Shared power bank anti-theft and state monitoring system based on AI multi-modal recognition

Abstract

The application discloses a shared power bank anti-theft and state monitoring system based on AI multi-modal recognition, relates to the technical field of shared technical services, and constructs a three-dimensional hardware fingerprint joint detection system of a PD protocol communication digital domain, a voltage ripple simulation domain and a chip response delay time sequence domain, and introduces an abnormal communication flow detection of a time sequence convolution network and a graph attention network hybrid architecture, realizes intelligent identification of various theft means such as protocol fraud, chip replacement and violent damage, extracts a digital domain feature vector, a simulation domain feature vector and a time sequence physical fingerprint vector respectively, fuses the digital-analog dual-domain fingerprint generation module into a dual-domain fingerprint vector F, outputs a fine-grained abnormal probability score S by the abnormal communication flow detection module, executes a hierarchical response, improves the robustness and adaptability of the anti-theft system, and provides reliable technical support for asset safe operation of the shared power bank industry.

Description

Technical Field

[0001] This invention relates to the field of shared technology service technology, specifically to a shared power bank anti-theft and status monitoring system based on AI multimodal recognition. Background Technology

[0002] Shared power banks, as a convenient mobile energy replenishment solution, have been widely deployed in public areas such as shopping malls, restaurants, transportation hubs, and tourist attractions. With the rapid expansion of the shared power bank market, security issues such as power bank theft, malicious replacement, and falsification of return status through protocol deception have become increasingly prominent, causing significant economic losses to operators.

[0003] Current anti-theft technologies for shared power banks mainly rely on mechanical locking structures and simple electrical continuity detection. For example, by installing electromagnetic locks or physical latches in each compartment of the charging cabinet and detecting changes in DC resistance or current between the contacts after the power bank is inserted to determine whether it has been returned properly, this approach has significant drawbacks: the mechanical structure is easily damaged by force or bypassed by prying it with foreign objects; the electrical detection methods are too simplistic, and attackers can use simple electronic simulators to generate false insertion signals, misleading the system into believing that the power bank has been safely stored, when in fact it has been illegally possessed.

[0004] Existing technologies attempt to detect abnormal vibrations by introducing single-mode sensors, such as accelerometers to detect abnormal vibrations and Hall effect sensors to detect magnetic attraction. However, these methods lack in-depth analysis of the complex interaction between the power bank and the charging cabinet, and fail to leverage the advantages of shared power banks. The fast charging protocol communication process performs identity authentication and behavior monitoring. Attackers can bypass traditional anti-theft mechanisms by replaying the protocol handshake packets of a normal power bank, simulating the response timing of the PD protocol chip, or even replacing the entire PD protocol chip to forge normal charging behavior.

[0005] Therefore, there is an urgent need for a shared power bank anti-theft and status monitoring system that can deeply integrate the digital communication characteristics of the PD protocol with the analog physical characteristics and use artificial intelligence models to identify abnormal behavior. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an AI-based multimodal recognition-based anti-theft and status monitoring system for shared power banks. This system constructs a three-dimensional hardware fingerprint joint detection system encompassing the digital domain of PD protocol communication, the analog domain of voltage ripple, and the temporal domain of chip response delay. It also introduces an abnormal communication traffic detection system using a hybrid architecture of temporal convolutional networks and graph attention networks, enabling intelligent identification of various theft methods such as protocol spoofing, chip replacement, and brute-force damage. By extracting feature vectors from the digital domain, analog domain, and temporal physical fingerprint vectors respectively, and fusing them into a dual-domain fingerprint vector F via a digital-analog dual-domain fingerprint generation module, the abnormal communication traffic detection module outputs a fine-grained abnormal probability score S, which executes a graded response. This improves the robustness and adaptability of the anti-theft system, providing reliable technical protection for the secure operation of assets in the shared power bank industry, and has broad market application prospects and significant economic benefits.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a shared power bank anti-theft and status monitoring system based on AI multimodal recognition, the system comprising: a PD protocol communication monitoring module, a voltage ripple acquisition circuit, a chip ID response delay measurement module, a digital-analog dual-domain fingerprint generation module, and an abnormal communication traffic detection module;

[0008] The PD protocol communication monitoring module is located inside the microcontroller of the shared power bank and the PD main control chip of the charging cabinet, and is used to capture the configuration channel CC line in real time. The protocol's communication data packets output digital domain feature vectors. ;

[0009] The voltage ripple acquisition circuit is connected in parallel to the battery terminal of the shared power bank and the VBUS line of the PD protocol chip. It includes an analog-to-digital converter and a bandpass filter, and is used to acquire voltage ripple signals during charging negotiation and discharging. It extracts the amplitude distribution, the position of the main peak of the spectrum, and the ripple noise floor of the ripple, and outputs an analog domain feature vector. ;

[0010] The chip ID response delay measurement module is integrated into the PD main control chip of the charging cabinet. It is used to send challenge-response commands to the PD protocol chip and measure the time delay from command issuance to response return, outputting a timing physical fingerprint vector. ;

[0011] The digital-analog dual-domain fingerprint generation module is used to receive digital domain feature vectors. , eigenvectors of the analog domain Time-series physical fingerprint vector , to transform the digital domain feature vector , eigenvectors of the analog domain Time-series physical fingerprint vector After joint embedding, a fixed-length dual-domain fingerprint vector is generated. ;

[0012] The abnormal communication traffic detection module is based on a hybrid architecture of temporal convolutional networks and graph attention networks. The input is a sequence of protocol communication flows during multiple return or lending processes. Each time step in the protocol communication flow sequence contains a corresponding digital domain feature vector. , eigenvectors of the analog domain Time-series physical fingerprint vector Output an anomaly probability score for the current communication behavior. And execute a tiered response.

[0013] Furthermore, the digital domain feature vector It includes the sequence encoding of communication data packets, the average time interval between adjacent communication data packets, and manufacturer-fixed field values;

[0014] The communication data packets include handshake protocol packet types, timing stamps for packet transmission and response, and vendor-defined extension messages.

[0015] Furthermore, the PD protocol communication monitoring module records the data packet sequence each time the shared power bank establishes a physical connection with the charging cabinet: Source_Capabilities data packet, Request data packet, Accept data packet, PS_RDY data packet, and simultaneously records the time interval between adjacent data packets. and number of retransmissions ;

[0016] The PD protocol communication monitoring module compares timing characteristics with the expected timing range in the standard protocol specification. When the time interval of the data packets... Deviation from preset threshold by more than 20% or number of retransmissions If the value is greater than 2, the PD protocol communication monitoring module generates a digital domain anomaly marker and incorporates the digital domain anomaly marker into the digital domain feature vector. middle.

[0017] Furthermore, the voltage ripple acquisition circuit continuously acquires the ripple voltage on the VBUS line at a sampling frequency of 1MHz. The power spectral density is obtained by fast Fourier transform. The voltage ripple acquisition circuit extracts the main peak frequency of the spectrum. Main peak amplitude and the average noise floor in the 100kHz-1MHz frequency band. As the feature vector of the simulation domain When the PD protocol chip in a shared power bank changes its internal switching frequency due to aging or replacement, the main peak frequency of the spectrum... offset >A preset threshold of 0.5MHz will trigger an alarm for an abnormality in the analog domain.

[0018] Furthermore, after the chip ID response delay measurement module sends a challenge command to the PD protocol chip of the shared power bank to read the chip's unique identifier, it uses a timer inside the PD main control chip of the charging cabinet to record the delay time. ,in, The time when the challenge command is sent, In response to the moment the instruction is received;

[0019] The chip ID response delay measurement module continuously measures the delay 10 times and calculates the mean delay and the standard deviation of the delay: Where k is the measurement sequence number;

[0020] The average delay With the aforementioned delay standard deviation Together they constitute the temporal physical fingerprint vector ;

[0021] The time-series physical fingerprint vector Sensitive to internal logic gate delays and temperature, any replacement of the PD protocol chip will affect the average delay. It produces a drift of more than ±5%.

[0022] Furthermore, the digital-analog dual-domain fingerprint generation module generates a dual-domain fingerprint vector. The steps are as follows:

[0023] Receive the digital domain feature vector from the PD protocol communication monitoring module. ,in: , for The protocol encodes the sequence of communication data packets. This represents the average time interval between adjacent communication data packets. Manufacturer-defined field values;

[0024] Receive the analog domain feature vector from the voltage ripple acquisition circuit. ;

[0025] Receive the timing physical fingerprint vector from the chip ID response delay measurement module. ;

[0026] The digital domain feature vector The simulated domain feature vector and the time-series physical fingerprint vector Concatenate into a joint feature vector The dual-domain fingerprint vector is obtained by performing nonlinear mapping through a pre-trained lightweight residual network. ,in, The weight matrix is ​​a learnable matrix. , Here, ReLU is the bias vector, and ReLU is the linear rectified activation function.

[0027] The dual-domain fingerprint vector The digital-analog dual-domain fingerprint generation module generates the dual-domain fingerprint vector with 128 dimensions. Shared power banks with factory fingerprint storage in the cloud A cosine similarity comparison is performed, and a fingerprint mismatch is determined when the cosine similarity is less than 0.85.

[0028] Furthermore, the abnormal communication traffic detection module consists of a temporal convolutional network and a graph attention network connected in parallel, which are then fused to output an abnormal probability score. The detection steps of the abnormal communication traffic detection module are as follows:

[0029] Continuous Protocol stream sequence during the next return process Input a temporal convolutional network, where each Includes digital domain feature vectors extracted from a single interaction , eigenvectors of the analog domain Time-series physical fingerprint vector Temporal convolutional networks output temporal dependency features ;

[0030] The relationships between different data packets within the same protocol stream are constructed as a graph structure, where each node represents a protocol packet and the edges represent temporal order. A graph attention network is used to calculate the attention weight of each node and the results are aggregated to obtain the graph structure features. ;

[0031] The time-dependent features With the graph structure features After splicing, an anomaly probability score is output through a fully connected layer. ,in, For the Sigmoid function, This is the weight matrix of the fully connected layer. For bias, This represents a vector concatenation operation. The value range is [0, 1].

[0032] Furthermore, the abnormal communication traffic detection module is based on anomaly probability scoring. Implementing a tiered response includes:

[0033] when When a low-level anomaly response is executed, a dynamic challenge code is sent from the PD main control chip of the charging cabinet to the microcontroller of the shared power bank, requiring the shared power bank to return a correct response within a preset 30 seconds using the built-in encryption unit. If the returned response is incorrect, the response is upgraded to a medium-level anomaly response.

[0034] when At this time, a moderate anomaly response is executed, cutting off the output MOSFET inside the shared power bank and driving the audible and visual alarm on the charging cabinet to sound a buzzer and flash a red light, while simultaneously setting the current anomaly probability score. and the dual-domain fingerprint vector output by the digital-analog dual-domain fingerprint generation module Upload to the cloud management platform;

[0035] when When a high-level anomaly occurs, a lock command is sent to the battery management IC of the shared power bank to disable its charging and discharging functions. The shared power bank's GPS location is then reported to the cloud management platform in real time, and the shared power bank is marked as a stolen device.

[0036] Furthermore, each shared power bank undergoes a registration process before leaving the factory:

[0037] Insert the shared power bank into the dedicated registration cabinet and perform 100 standard tests. The handshake, PD protocol communication monitoring module, voltage ripple acquisition circuit and chip ID response delay measurement module each acquired 100 sets of digital domain feature vectors. , eigenvectors of the analog domain Time-series physical fingerprint vector ;

[0038] The digital-analog dual-domain fingerprint generation module calculates the average of 100 sets of dual-domain fingerprint vectors as the factory fingerprint. and factory fingerprint Simultaneously stored in the cloud management platform database and the encrypted non-volatile memory of the shared power bank;

[0039] On each subsequent return, the digital-analog dual-domain fingerprint generation module calculates the current dual-domain fingerprint vector in real time. and the current dual-domain fingerprint vector With the factory fingerprint A cosine similarity comparison is performed. If the cosine similarity is lower than 0.85, the abnormal communication traffic detection module is triggered to perform in-depth analysis.

[0040] Compared with existing technologies, this AI-based multimodal recognition-based anti-theft and status monitoring system for shared power banks has the following advantages:

[0041] I. This invention improves the identity authentication security and anti-attack capability of shared power banks by constructing a digital-analog dual-domain fingerprint joint detection mechanism. The digital domain feature vector extracted by the PD protocol communication monitoring module contains high-level semantic information such as protocol packet sequence encoding, average data packet time interval, and manufacturer fixed field values. The analog domain feature vector extracted by the voltage ripple acquisition circuit reflects the unique switching frequency and noise floor characteristics of the PD protocol chip inside the power bank. The timing physical fingerprint vector output by the chip ID response delay measurement module describes the physical non-cloning characteristics of the chip-level logic gate delay. The dual-domain fingerprint vector generated by the fusion of the three has uniqueness and non-replicability, solving the security risks of protocol spoofing, replay attacks, and chip cloning.

[0042] Second, this invention introduces an abnormal communication traffic detection module based on a hybrid architecture of temporal convolutional networks and graph attention networks. This module enables dynamic intelligent analysis of protocol flow behavior during the return and lending of shared power banks, reducing false alarm rates and improving the ability to identify complex theft behaviors. The abnormal communication traffic detection module can not only extract temporal dependency features from protocol flow sequences of multiple consecutive interactions, but also capture the implicit patterns of correlation between different data packets within the same protocol flow through graph attention networks. This results in the output of fine-grained anomaly probability scores, which can be used to distinguish between protocol retransmission features caused by mechanical jitter during normal user plugging and unplugging and stable forged flow features generated by attackers using signal simulators.

[0043] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from an examination of the following, or may be learned from the practice of the invention. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0045] Figure 1 This is a data flow diagram of a shared power bank anti-theft and status monitoring system based on AI multimodal recognition.

[0046] Figure 2This is a block diagram of the module composition of the shared power bank anti-theft and status monitoring system based on AI multimodal recognition in an embodiment of the present invention;

[0047] Figure 3 This is a flowchart of voltage ripple acquisition and analog domain feature extraction in an embodiment of the present invention. Detailed Implementation

[0048] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0049] In one basic embodiment of the present invention, an AI-based multimodal recognition-based anti-theft and status monitoring system for shared power banks is provided. This system aims to address the security deficiencies caused by relying solely on mechanical structures or single electrical detection methods in existing technologies. Figure 2 As shown, Figure 2 This is a block diagram of the module composition of the AI ​​multimodal recognition-based shared power bank anti-theft and status monitoring system in this embodiment of the invention. The system includes: a PD protocol communication monitoring module, a voltage ripple acquisition circuit, a chip ID response delay measurement module, a digital-analog dual-domain fingerprint generation module, and an abnormal communication traffic detection module, wherein:

[0050] The PD protocol communication monitoring module is installed inside the microcontroller of the shared power bank and the PD main control chip of the charging cabinet, and is used to capture the configuration channel CC line in real time. The protocol's communication data packets output digital domain feature vectors. ;

[0051] The voltage ripple acquisition circuit is connected in parallel to the battery terminal of the shared power bank and the VBUS line of the PD protocol chip. It includes an analog-to-digital converter and a bandpass filter, and is used to acquire voltage ripple signals during charging negotiation and discharging, extract the amplitude distribution, the position of the main peak of the spectrum, and the ripple noise floor of the ripple, and output an analog domain feature vector. ;

[0052] The chip ID response delay measurement module is integrated into the PD main control chip of the charging cabinet. It is used to send challenge-response commands to the PD protocol chip and measure the time delay from command issuance to response return, and output a timing physical fingerprint vector. ;

[0053] The aforementioned digital-analog dual-domain fingerprint generation module is used to receive digital domain feature vectors. , eigenvectors of the analog domain Time-series physical fingerprint vector , to transform the digital domain feature vector , eigenvectors of the analog domain Time-series physical fingerprint vector After joint embedding, a fixed-length dual-domain fingerprint vector is generated. ;

[0054] The aforementioned abnormal communication traffic detection module is based on a hybrid architecture of temporal convolutional networks and graph attention networks. The input is a sequence of protocol communication flows during multiple return or borrowing processes. Each time step in the protocol communication flow sequence contains a corresponding digital domain feature vector. , eigenvectors of the analog domain Time-series physical fingerprint vector Output an anomaly probability score for the current communication behavior. And execute a tiered response.

[0055] The following is combined with Figure 1 The anti-theft and status monitoring system for shared power banks based on AI multimodal recognition, as described in this embodiment of the invention, is described in detail below. Figure 1 This is a data flow diagram of a shared power bank anti-theft and status monitoring system based on AI multimodal recognition.

[0056] In a preferred embodiment, the PD protocol communication monitoring module is deployed as a software functional module in the firmware of the microcontroller (MCU) built into the shared power bank and the PD main control chip of the charging cabinet, respectively. It captures and analyzes the communication between the shared power bank and the charging cabinet in real time via the CC line (…). Interaction on the configuration channel Protocol communication data packets.

[0057] When a user inserts a shared power bank into a slot in the charging cabinet, the mechanical contacts first connect the VBUS and GND lines, then the CC line establishes a physical connection. The charging cabinet's PD main control chip acts as the power supply (Source). The protocol specification initiates communication. The PD protocol communication monitoring module is designed in a listening mode, which copies and timestamps all data packets passing through the CC line without interfering with normal PD communication. The specific data acquisition process is as follows:

[0058] The PD protocol communication monitoring module first captures the Source_Capabilities data packet sent by the charging cabinet's PD main control chip. This data packet broadcasts the voltage and current levels that the cabinet can provide. The module then captures the Request data packet from the shared power bank's MCU, which contains the requested voltage / current level. Next, the module captures the Accept data packet from the cabinet, indicating acceptance of the request. Finally, it captures the PS_RDY data packet, signifying that power supply is ready and the charging process is about to begin. The module precisely records the time interval between any two adjacent data packets in these four consecutive packets; for example, the time difference between the completion of the Source_Capabilities transmission and the start of Request reception. The time difference from Request to Accept And the time difference from Accept to PS_RDY The PD protocol communication monitoring module is also responsible for monitoring retransmissions that occur during the initial establishment of a physical connection. For example, if the cabinet does not receive a valid Request data packet within a preset 100ms after sending Source_Capabilities, or if the received Request data packet fails to be verified, the PD protocol communication monitoring module records one retransmission. Continue monitoring until the communication link is stable or the number of retransmissions exceeds the threshold.

[0059] The PD protocol communication monitoring module generates a digital domain feature vector. Includes: Communication data packet sequence encoding ( ), the average time interval between adjacent data packets ( ),Right now Manufacturer fixed field values This field is a manufacturer-defined reserved field in the Source_Capabilities data packet. Different batches of power banks or different manufacturers' PD protocol chips embed specific identification codes in this field.

[0060] In a preferred embodiment, such as Figure 3 As shown, Figure 3 This is a flowchart of voltage ripple acquisition and analog domain feature extraction. The voltage ripple acquisition circuit is connected in parallel between the battery terminal of the shared power bank and the VBUS line of the PD protocol chip. It includes a high-speed analog-to-digital converter and an anti-aliasing bandpass filter. During the boost or buck process, the switching action of the DC-DC converter inside the shared power bank generates a small voltage ripple on the VBUS line, i.e. The ripple signal contains information about the chip's internal switching frequency, parasitic parameters, and load characteristics. Once the charging cabinet and shared power bank enter a stable charging or discharging state, the acquisition circuit starts. The bandpass filter first preprocesses the voltage signal on the VBUS line, with a passband range of 50kHz-2MHz, to filter out power frequency interference and extremely high frequency noise, retaining the effective ripple component within the PD protocol chip's operating frequency band. The analog-to-digital converter then samples the filtered ripple voltage at a fixed frequency of 1MHz. Continuous data acquisition is performed with a sampling time of 1 second to obtain multiple discrete sampling points. The acquired time-domain signal is converted to the frequency domain using a Fast Fourier Transform to obtain its power spectral density. Based on the power spectral density, the voltage ripple acquisition circuit further constructs an analog domain feature vector: ,in, This refers to the frequency corresponding to the main peak with the largest amplitude in the spectrum. The amplitude of this main peak reflects the intensity of the ripple. The power spectral density is the average value within the 100kHz-1MHz frequency band, representing the noise floor level of the circuit.

[0061] In addition to digital protocol characteristics and analog ripple characteristics, this system also introduces a timing-based physical fingerprint based on hardware physical characteristics. In a preferred embodiment, the chip ID response delay measurement module relies on a high-resolution timer integrated within the charging cabinet's PD main control chip. Specifically, after communication negotiation is completed, the charging cabinet's PD main control chip actively sends a specific challenge command to the shared power bank's PD protocol chip. In a preferred embodiment, this challenge command requests the reading of the unique chip identifier (UID) stored in the shared power bank's PD protocol chip's internal register. This command follows... The agreement The format is encapsulated to avoid conflicts with standard protocol instructions. The workflow of this chip ID response delay measurement module is as follows:

[0062] The PD main control chip in the rack starts its internal high-resolution timer at the same time as sending the challenge command, recording the moment of transmission. The instruction is driven by the physical layer, transmitted via the CC line, received, parsed, and executed by the shared power bank's PD protocol chip, and then the response data packet containing the UID is transmitted back via the CC line.

[0063] When the rack PD main control chip fully receives the last valid bit of the response command, it captures the current value of the timer and records it as the reception moment. .

[0064] The delay time of this interaction Because the signal transmission delay on the wire is relatively fixed and extremely small, The fluctuation of the value mainly depends on the response time and processing latency of the internal logic gate circuits of the PD protocol chip in the shared power bank. It is highly sensitive to the chip's manufacturing process, power supply voltage, and junction temperature. In order to construct a stable timing physical fingerprint, a single measurement is easily affected by random noise. Therefore, in a preferred embodiment, the chip ID response latency measurement module continuously performs the same challenge-response measurement 10 times during a complete charging or returning interaction. Each measurement is completed within a preset time interval to reduce interference from changes in the external environment.

[0065] After the measurement is completed, the chip ID response delay measurement module uses the 10 delay time values ​​obtained. Calculate its mean and standard deviation : Among them, the mean The standard deviation represents the typical response speed of the chip. Quantify the jitter of the response time, using a common temporal physical fingerprint vector: .

[0066] In a preferred embodiment, the digital-analog dual-domain fingerprint generation module runs on the main control processor of the charging cabinet. Its input is the feature vector output by the PD protocol communication monitoring module, the voltage ripple acquisition circuit, and the chip ID response delay measurement module. The output is a fixed-length, highly condensed dual-domain fingerprint vector F. Specifically, the fusion process is divided into feature splicing and nonlinear mapping.

[0067] For feature concatenation, the digital-analog dual-domain fingerprint generation module receives the digital domain feature vector from the PD protocol communication monitoring module. Receive analog domain feature vectors from voltage ripple acquisition circuit Receive timing physical fingerprint vector from chip ID response delay measurement module Since the three vectors have different dimensions and numerical ranges, the digital-analog dual-domain fingerprint generation module performs independent normalization processing on each feature component. In a preferred embodiment, the Min-Max normalization method is used to linearly map the value of each feature to the interval [0, 1]. The three normalized vectors are then concatenated end-to-end to form a high-dimensional joint feature vector. Its dimensions are 8.

[0068] For nonlinear mappings, linear concatenation cannot uncover the complex deep interactions between features of different modalities. The digital-analog dual-domain fingerprint generation module incorporates a lightweight residual network, which consists of two fully connected layers and one residual connection, jointly connecting the feature vectors. The input is first fed into the first fully connected layer, which contains 64 neurons. The computation process is as follows: ,in yes The weight matrix, It is a 64-dimensional bias vector, and ReLU is a linear rectified activation function, expressed as ReLU. This introduces nonlinearity into the model and suppresses negative values. The input is fed into the second fully connected layer, which contains 128 neurons. The operation process is as follows: ,in yes The weight matrix, It is a 128-dimensional bias vector, and the residual connection will combine the original joint feature vector. After passing through a linear transformation layer (making its dimension 128), and... The two domains are added together, and then subjected to a ReLU activation to obtain the final dual-domain fingerprint vector. ,in, It is a linear transformation matrix of the residual connections. Using residual networks can alleviate the vanishing gradient problem, making the model easier to train and enabling it to learn deeper feature representations. Ultimately, It is a 128-dimensional real vector, which serves as the digital identity identifier of the shared power bank in this interaction. It is temporarily stored in the cache of the cabinet for subsequent comparison and anomaly analysis.

[0069] In a preferred embodiment, the abnormal communication traffic detection module is deployed on a cloud management platform, based on a hybrid architecture of parallel temporal convolutional network (TCN) and graph attention network (GAT), and the input is the protocol communication flow sequence during multiple consecutive returns or loans of the same shared power bank. Each time step It is a composite data packet containing the first... Digital domain feature vector extracted from the second interaction , eigenvectors of the analog domain and temporal physical fingerprint vector The detection steps of this abnormal communication traffic detection module are as follows:

[0070] The communication stream sequence is fed in parallel into two network branches. In the TCN branch, the sequence... Treated as a time series, TCN uses stacked causal convolutional layers and dilated convolutional layers. Causal convolution ensures that the model relies only on information from the past and present when predicting the probability of anomalies at the current moment, while dilated convolution allows TCN to capture long-term dependencies across multiple return behaviors without increasing the number of parameters.

[0071] After multi-layer convolution and pooling, the TCN outputs a feature vector representing the temporal dependency of the entire sequence , the same communication flow sequence is sent to the GAT branch, in which each interaction is regarded as a node in the graph, and the connection relationships (edges) between nodes are constructed according to the natural time sequence, that is, if the -th interaction and the -th interaction are adjacent in time, a directed edge is established between these two nodes, and the direction is from to , a complete return-borrow sequence forms a chain-like graph structure, and when updating the feature of each node (that is, ), the information of its neighbor nodes is selectively aggregated through an attention mechanism, and the attention weight represents when calculating the new feature of node , the importance of the information of its neighbor node . This weight is automatically learned through a learnable attention network, and the output of GAT is a feature vector that integrates graph structure information .

[0072] The abnormal communication traffic detection module combines the output feature vectors of the two branches and to obtain a higher-dimensional joint feature , which is sent to a fully connected layer containing a single output neuron. This layer uses the Sigmoid activation function, and its output is the final abnormal probability score , and the calculation formula is: , wherein, is the Sigmoid function, which compresses the value of the linear output into the (0, 1) interval, so that it has a probability interpretation, the closer it is to 1, the higher the probability that the system determines that the current communication behavior sequence is abnormal.

[0073] In a preferred embodiment, the abnormal communication traffic detection module performs hierarchical response processing based on the abnormal probability score to balance security and user experience, which specifically includes:

[0074] Low-degree abnormal response (0.3 < S ≤ 0.6): the system does not interrupt the service, but performs a lightweight secondary authentication challenge. The PD main control chip of the cabinet sends a dynamic challenge code to the microcontroller of the shared power bank. Within 30 seconds, the power bank uses the internally integrated encryption unit to encrypt the challenge code with the pre-shared key and return the result. If the response is wrong or times out, the system will raise the abnormality level to moderate.

[0075] Moderate anomaly response (0.6 < S ≤ 0.85): The system shall implement physical isolation and on-site alarm measures. The PD main control chip of the cabinet first sends a cut-off signal to the driving circuit of the output MOS tube inside the power bank via the I2C or GPIO interface, immediately disconnecting its external discharging or internal charging circuit to disable the power bank. Meanwhile, the main control chip of the cabinet drives the installed acousto-optic alarm and LED to alert on-site workers. In addition, the current timestamp, anomaly probability score S and dual-domain fingerprint vector F are packaged into an alarm event and uploaded to the cloud management platform for recording and further analysis.

[0076] High anomaly response (S > 0.85): Hardware locking and tracking measures are implemented. The cabinet sends a special locking instruction to the battery management IC on the shared power bank via the 1-wire interface to disable the charging and discharging FETs. The system activates the GPS positioning integrated inside the power bank to obtain real-time latitude and longitude coordinates, and reports the position information to the cloud. The cloud platform marks the unique ID of the power bank as a stolen device and adds it to the global blacklist. Once any network-connected cabinet identifies this ID, it will directly refuse service and report the position.

[0077] In a preferred embodiment, to ensure the reference fingerprint of each power bank, the factory registration process of the shared power bank is performed:

[0078] An operator inserts the shared power bank to be registered into a special high-precision registration cabinet, and the cabinet runs a registration mode. The cabinet performs 100 times of standard USBPD handshakes and charging-discharging simulations with the power bank. During these 100 interactions, the PD protocol communication monitoring module, voltage ripple acquisition circuit and chip ID response delay measurement module are triggered repeatedly, and 100 sets of feature vectors D, A, and T are collected respectively. Then, the digital-analog dual-domain fingerprint generation module on the registration cabinet generates a temporary fingerprint for each interaction, and finally calculates the arithmetic mean of these 100 fingerprint vectors as the factory fingerprint of the device , then, are written to two positions simultaneously: uploaded and stored in a database of the cloud management platform through an encrypted channel; and written to an independent encrypted non-volatile memory on the shared power bank through a secure internal bus. In each subsequent normal return service, the cabinet calculates in real time and compares it with the stored in the cloud. Only when the cosine similarity is not lower than 0.85 and no other high-risk anomalies are triggered, the return is determined to be valid, and the order is settled for the user. Any mismatch below this threshold will activate the above-mentioned abnormal communication traffic detection module to conduct in-depth sequence behavior analysis, so as to confirm whether it is an accidental measurement error or a deliberate attack attempt.

[0079] In summary, the AI-based multimodal recognition-based anti-theft and status monitoring system for shared power banks proposed in this invention deeply integrates multidimensional features from the digital domain, analog domain, and temporal-physical domain, and combines them with an advanced hybrid neural network model to construct a three-dimensional intelligent anti-theft monitoring system that ranges from single-judgment to behavior analysis, providing reliable technical support for the safe operation of assets in the shared power bank industry.

[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A shared power bank anti-theft and status monitoring system based on AI multimodal recognition, characterized in that, The system includes: The PD protocol communication monitoring module is located inside the microcontroller of the shared power bank and the PD main control chip of the charging cabinet, and is used to capture the configuration channel CC line in real time. The protocol's communication data packets output digital domain feature vectors. ; The voltage ripple acquisition circuit is connected in parallel to the battery terminal of the shared power bank and the VBUS line of the PD protocol chip. It includes an analog-to-digital converter and a bandpass filter, and is used to acquire voltage ripple signals during charging negotiation and discharging. It extracts the amplitude distribution, the position of the main peak of the spectrum, and the ripple noise floor of the ripple, and outputs an analog domain feature vector. ; The chip ID response delay measurement module is integrated into the PD main control chip of the charging cabinet. It is used to send challenge-response commands to the PD protocol chip and measure the time delay from command issuance to response return, outputting a timing physical fingerprint vector. ; The digital-analog dual-domain fingerprint generation module is used to receive digital domain feature vectors. , eigenvectors of the analog domain Time-series physical fingerprint vector , to transform the digital domain feature vector , eigenvectors of the analog domain Time-series physical fingerprint vector After joint embedding, a fixed-length dual-domain fingerprint vector is generated. ; The abnormal communication traffic detection module is based on a hybrid architecture of temporal convolutional networks and graph attention networks. The input is a sequence of protocol communication flows during multiple return or lending processes. Each time step in the protocol communication flow sequence contains a corresponding digital domain feature vector. , eigenvectors of the analog domain Time-series physical fingerprint vector Output an anomaly probability score for the current communication behavior. And execute a tiered response.

2. The shared power bank anti-theft and status monitoring system based on AI multimodal recognition according to claim 1, characterized in that, The digital domain feature vector It includes the sequence encoding of communication data packets, the average time interval between adjacent communication data packets, and manufacturer-fixed field values; The communication data packets include handshake protocol packet types, timing stamps for packet transmission and response, and vendor-defined extension messages.

3. The shared power bank anti-theft and status monitoring system based on AI multimodal recognition according to claim 1, characterized in that, The PD protocol communication monitoring module records the data packet sequence each time a shared power bank establishes a physical connection with the charging cabinet: Source_Capabilities data packet, Request data packet, Accept data packet, PS_RDY data packet, and simultaneously records the time interval between adjacent data packets. and number of retransmissions ; The PD protocol communication monitoring module compares timing characteristics with the expected timing range in the standard protocol specification. When the time interval of the data packets... Deviation from preset threshold by more than 20% or number of retransmissions If the value is greater than 2, the PD protocol communication monitoring module generates a digital domain anomaly marker and incorporates the digital domain anomaly marker into the digital domain feature vector. middle.

4. The shared power bank anti-theft and status monitoring system based on AI multimodal recognition according to claim 1, characterized in that, The voltage ripple acquisition circuit continuously acquires the ripple voltage on the VBUS line at a sampling frequency of 1MHz. The power spectral density is obtained by fast Fourier transform. The voltage ripple acquisition circuit extracts the main peak frequency of the spectrum. Main peak amplitude and the average noise floor in the 100kHz-1MHz frequency band. As the feature vector of the simulation domain When the switching frequency inside the PD protocol chip in the shared power bank changes, the main peak frequency of the spectrum... offset >A preset threshold of 0.5MHz will trigger an alarm for an abnormality in the analog domain.

5. The shared power bank anti-theft and status monitoring system based on AI multimodal recognition according to claim 1, characterized in that, After the chip ID response delay measurement module sends a challenge command to the PD protocol chip of the shared power bank to read the chip's unique identifier, it uses a timer inside the PD main control chip of the charging cabinet to record the delay time. ,in, The time when the challenge command is sent, In response to the moment the instruction is received; The chip ID response delay measurement module continuously measures the delay 10 times and calculates the mean delay and the standard deviation of the delay: Where k is the measurement sequence number; The average delay With the aforementioned delay standard deviation Together they constitute the temporal physical fingerprint vector .

6. The shared power bank anti-theft and status monitoring system based on AI multimodal recognition according to claim 1, characterized in that, The digital-analog dual-domain fingerprint generation module generates a dual-domain fingerprint vector. The steps are as follows: Receive the digital domain feature vector from the PD protocol communication monitoring module. ,in: , for The protocol encodes the sequence of communication data packets. This represents the average time interval between adjacent communication data packets. Manufacturer-defined field values; Receive the analog domain feature vector from the voltage ripple acquisition circuit. ; Receive the timing physical fingerprint vector from the chip ID response delay measurement module. ; The digital domain feature vector The simulated domain feature vector and the time-series physical fingerprint vector Concatenate into a joint feature vector The dual-domain fingerprint vector is obtained by performing nonlinear mapping through a pre-trained lightweight residual network. ,in, The weight matrix is ​​a learnable matrix. , Here, ReLU is the bias vector, and ReLU is the linear rectified activation function. The dual-domain fingerprint vector The digital-analog dual-domain fingerprint generation module generates the dual-domain fingerprint vector with 128 dimensions. Shared power banks with factory fingerprint storage in the cloud A cosine similarity comparison is performed, and a fingerprint mismatch is determined when the cosine similarity is less than 0.

85.

7. The shared power bank anti-theft and status monitoring system based on AI multimodal recognition according to claim 1, characterized in that, The abnormal communication traffic detection module consists of a temporal convolutional network and a graph attention network connected in parallel, which are then fused to output an abnormal probability score. The detection steps of the abnormal communication traffic detection module are as follows: Continuous Protocol stream sequence during the next return process Input a temporal convolutional network, where each Includes digital domain feature vectors extracted from a single interaction , eigenvectors of the analog domain Time-series physical fingerprint vector Temporal convolutional networks output temporal dependency features ; The relationships between different data packets within the same protocol stream are constructed as a graph structure, where each node represents a protocol packet and the edges represent temporal order. A graph attention network is used to calculate the attention weight of each node and the results are aggregated to obtain the graph structure features. ; The time-dependent features With the graph structure features After splicing, an anomaly probability score is output through a fully connected layer. ,in, For the Sigmoid function, This is the weight matrix of the fully connected layer. For bias, This represents a vector concatenation operation. The value range is [0, 1].

8. The shared power bank anti-theft and status monitoring system based on AI multimodal recognition according to claim 1, characterized in that, The abnormal communication traffic detection module is based on anomaly probability scoring. Implementing a tiered response includes: when When a low-level anomaly response is executed, a dynamic challenge code is sent from the PD main control chip of the charging cabinet to the microcontroller of the shared power bank, requiring the shared power bank to return a correct response within a preset 30 seconds using the built-in encryption unit. If the returned response is incorrect, the response is upgraded to a medium-level anomaly response. when At this time, a moderate anomaly response is executed, cutting off the output MOSFET inside the shared power bank and driving the audible and visual alarm on the charging cabinet to sound a buzzer and flash a red light, while simultaneously setting the current anomaly probability score. and the dual-domain fingerprint vector output by the digital-analog dual-domain fingerprint generation module Upload to the cloud management platform; when When a high-level anomaly occurs, a lock command is sent to the battery management IC of the shared power bank to disable its charging and discharging functions. The shared power bank's GPS location is then reported to the cloud management platform in real time, and the shared power bank is marked as a stolen device.

9. The shared power bank anti-theft and status monitoring system based on AI multimodal recognition according to claim 1, characterized in that, Each shared power bank undergoes a registration process before leaving the factory: Insert the shared power bank into the dedicated registration cabinet and perform 100 standard tests. The handshake, PD protocol communication monitoring module, voltage ripple acquisition circuit and chip ID response delay measurement module each acquired 100 sets of digital domain feature vectors. , eigenvectors of the analog domain Time-series physical fingerprint vector ; The digital-analog dual-domain fingerprint generation module calculates the average of 100 sets of dual-domain fingerprint vectors as the factory fingerprint. and the factory fingerprint Simultaneously stored in the cloud management platform database and the encrypted non-volatile memory of the shared power bank; On each subsequent return, the digital-analog dual-domain fingerprint generation module calculates the current dual-domain fingerprint vector in real time. and the current dual-domain fingerprint vector With the factory fingerprint A cosine similarity comparison is performed. If the cosine similarity is lower than 0.85, the abnormal communication traffic detection module is triggered to perform in-depth analysis.

Images