instruction sequence based on the loss triggering scene semantic markers and post-loss environmental interaction feature set, and pushing the earphone retrieval operation instruction sequence to the user terminal, including: The lost event interpretation feature vector is input into the lost time regressor, which is composed of stacked fully connected layers. The regressor outputs a fine-tuned time offset relative to the start timestamp of the candidate lost time interval of the earphone by transforming layer by layer. The start timestamp of the interval is added to the fine-tuned time offset to obtain the lost time of the target earphone. The lost event explanation feature vector is input into the scene semantic classifier. The scene semantic classifier uses a fully connected mapping layer and a scene semantic classification output layer to map the lost event explanation feature vector to a preset scene semantic label set. Scene semantic labels whose output probabilities satisfy a preset probability threshold are selected as lost trigger scene semantic labels. The lost event interpretation feature vector is input into the motion sequence decoder. The motion sequence decoder uses the lost event interpretation feature vector as the initial hidden state, generates motion mode labels in each decoding step, and uses the motion mode labels output in the decoding step as the input of the next decoding step, gradually generating multiple consecutive motion mode labels, which are arranged in the order of generation to form a motion mode transition sequence before loss. The lost event interpretation feature vector is input into the environmental interaction reconstruction decoder. The environmental interaction reconstruction decoder initializes the environmental state vector according to the lost event interpretation feature vector, and gradually predicts the combination of object category identifiers and interaction action identifiers in the user's surrounding environment after the loss, generating an environmental interaction feature set containing multiple pairs of object category identifiers and interaction action identifiers. The target earphone loss time, semantic markers of the loss triggering scene, motion mode transition sequence before loss, and environmental interaction feature set after loss are combined into an earphone loss spatiotemporal tracing record; The semantic tags of the lost triggering scenario are input into the preset retrieval strategy knowledge base. The preset retrieval strategy knowledge base stores the mapping from scenario semantic tags to retrieval operation instruction templates. The corresponding retrieval operation instruction templates are extracted by tag matching. The retrieval operation instruction templates include audio prompt instruction templates and visual guidance instruction templates. Extract object category identifier sequences from the set of environmental interaction features after loss, determine the environmental object paths that the user may encounter after losing the headphones, and retrieve the corresponding spatial coordinates from the pre-stored environmental spatial layout map based on the environmental object paths to generate a sequence of object coordinates along the path. Using the sequence of object coordinates along the path as navigation intermediate points, and combining it with the current location coordinates of the user's terminal, call the path planning engine to generate navigation path instructions that start from the current location and pass through the navigation intermediate points in sequence. Combine the navigation path instructions with the audio prompt instruction template and the visual guidance instruction template to generate an initial retrieval operation instruction sequence. In the initial retrieval operation instruction sequence, each instruction entry is appended with semantic tags indicating the time when the target earphone was lost and the scene that triggered the loss. These tags are used by the terminal to present information on the time of loss and scene prompts. The initial retrieval operation instruction sequence with the appended information is then encapsulated into a message data packet, which is pushed to the user terminal via a wireless communication network.

6. The method for analyzing headphone loss correlation in user mobility scenarios using a deep learning model as described in claim 1, characterized in that, The method further includes: After the sequence of instructions for finding the earphone is pushed to the user terminal, the system receives feedback behavior data returned by the user terminal during the retrieval process in real time. The feedback behavior data includes data on changes in the terminal's motion trajectory and records of the user's response to the retrieval instructions. Based on the terminal motion trajectory change data, extract the user's actual movement direction sequence and user's actual movement speed sequence. At the same time, based on the response operation record, extract the user's reaction time to audio prompts and the user's interface touch status to visual guidance instructions. The user's actual movement direction sequence is compared with the motion mode transition sequence before the loss to calculate the direction deviation pattern, and the user's actual movement speed sequence is compared with the pre-stored object-finding speed baseline to generate the movement efficiency deviation pattern. Based on the direction deviation mode and the movement efficiency deviation mode, the degree of execution deviation of the current retrieval operation instruction sequence is determined. When the degree of execution deviation meets the instruction adjustment trigger condition, an adjustment demand signal is generated. In response to the adjustment demand signal, unused environmental object path branches are re-selected from the lost environmental interaction feature set, and the re-selected environmental object path branches are used as alternative navigation intermediate points. The alternative navigation intermediate points are updated based on the current user location to generate correct navigation path instructions. At the same time, the timing of audio prompts and the display position of visual guidance instructions are adjusted based on the reaction time and interface touch status. The audio prompts after correcting the navigation path, adjusting the timing of the reminders, and adjusting the visual guidance instructions after adjusting the display position are integrated into an updated retrieval operation instruction sequence. This updated retrieval operation instruction sequence replaces the original earphone retrieval operation instruction sequence and is pushed to the user terminal again. Subsequent feedback behavior data is monitored until the user terminal returns a retrieval completion signal.

7. The method for analyzing headphone loss in user mobility scenarios using a deep learning model as described in claim 1, characterized in that, The method further includes: After generating the spatiotemporal source tracing record of the lost earphone, all object category identifiers are extracted from the set of environmental interaction features after the loss, and the object category identifiers are input into the environmental object association network. The environmental object association network outputs a set of neighboring object category identifiers that have co-occurrence relationships with each object category identifier by querying a pre-built environmental object co-occurrence knowledge graph. Based on the set of neighboring object category identifiers and the set of interaction action identifiers in the set of lost environmental interaction features, an environmental object state transition probability graph is constructed with object category identifiers as nodes and co-occurrence relationships and interaction action relationships as edges. The environmental object state transition probability graph is then transformed into a weighted directed transition graph, where nodes are environmental objects, the direction of the directed edges represents the possible transition directions between objects, and the weight of the directed edges represents the magnitude of the transition probability determined by the interaction action identifier. Starting from the object node that the headphones last contacted, a weight-based path expansion search is performed on the weighted directed transition graph. The search depth is a preset number of expansion steps. Multiple possible environmental object transfer paths are generated. The object nodes on each possible environmental object transfer path are mapped to a pre-stored environmental spatial layout graph. The physical spatial coordinates of each object node are obtained and connected in the transfer order to obtain the object transfer trajectory projection. For each object transfer trajectory projection, the time distribution of trajectory points is estimated based on the motion inertia characteristics and the user's walking speed range, generating a spatiotemporal search area. All spatiotemporal search areas are then combined to obtain the spatiotemporal coverage of the area where the headphones may be left behind. The spatiotemporal coverage area of ​​the area where the earphone may be left is converted into a set of geofence coordinates and embedded into the visual guidance instruction template of the earphone retrieval operation instruction sequence. This replaces the original sequence of coordinates of the objects along the way, generates enhanced visual guidance instructions, and updates the enhanced visual guidance instructions to the user terminal, so that the user terminal highlights the area corresponding to the set of geofence coordinates when displaying the navigation map.

8. The method for analyzing headphone loss in user mobility scenarios using a deep learning model as described in claim 1, characterized in that, The method further includes: After generating the candidate time interval set of headphone loss, a historical headphone loss event record library is obtained. The historical headphone loss event record library contains multiple historical loss records. Each historical loss record consists of a historical loss interval, a historical signal attenuation marker chain, and a historical loss scene semantic marker. For each candidate time interval of headphone loss in the set of candidate time intervals of headphone loss, the corresponding signal attenuation marker chain is compared with the historical signal attenuation marker chain of each historical loss record in the historical headphone loss event record library to obtain a signal attenuation similarity sequence. At the same time, the position semantic marker with the highest frequency in the scene snapshot subsequence corresponding to the candidate time interval of headphone loss is used as the dominant scene marker of the current interval, and compared with the historical loss scene semantic marker of the historical loss record to obtain a scene overlap sequence. The signal attenuation similarity sequence and the scene overlap sequence are weighted and fused one by one to obtain a comprehensive historical matching degree sequence. The maximum comprehensive historical matching degree value in the comprehensive historical matching degree sequence is selected as the historical matching degree score of the candidate time interval of the headphone loss. When the historical matching degree score of any candidate time interval of the headphone loss is lower than the preset historical matching threshold, the candidate time interval of the headphone loss is marked as a historical deviation interval. For candidate time intervals of headphone loss marked as historical deviation intervals, extract the frequency of changes in wearing status markers and the amplitude of fluctuations in signal strength markers within a specified time window before and after the occurrence of the jump event. If the frequency of changes and the amplitude of fluctuations both fall within the feature range of the false triggering mode, then the historical deviation interval is further marked as a false triggering interval; otherwise, it is retained as an undetermined interval. Remove all candidate time intervals marked as false trigger intervals from the candidate time interval set for headphone loss, while retaining intervals with historical matching scores not lower than the historical matching threshold and undetermined intervals, and generate a set of candidate time intervals for headphone loss that has been filtered by historical verification. This set is used as the updated set of candidate time intervals for headphone loss, which will be used to construct the spatiotemporal correlation graph structure and generate the feature vector for explaining the loss event.

9. The method for analyzing headphone loss in user mobility scenarios using a deep learning model as described in claim 1, characterized in that, The method further includes: During the execution of the headphone retrieval operation command sequence on the user terminal, a continuous video image frame sequence of the user's surrounding environment is acquired by the user terminal's camera device and input into a pre-constructed headphone visual detection network. The pre-constructed headphone visual detection network includes a feature extraction backbone network and a detection output network. The feature extraction backbone network performs multi-layer convolution and downsampling processing on each frame of the environmental image to generate a multi-scale feature map set. The detection output network predicts the headphone target presence confidence and target bounding box coordinates position by position on the multi-scale feature map set. When the confidence level of the detected headphone target in any frame exceeds the preset detection confidence condition, the pixel region of the headphone target is extracted from the target bounding box coordinates of that frame, and the acquisition timestamp of that frame is recorded. Based on the pixel region of the headphone target, using the camera intrinsic parameter matrix of the user terminal and the attitude angle data of the user terminal when the frame was acquired, the center coordinates of the pixel region are back-projected to the camera coordinate system, and then transformed to the world coordinate system to obtain the estimated three-dimensional coordinates of the headphone target in the world coordinate system. Combining the current terminal positioning coordinates provided by the user terminal's global positioning system module, the estimated three-dimensional coordinates are fused and corrected to obtain the geographic coordinates of the earphone target. The geographic coordinates of the earphone target are compared with the navigation destination coordinates in the current earphone retrieval operation command sequence to evaluate the distance difference. When the distance difference is greater than the preset difference tolerance distance, a destination update command is generated. Based on the destination update instruction, the geographical coordinates of the target earphone are used as the new navigation destination. The walking path from the user's current location to the new navigation destination is replanned, and a corrected navigation path is generated. The corrected navigation path replaces the navigation path instruction in the original earphone retrieval operation instruction sequence, while retaining the original audio prompt instruction and visual guidance instruction. A corrected earphone retrieval operation instruction sequence is generated and pushed to the user terminal in real time. The navigation guidance on the user terminal interface is updated to guide the user directly to the detected earphone target location.

10. A system for analyzing the correlation between headphone loss and user mobility scenarios using a deep learning model, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the headphone loss correlation analysis method in user mobility scenarios incorporating a deep learning model, as described in any one of claims 1 to 9, by executing the machine-executable instructions.

Images