Intelligent door lock abnormity identification method and system based on big data

Abstract

The invention discloses an intelligent door lock abnormity identification method and system based on big data, and relates to the technical field of big data analysis, and the method comprises the following steps: collecting video stream data and fingerprint pressing time sequence data of a user in an unlocking process; fingerprint feature matching is carried out, and a legal user identity identifier is determined; after fingerprint feature matching is passed, performing visual analysis on the video stream data, and extracting a finger pressing dynamic feature vector; an individual historical behavior file of the legal user is constructed and continuously updated; the environment interaction feature vector and the finger pressing dynamic feature vector are compared with historical feature vectors, and an identification result that whether the current unlocking behavior is abnormal or not is output through an abnormity identification model; if the recognition result is abnormal, the door opening action is refused to be executed, and if the recognition result is normal, the door opening action is executed. The technical problem that the safety and the user experience are difficult to consider at the same time due to the fact that an existing intelligent door lock is difficult to effectively recognize complex abnormal scenes is solved.

Description

Technical Field

[0001] This application relates to the field of big data analytics technology, specifically to a method and system for identifying anomalies in smart door locks based on big data. Background Technology

[0002] With the rapid development of smart home technology, the security and convenience of smart door locks have received widespread attention. Traditional smart door locks mainly rely on single biometric identification technologies, such as fingerprint recognition, facial recognition, or password verification, for identity authentication.

[0003] However, traditional identification methods are susceptible to forgery attacks in practical applications. Single biometric features are easily copied or deceived, and they are difficult to effectively identify complex and abnormal scenarios such as coerced unlocking and tailgating intrusion. They also cannot distinguish between the unlocking behavior of legitimate users in abnormal states and normal unlocking behavior. Summary of the Invention

[0004] This application provides a method and system for identifying anomalies in smart door locks based on big data, which solves the technical problem that existing smart door locks are unable to effectively identify complex abnormal scenarios, resulting in a difficulty in balancing security and user experience.

[0005] The technical solution to the above-mentioned technical problems in this application is as follows:

[0006] Firstly, this application provides a method for identifying anomalies in smart door locks based on big data, the method comprising:

[0007] The smart door lock uses sensors to collect real-time video stream data and fingerprint press timing data during the unlocking process.

[0008] Fingerprint feature matching is performed based on the fingerprint press timing data to determine the legitimate user identity of the current operator.

[0009] After the fingerprint feature matching is successful, visual analysis is performed on the video stream data to extract the environmental interaction feature vector, and mechanical-time analysis is performed on the fingerprint pressing time sequence data to extract the finger pressing dynamic feature vector.

[0010] Based on the environmental interaction feature vectors and finger pressing dynamic feature vectors extracted from the historical unlocking process of legitimate users, a personal historical behavior profile of each legitimate user is constructed and continuously updated in the cloud big data platform.

[0011] The environmental interaction feature vector and the finger pressing dynamic feature vector extracted during the current unlocking process are compared with the historical feature vector corresponding to the legitimate user identity obtained from the personal historical behavior file. A comprehensive judgment is made through a pre-trained anomaly recognition model, and the recognition result of whether the current unlocking behavior is abnormal is output.

[0012] If the identification result is abnormal, the door opening action will be refused; if the identification result is normal, the door opening action will be performed.

[0013] Secondly, this application provides a smart door lock anomaly recognition system based on big data, including:

[0014] The information acquisition module is used to collect video stream data and fingerprint pressing timing data of the user in real time during the unlocking process through sensors on the smart door lock;

[0015] The identity verification module is used to perform fingerprint feature matching based on the fingerprint pressing timing data to determine the legitimate user identity identifier corresponding to the current operator;

[0016] The feature extraction module is used to perform visual analysis on the video stream data after the fingerprint feature matching is successful, extract the environmental interaction feature vector from it, and perform mechanical-temporal analysis on the fingerprint pressing time series data to extract the finger pressing dynamic feature vector from it.

[0017] The information update module is used to build and continuously update the personal historical behavior profile of each legitimate user in the cloud big data platform based on the environmental interaction feature vector and finger pressing dynamic feature vector extracted by the legitimate user in the historical unlocking process.

[0018] The result recognition module is used to compare the environmental interaction feature vector and the finger pressing dynamic feature vector extracted in the current unlocking process with the historical feature vector corresponding to the legitimate user identity obtained from the personal historical behavior file, and to make a comprehensive judgment through a pre-trained anomaly recognition model to output the recognition result of whether the current unlocking behavior is abnormal.

[0019] The action execution module is used to refuse to execute the door opening action if the recognition result is abnormal, and to execute the door opening action if the recognition result is normal.

[0020] This application provides one or more technical solutions, which have at least the following technical effects or advantages:

[0021] This application provides a smart lock anomaly identification method and system based on big data. First, sensors on the smart lock collect real-time video stream data and fingerprint press timing data during the unlocking process, achieving simultaneous acquisition of multimodal information. Second, fingerprint feature matching is used to determine the legitimate user's identity, ensuring accurate authentication while providing an identity anchor for personalized behavior profiles. Third, environmental interaction feature vectors are extracted through visual analysis, and dynamic feature vectors of finger presses are extracted through biomechanical-temporal analysis, characterizing the dynamic characteristics of unlocking behavior from both environmental context and biomechanical dimensions, effectively capturing subtle behavioral differences that are difficult to detect with traditional single biometric identification. Then, a personal historical behavior profile is built and continuously updated based on a cloud-based big data platform, using Gaussian mixture model clustering to mine typical user behavior patterns, achieving dynamic evolution and refined modeling of user behavior profiles. Finally, a pre-trained anomaly identification model compares the similarity and probability of current unlocking behavior with historical behavior patterns, improving the ability to identify complex anomaly scenarios such as coercive unlocking, tailgating, and spoofing attacks, ensuring the security of the smart lock while also considering the normal unlocking experience for legitimate users.

[0022] Through the above technical solutions, this application achieves in-depth verification of legitimate user identities and effective defense against potential security threats, thereby improving the accuracy of identifying abnormal behavior in smart door locks. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the intelligent door lock anomaly identification method based on big data provided in the embodiments of this application;

[0025] Figure 2 This is a schematic diagram of the structure of the intelligent door lock anomaly recognition system based on big data provided in the embodiments of this application.

[0026] The components represented by each number in the attached diagram are explained below:

[0027] Information collection module 11, identity verification module 12, feature extraction module 13, information update module 14, result recognition module 15, action execution module 16. Detailed Implementation

[0028] This application provides a method and system for identifying anomalies in smart door locks based on big data, which addresses the technical problem that existing smart door locks are unable to effectively identify complex abnormal scenarios, resulting in a difficulty in balancing security and user experience.

[0029] Example 1, as Figure 1 As shown in the figure, this application provides a smart door lock anomaly identification method based on big data, including:

[0030] S10: Real-time acquisition of video stream data and fingerprint press timing data of the user during the unlocking process through sensors on the smart door lock;

[0031] In this embodiment, the smart lock integrates a multimodal sensor array, including a high-definition camera for capturing visual information of the unlocking scene and a combination of a capacitive fingerprint sensor and a pressure sensor for recording the fingerprint pressing process. The high-definition camera and the fingerprint sensor are aligned at the millisecond level through the lock's built-in time synchronization module to form a multimodal data stream with time-series correlation, and are uploaded to the edge computing node in real time through an encrypted channel for preprocessing.

[0032] Specifically, step S10 in the method includes:

[0033] The high-definition infrared camera deployed on the smart door lock continuously collects video stream data of the entire process from when the user enters the camera's field of view to when the user completes the fingerprint pressing action at a first preset sampling frequency;

[0034] The pressure-type fingerprint sensor deployed on the smart door lock synchronously collects the pressure value, contact area coordinate sequence, and temperature value of the user's finger during the fingerprint unlocking process at a second preset sampling frequency, generating fingerprint pressing time sequence data that changes over time.

[0035] In this embodiment, firstly, a high-definition infrared camera continuously collects video stream data at a first preset sampling frequency of 30 frames per second to ensure that the dynamic picture of the user's entire process from approaching the door lock, raising their hand to prepare, pressing their finger to complete the unlocking is clearly captured.

[0036] Secondly, the pressure-type fingerprint sensor synchronously collects pressure values, contact area coordinate sequences, and temperature values ​​at a second preset sampling frequency of 1000 Hz per second. The pressure value reflects the curve of the force change of the finger pressing, the contact area coordinate sequence records the dynamic displacement trajectory of the fingerprint and the sensor contact surface, and the temperature value is used to help determine whether it is a live, real finger contact. The three together constitute high temporal resolution fingerprint pressing timing data.

[0037] Furthermore, the data acquisition from the two sensors is synchronized through a hardware timestamp module in the door lock's main control chip, ensuring that the video frames and fingerprint press data are aligned with millisecond-level precision.

[0038] S20: Perform fingerprint feature matching based on the fingerprint pressing timing data to determine the legitimate user identity identifier corresponding to the current operator;

[0039] In this embodiment, the fingerprint feature matching process employs a multi-level verification mechanism. First, a liveness detection is performed on the pressure value sequence in the fingerprint pressing time series data. By analyzing the fluctuation characteristics of the pressure curve and the dynamic change range of the temperature value, static attacks using counterfeit materials such as silicone fingerprint films and conductive gels are excluded. After the liveness detection is passed, fingerprint ridge detail feature points, including endpoints, bifurcation points, and ridge flow information, are extracted from the fingerprint pressing time series data to form a fingerprint feature template.

[0040] The extracted fingerprint feature template is compared with the legitimate user fingerprint database pre-stored in the local security chip of the smart lock. The Euclidean distance and angular deviation between feature points are calculated, and geometric consistency verification based on the RANSAC algorithm is used to select the candidate user with the highest matching degree.

[0041] Furthermore, if the highest matching score exceeds the preset threshold, the legitimate user identity of the current operator is determined, and the subsequent multi-dimensional behavior analysis process is triggered; if the matching fails or the liveness detection fails, it is directly judged as an abnormal unlocking behavior, the door opening action is refused, and a local alarm is triggered.

[0042] S30: After the fingerprint feature matching is successful, the video stream data is subjected to visual analysis to extract the environmental interaction feature vector, and the fingerprint pressing time sequence data is subjected to mechanical-time analysis to extract the finger pressing dynamic feature vector.

[0043] In this embodiment of the application, after the fingerprint feature matching is successful, visual analysis and mechanical-time analysis are performed in parallel to deeply characterize the dynamic characteristics of the unlocking behavior from two dimensions: environmental context and biomechanics, and extract environmental interaction feature vectors and finger pressing dynamic feature vectors.

[0044] Among them, the environmental interaction feature vector quantifies the user's unique body posture, hand holding state, and unlocking action sequence, reflecting each person's inherent behavioral inertia; the finger pressing dynamic feature vector reflects the user's behavioral inertia.

[0045] The video stream data is subjected to visual analysis to extract environmental interaction feature vectors, including:

[0046] By deploying a lightweight human pose estimation model on the edge of the door lock, the video stream data is analyzed frame by frame to extract the user's body key point coordinate sequence. Based on the body key point coordinate sequence, the relative position features of the user's body pose and the door lock, as well as the continuous action sequence features of the user during the unlocking process, are calculated.

[0047] By deploying a lightweight target detection model on the edge of the door lock, the video stream data is analyzed frame by frame to detect the user's hand area and the state of the hand holding objects, and the user's hand holding object state features are extracted from it.

[0048] The hand holding state features, the body posture and the relative position features of the door lock, and the continuous action sequence features during the unlocking process are encoded and fused to generate an environmental interaction feature vector.

[0049] In this embodiment, firstly, the lightweight human pose estimation model adopts an improved OpenPose architecture based on the MobileNetV3 backbone network to analyze the video stream data frame by frame, extracting the coordinate sequences of 18 body key points, including the head, shoulders, elbows, wrists, hips, knees, and ankles. Based on the key point coordinates, the angle between the user's torso and the door lock plane, the horizontal distance between the user's standing position and the door lock, and the vertical height difference are calculated to form the relative positional features of the body pose and the door lock. Simultaneously, by analyzing the displacement trajectory and velocity changes of the key point coordinates between consecutive frames, the continuous action sequence features of the user during the process of approaching the door lock, raising their hand to prepare, pressing their finger, and completing the unlocking are extracted, including action smoothness indicators and action rhythm patterns.

[0050] Secondly, the lightweight object detection model employs a pruned and optimized version of the YOLOv5s architecture to perform real-time object detection on video stream data, focusing on identifying the user's hand area and the state of the object they are holding. After the model outputs the bounding box coordinates of the hand, a fine-grained classification network is used to determine whether the hand is holding an object, the type of the object, and the relative position of the object to the door lock, forming a hand-holding state feature. This feature can distinguish between normal unlocking with bare hands and abnormal situations such as forced unlocking with an object under duress or hiding dangerous objects in a tailgating scenario.

[0051] Furthermore, the hand holding state features, the body posture and the relative position features of the door lock, and the continuous action sequence features during the unlocking process are encoded and fused.

[0052] Specifically, an attention-based feature fusion network is used to adaptively weight features of different dimensions. Body posture and relative position features are encoded as 64-dimensional vectors, continuous action sequence features are encoded as 128-dimensional vectors through a temporal convolutional network, and hand holding state features are encoded as 32-dimensional vectors. Finally, they are concatenated to form a 224-dimensional environmental interaction feature vector.

[0053] Furthermore, the construction process of the lightweight human pose estimation model and the lightweight object detection model includes:

[0054] The construction process of the lightweight human pose estimation model includes:

[0055] A lightweight convolutional neural network is used as the backbone network, and an initial human pose estimation model is constructed by combining a key point detection head.

[0056] Collect human image data under different lighting conditions, distances, and postures in smart door lock usage scenarios, and annotate the key points of the human body in the human image data to form a human posture training dataset.

[0057] The initial human pose estimation model is trained using the human pose training dataset, and the trained human pose estimation model is compressed and optimized using model pruning and parameter quantization techniques to obtain a lightweight human pose estimation model.

[0058] The construction process of the lightweight target detection model includes:

[0059] A lightweight object detection network is used as the basic framework to construct the initial object detection model;

[0060] Collect image data containing hands, various handheld objects, and door handles in smart door lock usage scenarios, and label the target categories and locations in the image data to form a target detection training dataset;

[0061] The initial object detection model is trained using the object detection training dataset, and the trained object detection model is compressed and optimized using channel pruning and parameter quantization techniques to obtain a lightweight object detection model.

[0062] In this embodiment, the lightweight human pose estimation model is first constructed using MobileNetV3-Small as the backbone network. This network combines depthwise separable convolutions with a lightweight attention mechanism, compressing the number of parameters to less than one-tenth of the traditional ResNet50 while maintaining high feature extraction capabilities. The keypoint detection head adopts a multi-stage heatmap regression structure, first predicting the center position and scale of the human body, then progressively refining the spatial distribution of keypoints in each part, and finally outputting a two-dimensional coordinate confidence map of 18 keypoints.

[0063] Specifically, to adapt to the resource constraints of the edge computing environment of smart door locks, 150,000 human images were collected under 12 typical lighting conditions, including strong daylight, infrared illumination at night, backlighting silhouettes, and sidelighting penumbra, within a range of 0.3 to 1.5 meters from the door lock. The LabelMe tool was used to annotate keypoint locations at the pixel level. The model was trained using the AdamW optimizer with an initial learning rate of 0.001 and a cosine annealing strategy. A total of 300 training epochs were performed, with a batch size of 32. After training, L1-norm-based structured pruning was used to remove redundant convolutional kernels, controlling channel sparsity to within 30%. INT8 quantization was then performed to further reduce the model size by 75%. The resulting lightweight human pose estimation model achieved an inference latency of less than 50 milliseconds on the door lock's embedded processor, with a keypoint localization accuracy exceeding 90%.

[0064] Secondly, the lightweight object detection model is constructed using YOLOv5s as the baseline architecture, replacing the C3 module in its backbone network with the channel shuffling unit proposed by ShuffleNetV2, and simplifying the feature pyramid network in the neck part into a bidirectional two-scale fusion structure.

[0065] Specifically, to address the unique needs of smart door lock scenarios, a total of 80,000 images were collected, encompassing eight hand states: empty hand, holding a key, holding a mobile phone, holding a package, and holding a dangerous object. Environmental targets such as door handles and obstructions were also included. The CVAT platform was used for bounding box and category labeling. During training, Mosaic data augmentation and MixUp techniques were introduced to improve the model's generalization ability, and the CIoU loss function was employed to optimize localization accuracy. In the model compression stage, feature channels with contributions below a threshold were pruned using channel importance scoring. Knowledge distillation techniques were used to transfer the detection capabilities of the original model to the compressed lightweight network. Ultimately, the model's parameter count was reduced to 15% of the original YOLOv5s, achieving a real-time detection speed of 25 frames per second on door lock devices. The average accuracy for hand region detection and the accuracy for object-holding state classification both exceeded 90%.

[0066] Specifically, a biomechanical-temporal analysis is performed on the fingerprint press timing data to extract the dynamic feature vector of finger press, including:

[0067] Curve fitting was performed on the data segment of the pressure holding phase in the fingerprint pressing timing data, and the stress relaxation coefficient and creep coefficient were extracted from it.

[0068] Spectral analysis was performed on the pressure value sequence in the fingerprint pressing time series data to extract the pulse wave modulation energy features corresponding to the 0.8 Hz to 2 Hz frequency band and the muscle micro-tremor energy features corresponding to the 5 Hz to 20 Hz frequency band.

[0069] Calculate the fractal dimension of the pressure center trajectory based on the contact area coordinate sequence in the fingerprint pressing time sequence data;

[0070] The temperature value sequence in the fingerprint pressing time series data is analyzed, the temperature change rate curve is extracted, and the heat conduction time constant from the initial value to the stable value is calculated;

[0071] The stress relaxation coefficient, creep coefficient, pulse wave modulation energy characteristics, muscle micro-tremor energy characteristics, fractal dimension of the pressure center trajectory, temperature change rate curve, and heat conduction time constant are fused in multiple dimensions to generate a dynamic feature vector of finger pressing.

[0072] In this embodiment, firstly, curve fitting based on the Voigt viscoelastic model is performed on the data segment of the pressure holding phase. This model regards the finger skin-sensor contact system as a parallel combination of spring and damper. The elastic modulus and viscosity coefficient are obtained by fitting with nonlinear least squares method. Then, the stress relaxation coefficient is calculated to characterize the rate of pressure decay, and the creep coefficient is calculated to characterize the viscous characteristics of continuous expansion of contact area under constant pressure.

[0073] Specifically, after the user completes the initial press, a pressure holding phase of about 500 milliseconds is entered. The pressure value sequence of 1000 sampling points per second within this interval is collected. The goodness of fit R² must be greater than 0.95 to be considered effective. The two coefficients together reflect the biomechanical properties of the user's finger soft tissue, which are unique to each individual and difficult to be replicated by counterfeit materials.

[0074] Secondly, the pressure value sequence was subjected to fast Fourier transform to obtain the spectral distribution, and the pulse wave modulation energy features corresponding to the 0.8 Hz to 2 Hz frequency band and the muscle microtremor energy features corresponding to the 5 Hz to 20 Hz frequency band were extracted.

[0075] The capillary network distributed under the skin of a real finger, along with the rhythmic changes in the blood vessel volume caused by the periodic beating of the heart, creates physiological fluctuations that slightly modulate the contact pressure between the finger and the sensor, forming a pulse wave signal superimposed on the pressure curve. The human resting heart rate is typically 60 to 120 beats per minute, corresponding to a frequency of 1 to 2 Hz. Considering the slight heart rate fluctuations that may occur during pressing, extending the frequency band to 0.8 to 2 Hz fully covers the normal human pulse wave frequency range. Fingerprint films, being inert materials, lack an internal blood circulation system. The pressure curve generated during pressing lacks corresponding periodic modulation energy within this frequency band. Therefore, extracting the energy characteristics of this frequency band can effectively distinguish between real fingers and counterfeit fingerprint films.

[0076] The 5 Hz to 20 Hz frequency band corresponds to the physiological micro-vibrations generated by the coordinated contraction of the agonist and antagonist muscles of the fingers. The actual pressing action of a finger is supported by bones, driven by muscles, and continuously controlled by the nervous system. Even during the seemingly stable pressure-holding phase, muscle fibers produce unconscious physiological micro-vibrations, primarily concentrated in the 5 Hz to 20 Hz range. When a user uses a fingerprint membrane to unlock, the attacker's finger muscles are in a passively pressed state, lacking the muscle micro-vibrations generated by actively maintaining pressure during the actual pressing process, or producing abnormally frequent tremors due to tension. Therefore, extracting the energy characteristics of this frequency band can capture the unique physiological signals generated by the neuromuscular control system.

[0077] Furthermore, the two-dimensional motion trajectory of the pressure center point is calculated based on the coordinate sequence of the contact area, and the fractal dimension of this trajectory is calculated using box counting. During normal unlocking, the user's finger will produce slight unconscious slippage during the pressure holding phase, and its trajectory exhibits typical random fractal characteristics, with a fractal dimension typically between 1.2 and 1.6. In contrast, when a mechanical prosthesis or rigid mold is pressed, the trajectory of the pressure center is almost stationary or exhibits regular mechanical vibration, and the fractal dimension deviates significantly from this range. This can effectively distinguish between living organisms and non-biological materials.

[0078] Next, the temperature value sequence is analyzed. First, a sliding window is used for smoothing and noise reduction. Then, the first-order difference is calculated to obtain the temperature change rate curve. The transition interval from the initial contact value to the stable plateau period is identified. An exponential decay model is used to fit the heat conduction time constant. This parameter reflects the heat exchange dynamics between the real finger tissue and the sensor surface. It is closely related to individual physiological parameters such as finger blood flow and subcutaneous fat thickness. The normal range is usually between 200 milliseconds and 800 milliseconds. However, preheated fake materials often show an excessively short time constant or a monotonically decreasing trend.

[0079] Finally, after standardizing the seven-dimensional features, including stress relaxation coefficient, creep coefficient, pulse wave modulation energy characteristics, muscle microtremor energy characteristics, fractal dimension of pressure center trajectory, statistical moment characteristics of temperature change rate curve, and heat conduction time constant, they are mapped to a 128-dimensional finger pressure dynamic feature vector through a fully connected network. This vector encodes the user's unique biomechanical fingerprint.

[0080] Specifically, curve fitting is performed on the data segment of the pressure holding phase in the fingerprint pressing time sequence data, and the stress relaxation coefficient and creep coefficient are extracted from it, including:

[0081] Identify the start and end times of the pressure holding phase from the fingerprint press timing data, and extract the pressure value subsequence and contact area coordinate subsequence corresponding to the pressure holding phase;

[0082] An exponential function is fitted to the subsequence of pressure values ​​during the pressure holding phase, and the stress relaxation coefficient is calculated based on the stress relaxation time constant obtained from the fitting. The stress relaxation coefficient is negatively correlated with the stress relaxation time constant.

[0083] Based on the contact area coordinate subsequence during the pressure holding phase, the contact area at each moment is calculated, and an area growth curve of the contact area over time is generated.

[0084] The area growth curve is fitted with a logarithmic function, and the creep coefficient is calculated based on the logarithmic growth coefficient obtained from the fitting. The creep coefficient is positively correlated with the logarithmic growth coefficient.

[0085] In this embodiment, the identification of the pressure holding phase is first based on the determination of the turning point of the dynamic characteristics of the pressing action. Specifically, the rate of change of pressure value is monitored in real time from the fingerprint pressing time sequence data. When the rate of pressure increase suddenly drops from a positive value to close to zero and then continues to maintain a low fluctuation state, it is determined as the start time of the pressure holding phase; when the pressure value begins to decrease significantly or the contact area shrinks sharply, it is determined as the end time of the pressure holding phase.

[0086] Secondly, a curve fitting based on a double exponential function is performed on the pressure value subsequence during the pressure holding phase. The fitting formula is P(t) = P0 + A1exp(-t / τ1) + A2exp(-t / τ2), where P0 is the steady-state pressure value, and τ1 and τ2 are the time constants for rapid and slow stress relaxation, respectively. The stress relaxation coefficient is defined as S = 1 / (w1τ1 + w2τ2), where w1 and w2 are coefficients obtained by normalization based on amplitude weights. The larger the coefficient, the faster the pressure decays, reflecting the stronger the viscoelasticity of the finger soft tissue.

[0087] Furthermore, based on the contact area coordinate subsequence during the pressure holding phase, the convex hull algorithm is used to calculate the contact area at each sampling time, generating an area growth curve that shows the area changing over time. Normal biological tissues under constant pressure will experience a slow and continuous expansion of the contact area due to stress relaxation, a process that follows a logarithmic growth law. The area growth curve is then expressed as A(t) = A0 + B. The form of a logarithmic function is used for fitting, where C is the logarithmic growth coefficient, and the creep coefficient is defined as Cr=B·C / A0. This coefficient is positively correlated with the logarithmic growth coefficient and characterizes the flow deformation characteristics of the material under continuous load.

[0088] For example, assuming a legitimate user presses their fingerprint and the pressure is maintained for 600 milliseconds, the collected pressure value subsequence, after double exponential fitting, yields τ1 = 45 milliseconds and τ2 = 320 milliseconds, corresponding to amplitudes A1 = 2.3N and A2 = 0.8N. After normalization and weighting, the stress relaxation coefficient S = 0.015 milliseconds⁻¹ is obtained. During the same period, the contact area increases from an initial 180 square millimeters to 215 square millimeters. Logarithmic fitting yields a logarithmic growth coefficient C = 0.12 and a creep coefficient Cr = 0.023. These parameters fall within the 95% confidence interval of the user's historical unlock data distribution, and are therefore considered normal biometric behavior.

[0089] S40: Based on the environmental interaction feature vector and finger pressing dynamic feature vector extracted by legitimate users during the historical unlocking process, a personal historical behavior profile of each legitimate user is constructed and continuously updated in the cloud big data platform.

[0090] In this embodiment, the cloud-based big data platform adopts a distributed storage architecture, managing the feature data uploaded by each smart lock terminal by partition according to user identifier. Each legitimate user's personal historical behavior profile contains multi-dimensional behavioral pattern information, specifically including the temporal sequence of environmental interaction feature vectors and the statistical distribution parameters of finger pressing dynamic feature vectors.

[0091] Specifically, step S40 in the method includes:

[0092] A unique identity is established for each legitimate user, and a dedicated storage space associated with the identity is created in the cloud vector database. The legitimate user refers to a user who has completed registration in the smart door lock system and has been authorized to unlock the door.

[0093] The historical environmental interaction feature vector and historical pressing dynamic feature vector extracted from each verified normal unlocking event are stored as historical feature vectors in the dedicated storage space.

[0094] Multiple historical feature vectors in the dedicated storage space are clustered using a Gaussian mixture model to obtain K cluster centers, where each cluster center corresponds to a typical behavior pattern.

[0095] Calculate the frequency of each cluster center, take the cluster center with the highest frequency as the main pattern vector, and take the remaining cluster centers as the secondary pattern vectors;

[0096] Based on the primary pattern vector and the secondary pattern vector, a personal historical behavior profile of each legitimate user is constructed and continuously updated in the cloud-based big data platform.

[0097] In this embodiment, firstly, a globally unique 128-bit UUID identity identifier is assigned to each legitimate user. This identifier is bound to the user's biometric template and door lock authorization information. In a cloud-based vector database, a dedicated storage space associated with the UUID is created using a high-performance vector retrieval engine such as Milvus. The storage space is divided into a hot data area and a cold data area according to time series. Feature vectors from the past 90 days are stored in the hot data area to support fast real-time comparison, while historical data is automatically migrated to the cold data area and stored using Parquet columnar compression.

[0098] Specifically, after each successful unlock verification, the smart lock terminal uploads the environmental interaction feature vector and the dynamic feature vector of finger pressing to the cloud via a TLS encrypted channel. Upon receiving the data, the cloud first performs data integrity verification and outlier filtering, removing obvious outliers caused by sensor malfunctions. Then, it writes the verified historical feature vectors, using timestamps as keys, into the user's dedicated storage space. Each feature vector is accompanied by contextual metadata, including the unlocking time, lock device identifier, ambient light conditions, temperature and humidity readings, etc., forming a complete unlocking behavior record.

[0099] Furthermore, an online learning mechanism is employed to perform dynamic clustering analysis on historical feature vectors in the storage space. During the initialization phase, when the number of historical samples reaches 100, the first Gaussian mixture model clustering is performed. The optimal number of clusters K is automatically determined using the Bayesian information criterion. Typically, K is set between 3 and 7 to cover the main behavioral pattern variations of users.

[0100] As new data is continuously added, a mini-batch expectation-maximization algorithm is used to incrementally update the Gaussian mixture model. Model parameter optimization is triggered every 50 new samples to avoid the computational overhead of full retraining. The cluster centers with the highest frequency of occurrence, i.e., those with the highest prior probability π, are selected. max The cluster centers are labeled as the user's primary pattern vectors, representing their most typical and stable behavioral characteristics; the remaining cluster centers serve as secondary pattern vectors, corresponding to behavioral variations in different scenarios, such as pressing while wearing gloves in winter, unlocking with one hand while holding an object, or unlocking quickly in a hurry.

[0101] Ultimately, each legitimate user's personal historical behavior profile is persistently stored in a structured format, including a primary pattern vector, a set of secondary pattern vectors, the prior probability distribution of each pattern, the most recent update timestamp, and a data quality score. The profile supports versioning, retaining the 10 most recent historical versions for anomaly tracking. An automatic archiving strategy is also configured: for user profiles that have not exhibited unlocking activity for 180 consecutive days, they are moved to a low-frequency storage tier to reduce operational costs, while retaining rapid recovery capabilities to handle user reactivation scenarios.

[0102] S50: The environmental interaction feature vector and the finger pressing dynamic feature vector extracted during the current unlocking process are compared with the historical feature vector corresponding to the legitimate user identity obtained from the personal historical behavior file, and a comprehensive judgment is made through a pre-trained anomaly recognition model to output the recognition result of whether the current unlocking behavior is abnormal.

[0103] In this embodiment, the environmental interaction feature vector and the finger pressing dynamic feature vector are compared with the historical feature vector corresponding to the legitimate user identity, and the cosine similarity is calculated. Based on the anomaly recognition model, it is determined whether the current unlocking behavior is abnormal.

[0104] Specifically, step S50 in the method includes:

[0105] Based on the legitimate user identity identifier, the historical environment interaction feature vector and historical pressure dynamic feature vector associated with the legitimate user identity identifier are retrieved from the personal historical behavior archive of the cloud big data platform.

[0106] Calculate the first cosine similarity between the environmental interaction feature vector extracted during the current unlocking process and the historical environmental interaction feature vector;

[0107] Calculate the second cosine similarity between the dynamic feature vector of finger pressing extracted during the current unlocking process and the dynamic feature vector of finger pressing in the past process;

[0108] The first cosine similarity, the second cosine similarity, and the temporal context information of the current unlocking process are concatenated into a fused feature vector;

[0109] The fused feature vector is input into a pre-trained anomaly detection model, and the probability value that the current unlocking behavior is abnormal is calculated by the anomaly detection model.

[0110] The probability value is compared with a preset anomaly determination threshold. If the probability value exceeds the anomaly determination threshold, the current unlocking behavior is determined to be abnormal. If the probability value does not exceed the anomaly determination threshold, the current unlocking behavior is determined to be normal.

[0111] In this embodiment, firstly, based on the timestamp of the current unlocking time, a subset of historical features related to the time period is extracted from the personal historical behavior file. Then, an approximate nearest neighbor search algorithm is used to retrieve the K most similar historical samples to the current vector in the 128-dimensional feature space. The first cosine similarity between the current environment interaction feature vector and the historical environment interaction feature vector is calculated, with the formula S1=(vc·vh) / (||vc|| ||vh||), where vc represents the current environment interaction feature vector and vh represents the historical environment interaction feature vector.

[0112] Similarly, the second cosine similarity S2 between the current finger pressing dynamic feature vector and the historical pressing dynamic feature vector is calculated. Considering the natural temporal evolution of user behavior, a time decay factor λ=exp(-Δt / τ) is introduced for historical samples older than 90 days, where Δt is the number of days since the sample and τ is the decay time constant of 30 days, giving recent samples a higher weight in the similarity calculation.

[0113] Furthermore, the first cosine similarity S1, the second cosine similarity S2, and the temporal context information of the current unlocking process are concatenated to form a fused feature vector. The temporal context information specifically includes: the time interval between the current moment and the user's last unlocking, the unlocking frequency for the day, the matching degree between the current time period and the user's historical active time periods, a binary identifier indicating whether it is a holiday or a weekday, and the deviation of ambient light and temperature from the user's historical averages. This fused feature vector has 12 dimensions, comprehensively reflecting the current unlocking behavior in terms of biometric consistency, behavioral pattern coherence, and contextual rationality.

[0114] Furthermore, the pre-trained anomaly detection model takes the aforementioned 12-dimensional fused feature vector as input and outputs a probability value P∈[0,1] indicating that the current unlocking behavior is an anomaly. This probability value is then compared with a preset anomaly judgment threshold θ, which is dynamically adjustable based on the security level of the application scenario. For example, θ is set to 0.3 in high-security mode and 0.7 in convenience-first mode to reduce interference with normal user operation. In the default balanced mode, θ is set to 0.5, balancing security and user experience.

[0115] Specifically, when the probability value P exceeds the threshold θ, the current unlocking behavior is determined to be abnormal, triggering a tiered response mechanism: for mild abnormalities, such as 0.5≤P<0.7, the user is required to perform secondary verification, such as entering a backup password or undergoing facial recognition verification; for moderate abnormalities, such as 0.7≤P<0.9, the fingerprint unlocking method is immediately locked, retaining only the mechanical key or the administrator's remote authorization channel; for severe abnormalities, such as P≥0.9, a real-time alarm is triggered, pushing intrusion warning information to the user's bound mobile terminal and automatically reporting to the cloud security center for global risk marking.

[0116] Furthermore, the construction process of the anomaly detection model includes:

[0117] Collect historical unlocking data generated by legitimate users in normal unlocking scenarios, and extract the corresponding environmental interaction feature vector and finger pressing dynamic feature vector from the historical unlocking data as positive sample raw data;

[0118] Collect unlocking data generated under simulated attack scenarios, and extract the corresponding environmental interaction feature vector and finger pressing dynamic feature vector from the simulated attack unlocking data as negative sample raw data. The simulated attack scenarios include at least one or more of the following: unlocking with fingerprint film, unlocking under duress, and tailgating intrusion.

[0119] For each set of positive sample raw data and negative sample raw data, the historical environment interaction feature vector and historical pressing dynamic feature vector of the legitimate user corresponding to the unlocking operation are retrieved from the personal historical behavior archive of the cloud big data platform.

[0120] Calculate the cosine similarity between the environmental interaction feature vector in each set of positive sample original data or negative sample original data and the historical environmental interaction feature vector, and use it as the first similarity feature;

[0121] Calculate the cosine similarity between the finger pressing dynamic feature vector in each set of positive sample original data or negative sample original data and the historical pressing dynamic feature vector, and use it as the second similarity feature;

[0122] Obtain the temporal context information corresponding to the original data of each group of positive or negative samples as contextual features;

[0123] The first similarity feature, the second similarity feature, and the context feature are concatenated to form a training sample fusion feature vector set, and the corresponding normal label or abnormal label is used as the label of the training sample to form a training sample supervision label set.

[0124] The logistic regression model is trained using the fusion feature vector set of the training samples and the supervision label set of the training samples. The parameters of the logistic regression model are optimized by maximizing the likelihood function to obtain the trained anomaly recognition model.

[0125] In this embodiment, firstly, the positive sample data comes from the normal unlocking records of legitimate users over the past 12 months, retaining over 500,000 valid unlocking events after data cleaning. Negative sample data is obtained through multi-dimensional simulated attack experiments, specifically including: spoofing attacks using high-precision silicone fingerprint films, conductive gel fingerprint films, etc., to forge biometric features; requiring subjects to press the fingerprint quickly or excessively in unnatural postures under simulated coercion scenarios; and unauthorized personnel attempting to unlock the fingerprint immediately after closely following a legitimate user in a tailgating scenario. A total of over 80,000 negative sample data entries were collected.

[0126] During the feature engineering phase, for each set of raw data, the user's historical behavior profile is first retrieved from the cloud based on the user's identity, and the main pattern vector and secondary pattern vector sets are extracted as comparison benchmarks. The cosine similarity S1m between the current environment interaction feature vector and the main pattern vector, the cosine similarity S1s between the current vector and the nearest secondary pattern vector, and the S2m and S2s corresponding to the finger-pressing dynamic feature vector are calculated. Simultaneously, the Mahalanobis distance between the current vector and each cluster center is calculated to measure the degree of distribution deviation in the feature space. In addition to the aforementioned dimensions, the temporal context features further incorporate the KL divergence between the unlocking time and the distribution of the user's historical active time periods to quantify the degree of deviation from the user's daily routine.

[0127] Specifically, Mahalanobis distance is a statistical distance metric that takes into account the correlation between feature dimensions. Compared with Euclidean distance, Mahalanobis distance can effectively eliminate the influence of differences in feature dimensions and correlation structure, and is more suitable for anomaly detection in high-dimensional feature spaces.

[0128] Furthermore, the aforementioned features are concatenated to form an 18-dimensional training sample fusion feature vector, which includes 4-dimensional similarity features, 3-dimensional distance features, 8-dimensional temporal context features, and 3-dimensional statistical derived features. A stratified sampling strategy is used to construct the training and validation sets, ensuring that the ratio of positive to negative samples is approximately 5:1 to closely approximate the actual attack incidence rate. A logistic regression model with L2 regularization is selected as the base classifier, and the regularization coefficient is determined to be 0.01 through 5-fold cross-validation to prevent overfitting on sparse attack samples.

[0129] The model was trained using a stochastic gradient descent optimizer with an initial learning rate of 0.05, which decayed exponentially with each iteration. Furthermore, to enhance the model's adaptability to novel attack patterns, an online learning mechanism was introduced. This involved adding newly validated unlocking data and manually reviewed anomaly cases to the training set monthly, triggering fine-tuning of model parameters to maintain the model's timeliness and detection sensitivity.

[0130] Ultimately, the deployed anomaly detection model is serialized and stored in a lightweight format on a cloud-based inference service, supporting millisecond-level response latency. The model's output probability values ​​are mapped to reliable risk scores after being scaled and calibrated by Platt. Combined with a dynamic threshold strategy and a hierarchical response mechanism, this forms a complete closed loop from feature extraction, document retrieval, similarity calculation to anomaly detection.

[0131] S60: If the identification result is abnormal, the door opening action is refused; if the identification result is normal, the door opening action is performed.

[0132] In this embodiment of the application, if the identification result is abnormal, it is determined that there may be risks such as fingerprint forgery or coercion to open the door, and the door opening action is refused, thereby achieving deep verification of the legitimate user's identity and effective defense against potential security threats.

[0133] If the recognition result is normal, it means that the environmental interaction characteristics and finger pressing dynamic characteristics of the current unlocking behavior are highly consistent with the typical patterns in the user's historical behavior profile. The biomechanical fingerprint verification is successful, the identity is confirmed, the smart door lock executes the door opening action, and allows the user to enter normally.

[0134] In summary, compared to existing technologies, this application effectively distinguishes the essential differences between real fingers and counterfeit materials in terms of pulse wave modulation and muscle micro-tremors through multi-band spectral analysis of fingerprint pressing time-series data. Secondly, by calculating the fractal dimension of the pressure center trajectory, it demonstrates the significant difference between the unique random motion patterns of living organisms and the rigid motion characteristics of mechanical devices.

[0135] In summary, the embodiments of this application have at least the following technical effects:

[0136] This application provides a big data-based method for anomaly identification in smart locks. First, sensors on the smart lock collect real-time video stream data and fingerprint press timing data during the unlocking process, enabling simultaneous acquisition of multimodal information. Second, fingerprint feature matching is used to determine the legitimate user's identity, ensuring accurate authentication while providing an identity anchor for personalized behavior profiles. Third, environmental interaction feature vectors are extracted through visual analysis, and dynamic feature vectors of finger presses are extracted through biomechanical-temporal analysis, characterizing the dynamic characteristics of unlocking behavior from both environmental context and biomechanical dimensions, effectively capturing subtle behavioral differences that are difficult to detect with traditional single-biometric identification. Then, a personal historical behavior profile is built and continuously updated based on a cloud-based big data platform, using Gaussian mixture model clustering to mine typical user behavior patterns, achieving dynamic evolution and refined modeling of user behavior profiles. Finally, a pre-trained anomaly identification model compares the similarity and probability of current unlocking behavior with historical behavior patterns, improving the ability to identify complex anomaly scenarios such as coercive unlocking, tailgating, and spoofing attacks, ensuring the security of the smart lock while also considering the normal unlocking experience for legitimate users.

[0137] Through the above technical solutions, this application achieves in-depth verification of legitimate user identities and effective defense against potential security threats, thereby improving the accuracy of identifying abnormal behavior in smart door locks.

[0138] Example 2, as Figure 2 As shown, based on the same inventive concept as the big data-based smart lock anomaly identification method provided in Embodiment 1, this application also provides a big data-based smart lock anomaly identification system, including:

[0139] Information acquisition module 11 is used to collect video stream data and fingerprint pressing timing data of the user in real time during the unlocking process through sensors on the smart door lock;

[0140] The identity verification module 12 is used to perform fingerprint feature matching based on the fingerprint pressing timing data to determine the legitimate user identity identifier corresponding to the current operator;

[0141] The feature extraction module 13 is used to perform visual analysis on the video stream data after the fingerprint feature matching is successful, extract the environmental interaction feature vector from it, and perform mechanical-time analysis on the fingerprint pressing time series data to extract the finger pressing dynamic feature vector from it.

[0142] Information update module 14 is used to build and continuously update the personal historical behavior profile of each legitimate user in the cloud big data platform based on the environmental interaction feature vector and finger pressing dynamic feature vector extracted by the legitimate user in the historical unlocking process.

[0143] The result recognition module 15 is used to compare the environmental interaction feature vector and the finger pressing dynamic feature vector extracted in the current unlocking process with the historical feature vector corresponding to the legitimate user identity obtained from the personal historical behavior file, and to make a comprehensive judgment through a pre-trained anomaly recognition model, and output the recognition result of whether the current unlocking behavior is abnormal.

[0144] The action execution module 16 is used to refuse to execute the door opening action if the recognition result is abnormal, and to execute the door opening action if the recognition result is normal.

[0145] In one embodiment, the information acquisition module 11 is specifically used for:

[0146] The high-definition infrared camera deployed on the smart door lock continuously collects video stream data of the entire process from when the user enters the camera's field of view to when the user completes the fingerprint pressing action at a first preset sampling frequency;

[0147] The pressure-type fingerprint sensor deployed on the smart door lock synchronously collects the pressure value, contact area coordinate sequence, and temperature value of the user's finger during the fingerprint unlocking process at a second preset sampling frequency, generating fingerprint pressing time sequence data that changes over time.

[0148] Furthermore, in one embodiment, visual analysis is performed on the video stream data to extract environmental interaction feature vectors, including:

[0149] By deploying a lightweight human pose estimation model on the edge of the door lock, the video stream data is analyzed frame by frame to extract the user's body key point coordinate sequence. Based on the body key point coordinate sequence, the relative position features of the user's body pose and the door lock, as well as the continuous action sequence features of the user during the unlocking process, are calculated.

[0150] By deploying a lightweight target detection model on the edge of the door lock, the video stream data is analyzed frame by frame to detect the user's hand area and the state of the hand holding objects, and the user's hand holding object state features are extracted from it.

[0151] The hand holding state features, the body posture and the relative position features of the door lock, and the continuous action sequence features during the unlocking process are encoded and fused to generate an environmental interaction feature vector.

[0152] Furthermore, the construction process of the lightweight human pose estimation model and the lightweight object detection model includes:

[0153] The construction process of the lightweight human pose estimation model includes:

[0154] A lightweight convolutional neural network is used as the backbone network, and an initial human pose estimation model is constructed by combining a key point detection head.

[0155] Collect human image data under different lighting conditions, distances, and postures in smart door lock usage scenarios, and annotate the key points of the human body in the human image data to form a human posture training dataset.

[0156] The initial human pose estimation model is trained using the human pose training dataset, and the trained human pose estimation model is compressed and optimized using model pruning and parameter quantization techniques to obtain a lightweight human pose estimation model.

[0157] The construction process of the lightweight target detection model includes:

[0158] A lightweight object detection network is used as the basic framework to construct the initial object detection model;

[0159] Collect image data containing hands, various handheld objects, and door handles in smart door lock usage scenarios, and label the target categories and locations in the image data to form a target detection training dataset;

[0160] The initial object detection model is trained using the object detection training dataset, and the trained object detection model is compressed and optimized using channel pruning and parameter quantization techniques to obtain a lightweight object detection model.

[0161] Furthermore, in one embodiment, a biomechanical-temporal analysis is performed on the fingerprint press timing data to extract a dynamic feature vector of finger press, including:

[0162] Curve fitting was performed on the data segment of the pressure holding phase in the fingerprint pressing timing data, and the stress relaxation coefficient and creep coefficient were extracted from it.

[0163] Spectral analysis was performed on the pressure value sequence in the fingerprint pressing time series data to extract the pulse wave modulation energy features corresponding to the 0.8 Hz to 2 Hz frequency band and the muscle micro-tremor energy features corresponding to the 5 Hz to 20 Hz frequency band.

[0164] Calculate the fractal dimension of the pressure center trajectory based on the contact area coordinate sequence in the fingerprint pressing time sequence data;

[0165] The temperature value sequence in the fingerprint pressing time series data is analyzed, the temperature change rate curve is extracted, and the heat conduction time constant from the initial value to the stable value is calculated;

[0166] The stress relaxation coefficient, creep coefficient, pulse wave modulation energy characteristics, muscle micro-tremor energy characteristics, fractal dimension of the pressure center trajectory, temperature change rate curve, and heat conduction time constant are fused in multiple dimensions to generate a dynamic feature vector of finger pressing.

[0167] Furthermore, in one embodiment, curve fitting is performed on the data segment of the pressure holding phase in the fingerprint pressing timing data, and the stress relaxation coefficient and creep coefficient are extracted from it, including:

[0168] Identify the start and end times of the pressure holding phase from the fingerprint press timing data, and extract the pressure value subsequence and contact area coordinate subsequence corresponding to the pressure holding phase;

[0169] An exponential function is fitted to the subsequence of pressure values ​​during the pressure holding phase, and the stress relaxation coefficient is calculated based on the stress relaxation time constant obtained from the fitting. The stress relaxation coefficient is negatively correlated with the stress relaxation time constant.

[0170] Based on the contact area coordinate subsequence during the pressure holding phase, the contact area at each moment is calculated, and an area growth curve of the contact area over time is generated.

[0171] The area growth curve is fitted with a logarithmic function, and the creep coefficient is calculated based on the logarithmic growth coefficient obtained from the fitting. The creep coefficient is positively correlated with the logarithmic growth coefficient.

[0172] Furthermore, the information update module 14 is specifically used for:

[0173] A unique identity is established for each legitimate user, and a dedicated storage space associated with the identity is created in the cloud vector database. The legitimate user refers to a user who has completed registration in the smart door lock system and has been authorized to unlock the door.

[0174] The historical environmental interaction feature vector and historical pressing dynamic feature vector extracted from each verified normal unlocking event are stored as historical feature vectors in the dedicated storage space.

[0175] Multiple historical feature vectors in the dedicated storage space are clustered using a Gaussian mixture model to obtain K cluster centers, where each cluster center corresponds to a typical behavior pattern.

[0176] Calculate the frequency of each cluster center, take the cluster center with the highest frequency as the main pattern vector, and take the remaining cluster centers as the secondary pattern vectors;

[0177] Based on the primary pattern vector and the secondary pattern vector, a personal historical behavior profile of each legitimate user is constructed and continuously updated in the cloud-based big data platform.

[0178] Furthermore, the result recognition module 15 is specifically used for:

[0179] Based on the legitimate user identity identifier, the historical environment interaction feature vector and historical pressure dynamic feature vector associated with the legitimate user identity identifier are retrieved from the personal historical behavior archive of the cloud big data platform.

[0180] Calculate the first cosine similarity between the environmental interaction feature vector extracted during the current unlocking process and the historical environmental interaction feature vector;

[0181] Calculate the second cosine similarity between the dynamic feature vector of finger pressing extracted during the current unlocking process and the dynamic feature vector of finger pressing in the past process;

[0182] The first cosine similarity, the second cosine similarity, and the temporal context information of the current unlocking process are concatenated into a fused feature vector;

[0183] The fused feature vector is input into a pre-trained anomaly detection model, and the probability value that the current unlocking behavior is abnormal is calculated by the anomaly detection model.

[0184] The probability value is compared with a preset anomaly determination threshold. If the probability value exceeds the anomaly determination threshold, the current unlocking behavior is determined to be abnormal. If the probability value does not exceed the anomaly determination threshold, the current unlocking behavior is determined to be normal.

[0185] Furthermore, in one embodiment of the application, the process of constructing the anomaly detection model includes:

[0186] Collect historical unlocking data generated by legitimate users in normal unlocking scenarios, and extract the corresponding environmental interaction feature vector and finger pressing dynamic feature vector from the historical unlocking data as positive sample raw data;

[0187] Collect unlocking data generated under simulated attack scenarios, and extract the corresponding environmental interaction feature vector and finger pressing dynamic feature vector from the simulated attack unlocking data as negative sample raw data. The simulated attack scenarios include at least one or more of the following: unlocking with fingerprint film, unlocking under duress, and tailgating intrusion.

[0188] For each set of positive sample raw data and negative sample raw data, the historical environment interaction feature vector and historical pressing dynamic feature vector of the legitimate user corresponding to the unlocking operation are retrieved from the personal historical behavior archive of the cloud big data platform.

[0189] Calculate the cosine similarity between the environmental interaction feature vector in each set of positive sample original data or negative sample original data and the historical environmental interaction feature vector, and use it as the first similarity feature;

[0190] Calculate the cosine similarity between the finger pressing dynamic feature vector in each set of positive sample original data or negative sample original data and the historical pressing dynamic feature vector, and use it as the second similarity feature;

[0191] Obtain the temporal context information corresponding to the original data of each group of positive or negative samples as contextual features;

[0192] The first similarity feature, the second similarity feature, and the context feature are concatenated to form a training sample fusion feature vector set, and the corresponding normal label or abnormal label is used as the label of the training sample to form a training sample supervision label set.

[0193] The logistic regression model is trained using the fusion feature vector set of the training samples and the supervision label set of the training samples. The parameters of the logistic regression model are optimized by maximizing the likelihood function to obtain the trained anomaly recognition model.

Claims

1. A method for anomaly identification of smart door locks based on big data, characterized in that, The method includes: The smart door lock uses sensors to collect real-time video stream data and fingerprint press timing data during the unlocking process. Fingerprint feature matching is performed based on the fingerprint press timing data to determine the legitimate user identity of the current operator. After the fingerprint feature matching is successful, visual analysis is performed on the video stream data to extract the environmental interaction feature vector, and mechanical-time analysis is performed on the fingerprint pressing time sequence data to extract the finger pressing dynamic feature vector. Based on the environmental interaction feature vectors and finger pressing dynamic feature vectors extracted from the historical unlocking process of legitimate users, a personal historical behavior profile of each legitimate user is constructed and continuously updated in the cloud big data platform. The environmental interaction feature vector and the finger pressing dynamic feature vector extracted during the current unlocking process are compared with the historical feature vector corresponding to the legitimate user identity obtained from the personal historical behavior file. A comprehensive judgment is made through a pre-trained anomaly recognition model, and the recognition result of whether the current unlocking behavior is abnormal is output. If the identification result is abnormal, the door opening action will be refused; if the identification result is normal, the door opening action will be performed.

2. The method for anomaly identification of smart door locks based on big data according to claim 1, characterized in that, The smart lock uses sensors to collect real-time video stream data and fingerprint press timing data during the unlocking process, including: The high-definition infrared camera deployed on the smart door lock continuously collects video stream data of the entire process from when the user enters the camera's field of view to when the user completes the fingerprint pressing action at a first preset sampling frequency; The pressure-type fingerprint sensor deployed on the smart door lock synchronously collects the pressure value, contact area coordinate sequence, and temperature value of the user's finger during the fingerprint unlocking process at a second preset sampling frequency, generating fingerprint pressing time sequence data that changes over time.

3. The method for anomaly identification of smart door locks based on big data according to claim 1, characterized in that, Visual analysis is performed on the video stream data to extract environmental interaction feature vectors, including: By deploying a lightweight human pose estimation model on the edge of the door lock, the video stream data is analyzed frame by frame to extract the user's body key point coordinate sequence. Based on the body key point coordinate sequence, the relative position features of the user's body pose and the door lock, as well as the continuous action sequence features of the user during the unlocking process, are calculated. By deploying a lightweight target detection model on the edge of the door lock, the video stream data is analyzed frame by frame to detect the user's hand area and the state of the hand holding objects, and the user's hand holding object state features are extracted from it. The hand holding state features, the body posture and the relative position features of the door lock, and the continuous action sequence features during the unlocking process are encoded and fused to generate an environmental interaction feature vector.

4. The method for anomaly identification of smart door locks based on big data according to claim 3, characterized in that, The construction process of lightweight human pose estimation models and lightweight object detection models includes: The construction process of the lightweight human pose estimation model includes: A lightweight convolutional neural network is used as the backbone network, and an initial human pose estimation model is constructed by combining a key point detection head. Collect human image data under different lighting conditions, distances, and postures in smart door lock usage scenarios, and annotate the key points of the human body in the human image data to form a human posture training dataset. The initial human pose estimation model is trained using the human pose training dataset, and the trained human pose estimation model is compressed and optimized using model pruning and parameter quantization techniques to obtain a lightweight human pose estimation model. The construction process of the lightweight target detection model includes: A lightweight object detection network is used as the basic framework to construct the initial object detection model; Collect image data containing hands, various handheld objects, and door handles in smart door lock usage scenarios, and label the target categories and locations in the image data to form a target detection training dataset; The initial object detection model is trained using the object detection training dataset, and the trained object detection model is compressed and optimized using channel pruning and parameter quantization techniques to obtain a lightweight object detection model.

5. The method for anomaly identification of smart door locks based on big data according to claim 1, characterized in that, A biomechanical-temporal analysis is performed on the fingerprint press timing data to extract the dynamic feature vector of finger press, including: Curve fitting was performed on the data segment of the pressure holding phase in the fingerprint pressing timing data, and the stress relaxation coefficient and creep coefficient were extracted from it. Spectral analysis was performed on the pressure value sequence in the fingerprint pressing time series data to extract the pulse wave modulation energy features corresponding to the 0.8 Hz to 2 Hz frequency band and the muscle micro-tremor energy features corresponding to the 5 Hz to 20 Hz frequency band. Calculate the fractal dimension of the pressure center trajectory based on the contact area coordinate sequence in the fingerprint pressing time sequence data; The temperature value sequence in the fingerprint pressing time series data is analyzed, the temperature change rate curve is extracted, and the heat conduction time constant from the initial value to the stable value is calculated; The stress relaxation coefficient, creep coefficient, pulse wave modulation energy characteristics, muscle micro-tremor energy characteristics, fractal dimension of the pressure center trajectory, temperature change rate curve, and heat conduction time constant are fused in multiple dimensions to generate a dynamic feature vector of finger pressing.

6. The method for anomaly identification of smart door locks based on big data according to claim 5, characterized in that, Curve fitting is performed on the data segment of the pressure holding phase in the fingerprint press timing data, and the stress relaxation coefficient and creep coefficient are extracted from it, including: Identify the start and end times of the pressure holding phase from the fingerprint press timing data, and extract the pressure value subsequence and contact area coordinate subsequence corresponding to the pressure holding phase; An exponential function is fitted to the subsequence of pressure values ​​during the pressure holding phase, and the stress relaxation coefficient is calculated based on the stress relaxation time constant obtained from the fitting. The stress relaxation coefficient is negatively correlated with the stress relaxation time constant. Based on the contact area coordinate subsequence during the pressure holding phase, the contact area at each moment is calculated, and an area growth curve of the contact area over time is generated. The area growth curve is fitted with a logarithmic function, and the creep coefficient is calculated based on the logarithmic growth coefficient obtained from the fitting. The creep coefficient is positively correlated with the logarithmic growth coefficient.

7. The method for anomaly identification of smart door locks based on big data according to claim 1, characterized in that, Based on the environmental interaction feature vectors and finger pressure dynamic feature vectors extracted from the historical unlocking process of legitimate users, a personal historical behavior profile for each legitimate user is constructed and continuously updated in a cloud-based big data platform, including: A unique identity is established for each legitimate user, and a dedicated storage space associated with the identity is created in the cloud vector database. The legitimate user refers to a user who has completed registration in the smart door lock system and has been authorized to unlock the door. The historical environmental interaction feature vector and historical pressing dynamic feature vector extracted from each verified normal unlocking event are stored as historical feature vectors in the dedicated storage space. Multiple historical feature vectors in the dedicated storage space are clustered using a Gaussian mixture model to obtain K cluster centers, where each cluster center corresponds to a typical behavior pattern. Calculate the frequency of each cluster center, take the cluster center with the highest frequency as the main pattern vector, and take the remaining cluster centers as the secondary pattern vectors; Based on the primary pattern vector and the secondary pattern vector, a personal historical behavior profile of each legitimate user is constructed and continuously updated in the cloud-based big data platform.

8. The method for anomaly identification of smart door locks based on big data according to claim 1, characterized in that, The environmental interaction feature vector and the finger pressing dynamic feature vector extracted during the current unlocking process are compared with the historical feature vector corresponding to the legitimate user identity obtained from the personal historical behavior profile. A comprehensive judgment is then made using a pre-trained anomaly detection model to output a result indicating whether the current unlocking behavior is abnormal, including: Based on the legitimate user identity identifier, the historical environment interaction feature vector and historical pressure dynamic feature vector associated with the legitimate user identity identifier are retrieved from the personal historical behavior archive of the cloud big data platform. Calculate the first cosine similarity between the environmental interaction feature vector extracted during the current unlocking process and the historical environmental interaction feature vector; Calculate the second cosine similarity between the dynamic feature vector of finger pressing extracted during the current unlocking process and the dynamic feature vector of finger pressing in the past process; The first cosine similarity, the second cosine similarity, and the temporal context information of the current unlocking process are concatenated into a fused feature vector; The fused feature vector is input into a pre-trained anomaly detection model, and the probability value that the current unlocking behavior is abnormal is calculated by the anomaly detection model. The probability value is compared with a preset anomaly determination threshold. If the probability value exceeds the anomaly determination threshold, the current unlocking behavior is determined to be abnormal. If the probability value does not exceed the anomaly determination threshold, the current unlocking behavior is determined to be normal.

9. The method for anomaly identification of smart door locks based on big data according to claim 1, characterized in that, The process of building an anomaly detection model includes: Collect historical unlocking data generated by legitimate users in normal unlocking scenarios, and extract the corresponding environmental interaction feature vector and finger pressing dynamic feature vector from the historical unlocking data as positive sample raw data; Collect unlocking data generated under simulated attack scenarios, and extract the corresponding environmental interaction feature vector and finger pressing dynamic feature vector from the simulated attack unlocking data as negative sample raw data. The simulated attack scenarios include at least one or more of the following: unlocking with fingerprint film, unlocking under duress, and tailgating intrusion. For each set of positive sample raw data and negative sample raw data, the historical environment interaction feature vector and historical pressing dynamic feature vector of the legitimate user corresponding to the unlocking operation are retrieved from the personal historical behavior archive of the cloud big data platform. Calculate the cosine similarity between the environmental interaction feature vector in each set of positive sample original data or negative sample original data and the historical environmental interaction feature vector, and use it as the first similarity feature; Calculate the cosine similarity between the finger pressing dynamic feature vector in each set of positive sample original data or negative sample original data and the historical pressing dynamic feature vector, and use it as the second similarity feature; Obtain the temporal context information corresponding to the original data of each group of positive or negative samples as contextual features; The first similarity feature, the second similarity feature, and the context feature are concatenated to form a training sample fusion feature vector set, and the corresponding normal label or abnormal label is used as the label of the training sample to form a training sample supervision label set. The logistic regression model is trained using the fusion feature vector set of the training samples and the supervision label set of the training samples. The parameters of the logistic regression model are optimized by maximizing the likelihood function to obtain the trained anomaly recognition model.

10. A smart door lock anomaly detection system based on big data, characterized in that: The method for performing the big data-based smart door lock anomaly identification method according to any one of claims 1-9 includes: The information acquisition module is used to collect video stream data and fingerprint pressing timing data of the user in real time during the unlocking process through sensors on the smart door lock; The identity verification module is used to perform fingerprint feature matching based on the fingerprint pressing timing data to determine the legitimate user identity identifier corresponding to the current operator; The feature extraction module is used to perform visual analysis on the video stream data after the fingerprint feature matching is successful, extract the environmental interaction feature vector from it, and perform mechanical-temporal analysis on the fingerprint pressing time series data to extract the finger pressing dynamic feature vector from it. The information update module is used to build and continuously update the personal historical behavior profile of each legitimate user in the cloud big data platform based on the environmental interaction feature vector and finger pressing dynamic feature vector extracted by the legitimate user in the historical unlocking process. The result recognition module is used to compare the environmental interaction feature vector and the finger pressing dynamic feature vector extracted in the current unlocking process with the historical feature vector corresponding to the legitimate user identity obtained from the personal historical behavior file, and to make a comprehensive judgment through a pre-trained anomaly recognition model to output the recognition result of whether the current unlocking behavior is abnormal. The action execution module is used to refuse to execute the door opening action if the recognition result is abnormal, and to execute the door opening action if the recognition result is normal.

Images