3c leasing equipment life cycle intelligent management and control method and system of multi-modal internet of things

Abstract

The application relates to the field of life cycle management, and discloses a 3C rental equipment life cycle intelligent management method and system of a multi-modal Internet of Things, which comprises the following steps: performing benchmark state anchoring when a user handover event is detected, obtaining standardized test data generated by a current user performing a standardized operation test to update a health benchmark baseline; based on the health benchmark baseline, obtaining a touch operation sequence and a device physical response sequence of the current user within a preset time window, inputting the touch operation sequence and the device physical response sequence into a double-branch cross attention network, and then extracting cross-modal correlation features; outputting the cumulative loss contribution value of each key component according to the cross-modal correlation features, triggering an abnormal responsibility determination process, collecting inherent physical feature vectors of the device and real-time operation behavior feature vectors of the user; and calculating a deviation degree through feature space decoupling comparison and outputting a determination result, so that the health state tracking of the rental equipment in the alternating use of multiple users is realized.

Description

Technical Field

[0001] This invention relates to the field of lifecycle management, specifically to a method and system for intelligent lifecycle management of 3C rental equipment in the context of multimodal Internet of Things (IoT). Background Technology

[0002] With the increasing popularity of the sharing economy and rental services in the consumer electronics sector, the rental business of 3C devices such as tablets and smartphones continues to grow. During the rental period, these devices are typically transferred between multiple users, whose operating habits and usage intensity vary significantly. This presents a challenge for rental platforms in managing the health status of the devices.

[0003] Existing technologies already include monitoring solutions for equipment operating status. For example, some solutions collect multi-source data such as vibration, temperature, and pressure from sensors, and combine this data with random forest models and long short-term memory networks to predict equipment failures. These solutions primarily focus on changes in the physical state of the equipment itself, using historical operating data to build trend models and determine whether the equipment has any abnormalities or failure risks. Other solutions monitor user touch event sequences and head posture videos to assess the user's operational intensity and fatigue level, and provide usage time warnings.

[0004] However, all of the above solutions have the following shortcomings:

[0005] Firstly, modeling the health status of equipment is usually based on continuous use by a single user. When the equipment is frequently used by different users, the monitoring baseline is prone to shift due to user switching, resulting in distortion of the benchmark for anomaly detection and making it difficult to accurately determine the cumulative degree of equipment wear and tear.

[0006] Secondly, existing solutions have weak attribution analysis capabilities for equipment wear and tear. When abnormal equipment status or component performance degradation is detected, it is impossible to effectively distinguish whether the wear and tear is caused by the aging process of the equipment itself or by the current user's operation. In the leasing scenario, this is directly related to the calculation of equipment depreciation and the division of user responsibility, which is a prominent problem faced by leasing platforms in actual operation.

[0007] Third, existing solutions do not delve deeply enough into the relationship between touch operation and device physical response, and fail to fully establish the mapping relationship between user operation patterns and the wear and tear of specific components, thus making it difficult to achieve targeted device health prediction.

[0008] Therefore, it is necessary to provide a lifecycle management solution for leased equipment that can adapt to multi-user alternating use scenarios and has the functions of equipment health tracking and abnormal responsibility determination. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for intelligent lifecycle management of 3C rental equipment in multimodal Internet of Things, thereby solving at least one of the above-mentioned technical problems.

[0010] The objective of this invention can be achieved through the following technical solutions:

[0011] A multimodal IoT-based intelligent lifecycle management method for 3C rental devices includes the following steps:

[0012] S1. In response to a detected user handover event, perform a baseline state anchoring operation, wherein the baseline state anchoring operation includes acquiring standardized test data generated when the current user performs an operation according to a preset standardized operation test, and updating the health baseline of the rental equipment based on the standardized test data;

[0013] S2. Based on the health baseline, obtain the current user's touch operation sequence within a preset time window and the rental device's physical response sequence within the time window. Input the touch operation sequence and the device physical response sequence into a dual-branch cross-attention network. The dual-branch cross-attention network extracts the cross-modal correlation features between the touch operation sequence and the device physical response sequence through a cross-attention mechanism, and outputs the cumulative loss contribution value of each key component of the rental device within the time window based on the cross-modal correlation features.

[0014] S3. Trigger an anomaly determination process based on the accumulated loss contribution value. The anomaly determination process includes collecting the inherent physical feature vector of the leased equipment and the real-time operation behavior feature vector of the current user, calculating the deviation between the inherent physical feature vector of the equipment and the real-time operation behavior feature vector through feature space decoupling comparison, and outputting an anomaly attribution determination result based on the deviation. The anomaly attribution determination result is used to characterize the responsible party for the accumulated loss contribution value as either the aging of the leased equipment itself or the operation behavior of the current user.

[0015] A further approach is to detect the user handover event by including:

[0016] Monitor the touch event stream of the rental device and obtain the silent interval duration between adjacent touch events in the touch event stream;

[0017] Monitor the inertial measurement data of the rental equipment, and determine the attitude change state of the rental equipment based on the inertial measurement data;

[0018] When the silence interval exceeds a preset silence duration threshold and the posture change state meets a preset storage posture condition, the rental device is determined to be in a pending handover state. In the pending handover state, the network connection status of the rental device is monitored. When the network connection status changes and a touch event reappears in the touch event stream, the user handover event is determined to have occurred.

[0019] The preset standardized operation test includes a preset set of standard touch action sequences, which include standard sliding trajectory operations and standard click pressure operations. The standardized test data includes touch response delay data and trajectory tracking deviation data of the rental device when performing the standard sliding trajectory operation, and touch pressure response curve data when performing the standard click pressure operation.

[0020] A further approach, updating the health baseline of the leased equipment based on the standardized test data, includes:

[0021] The touch response delay data, the trajectory tracking deviation data, and the touch pressure response curve data are compared with the historical benchmark records of the rental equipment to calculate the drift of each data dimension.

[0022] When the drift amount is less than the preset drift tolerance threshold, the standardized test data is weighted and fused into the historical benchmark record to generate an updated health benchmark baseline.

[0023] When the drift amount is greater than or equal to the drift tolerance threshold, a device abnormality alarm is triggered, and the update operation of the health baseline is suspended until manual confirmation is completed.

[0024] In a further embodiment, the touch operation sequence includes a coordinate sequence of touch events, a sequence of touch pressure values, and a time interval sequence between adjacent touch events; the device physical response sequence includes the accelerometer triaxial output sequence, the gyroscope triaxial angular velocity sequence, the battery current sampling sequence, and the screen driver chip feedback signal sequence of the rental device.

[0025] The dual-branch cross-attention network includes a first encoding branch and a second encoding branch. The first encoding branch is used to encode the touch operation sequence into an operation temporal latent vector, and the second encoding branch is used to encode the device physical response sequence into a physical response temporal latent vector. The cross-attention mechanism is used to calculate cross-modal attention weights using the operation temporal latent vector as a query vector and the physical response temporal latent vector as a key value vector, and to generate a fused latent vector based on the cross-modal attention weights. The fused latent vector is mapped to the cumulative loss contribution value through a fully connected regression layer.

[0026] In a further embodiment, the cumulative loss contribution value includes the screen module loss contribution value, the battery module loss contribution value, and the audio module loss contribution value.

[0027] The screen module loss contribution value characterizes the cumulative loss caused by the touch operation sequence to the screen touch module of the rental equipment;

[0028] The battery module loss contribution value represents the cumulative loss caused to the battery module of the rental equipment by the touch operation sequence and the corresponding device operating load.

[0029] The audio module loss contribution value characterizes the cumulative loss of the audio module caused by the audio output load of the rental device during the execution of the touch operation sequence.

[0030] A further approach involves collecting the inherent physical feature vectors of the leased equipment, including:

[0031] The impedance spectrum characteristics and frequency response offset of the touch screen driver chip of the rental equipment are collected; the zero-bias stability and random walk coefficient of the inertial measurement unit of the rental equipment are collected; and the resonant frequency offset of the speaker of the rental equipment is collected.

[0032] The impedance spectrum characteristics, the frequency response offset, the zero-bias stability, the random walk coefficient, and the resonant frequency offset are concatenated to form the inherent physical characteristic vector of the device;

[0033] The feature vector of the current user's real-time operation behavior collected includes:

[0034] Obtain the real-time touch event sequence of the current user within the preset acquisition window, and extract the sliding trajectory curvature features, touch pressure distribution features, and touch interval rhythm features based on the real-time touch event sequence;

[0035] Obtain the real-time device grip posture sequence of the current user within the preset acquisition window, and extract grip tilt angle stability features and grip position offset features based on the real-time device grip posture sequence;

[0036] The sliding trajectory curvature feature, the touch pressure distribution feature, the touch interval rhythm feature, the grip tilt angle stability feature, and the grip position offset feature are concatenated to form the real-time operation behavior feature vector.

[0037] A further approach involves calculating the deviation between the device's inherent physical feature vector and the real-time operational behavior feature vector using a feature space decoupling ratio, including:

[0038] The inherent physical feature vector of the device is projected onto a preset orthogonal subspace of physical features to obtain a physical subspace projection vector. The real-time operation behavior feature vector is projected onto a preset orthogonal subspace of behavior features to obtain a behavior subspace projection vector.

[0039] The physical feature orthogonal subspace and the behavioral feature orthogonal subspace are orthogonal to each other in the feature space. The Mahalanobis distance between the projection vector of the physical subspace and the projection vector of the behavioral subspace is calculated, and the Mahalanobis distance is used as the deviation.

[0040] The abnormal attribution determination result output based on the deviation includes:

[0041] When the deviation is less than a preset first deviation threshold, the anomaly attribution result is determined to be mainly due to the aging of the equipment itself.

[0042] When the deviation is greater than a preset second deviation threshold, the anomaly attribution determination result is determined to be dominated by the current user operation behavior, and the second deviation threshold is greater than the first deviation threshold;

[0043] When the deviation is between the first deviation threshold and the second deviation threshold, the anomaly attribution determination result is determined to be dominated by mixed factors, and the device aging ratio and the user operation behavior ratio are output.

[0044] In a further embodiment, the method also includes:

[0045] Step S7: Obtain the cumulative loss contribution value and the abnormal attribution determination result, and generate the dynamic depreciation coefficient of the leased equipment based on the cumulative loss contribution value and the abnormal attribution determination result;

[0046] Step S8: Adjust the current rental pricing strategy and maintenance scheduling strategy of the leased equipment according to the dynamic depreciation factor.

[0047] A multimodal perception-based intelligent lifecycle management and control system for leased equipment is provided, wherein the system is used to execute the multimodal perception-based intelligent lifecycle management and control method for leased equipment.

[0048] The beneficial effects of this invention are:

[0049] (1) This invention detects user handover events and performs baseline state anchoring operations. It updates the health baseline based on standardized operation tests each time a user switches, so that the monitoring baseline can be dynamically adjusted to follow the flow of the device between different users. This solves the problem of mismatch of fixed baselines in multi-user alternating use scenarios and helps to improve the accuracy of device health status tracking.

[0050] (2) This invention collects touch operation sequences and device physical response sequences, uses a dual-branch cross-attention network to extract cross-modal correlation features between the two, and outputs the cumulative loss contribution value of each key component accordingly. It can establish a mapping relationship between user operation mode and the loss of specific components, so that the device health prediction is related to the user's actual usage behavior, and provides a basis for the equipment maintenance decision of the rental platform.

[0051] (3) When the cumulative loss contribution value is detected and the judgment condition is triggered, the present invention collects the inherent physical feature vector of the equipment and the real-time operation behavior feature vector of the current user, and performs decoupling comparison in the feature space to calculate the deviation between the two. Based on this, the abnormal attribution judgment result is output, which can distinguish the responsibility for equipment loss as being dominated by its own aging or by user operation behavior. This provides technical support for equipment depreciation calculation and user responsibility division in the leasing scenario and helps to improve the intelligence level of leasing operation. Attached Figure Description

[0052] The invention will now be further described with reference to the accompanying drawings.

[0053] Figure 1 This is a diagram illustrating the method steps of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Example 1:

[0056] Please see Figure 1 As shown in the illustration, this embodiment will describe in detail each step of the IoT-based 3C rental device full lifecycle monitoring method provided in this application. The method is illustrated using a high-performance laptop and its accompanying portable projector rented in a conference scenario as a specific application example. However, those skilled in the art will understand that this method is also applicable to various types of 3C rental devices such as smartphones, tablets, servers, VR headsets, and network routers.

[0057] Step S1: Multi-source sensor data acquisition and preprocessing;

[0058] This step aims to acquire multi-dimensional raw data that comprehensively reflects the operating status of the equipment in real time and accurately by using multi-source sensors and built-in smart hardware interfaces deployed on 3C rental equipment.

[0059] Specifically, take a rented high-performance laptop as an example. This device has multiple built-in or external sensing units:

[0060] Temperature sensing: The device temperature data is collected through the CPU core temperature sensor (such as DTS digital temperature sensor) on the motherboard and the NTC thermistor in the battery area. The sampling frequency is set to 1Hz.

[0061] Power consumption and current sensing: The real-time power consumption (unit: W) and battery charging and discharging current (unit: mA) of the whole machine are read through the battery management chip BMS and the motherboard power management unit PMU, and the sampling frequency is set to 10Hz.

[0062] Vibration and Attitude Sensing: The device's built-in six-axis inertial measurement unit (IMU), which includes a three-axis accelerometer and a three-axis gyroscope, collects the vibration acceleration (unit: m / s²) of the device during transportation or use. 2 The sampling frequency is set to 100Hz to monitor whether the device has been dropped or subjected to abnormal impacts.

[0063] Location data: The device's real-time latitude and longitude coordinates are obtained through its built-in Wi-Fi scanning module and GPS global positioning system chip, combined with Wi-Fi fingerprint positioning or satellite positioning technology, with an update frequency set to 1Hz. In indoor scenarios, Bluetooth AoA (Aspect-of-Arrival) positioning base stations can be used to further improve accuracy.

[0064] The raw digital signals output by the aforementioned sensors and chips are transmitted to the cloud server via the device's built-in IoT communication module, such as NB-IoT, 4G / 5G cellular module, or Wi-Fi module. Before transmission, the device-side embedded management agent performs preliminary data preprocessing. Preprocessing includes: applying median filtering (window size 5) to the acceleration data to remove impulse noise; and performing maximum-minimum normalization on all data to eliminate the influence of different sensor data due to different units, mapping the numerical range of all feature data to the [0, 1] interval. The preprocessed data is sent to the cloud server in the form of JSON data packets via the Transport Layer Security (TLS) protocol. Each data packet contains a unique device identifier (such as IMEI or MAC address), tenant ID, timestamp (accurate to milliseconds), and normalized temperature, power consumption, vibration acceleration, and location data.

[0065] Through the multi-source sensor fusion and preprocessing described above for 3C devices, this method can provide a high signal-to-noise ratio and standardized data foundation for subsequent analysis. In particular, the monitoring of power consumption and attitude can effectively capture the unique fault precursors of 3C devices, such as battery aging and accidental drops, comprehensively improving the expressive power of state data.

[0066] Step S2: Multidimensional feature extraction based on CNN;

[0067] This step aims to extract deep features from the preprocessed real-time status data and construct a multi-dimensional feature vector set that can characterize the essential operating state of 3C rental equipment.

[0068] Specifically, after receiving data streams from multiple rental devices, the cloud server inputs them into a pre-trained convolutional neural network (CNN) feature extraction model. The input layer of this CNN model receives a tensor of shape (T, C), where T represents the length of the time window, for example, containing the past 60 sampling points, corresponding to 60 seconds of data, and C represents the number of data channels, which is 5 in this embodiment, namely temperature, power consumption, X-axis acceleration, Y-axis acceleration, and Z-axis acceleration.

[0069] The model contains three parallel feature extraction branches:

[0070] Temperature and Power Consumption Branch: Temperature and power consumption data are combined and input into a subnetwork containing two one-dimensional convolutional layers. The first layer has a kernel size of 5 and 32 output channels; the second layer has a kernel size of 3 and 64 output channels. This branch is used to extract the coupling pattern between the device's thermal state and energy consumption changes, such as the anomalous trend of a sudden temperature rise under high power consumption.

[0071] The three-axis acceleration branch takes X, Y, and Z axis acceleration data as a three-channel input, passes it through a convolutional layer with a kernel size of 3 and 32 output channels, followed by a ReLU activation function and a max-pooling layer with a stride of 2. This branch is used to extract the motion impact features of the device in space, such as short-term, high-frequency events like free fall and collisions.

[0072] The feature maps output from the three branches are concatenated in a fusion layer. This concatenated feature is then fed into a feature fusion network consisting of two fully connected layers with 128 and 64 neurons, respectively; a Dropout operation is introduced between the two layers with a dropout rate of 0.25. Finally, the network outputs a 64-dimensional feature vector. This feature vector, along with the current location data (longitude, latitude) and rental period (e.g., weekday / weekend, daytime / nighttime encoding), constitutes a multi-dimensional feature vector set.

[0073] By using deep feature fusion of CNN models, this method can automatically identify physical coupling anomalies of power consumption and temperature in 3C devices as well as the risk of physical damage from acceleration impacts, effectively solving the limitation of traditional IT operations and maintenance relying solely on single log data for fault diagnosis.

[0074] Step S3: Anomaly state identification based on random forest and spatiotemporal clustering;

[0075] This step aims to utilize the multi-dimensional feature vector set constructed in the preceding steps to accurately determine whether the leased equipment is in an abnormal state through a triple mechanism of dynamic threshold triggering, random forest model classification, and spatiotemporal clustering analysis.

[0076] Dynamic statistical threshold prediction:

[0077] Before inputting the feature vectors into the random forest model, this method first introduces a dynamic statistical feature set for predicting anomaly warning points. Specifically, the cloud server maintains a data cache based on a sliding time window for each type of feature data for each device, such as vibration feature values ​​and temperature feature values. The window length is set to 10 minutes. Within each window, the system calculates the mean μ and standard deviation σ of that type of feature data in real time, thereby constructing a dynamic statistical feature set.

[0078] When a newly acquired feature value x new Upon arrival, the system calculates the degree of deviation using a standardized scoring formula: z=(x new -μ) / σ. The system presets an anomaly trigger threshold, for example, z > 4.0 or z < -4.0. If the currently calculated z value exceeds this threshold, the data point is determined to be an anomaly warning point. This anomaly warning result is used as a binary label, 1 for an anomaly warning and 0 for no warning, as prior judgment information, and is input into the subsequent random forest model along with other feature vectors.

[0079] Random forest classification and spatiotemporal clustering confirmation:

[0080] First, a multi-dimensional feature vector set incorporating dynamic anomaly warning labels is input into a pre-trained random forest model. This model consists of 500 decision trees and is trained using a training dataset containing historical operational data and corresponding status labels (normal, mild anomaly, severe anomaly) of leased equipment at various usage stages, such as startup, heavy load, idle, and shutdown.

[0081] The random forest model, based on the input feature data, outputs the predicted probability of a device's state belonging to each preset category through voting among all decision trees. For example, if the output is: normal probability 10%, mildly abnormal probability 20%, and severely abnormal probability 70%, then the system will mark that time point as severely abnormal.

[0082] To further filter out sporadic false alarms and identify anomalous events with actual physical significance, this method introduces a spatiotemporal clustering algorithm. The system aggregates data points marked as anomalous based on their occurrence time (i.e., consecutive time points) and spatial location (GPS coordinates), using a density-based clustering algorithm such as DBSCAN. The neighborhood radius Eps of the DBSCAN algorithm is set to 50 meters, and the minimum sample size MinPts is set to 5 data points. This algorithm groups spatially adjacent and temporally consecutive anomalous points into the same cluster.

[0083] Cluster analysis identifies whether abnormal behavior exists in a group setting (e.g., multiple devices malfunctioning simultaneously in the same area), a region (e.g., devices in a specific geographical location consistently malfunction), or a phase (e.g., devices continuously malfunctioning for 10-15 minutes after startup), forming an abnormal pattern within a spatially continuous area or time period. Finally, the system compares this identified abnormal pattern with common behavioral patterns of the devices in historical data, calculating the similarity. If the similarity is below a preset first threshold, for example, <40%, the abnormal pattern is considered rare and atypical, and thus ultimately marked as a confirmed abnormal state. This confirmation result is then passed to step S4.

[0084] By combining dynamic thresholding, random forest, and spatiotemporal clustering, this method achieves anomaly identification from point to surface. Dynamic thresholding ensures the model's adaptability to changes in operating conditions, random forest provides powerful nonlinear classification capabilities, and spatiotemporal clustering can confirm anomalies from a spatiotemporal dimension, eliminating isolated noise points and identifying genuine group or regional risks, thereby enhancing the accuracy of anomaly detection.

[0085] Step S4: Fault probability prediction based on LSTM and baseline model;

[0086] This step aims to predict the probability of future failures after confirming that the equipment is in an abnormal state, by combining the anomaly detection results, historical operating data, and baseline models. The specific implementation process of this step is as follows:

[0087] Step 1: Construct a time-series input matrix containing abnormal state identifiers;

[0088] First, for example, the historical operating data of the leased equipment is sliced ​​using a fixed time window of the past 60 minutes. To construct a multi-dimensional time-series input matrix, the confirmed anomaly detection results are treated as an explicit state variable and combined with other multi-dimensional features such as vibration characteristics, temperature characteristics, and pressure characteristics.

[0089] Each row of the time series input matrix corresponds to a time segment (e.g., every second), and each column corresponds to a specific feature dimension, including but not limited to: vibration feature values, temperature feature values, pressure feature values, and abnormal state labels (0 / 1). The row and column structure of the matrix ensures the sequential nature of the time series and the integrity of the feature dimensions, providing standardized data input for time series modeling of the LSTM model.

[0090] Step 2: Input the time series input matrix into the LSTM model for feature extraction;

[0091] The multidimensional temporal input matrix constructed in the first step is input into a pre-built and trained Long Short-Term Memory (LSTM) network model. This LSTM model contains two layers of LSTM units, with 128 units in the first layer and 64 units in the second layer. Dropout of 0.3 is set between the two layers to prevent overfitting.

[0092] LSTM models, through their unique gating mechanisms (input gate, forget gate, output gate) and memory units, automatically learn and extract long-term dynamic features contained in the input matrix. This process does not rely on manually set statistical rules, but rather captures the evolution trend and dependencies of device operating data over a long period of time through the recurrent connections within the model, ultimately generating a hidden state vector that can represent the current operating state of the device.

[0093] Step 3: Integrate the baseline model comparison results to form a comprehensive feature vector;

[0094] While the LSTM model is extracting temporal features, the system simultaneously calls a preset baseline model. This baseline model records the health parameter ranges of the rental equipment of this model at different operating stages (e.g., operating time 0-500 hours, 500-1000 hours), such as the normal range of vibration frequency, the normal range of operating temperature, and the normal fluctuation range of pressure.

[0095] The system compares the device operating parameters within the current time window with the corresponding health parameter intervals in the baseline model, calculating the deviation of each parameter from the health interval, for example, using Mahalanobis distance or standardized residuals. This deviation calculation result serves as a quantified health deviation feature.

[0096] Subsequently, the system concatenates or fuses the hidden state vector (representing temporal dynamic features) output by the LSTM model in the second step with the health deviation features (representing physical deviation from the health benchmark) generated here to form a comprehensive feature vector.

[0097] Step 4: Predict the failure probability based on the comprehensive feature vector;

[0098] The combined feature vector generated in the third step is input into the fully connected output layer at the top of the LSTM model. This output layer contains one neuron and uses the sigmoid function as the activation function.

[0099] The Sigmoid function maps the linear output of a fully connected layer to the interval [0, 1]. This output value is then quantified as the probability of device failure during the prediction period (e.g., the next 24 hours). This process can be represented as:

[0100] Failure probability = Sigmoid(W * comprehensive feature vector + b);

[0101] Where W is the weight matrix of the output layer and b is the bias term, both of which are optimized during the model training phase using the backpropagation algorithm.

[0102] This probability value combines the temporal evolution of the equipment's own operating status (captured by LSTM) and the degree of physical deviation from the health benchmark (provided by the baseline model), thereby achieving a combination of data-driven and knowledge-driven approaches, which can improve the accuracy and interpretability of fault prediction.

[0103] Step 5: Output and apply the fault prediction results;

[0104] Finally, the LSTM model outputs a failure probability value in the range [0, 1]. The system compares this probability value with a preset risk level threshold; for example, low risk < 0.3, medium risk 0.3-0.7, and high risk > 0.7, to trigger different levels of warnings or maintenance recommendations.

[0105] For example, if the calculated failure probability value is 0.85, the system determines that the equipment is in a high-risk state, automatically generates an early warning message containing the failure probability value, main deviation parameters and suggested maintenance actions, and pushes it to the management platform to provide direct and quantitative basis for the operation and maintenance personnel's decision-making.

[0106] Step S5: Trajectory deviation warning based on Kalman filtering and semantic fence;

[0107] This step aims to predict the future location of 3C devices based on location data and perform spatial compliance checks through dynamic semantic fences to provide early warnings of trajectory deviations; for example, to prevent devices from being taken out of the rental area.

[0108] The real-time acquired location data (latitude and longitude) is input into a Kalman filter (using a constant velocity model) and outputs a set of trajectory point predictions for multiple future moments (e.g., the next 10 seconds).

[0109] The cloud server dynamically generates semantic fences for devices based on high-precision map data. Taking a rental park as an example:

[0110] Constructing coarse boundaries: Calculate the bounding polygon of the active range based on the device's historical location data.

[0111] Extracting spatial constraint factors: Overlay park DEM data and building outline images to extract spatial constraint factors such as office buildings, walls, and water areas.

[0112] Identify restricted areas: Automatically mark the interior of buildings (non-rental service areas), the exterior of the park perimeter wall, etc., as inaccessible areas.

[0113] Generate restricted polygons: Analyze obstacle boundaries, combine historical travel trajectories, generate restricted polygons that allow device movement, and integrate the internal restricted areas as nested holes.

[0114] The system performs topological analysis on the trajectory points predicted by Kalman filtering and semantic fences. If a predicted trajectory point falls into a hole or exceeds the main boundary, a trajectory deviation warning notification is immediately generated.

[0115] Steps S6-S7: Visual monitoring and analysis based on heatmaps;

[0116] The system calculates the shortest Euclidean distance between the predicted offset point and the fence boundary as the degree of deviation and correlates it with the failure probability value. Two types of dynamic layers are generated on the management platform using heatmap rendering technology:

[0117] Trajectory Deviation Distribution Heatmap: Shows the degree to which each device is currently deviating from the safe zone.

[0118] Failure probability distribution heatmap: Shows the clustering areas of equipment with high failure risk.

[0119] Maintenance personnel can intuitively see which office building areas have high equipment heat dissipation risks, or which entrances and exits have a tendency for equipment to be illegally taken out, achieving visualized intelligent management and control throughout the entire life cycle.

[0120] Example 2:

[0121] This application also provides an IoT-based full lifecycle monitoring system for leased equipment, used to perform the above-described methods. The system includes:

[0122] The acquisition module is used to acquire real-time status data through multi-source sensors.

[0123] The extraction module is used to extract features from the real-time status data to obtain a multi-dimensional feature vector set.

[0124] The anomaly detection module is used to input the multidimensional feature vector set into the random forest model and combine it with the spatiotemporal clustering algorithm to determine whether the rental equipment is in an abnormal state.

[0125] The fault diagnosis module, when judged to be in an abnormal state, combines the abnormal detection results, historical operating data and the preset baseline model to predict the fault of the leased equipment and obtain the fault probability value.

[0126] The location prediction module predicts the future location of the rental equipment based on location data, sets electronic fence boundaries, and generates trajectory deviation warning notifications.

[0127] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for intelligent lifecycle management and control of 3C rental equipment using multimodal IoT, characterized in that, Includes the following steps: S1. In response to a user handover event, perform a baseline state anchoring operation, obtain standardized test data generated when the current user performs a test according to a preset standardized operation, and update the health baseline of the leased equipment based on the standardized test data. S2. Based on the health baseline, obtain the current user's touch operation sequence within a preset time window and the rental device's physical response sequence within the time window. Input the touch operation sequence and the device physical response sequence into a dual-branch cross-attention network, extract cross-modal correlation features through the cross-attention mechanism, and output the cumulative loss contribution value of each key component based on the cross-modal correlation features. S3. Trigger the anomaly determination process based on the accumulated loss contribution value, collect the inherent physical feature vector of the leased equipment and the real-time operation behavior feature vector of the current user, calculate the deviation between the two through feature space decoupling comparison, and output the anomaly attribution determination result based on the deviation. The anomaly attribution determination result is used to characterize whether the responsible party is the aging of the leased equipment itself or the operation behavior of the current user.

2. The method for intelligent lifecycle management of 3C rental equipment in a multimodal Internet of Things according to claim 1, characterized in that, Detecting the user handover event includes: Monitor the touch event stream of the rental device and obtain the silent interval duration between adjacent touch events in the touch event stream; Monitor the inertial measurement data of the rental equipment, and determine the attitude change state of the rental equipment based on the inertial measurement data; When the silence interval exceeds a preset silence duration threshold and the posture change state meets a preset storage posture condition, the rental device is determined to be in a pending handover state. In the pending handover state, the network connection status of the rental device is monitored. When the network connection status changes and a touch event reappears in the touch event stream, the user handover event is determined to have occurred. The preset standardized operation test includes a preset set of standard touch action sequences, which include standard sliding trajectory operations and standard click pressure operations. The standardized test data includes touch response delay data and trajectory tracking deviation data of the rental device when performing the standard sliding trajectory operation, and touch pressure response curve data when performing the standard click pressure operation.

3. The method for intelligent lifecycle management of 3C rental equipment in a multimodal Internet of Things according to claim 2, characterized in that, Updating the health baseline based on the standardized test data includes: The touch response delay data, the trajectory tracking deviation data, and the touch pressure response curve data are compared with the historical benchmark records of the rental equipment to calculate the drift of each data dimension. When the drift amount is less than the preset drift tolerance threshold, the standardized test data is weighted and fused into the historical benchmark record to generate an updated health benchmark baseline. When the drift amount is greater than or equal to the drift tolerance threshold, a device abnormality alarm is triggered, and the update operation of the health baseline is suspended until manual confirmation is completed.

4. The method for intelligent lifecycle management of 3C rental equipment in a multimodal Internet of Things according to claim 1, characterized in that, The touch operation sequence includes a coordinate sequence of touch events, a sequence of touch pressure values, and a time interval sequence between adjacent touch events. The device physical response sequence includes the accelerometer triaxial output sequence, the gyroscope triaxial angular velocity sequence, the battery current sampling sequence, and the screen driver chip feedback signal sequence of the rental device. The dual-branch cross-attention network includes a first encoding branch and a second encoding branch. The first encoding branch is used to encode the touch operation sequence into an operation temporal latent vector, and the second encoding branch is used to encode the device physical response sequence into a physical response temporal latent vector. The cross-attention mechanism is used to calculate cross-modal attention weights using the operation temporal latent vector as a query vector and the physical response temporal latent vector as a key value vector, and to generate a fused latent vector based on the cross-modal attention weights. The fused latent vector is mapped to the cumulative loss contribution value through a fully connected regression layer.

5. The method for intelligent lifecycle management of 3C rental equipment in a multimodal Internet of Things according to claim 4, characterized in that, The cumulative loss contribution value includes the screen module loss contribution value, the battery module loss contribution value, and the audio module loss contribution value. The screen module loss contribution value characterizes the cumulative loss caused by the touch operation sequence to the screen touch module of the rental equipment; The battery module loss contribution value represents the cumulative loss caused to the battery module of the rental equipment by the touch operation sequence and the corresponding device operating load. The audio module loss contribution value characterizes the cumulative loss of the audio module caused by the audio output load of the rental device during the execution of the touch operation sequence.

6. The method for intelligent lifecycle management of 3C rental equipment in a multimodal Internet of Things according to claim 5, characterized in that, The acquisition of the device's inherent physical feature vector includes: acquiring the impedance spectrum characteristics and frequency response offset of the touch screen driver chip, the zero-bias stability and random walk coefficient of the inertial measurement unit, and the resonant frequency offset of the speaker, and concatenating them into the device's inherent physical feature vector; The acquisition of the real-time operation behavior feature vector includes: extracting the sliding trajectory curvature features, touch pressure distribution features, and touch interval rhythm features from the real-time touch event sequence, as well as the grip tilt angle stability features and grip position offset features from the real-time device grip posture sequence, and concatenating them into the real-time operation behavior feature vector.

7. The method for intelligent lifecycle management of 3C rental equipment in a multimodal Internet of Things according to claim 6, characterized in that, The deviation is calculated by using the feature space decoupling ratio, including: The inherent physical feature vector of the device is projected onto a preset physical feature orthogonal subspace, and the real-time operation behavior feature vector is projected onto a preset behavior feature orthogonal subspace. The physical feature orthogonal subspace and the behavior feature orthogonal subspace are orthogonal to each other. The Mahalanobis distance between the two projection vectors is calculated as the deviation. The abnormal attribution determination result output based on the deviation degree includes: when the deviation degree is less than the first deviation degree threshold, it is determined that the device itself is aging-driven; when the deviation degree is greater than the second deviation degree threshold, it is determined that the current user operation behavior is driven by the deviation degree; when the deviation degree is between the two, it is determined that a mixture of factors is driven by the deviation degree and the respective proportions are output.

8. The method for intelligent lifecycle management of 3C rental equipment in a multimodal Internet of Things according to claim 1, characterized in that, The method further includes: Step S7: Obtain the cumulative loss contribution value and the abnormal attribution determination result, and generate the dynamic depreciation coefficient of the leased equipment based on the cumulative loss contribution value and the abnormal attribution determination result; Step S8: Adjust the current rental pricing strategy and maintenance scheduling strategy of the leased equipment according to the dynamic depreciation factor.

9. A smart lifecycle management system for leased equipment based on multimodal perception, characterized in that: The system is used to execute the multimodal Internet of Things (IoT) intelligent lifecycle management method for 3C rental equipment as described in any one of claims 1 to 8.

Images