Non-invasive vibration behavior identification methods, devices and storage media
By collecting and extracting multi-dimensional structural vibration features using a triaxial accelerometer, and combining Fisher ratio and Pareto optimal selection, a weighted normalized Euclidean distance is calculated. This solves the problems of high false alarm rate and cumbersome deployment in existing technologies, and achieves high-precision, low-power refrigerator door opening and closing behavior recognition.
Patent Information
- Application Number
- CN202611229238.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-13
- Publication Date
- 2026-11-10
AI Technical Summary
Existing vibration sensors cannot adapt to the differences in vibration characteristics of different installation locations and equipment models when identifying refrigerator door opening and closing behavior. They are difficult to distinguish between door closing actions and daily environmental interference, resulting in a high false alarm rate and cumbersome deployment.
Vibration signals are collected using a triaxial accelerometer, and multi-dimensional physically interpretable structural vibration features are extracted. The optimal feature subset is selected to form a feature vector, and a weighted normalized Euclidean distance is calculated for discrimination. Combined with Fisher ratio ranking and Pareto optimal selection, the optimal feature subset is adaptively matched to achieve non-invasive, configuration-free, interference-resistant, and ultra-low power consumption identification.
While ensuring privacy and security, it achieves high-precision refrigerator door opening and closing behavior recognition, reduces false alarm rate, simplifies deployment process, and runs stably on low-power microcontroller, providing reliable support for home activity monitoring.
Smart Images

Figure CN123069019A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of refrigerator vibration recognition technology, and in particular to a non-invasive vibration behavior recognition method, device and storage medium. Background Technology
[0002] As society ages, the number of elderly people living alone continues to grow, making their home safety monitoring increasingly urgent. Indirectly assessing the daily activities of the elderly by monitoring the usage patterns of various home appliances is a key technological direction in the field of home-based elderly care monitoring. Refrigerators, as one of the most frequently used and periodically cyclical home appliances, can effectively reflect the elderly's meal-taking and eating habits through their door-opening and closing behavior. Therefore, indirectly monitoring the daily activities of elderly people living alone by identifying refrigerator door opening and closing events is gradually becoming a research hotspot in this field.
[0003] Traditional sensor solutions mainly include reed switch sensors and commercial vibration sensors. Reed switch sensors are relatively inexpensive, but they require precise alignment between the refrigerator door and the refrigerator body, which not only damages the appearance of the appliance but also makes installation and maintenance more complicated.
[0004] Furthermore, traditional commercial vibration sensors mostly use fixed amplitude thresholds to determine door opening and closing. However, the vibration characteristics of different refrigerator models vary significantly, and the vibration transmission characteristics of the same refrigerator also differ depending on its installation location. Fixed thresholds cannot accommodate these differences. More importantly, such solutions struggle to effectively distinguish between vibrations generated by opening and closing the refrigerator door and vibrations caused by everyday disturbances such as a robot vacuum colliding with the refrigerator, a pet jumping and landing, or accidental bumps. In practical applications, the false alarm rate generally exceeds 20%. In summary, the current field lacks a systematic technical framework that simultaneously ensures privacy protection, easy deployment without configuration, strong environmental interference resistance, and ultra-low power consumption. It is difficult to achieve high-precision, non-intrusive refrigerator door opening and closing behavior recognition on a single low-power microcontroller.
[0005] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention
[0006] The main purpose of this application is to provide a non-invasive vibration behavior recognition method, device and storage medium, which aims to solve the technical problem that existing vibration sensors use fixed amplitude thresholds to recognize refrigerator door opening and closing, which cannot adapt to the differences in vibration characteristics of different installation locations and device models, and cannot distinguish between door closing actions and daily environmental interference, resulting in a high false alarm rate.
[0007] To achieve the above objectives, this application proposes a non-invasive vibration behavior recognition method, the method comprising: Vibration signals are collected by a triaxial accelerometer attached to a certain installation position on the outer wall of the refrigerator. When the vibration amplitude exceeds a preset trigger threshold, the microcontroller is awakened and the triaxial acceleration timing data of a single vibration event is recorded. Multi-dimensional physically interpretable structural vibration features are extracted from the triaxial acceleration time series data. These structural vibration features include total energy features, peak values for each axis, power attenuation ratio features, power proportion features for each axis, zero-crossing rate features for each axis, and normalized peak time features for each axis. According to predetermined screening criteria, an optimal feature subset is selected from the multi-dimensional structural vibration features. A feature vector is constructed based on the optimal feature subset. The screening criteria for the optimal feature subset are predetermined based on the installation location, and the optimal feature subsets corresponding to different installation locations are different. The optimal feature subset is determined as follows: the discrimination index of the structural vibration features in each dimension is calculated. The discrimination index is used to distinguish between door closing events and interference events. The features are sorted from high to low according to the discrimination index. Multiple candidate feature subsets are constructed based on the sorting results. The Pareto optimal feature subset is selected from the candidate feature subset with the false alarm rate and inference delay as optimization objectives. Calculate the weighted normalized Euclidean distance between the feature vector and the door closing event feature template; If the weighted normalized Euclidean distance is less than the preset classification threshold, it is determined to be a refrigerator door closing event; otherwise, it is determined to be an environmental interference event.
[0008] In one embodiment, the step of calculating the discrimination index of structural vibration characteristics in each dimension includes: Obtain the values of each sample in the door closing event sample set for the structural vibration characteristics in each dimension, and calculate the first mean and first variance of the structural vibration characteristics in each dimension in the door closing event sample set; Obtain the values of each sample in the disturbance event sample set for the structural vibration characteristics in each dimension, and calculate the second mean and second variance of the structural vibration characteristics in each dimension in the disturbance event sample set; For each dimension of structural vibration characteristics, the square of the difference between the first mean and the second mean is calculated. The square is then divided by the sum of the first variance and the second variance. The resulting ratio is used as the discrimination index of the structural vibration characteristics of that dimension. The larger the value of the discrimination index, the stronger the ability of the structural vibration characteristics of that dimension to distinguish between door closing events and disturbance events.
[0009] In one embodiment, the step of constructing multiple candidate feature subsets based on the ranking results, and selecting the Pareto optimal feature subset from the candidate feature subsets with false positive rate and inference latency as optimization objectives, includes: Based on the sorting results of the structural vibration features of each dimension from high to low according to the discrimination index, the structural vibration features of the first number of dimensions ranked first are selected to form the first candidate subset, the structural vibration features of the second number of dimensions ranked second are selected to form the second candidate subset, and so on, until all structural vibration features of all dimensions are selected to form the final candidate subset, thus obtaining multiple candidate feature subsets. For each candidate feature subset, the false alarm rate for distinguishing between door closing events and interference events based on the candidate feature subset is tested, and the inference latency required for performing feature extraction and distance calculation on the microcontroller based on the candidate feature subset is tested. Among all candidate feature subsets, a subset of candidate features whose inference latency does not exceed a preset latency budget is selected, and the subset of candidate features with the lowest false alarm rate is selected as the optimal feature subset.
[0010] In one embodiment, the step of pre-determining the selection criteria for the optimal feature subset based on the installation location, and ensuring that the optimal feature subsets corresponding to different installation locations are different, includes: When the installation location is the side wall opposite the refrigerator hinge, the optimal feature subset is the top ten structural vibration features in the order of the discrimination index from high to low. When the installation location is on the side wall near the refrigerator hinge, the optimal feature subset is the top eleven dimensions of structural vibration features after sorting according to the discrimination index from high to low. When the installation location is the top surface of the refrigerator, the optimal feature subset is the top seven dimensions of structural vibration features after sorting according to the discrimination index from high to low.
[0011] In one embodiment, the step of calculating the weighted normalized Euclidean distance between the feature vector and the door closing event feature template includes: For each dimension of the structural vibration feature in the feature vector, the feature value of the structural vibration feature in that dimension is subtracted from the template value of the corresponding dimension of the structural vibration feature in the door closing event feature template, and the difference is taken as the difference of the structural vibration feature in that dimension. The squared value of the difference is then multiplied by the component value of the weight vector corresponding to the structural vibration feature of the dimension to obtain the weighted squared distance of the structural vibration feature of the dimension. The weighted squared distances of the structural vibration features across all dimensions are summed, and the square root of the summation is taken. The resulting square root value is the weighted normalized Euclidean distance between the feature vector and the door closing event feature template.
[0012] In one embodiment, before the step of calculating the weighted normalized Euclidean distance between the feature vector and the door closing event feature template, the method further includes: In response to a user-triggered learning mode command, the microcontroller is controlled to enter a self-learning calibration mode; In the self-learning calibration mode, door closing vibration event samples corresponding to multiple door closing actions performed by the user and interference vibration event samples corresponding to multiple typical interference actions performed by the user are collected sequentially. Based on each vibration event sample collected, the feature vector corresponding to the optimal feature subset in the vibration event sample is extracted, and the mean and standard deviation of the structural vibration features of each dimension in the door closing event sample set, as well as the mean and standard deviation of the structural vibration features of each dimension in the interference event sample set, are calculated online in a preset recursive manner. After collecting all vibration event samples, the discrimination index of structural vibration features in each dimension is calculated based on the mean and standard deviation of the structural vibration features of the door closing event sample set and the interference event sample set in each dimension. The discrimination index is then normalized to generate the weight vector required for the weighted normalized Euclidean distance calculation. At the same time, the door closing event feature template is generated based on the mean of the structural vibration features of the door closing event sample set in each dimension.
[0013] In one embodiment, the typical interference actions include vibrations caused by colliding with the refrigerator, vibrations caused by jumping and landing, and vibrations caused by accidentally bumping the refrigerator; the door closing vibration event samples are collected five times, and the interference vibration event samples are collected three times.
[0014] In one embodiment, the microcontroller is a low-power microcontroller that supports deep sleep mode, and the method further includes: During periods when no vibration signal is detected, the microcontroller is controlled to maintain a deep sleep mode, in which the current consumption is less than microamperes; The microcontroller is activated to perform vibration signal acquisition, feature extraction, feature filtering, and weighted normalized Euclidean distance calculation operations only when the vibration amplitude detected by the triaxial accelerometer exceeds the preset trigger threshold. After determining whether a single vibration event is a door closing event or an environmental disturbance event, the microcontroller is controlled to return to the deep sleep mode.
[0015] In addition, to achieve the above objectives, this application also proposes a non-invasive vibration behavior recognition device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the non-invasive vibration behavior recognition method described above.
[0016] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the non-invasive vibration behavior recognition method described above.
[0017] One or more technical solutions proposed in this application have at least the following technical effects: The technical solution of this application collects vibration signals using a triaxial accelerometer attached to a specific installation location on the outer wall of a refrigerator. When the vibration amplitude exceeds a preset trigger threshold, a microcontroller is activated to record the triaxial acceleration time-series data of a single vibration event. Multi-dimensional physically interpretable structural vibration features are extracted from the triaxial acceleration time-series data. These structural vibration features include total energy characteristics, peak values for each axis, power attenuation ratio characteristics, power proportion characteristics for each axis, zero-crossing rate characteristics for each axis, and normalized peak time characteristics for each axis. Based on predetermined screening conditions, an optimal feature subset is selected from the multi-dimensional structural vibration features. A feature vector is constructed based on this optimal feature subset. The selection of the optimal feature subset... The selection criteria are predetermined based on the installation location, and the optimal feature subsets corresponding to different installation locations are different. The optimal feature subsets are determined as follows: calculate the discriminant index of the structural vibration features in each dimension. The discriminant index is used to distinguish between door closing events and interference events. Sort the discriminant index from high to low, construct multiple candidate feature subsets based on the sorting results, and select the Pareto optimal feature subset from the candidate feature subsets with the false alarm rate and inference delay as optimization objectives. Calculate the weighted normalized Euclidean distance between the feature vector and the door closing event feature template. If the weighted normalized Euclidean distance is less than a preset classification threshold, it is determined to be a refrigerator door closing event; otherwise, it is determined to be an environmental interference event.
[0018] The technical solution of this application solves the technical problems of existing fixed threshold solutions that cannot adapt to different installation locations and equipment models, have difficulty distinguishing between door closing actions and daily environmental interference, have high false alarm rates, and are cumbersome to deploy, while ensuring home privacy and security. It realizes a non-intrusive, configuration-free, interference-resistant, and ultra-low power refrigerator door opening and closing behavior recognition system that can run stably on a single low-power microcontroller, providing reliable technical support for home activity monitoring of elderly people living alone. Attached Figure Description
[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 This is a flowchart illustrating the first embodiment of the non-invasive vibration behavior recognition method of this application; Figure 2 This is a detailed step diagram of step S30 in one embodiment; Figure 3 This is a schematic diagram of another detailed step of step S30 in one embodiment; Figure 4 This is a detailed step diagram of step S40 in one embodiment; Figure 5 This is a flowchart illustrating the second embodiment of the non-invasive vibration behavior recognition method of the present invention; Figure 6 This is a flowchart illustrating the third embodiment of the non-invasive vibration behavior recognition method of the present invention; Figure 7 This is a schematic diagram of the device structure of the hardware operating environment involved in the non-invasive vibration behavior recognition method in this application embodiment.
[0022] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0023] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.
[0024] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.
[0025] In the field of refrigerator door opening and closing behavior recognition and home appliance usage monitoring, the relevant technical solutions mainly follow three technical paths: visual acoustic monitoring, radio frequency sensing, and traditional sensor detection. However, all of them have inherent defects and cannot meet the multiple requirements of privacy protection, ease of deployment, environmental robustness, and low power consumption operation.
[0026] While visual-acoustic solutions offer high recognition accuracy, they inherently present serious privacy and ethical issues in private home settings. Furthermore, their large data throughput and high computational demands on edge nodes make them difficult to implement in practice. Radio frequency (RF) sensing solutions, while enabling contactless monitoring, are susceptible to indoor multipath fading and dynamic interference from people walking around and pets, resulting in high false triggering frequencies and poor environmental robustness. In traditional reed switch and commercial vibration sensor solutions, reed switches require precise alignment and installation, which can damage the integrity of the appliance's appearance. Existing commercial vibration sensors mostly use fixed amplitude thresholds to determine door opening and closing. Since the vibration characteristics of different refrigerator models vary significantly, and the vibration transmission characteristics of the same refrigerator also differ at different installation locations, fixed threshold solutions cannot adapt to these differences. They also struggle to distinguish between refrigerator door opening and closing vibrations and everyday interference such as robot vacuum cleaner collisions, pet jumping, and accidental bumps, leading to a generally high false alarm rate in real-world scenarios.
[0027] Comprehensive analysis reveals that the core dilemma faced by the above-mentioned technical approaches lies in the fact that existing solutions all use a single dimension or single-point threshold for judgment, lack the ability to perform multi-dimensional physical modeling and adaptive analysis of vibration signals, cannot maintain stable recognition accuracy and anti-interference capability under different installation locations and equipment models, and cannot achieve high-precision, configuration-free door opening and closing behavior recognition on resource-constrained low-power microcontrollers.
[0028] To address the aforementioned shortcomings, this application proposes a non-invasive vibration behavior recognition method. The core inventive concept of this method lies in: utilizing the structural vibration signals generated by the opening and closing of a refrigerator door, collecting vibration data through a non-invasively attached triaxial accelerometer, and constructing a 14-dimensional physically interpretable structural vibration feature set comprising six categories: total energy, peak values for each axis, power attenuation ratio, power proportion for each axis, zero-crossing rate for each axis, and normalized peak time for each axis. This establishes a multi-dimensional physical representation system for the vibration signal, enabling the door-closing vibration to be distinguishable from various environmental disturbances in a high-dimensional space. To address the significant differences in vibration transmission characteristics at different installation locations, a Fischer method is employed. A two-stage feature selection method combining r-r ranking and Pareto optimal selection adaptively matches the optimal feature subsets (corresponding to 10-dimensional, 11-dimensional, and 7-dimensional features respectively) based on the installation location (side wall opposite the hinge, side wall near the hinge, and top of the refrigerator), achieving an optimal balance between recognition accuracy and inference latency. Furthermore, an eight-sample online self-learning calibration mechanism based on the Welford recursive algorithm is introduced. Users only need to complete five door-closing actions and three typical interference actions for the system to automatically generate a door-closing event feature template and the weight vector required for weighted distance calculation, enabling plug-and-play deployment with zero professional configuration. Based on the above feature extraction and adaptive selection, the system accurately distinguishes between door-closing events and environmental interference by calculating the weighted normalized Euclidean distance between the current event feature vector and the door-closing template and comparing it with a classification threshold.
[0029] Through the above-mentioned technical means, this application solves the technical problems of existing fixed threshold schemes that cannot adapt to different installation locations and equipment models, have difficulty distinguishing between door closing actions and daily environmental interference, have high false alarm rates, and are cumbersome to deploy, while ensuring home privacy and security. It realizes a non-intrusive, configuration-free, anti-interference, and ultra-low power consumption refrigerator door opening and closing behavior recognition system that can run stably on a single low-power microcontroller, providing reliable technical support for home activity monitoring of elderly people living alone.
[0030] Based on this, the embodiments of this application provide a non-invasive vibration behavior recognition method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the non-invasive vibration behavior recognition method of this application. In this embodiment, the non-invasive vibration behavior recognition method includes steps S10 to S50: Step S10: Vibration signals are collected by a triaxial accelerometer attached to a certain installation position on the outer wall of the refrigerator. When the vibration amplitude exceeds a preset trigger threshold, the microcontroller is woken up and the triaxial acceleration timing data of a single vibration event is recorded. In this embodiment, vibration signals are acquired and event-activated. Specifically, a triaxial accelerometer is attached to a specific location on the outer wall of the refrigerator. This triaxial accelerometer can simultaneously sense vibration acceleration signals in three orthogonal directions: the x-axis, y-axis, and z-axis. The sensor can be attached using double-sided adhesive or a magnetic structure, allowing the entire sensing node to be non-invasively fixed to the surface of the refrigerator's outer wall without requiring any drilling or modification to the refrigerator body. Different areas of the refrigerator's outer wall have different installation location attributes. These locations include the side wall opposite the refrigerator hinge, the side wall near the refrigerator hinge, and the top surface of the refrigerator. Different installation locations correspond to different vibration transmission paths and vibration attenuation characteristics.
[0031] During operation, the triaxial accelerometer continuously monitors vibration signals from the refrigerator's outer wall. The microcontroller is normally in deep sleep mode, where most of its functional modules are disabled, with only the minimum necessary circuitry for detecting vibration wake-up signals remaining in standby mode. When the refrigerator door opens or closes, or when external interference causes vibration, the vibration energy is transmitted through the refrigerator's outer shell to the triaxial accelerometer. The sensor converts the vibration signal into an electrical signal, which is then amplified and filtered by the internal signal conditioning circuitry, which calculates the current vibration amplitude in real time. This amplitude is compared to a preset trigger threshold. This preset threshold is a value pre-set during system initialization based on the sensor's sensitivity and background noise level; its purpose is to filter out weak vibration noise from the environment and prevent the microcontroller from being frequently woken up by meaningless minor vibrations.
[0032] When the vibration amplitude exceeds the preset trigger threshold, the triaxial accelerometer sends a wake-up signal to the microcontroller via an interrupt pin. Upon receiving this wake-up signal, the microcontroller switches from deep sleep mode to working mode and begins subsequent data acquisition operations. After the microcontroller is awakened, the system records triaxial acceleration time-series data within a continuous time window starting from the vibration trigger moment. This time-series data includes the dynamic change curves of the acceleration values in the x, y, and z axes over time throughout the entire process of the vibration event from its onset to decay. This time-series data is the foundational data source for all subsequent feature extraction and recognition judgments, and its completeness directly affects the accuracy of the entire recognition method. After completing the time-series data recording for a single vibration event, the microcontroller immediately enters the subsequent feature extraction process, while the triaxial accelerometer continues to monitor the next round of vibration signals, waiting for the next vibration event exceeding the preset trigger threshold to occur.
[0033] Specifically, a total of [number] samples were collected during a single complete vibration event. N The number of triaxial acceleration sampling points, the first k The acceleration vector of each sampling point is ,in They are respectively x , y , z Axis at the k The acceleration value at any given moment. Through the event-driven wake-up mechanism described above, the system remains in an ultra-low-power sleep state most of the time, only operating briefly when a vibration event occurs, thus achieving the goal of long-term maintenance-free deployment.
[0034] Step S20: Extract multi-dimensional physically interpretable structural vibration features from the triaxial acceleration time series data. The structural vibration features include total energy features, peak values for each axis, power attenuation ratio features, power proportion features for each axis, zero-crossing rate features for each axis, and normalized peak time features for each axis. After acquiring triaxial acceleration time-series data, multi-dimensional, physically interpretable structural vibration features were extracted from the raw time-series data. The core idea behind this process is that the vibration generated by the closing of a refrigerator door and the vibration generated by various environmental disturbances have different vibration propagation characteristics at the physical level. By extracting multi-dimensional features that characterize these physical differences, a foundation can be laid for accurate classification later. A total of six types of structural vibration features were extracted, comprising fourteen feature dimensions, denoted as M1 to M... 14 Each type of feature corresponds to a specific physical property of the vibration signal.
[0035] First, extract the total energy feature, denoted as M1: ; This feature is obtained by integrating the square of the triaxial acceleration time series data on the time axis, and its physical meaning is the total vibration energy contained in a single vibration event. The closing action involves the collision between the refrigerator door and the refrigerator body, which usually produces a relatively strong impact vibration, so the total energy characteristic value is relatively large; while slight disturbances, such as a pet lightly touching the refrigerator, have a relatively small total energy characteristic value.
[0036] Next, peak features for each axis are extracted and denoted as M2, M3, and M4, respectively. These features are obtained by extracting the maximum absolute value of the acceleration time series data for the x-axis, y-axis, and z-axis, respectively, to obtain the peak acceleration values for the three axes: ; The peak values along different axes reflect the distribution of vibration energy in three spatial directions. When a refrigerator door closes, the vibration energy is mainly concentrated in the direction perpendicular to the door panel due to the door rotating around the hinge and impacting the refrigerator. However, the vibration generated by a robot vacuum cleaner colliding with a refrigerator may have a larger component in the horizontal direction. This difference can be quantified by the peak characteristics of each axis.
[0037] Then, the power attenuation ratio feature is extracted and denoted as M5. This feature divides the time series data of a single vibration event into two parts, the front segment and the back segment, according to the peak time. The energy of the front segment and the energy of the back segment are calculated separately, and the ratio of the energy of the back segment to the energy of the front segment is used as the power attenuation ratio.
[0038] First, define instantaneous power: The sampling number corresponding to its maximum value is ,but: ; The impact vibration generated by closing a door has the characteristic of rapid decay, and its power attenuation is relatively small; while the vibration generated by the continuous collision of the robot vacuum cleaner or the continuous jumping of a pet often decays more slowly or even exhibits the characteristics of multiple impacts, and its power attenuation is relatively large.
[0039] Next, the power proportion features of each axis are extracted and denoted as M6, M7, and M8, respectively. These features are used to calculate the power proportion of the total power along the x-axis, y-axis, and z-axis. ; This feature reflects the distribution pattern of vibration energy in three spatial axes. Vibrations from different sources have different axial energy distribution fingerprints due to their different excitation directions and conduction paths.
[0040] Then extract the zero-crossing rate features for each axis, denoted as M9 and M, respectively. 10 M 11 This feature counts the number of times the acceleration time series data for each axis crosses the zero level per unit time, and defines an indicator function. The value is 1 if the condition is true, and 0 otherwise. ; Zero-crossing rate reflects the oscillation frequency characteristics of a vibration signal. Door-closing impact signals usually contain rich high-frequency components and have a high zero-crossing rate; while low-frequency interference such as accidental bumps to the human body has a relatively low zero-crossing rate.
[0041] Finally, the normalized peak time features for each axis are extracted and denoted as M. 12 M 13 M 14 This feature records the position of the peak acceleration of each axis relative to the start time of the vibration event, and normalizes it using the total duration of the vibration event. Let be the sampling point numbers with the largest absolute values of acceleration for each axis, then: ; The peak values appear sequentially along different axes, and this time series reflects the propagation sequence of vibration energy in different directions, serving as an important physical clue for distinguishing different vibration sources. These six characteristics comprehensively characterize the vibration signal from six dimensions: energy, amplitude, attenuation, distribution, frequency, and time, forming a multidimensional feature space with clear physical meaning.
[0042] Step S30: According to the predetermined screening conditions, the optimal feature subset is selected from the multi-dimensional structural vibration features. A feature vector is constructed based on the optimal feature subset. The screening conditions for the optimal feature subset are predetermined according to the installation position, and the optimal feature subsets corresponding to different installation positions are different. The optimal feature subset is determined by: calculating the discrimination index of the structural vibration features of each dimension. The discrimination index is used to distinguish between door closing events and interference events. The features are sorted from high to low according to the discrimination index. Multiple candidate feature subsets are constructed according to the sorting results. The Pareto optimal feature subset is selected from the candidate feature subset with the false alarm rate and inference delay as optimization objectives. After extracting multi-dimensional structural vibration features, the optimal feature subset is selected. Because there are many extracted feature dimensions, and the vibration transmission characteristics differ significantly across installation locations, not all features have the same discriminative power across all locations. Therefore, it is necessary to select the optimal feature subset based on the specific installation location to achieve the best balance between recognition accuracy and computational cost. This optimal feature subset is determined offline before system deployment and, once determined, is embedded in the microcontroller's firmware. During online runtime, the selection result can be directly invoked.
[0043] The process of determining the optimal feature subset begins with calculating the discriminative index for each dimension of the feature. This calculation requires the prior collection of two types of sample data: vibration samples from refrigerator door closing events and vibration samples from various environmental disturbance events. For each dimension of structural vibration features, the mean μ_close,i and variance σ_close,i² of that feature in the door closing sample set, and the mean μ_disturb,i and variance σ_disturb,i² of that feature in the disturbance sample set are calculated. Then, for each dimension of the feature, its discriminative index is calculated according to the following Fisher ratio formula: ; Where μ_close,i and σ_close,i² are the mean and variance of the door closing event samples on the i-th feature, respectively; μ_disturb,i and σ_disturb,i² are the mean and variance of the interference event samples on the i-th feature, respectively; and ε= To prevent division by zero, the normalized weight vector is: : The higher the value of the discrimination index, the greater the inter-class distance relative to the intra-class dispersion of that feature between closing events and interference events, meaning the stronger the discriminative power of that feature. Through the above calculations, the discrimination index values for each of the various features can be obtained.
[0044] Then, all features are sorted according to their discriminative power index from highest to lowest. After sorting, multiple candidate feature subsets are constructed based on the sorting results. Specifically, the first candidate subset is constructed by selecting the top-ranked features of the first dimension, followed by the first and second-ranked features of the second dimension, and so on, adding one dimension each time, until all dimensions are selected to form the final candidate subset. This results in a total number of candidate feature subsets equal to the total number of features, with each candidate subset containing a certain number of top-ranked features.
[0045] Next, the performance of each candidate feature subset is evaluated. The evaluation includes two dimensions: the first is the false alarm rate, which is the probability of misjudging door closing events and interference events using this candidate subset; the second is the inference latency, which is the time required for feature extraction and distance calculation on the microcontroller using this candidate subset. The false alarm rate is obtained by testing on a validation set, and the inference latency is obtained by actual operation measurement on the target microcontroller. From all candidate feature subsets, firstly, candidate subsets whose inference latency does not exceed a preset latency budget are selected. Then, the candidate subset with the lowest false alarm rate among these latency-constrained subsets is selected as the optimal feature subset for that installation location. Through this two-stage selection process, the final optimal feature subset ensures both recognition accuracy and meets the real-time requirements of hardware resources.
[0046] Furthermore, the step of pre-determining the selection criteria for the optimal feature subset based on the installation location, and ensuring that the optimal feature subsets corresponding to different installation locations are different, includes: When the installation location is the side wall opposite the refrigerator hinge, the optimal feature subset is the top ten structural vibration features in the order of the discrimination index from high to low. When the installation location is on the side wall near the refrigerator hinge, the optimal feature subset is the top eleven dimensions of structural vibration features after sorting according to the discrimination index from high to low. When the installation location is the top surface of the refrigerator, the optimal feature subset is the top seven dimensions of structural vibration features after sorting according to the discrimination index from high to low.
[0047] In this embodiment, the selection criteria for the optimal feature subset are predetermined based on the installation location, and the optimal feature subsets corresponding to different installation locations are different. For the specific application of a refrigerator, different regions of its outer wall exhibit significantly different vibration transmission characteristics, and these differences stem from the non-uniformity of the refrigerator's internal structure. The three typical installation locations—the side wall opposite the refrigerator hinge, the side wall near the hinge, and the top surface of the refrigerator—represent three different vibration propagation paths and boundary conditions, thus their corresponding optimal feature subsets differ in the number of feature dimensions. After the above two-stage selection process, under the constraint that the microcontroller inference delay does not exceed the preset delay budget, optimal feature subsets of different sizes are selected for each of the three installation locations.
[0048] When the installation location is the side wall opposite the refrigerator hinge, the optimal feature subset consists of the top ten structural vibration features ranked from highest to lowest according to their discriminative power. The side wall opposite the refrigerator hinge is located at the far end of the refrigerator door's opening direction. When the door closes, it rotates around the hinge and eventually impacts the refrigerator body. Vibration energy is transmitted from the impact point through the door panel, door frame, and hinge structure to the side wall opposite the hinge. This transmission path crosses multiple structural interfaces between the door and the refrigerator body, and the vibration signal undergoes complex reflections and refractions during transmission. Therefore, the discriminative power of each feature dimension is relatively dispersed. Through the Pareto optimal selection process, the top ten features are sufficient to cover the main differences between door-closing vibration and environmental interference in terms of energy, peak value, attenuation, axial distribution, frequency, and time. While adding more feature dimensions can slightly reduce the false alarm rate, the increase in inference latency exceeds the preset latency budget. Therefore, the top ten features are the optimal choice for the installation location opposite the hinge.
[0049] When the installation location is the side wall near the refrigerator hinge, the optimal feature subset consists of the top eleven dimensions of structural vibration features, sorted from highest to lowest according to the discriminative index. The side wall near the hinge is adjacent to the hinge installation location. When the door closes, the vibration energy generated by the impact is directly transmitted to the outer wall of the refrigerator body in this area through the hinge structure. The transmission path is short and there are few structural interfaces, so the vibration signal retains relatively rich original feature information when it reaches the sensor, especially with minimal attenuation of high-frequency components. This transmission characteristic allows the side wall near the hinge to retain more effective feature information in more dimensions. These additional feature dimensions each contribute to different degrees of discriminative ability when distinguishing between door closing events and interference events. Through the Pareto optimal selection process, it was found that selecting eleven dimensions achieves the optimal balance between false alarm rate and inference delay. Although the transmission path of the side wall near the hinge is short, due to its proximity to the vibration source, this location is more susceptible to the influence of multiple vibration modes generated at different stages of the door opening and closing process. Therefore, one more feature dimension is needed than that of the side wall opposite the hinge to comprehensively characterize this complex vibration mode.
[0050] When the installation location is the top surface of the refrigerator, the optimal feature subset consists of the top seven structural vibration features ranked from highest to lowest according to their discriminative power. The top surface of the refrigerator is far from the door impact area; vibration energy travels from the impact point to the top surface through the entire three-dimensional structure of the refrigerator, resulting in the longest transmission path and the most structural interfaces involved. High-frequency vibration components attenuate significantly during long-distance transmission, leading to lower signal-to-noise ratios and weaker discriminative power for some features at the top surface location. Pareto optimality screening revealed that, at the refrigerator top installation location, only the top seven features possess sufficient inter-class discriminative power; the remaining dimensions have low discriminative power. Increasing these low-discriminative features not only fails to significantly reduce the false alarm rate but also increases inference latency and may introduce noise interference. Therefore, the top seven features are the optimal choice for the refrigerator top installation location. The number of optimal feature dimensions corresponding to the three installation locations mentioned above indicates that there is a clear correspondence between the installation location and the optimal feature subset. This correspondence is obtained by independently optimizing each installation location through the two-stage screening process in step S30, ensuring that the best balance between recognition performance and resource overhead can be achieved under different installation conditions.
[0051] Step S40: Calculate the weighted normalized Euclidean distance between the feature vector and the door closing event feature template; After selecting the optimal feature subset and constructing a feature vector based on it, the weighted normalized Euclidean distance between this feature vector and the door-closing event feature template is calculated. This process is the core decision-making step for determining whether the current vibration event is a refrigerator door-closing event. Before calculating the weighted normalized Euclidean distance, two key parameters need to be generated through an online self-learning calibration process: the door-closing event feature template and the weight vector.
[0052] Specifically, during the generation of the door closing event feature template, after the user triggers the learning mode command, the microcontroller enters a self-learning calibration mode. In this mode, the user sequentially performs five door closing actions. Each time the system detects a vibration event generated by the door closing action, it extracts the feature vector corresponding to the optimal feature subset of that event. After collecting the feature vectors of all five door closing samples, the system calculates the mean of these five samples for the structural vibration features of each dimension, and uses this mean vector as the door closing event feature template. This template represents the typical vibration characteristic pattern of the refrigerator's door closing action at this installation position.
[0053] During the generation of the weight vector, based on the self-learning calibration mode, in addition to performing five door-closing actions, the user also needs to perform three typical disturbance actions. These typical disturbance actions include vibrations caused by the robot vacuum colliding with the refrigerator, vibrations caused by a pet jumping and landing, and vibrations caused by a person accidentally bumping into the refrigerator. After each vibration event caused by a disturbance action is detected, the feature vector corresponding to the optimal feature subset of that event is extracted. After collecting the feature vectors of all three disturbance samples, the mean and variance of the disturbance samples are calculated for the structural vibration features of each dimension. Combining the mean and variance of the previous door-closing samples, the discrimination index of each dimension feature is calculated according to the Fisher ratio calculation formula mentioned above, and the discrimination indices of all dimensions are normalized so that the sum of the weight values of each dimension equals a preset constant. The normalized discrimination index vector is the weight vector. In this weight vector, the feature dimension with stronger discrimination ability is assigned a larger weight, playing a more important role in subsequent distance calculations.
[0054] After the feature template and weight vector for the door closing event have been generated, the actual distance calculation is performed. For each dimension of the structural vibration feature in the feature vector of the current vibration event, firstly, the feature value of that dimension is calculated and subtracted from the template value of the corresponding dimension in the feature template of the door closing event, resulting in a difference. Then, the square of this difference is taken, and the squared value is multiplied by the component value of the weight vector corresponding to that dimension, resulting in the weighted squared distance for that dimension. The weighted squared distances of all dimensions are summed, and finally, the square root of the sum is taken. The square root value is the weighted normalized Euclidean distance between the current feature vector and the feature template of the door closing event, i.e.: ; like (Preset classification threshold) If the event is classified as a door closing event, it is considered a disturbance event; otherwise, it is classified as a disturbance event. The smaller the distance value, the closer the vibration characteristics of the current vibration event are to the typical characteristics of a door closing event, and the greater the probability that the current event is a door closing event.
[0055] Step S50: If the weighted normalized Euclidean distance is less than the preset classification threshold, it is determined to be a refrigerator door closing event; otherwise, it is determined to be an environmental interference event.
[0056] After calculating the weighted normalized Euclidean distance between the feature vector of the current vibration event and the feature template of the door closing event, the final event type determination is made based on the relationship between this distance value and a preset classification threshold. This preset classification threshold is a value determined through experimental statistics during the system design phase; its physical meaning is the optimal decision boundary for distinguishing between door closing events and environmental disturbance events. Ideally, the weighted normalized Euclidean distances of all door closing events should be distributed on one side of the threshold, while the weighted normalized Euclidean distances of all environmental disturbance events should be distributed on the other side, allowing the two types of events to be clearly distinguished in terms of distance.
[0057] The judgment logic is as follows: The calculated weighted normalized Euclidean distance is compared with a preset classification threshold. If the distance is less than the preset classification threshold, it indicates that the vibration characteristics of the current vibration event are sufficiently similar to the feature template of a door closing event in the weighted multidimensional space. That is, the current event highly matches the typical characteristics of a door closing action in all selected dimensions, including energy characteristics, peak characteristics, attenuation characteristics, axial distribution characteristics, frequency characteristics, and time characteristics. Therefore, the system determines the current vibration event to be a refrigerator door closing event. If the distance is greater than or equal to the preset classification threshold, it indicates that the vibration characteristics of the current vibration event are insufficiently similar to the feature template of a door closing event. That is, the current event differs significantly from the typical characteristics of a door closing action in the key feature dimensions. Therefore, the system determines the current vibration event to be an environmental disturbance event, rather than a refrigerator door closing event.
[0058] After a refrigerator door closing event is identified, subsequent data upload operations can be performed, sending information such as the time of the closing event to the upper-level monitoring data platform. This data is used to statistically analyze the frequency and timing of refrigerator use, indirectly reflecting the daily meal-taking and eating behaviors of elderly people living alone. If an environmental disturbance event is identified, no data upload operation is performed; the system simply records the event as a disturbance and then returns to sleep mode, avoiding misreporting normal daily disturbances as elderly activity. Regardless of the identification result, after completing the identification and processing of this vibration event, the microcontroller returns to deep sleep mode, waiting to be awakened again when the vibration amplitude exceeds a preset trigger threshold. Through this event-driven identification and sleep mechanism, the system can maintain ultra-low power consumption during long-term operation while ensuring accurate identification and classification of every valid vibration event.
[0059] Further, see Figure 2 , Figure 2 This is a detailed step diagram of step S30 in one embodiment, based on Figure 2 The step of calculating the discrimination index of structural vibration characteristics in each dimension includes S31~S32: Step S31: Obtain the values of each sample in the door closing event sample set in terms of structural vibration characteristics in each dimension, and calculate the first mean and first variance of the structural vibration characteristics in each dimension in the door closing event sample set. Step S32: Obtain the values of each sample in the interference event sample set in terms of structural vibration characteristics in each dimension, and calculate the second mean and second variance of the structural vibration characteristics in each dimension in the interference event sample set. Step S33: For the structural vibration characteristics of each dimension, calculate the square of the difference between the first mean and the second mean, divide the square by the sum of the first variance and the second variance, and use the resulting ratio as the discrimination index of the structural vibration characteristics of the dimension. The larger the value of the discrimination index, the stronger the ability of the structural vibration characteristics of the dimension to distinguish between door closing events and disturbance events.
[0060] In this embodiment, the discriminative index of the structural vibration characteristics in each dimension is the fundamental step in performing the optimal feature subset selection. The calculation of this discriminative index requires a sample set of door-closing events and a sample set of disturbance events as the data basis. By statistically analyzing the value distribution of the two types of samples on the same dimension feature, the effectiveness of this dimension feature in distinguishing between door-closing events and disturbance events is quantified.
[0061] First, the values of structural vibration features in each dimension of the sample set of closing events are obtained, and the first mean and first variance of each dimension feature in the closing event sample set are calculated based on these values. The closing event sample set is constructed as follows: After system deployment, the installer or user manually performs multiple refrigerator door closing actions. The vibration events generated by each closing action are collected by a triaxial accelerometer and processed as described above to obtain the multi-dimensional structural vibration feature vector of that closing event. The feature vectors corresponding to multiple closing actions are summarized to form the closing event sample set. For each dimension feature in this set, the arithmetic mean of the values of all closing samples in that dimension is calculated. This mean is the first mean, denoted as μ_close,i, where the subscript i represents the feature dimension index. At the same time, the dispersion of the values of all closing samples in that dimension is calculated, i.e., the first variance, denoted as σ_close,i². This variance reflects the stability of the closing action itself in terms of vibration features. If the value of a certain dimension of the door-closing action is very stable, then the first variance of that dimension is small, indicating that the feature has good consistency within the door-closing event, which is helpful in distinguishing it from interfering events.
[0062] Then, the values of structural vibration features in each dimension of the sample in the interference event sample set are obtained, and the second mean and second variance of each dimension feature in the interference event sample set are calculated based on these values. The construction method of the interference event sample set is similar to that of the door closing event sample set, the difference being that the objects collected are various typical daily environmental interference actions, including but not limited to vibrations caused by a robot vacuum cleaner colliding with a refrigerator, vibrations caused by a pet jumping and landing, and vibrations caused by a person accidentally bumping into a refrigerator. The above interference actions represent the three most frequent sources of non-closed door vibration in the home environment. The feature vectors corresponding to multiple interference actions are summarized to form the interference event sample set. For each dimension feature in this set, the arithmetic mean of the values of all interference samples in that dimension is calculated, and this mean is the second mean, denoted as μ_disturb,i. At the same time, the dispersion of the values of all interference samples in that dimension is calculated, i.e., the second variance, denoted as σ_disturb,i². If the values of a certain dimension feature are very dispersed within the interference events, the second variance of that dimension is large, indicating that the feature changes drastically between different types of interference, which is actually helpful in distinguishing it from the relatively concentrated values of the door closing event.
[0063] Next, for each dimension of structural vibration characteristics, the square of the difference between the first and second means is calculated, and this square is divided by the sum of the first and second variances. The resulting ratio is used as the discrimination index for that dimension's characteristics. The above calculation process uses the basic principle of the Fisher's ratio, and its formula is FR_i = (μ_close,i...). The formula is: μ_disturb,i)² / (σ_close,i² + σ_disturb,i² + ε), where ε is a very small positive constant to prevent overflow due to a zero denominator. The physical meaning of this formula is as follows: The numerator represents the inter-class distance between the closed sample and the interference sample on this dimension. The larger the distance, the more significant the difference in values between the two types of events on this dimension, and the better it is for distinguishing them. The denominator represents the sum of the intra-class dispersion of the closed sample and the interference sample on this dimension. The smaller the sum, the more concentrated the values within each type of event, and the better it is for distinguishing them. Therefore, the larger the numerator and the smaller the denominator, the larger the ratio, and the stronger the discriminative power of this dimension. The larger the discriminative index value calculated by the above steps, the stronger the ability of this dimension to distinguish between closed events and interference events, and the higher the priority of this dimension in subsequent feature ranking. The calculation result of this discrimination index is used throughout the entire feature selection process. It directly determines the quality of the input data for feature ranking and candidate subset construction in step S30 and is a key prerequisite for ensuring the scientific nature of the optimal feature subset selection result.
[0064] Further, see Figure 3 , Figure 3 This is a schematic diagram of another detailed step of step S30 in one embodiment, based on Figure 3 The step of constructing multiple candidate feature subsets based on the ranking results, and selecting the Pareto optimal feature subset from the candidate feature subsets with false positive rate and inference latency as optimization objectives, includes S34~S36: Step S34: According to the sorting results of the discrimination index of the structural vibration features of each dimension from high to low, select the first number of structural vibration features of the top-ranked dimensions to form the first candidate subset, select the second number of structural vibration features of the top-ranked dimensions to form the second candidate subset, and so on, until all dimensions of structural vibration features are selected to form the final candidate subset, thus obtaining multiple candidate feature subsets. Step S35: For each candidate feature subset, test the false alarm rate of distinguishing between door closing events and interference events based on the candidate feature subset, and test the inference latency required for performing feature extraction and distance calculation on the microcontroller based on the candidate feature subset; Step S36: In all candidate feature subsets, select the candidate feature subset whose inference delay does not exceed the preset delay budget, and select the candidate feature subset with the lowest false alarm rate as the optimal feature subset from the selected candidate feature subset.
[0065] In this embodiment, after calculating the discriminative index of structural vibration features in each dimension, multiple candidate feature subsets are constructed based on the ranking results, and the optimal feature subset is selected through multi-objective optimization. The core of this selection process lies in balancing two mutually restrictive optimization objectives: one is recognition accuracy, quantified by the false alarm rate; the other is computational efficiency, quantified by the inference latency required to perform feature extraction and distance calculation on the microcontroller. These two objectives are inherently competitive. Increasing the feature dimension usually improves recognition accuracy and reduces the false alarm rate, but it also increases inference latency; reducing the feature dimension can reduce inference latency, but it may sacrifice recognition accuracy. Therefore, a Pareto optimal selection method is needed to find the solution set that achieves the optimal balance between the false alarm rate and inference latency among all candidate feature subsets.
[0066] First, based on the ranking of the discriminative indices of structural vibration features across all dimensions from highest to lowest, multiple candidate feature subsets are constructed sequentially. The specific construction method is as follows: The discriminative indices of all calculated features across all dimensions are arranged in descending order, resulting in an ordered feature sequence. Then, starting with the first feature in this sequence, a different number of features from the top-ranking dimensions are selected to form candidate subsets. First, the features of the top-ranking first dimension are selected to form the first candidate subset, which contains only the feature with the strongest discriminative power. Then, the features of the top-ranking first and second dimensions are selected to form the second candidate subset, which contains the two features with the strongest discriminative power. This process continues, adding one dimension of features each time, until all dimensions are selected to form the final candidate subset. Through this method, a total number of candidate feature subsets equal to the total number of feature dimensions are obtained. Each candidate subset contains a certain number of top-ranking features, and each subsequent candidate subset adds one more dimension of features to the previous one. This nested candidate subset construction method ensures the monotonicity of feature selection. That is, as the feature dimension increases, the amount of information contained in the candidate subset continues to increase, providing a complete candidate space for subsequent Pareto optimal selection.
[0067] After constructing the candidate feature subsets, performance testing is performed on each subset. Performance testing includes two dimensions. The first dimension is the false positive rate test, which is conducted as follows: Pre-labeled validation set samples are input into the system. The system uses the current candidate feature subset to extract the feature vector of each sample and then performs discrimination based on door closing events and environmental interference. The discrimination results are compared with the true labels of the samples, and the proportion of incorrectly judged samples to the total number of validation set samples is counted. This proportion is the false positive rate of the candidate subset. The validation set samples should cover door closing events and various environmental interference events, and should be independent of the sample set used to calculate the discrimination index to ensure the objectivity of the test results. The second dimension is the inference latency test, which is conducted as follows: The feature extraction and distance calculation program corresponding to the current candidate feature subset is deployed on the target microcontroller. The complete inference process is run a preset number of times, and the average time required from the start of feature extraction to the completion of the weighted normalized Euclidean distance calculation is measured. This average time is the inference latency of the candidate subset. Inference latency needs to be measured in a real-world operating environment to reflect the microcontroller’s processing capability under real-world working conditions.
[0068] After performance testing of all candidate feature subsets, a subset with inference latency not exceeding a preset latency budget is selected. From this subset, the one with the lowest false positive rate is chosen as the optimal feature subset. The preset latency budget is a time limit pre-set during the system design phase based on the microcontroller's processing power and the real-time requirements of the application scenario. Its purpose is to ensure that the finally selected feature subset can run in real-time on the target hardware platform. This step first performs latency constraint screening, eliminating all candidate subsets whose inference latency exceeds the preset latency budget. While these eliminated subsets may have lower false positive rates, the excessive feature dimensions result in computational demands exceeding the microcontroller's real-time processing capabilities, making them unsuitable for practical deployment. Within the candidate subsets that meet the latency constraints, the system compares the false positive rates of each subset and selects the one with the lowest false positive rate as the optimal feature subset for that installation location. Through the above two-stage screening, the final optimal feature subset ensures that the recognition accuracy is as low as possible within an acceptable range, while also ensuring that the inference latency meets the real-time constraints of the hardware platform, thus achieving the optimal engineering balance between recognition performance and computing resources.
[0069] Further, see Figure 4 , Figure 4 This is a detailed step diagram of step S40 in one embodiment, based on the shown... Figure 4 The step of calculating the weighted normalized Euclidean distance between the feature vector and the door closing event feature template includes S41~S43: Step S41: For each dimension of the structural vibration feature in the feature vector, subtract the template value of the corresponding dimension of the structural vibration feature in the door closing event feature template from the feature value of the structural vibration feature of the dimension, and use the difference as the difference of the structural vibration feature of the dimension. Step S42: Squaring the difference, multiplying the squaring value by the component value of the weight vector corresponding to the structural vibration feature of the dimension, to obtain the weighted squared distance of the structural vibration feature of the dimension. Step S43: Sum the weighted squared distances of the structural vibration features of all dimensions, and take the square root of the summation result. The square root value is the weighted normalized Euclidean distance between the feature vector and the door closing event feature template.
[0070] In this embodiment, the calculation of weighted normalized Euclidean distance is the core decision-making basis for determining whether the current vibration event is a refrigerator door closing event. The calculation process involves three input parameters: the feature vector of the current vibration event, the door closing event feature template generated in advance through self-learning, and the weight vector. These three parameters work synergistically within the framework of weighted normalized Euclidean distance calculation. Through a three-step process of dimensional difference calculation, weighted square summation, and square root normalization, a distance value that quantifies the similarity between the current vibration event and the door closing event is finally output. The smaller the distance value, the closer the vibration characteristics of the current vibration event are to the door closing action.
[0071] First, for each dimension of the structural vibration feature in the aforementioned feature vector, the feature value of that dimension's structural vibration feature is subtracted from the template value of the corresponding dimension's structural vibration feature in the aforementioned door-closing event feature template. The resulting difference is taken as the difference of the structural vibration feature in that dimension. Taking the i-th dimension as an example, let the feature value of the current vibration event be f_i, and the template value of the corresponding dimension in the door-closing event feature template be t_i. Then, according to the calculation method, the difference d_i = f_i - t_i. The physical meaning of this difference is the degree of deviation between the current vibration event and the door-closing event in the i-th dimension. If the absolute value of this difference is large, it indicates that the current vibration event differs significantly from the typical characteristics of the door-closing action in that dimension, and this dimension provides strong evidence for distinguishing between the two types of events; conversely, if the absolute value of this difference is small, it indicates that the current vibration event is relatively close to the typical characteristics of the door-closing action in that dimension, and this dimension tends to support the judgment that the current event is a door-closing event. The above difference calculation process is performed one by one for each dimension in the optimal feature subset, resulting in a set of difference sequences with the same number of dimensions as the feature subset.
[0072] Then, the squared difference is calculated and multiplied by the component value of the weight vector corresponding to the feature dimension to obtain the weighted squared distance of that feature dimension. Taking the i-th feature dimension as an example, let the calculated difference be d_i, and the component value of the i-th dimension in the weight vector be w_i. Then, the weighted squared distance is p_i = w_i × d_i². This step involves two key operations. The first operation is to square the difference, which eliminates the influence of the positive or negative sign of the difference on the distance metric, so that both positive and negative deviations are converted into positive distance contributions. At the same time, the squaring operation amplifies the influence of larger differences, so that events with significant deviations in key dimensions can receive a greater penalty in the final distance value. The second operation is to multiply by the component value of the weight vector, which reflects the importance of each feature dimension in distinguishing between closed events and interference events. The weight vector is generated as described in step S40 by normalizing the discrimination index of each feature dimension. Dimensions with stronger discrimination ability are assigned larger weight values. Therefore, in the calculation of weighted squared distance, dimensions with strong discriminative power contribute more to the final distance value, while dimensions with weak discriminative power contribute less. This weighting mechanism ensures that the recognition process focuses on the most discriminative key feature dimensions, improving the accuracy of classification decisions.
[0073] Finally, the weighted squared distances of the structural vibration features across all dimensions are summed, and the square root of the sum is taken. The square root value is the weighted normalized Euclidean distance between the aforementioned feature vector and the aforementioned door-closing event feature template. Let N represent the number of dimensions in the optimal feature subset. The weighted normalized Euclidean distance D is calculated as D = √(Σ_{i=1}^{N} w_i × d_i²). The summation operation in this step combines the weighted squared distances of all dimensions into a single value, achieving a comprehensive convergence of multi-dimensional information. The square root operation after summation restores the distance dimension to the same dimension as the feature value, giving the distance value a clear physical scale meaning. The weighted normalized Euclidean distance D obtained through the above three steps is a non-negative real number, theoretically ranging from zero to positive infinity. When D approaches zero, it indicates that the current vibration event is completely consistent with the feature template of a door-closing event in all dimensions, and the confidence level of the current event being a door-closing event is the highest. As D gradually increases, it indicates that the deviation of the current vibration event from the feature template of a door-closing event in all dimensions gradually increases, the confidence level of the current event being a door-closing event gradually decreases, and the confidence level of the current event being an environmental disturbance event gradually increases. Finally, by comparing D with a preset classification threshold, binary classification of door-closing events and environmental disturbance events can be achieved. Compared with the traditional fixed threshold comparison method, the above-mentioned weighted normalized Euclidean distance calculation method has multi-dimensional joint judgment capability and adaptive weighting capability, and can achieve accurate event classification under complex home environmental disturbance conditions.
[0074] Furthermore, you can also view Figure 5 , Figure 5 This is a flowchart illustrating the second embodiment of the non-invasive vibration behavior recognition method of the present invention, based on the shown... Figure 5 Before the step of calculating the weighted normalized Euclidean distance between the feature vector and the door closing event feature template, steps S60~S90 are further included: Step S60: In response to a user-triggered learning mode command, control the microcontroller to enter self-learning calibration mode; Step S70: In the self-learning calibration mode, door closing vibration event samples corresponding to multiple door closing actions performed by the user, and interference vibration event samples corresponding to multiple typical interference actions performed by the user are collected sequentially. Step S80: Based on each collected vibration event sample, extract the feature vector corresponding to the optimal feature subset in the vibration event sample, and calculate the mean and standard deviation of the structural vibration features of each dimension in the door closing event sample set and the mean and standard deviation of the structural vibration features of each dimension in the interference event sample set online in a preset recursive manner. Step S90: After collecting all vibration event samples, calculate the discrimination index of structural vibration features in each dimension based on the mean and standard deviation of the structural vibration features of the door closing event sample set and the interference event sample set in each dimension, and normalize the discrimination index to generate the weight vector required for the weighted normalized Euclidean distance calculation. At the same time, generate the door closing event feature template based on the mean of the structural vibration features of the door closing event sample set in each dimension.
[0075] In this embodiment, the calculation of the weighted normalized Euclidean distance relies on two key parameters: a door closing event feature template and a weight vector. These two parameters are not pre-defined in the system but are dynamically generated based on the current refrigerator model and installation location through an online self-learning calibration process.
[0076] First, in response to a user-triggered learning mode command, the microcontroller enters self-learning calibration mode. This learning mode command can be triggered in various ways, such as via a physical button on the terminal node, a Bluetooth command sent via a companion mobile application, or a preset sequence of actions. Upon receiving the learning mode command, the microcontroller immediately pauses the normal vibration event monitoring and recognition process and switches to self-learning calibration mode. In this mode, the microcontroller disables its door-closing event and environmental interference detection functions, instead performing sample acquisition and parameter calculation. Simultaneously, it provides feedback to the user on the current learning status through indicator light flashing or a buzzer.
[0077] After entering the self-learning calibration mode, vibration event samples corresponding to multiple door-closing actions performed by the user, as well as interference vibration event samples corresponding to multiple typical interference actions performed by the user, are collected sequentially under this mode. The door-closing action samples are collected as follows: the user pushes and closes the refrigerator door according to normal usage habits, allowing the door to close naturally under the action of the door lock mechanism. The vibration event generated by each door-closing action is collected by a three-axis accelerometer. Through the aforementioned wake-up and recording process and feature extraction process, a multi-dimensional structural vibration feature vector for that door-closing action is generated and stored in the door-closing event sample set. Five door-closing actions are collected, covering different closing forces and speeds that may occur in daily use, ensuring the door-closing sample set is sufficiently representative. The interference action samples are collected as follows: the user sequentially performs three typical interference actions, including vibrations caused by a robot vacuum colliding with the refrigerator, vibrations caused by a pet jumping and landing, and vibrations caused by a person accidentally bumping into the refrigerator. These three interference actions represent the three most frequent sources of non-door-closing vibrations in the home environment. Each time a disturbance action is executed, a feature vector for that disturbance action is generated following the same data processing procedure as the door-closing sample, and then stored in the disturbance event sample set. The disturbance actions are collected three times, covering three different types of disturbance sources. The five door-closing samples and the three disturbance samples, totaling eight samples, constitute the entire data foundation for self-learning calibration.
[0078] After collecting all eight samples, based on each vibration event sample, the feature vector corresponding to the optimal feature subset mentioned above is extracted. The mean and standard deviation of the structural vibration features of each dimension in the door-closing event sample set, as well as the mean and standard deviation of the structural vibration features of each dimension in the disturbance event sample set, are calculated online using a preset recursive method. The preset recursive method employs the Welford recursive algorithm, whose core feature is that it can update statistics online without storing all historical data. Assuming there is already... n One sample, mean is The cumulative sum of squares is Join the n A new sample hour: ; ; ; ; all n After processing each sample, the mean is Unbiased variance is The standard deviation is Specifically, when receiving the feature vectors of the closing samples one by one, each time a new sample is received, the mean and variance of the current closing sample set are updated according to the Welford recursive formula. After the update is complete, the feature vector data of that sample can be discarded; it is not necessary to retain the complete data of all five closing samples. Similarly, the statistics of the interference samples are also updated using the same recursive method. For the closing event (the first 5 times), for each feature dimension... k : ; After each event is processed, the sample count is... n The increment is 1, continuing until all eight iterations are completed. The advantage of this recursive method is that microcontrollers have extremely limited storage resources. Storing all sample data for offline statistical calculations would consume a large amount of SRAM. However, with the recursive method, the system only needs to store the current sample count, mean, and variance to complete the statistical calculations for all samples, significantly reducing SRAM usage. Furthermore, the recursive method supports flexible expansion of the sample size. If users find the recognition accuracy insufficient during actual use, they can collect new samples at any time, and the system can update the statistics without reprocessing historical data.
[0079] After recursively calculating the mean and standard deviation of the closing event sample and the interference event sample, the discrimination index of the structural vibration characteristics in each dimension is calculated based on the mean and standard deviation of the structural vibration characteristics in each dimension of the closing event sample set and the interference event sample set. This discrimination index is then normalized to generate the weight vector required for the weighted normalized Euclidean distance calculation. Simultaneously, the closing event feature template is generated based on the mean of the structural vibration characteristics in each dimension of the closing event sample set. The generation method of the weight vector is consistent with the Fisher ratio calculation method used to determine the optimal feature subset, i.e., according to the formula FR_i = (μ_close,i... The discriminant index for each dimension of the feature is calculated using μ_disturb,i)² / (σ_close,i² + σ_disturb,i² + ε). Then, all discriminant indices are normalized so that the sum of the weights for each dimension equals a preset constant. The normalized discriminant index vector is the weight vector. The above-mentioned door-closing event feature template is generated by directly using the mean of the calculated door-closing event sample set across all dimensions as the template value. Through this online self-learning calibration process, only five door-closing actions and three interference actions from the user are required to automatically generate the door-closing event feature template and weight vector. The entire process requires no professional configuration and enables rapid adaptive deployment for different refrigerator models and installation locations.
[0080] Furthermore, you can also view Figure 6 , Figure 6 This is a flowchart illustrating the third embodiment of the non-invasive vibration behavior recognition method of the present invention, based on the shown... Figure 6 The microcontroller is a low-power microcontroller that supports deep sleep mode, and the method further includes steps S100~S120: Step S100: During the period when no vibration signal is detected, the microcontroller is controlled to maintain a deep sleep mode, wherein the current consumption in the deep sleep mode is less than microamperes. Step S110: Only when the vibration amplitude detected by the triaxial accelerometer exceeds the preset trigger threshold, the microcontroller is woken up to perform vibration signal acquisition, feature extraction, feature filtering and weighted normalized Euclidean distance calculation operations. Step S120: After determining whether a single vibration event is a door closing event or an environmental disturbance event, control the microcontroller to return to the deep sleep mode.
[0081] In this embodiment, the pre-configured complete algorithm for refrigerator door opening and closing behavior recognition relies on the power supply of the microcontroller for continuous operation. In home monitoring scenarios, terminal sensing nodes are typically deployed using battery power. If the microcontroller operates at full speed for extended periods, the battery power will be depleted within a few days, failing to meet the practical requirements for long-term maintenance-free deployment. To address these energy consumption constraints, the microcontroller used is a low-power microcontroller that supports deep sleep mode. This type of microcontroller has multiple power state switching capabilities and can dynamically adjust its power consumption level according to the workload, thereby achieving an optimal balance between energy consumption and processing performance.
[0082] This plays a fundamental role in the power management mechanism of the microcontroller. During periods when no vibration signal is detected, the system controls the microcontroller to maintain a deep sleep mode. This deep sleep mode is the lowest power consumption operating mode among all power states of the microcontroller. In this deep sleep mode, the microcontroller's central processing unit (CPU) core completely stops its clock supply and ceases all instruction fetching and decoding operations; the clock signals of most peripheral modules are cut off, including serial communication interfaces, general-purpose input / output ports, timers, and analog-to-digital converters, all of which are in a powered-on state; only the minimum necessary circuitry for responding to external wake-up signals remains powered on and ready to go. This minimum necessary circuitry typically includes an external interrupt trigger pin and a low-power wake-up timer. Because this deep sleep mode shuts down the power supply and clock of most functional modules, its current consumption is below the microampere level, with a typical measured value of approximately twelve microamperes. This microampere-level sleep current ensures that the terminal sensing node consumes only a very small amount of power most of the time, providing a fundamental physical layer guarantee for long-term maintenance-free deployment.
[0083] In the aforementioned low-power architecture, the microcontroller performs event-driven wake-up execution. Only when the vibration amplitude detected by the triaxial accelerometer exceeds the preset trigger threshold will the system wake up the microcontroller to perform vibration signal acquisition, feature extraction, feature filtering, and weighted normalized Euclidean distance calculation. This wake-up mechanism is a typical event-driven architecture, and its workflow is as follows: When the triaxial accelerometer detects an external vibration signal, its internal signal conditioning circuit amplifies and filters the vibration signal, and a comparator compares the processed signal amplitude with the preset trigger threshold. When the signal amplitude exceeds the preset trigger threshold, the accelerometer sends an edge-triggered signal to the microcontroller's external interrupt input pin through its interrupt output pin. Even in deep sleep mode, the microcontroller maintains its ability to detect the external interrupt pin. Once the edge-triggered signal is detected, the microcontroller's power management unit immediately initiates the wake-up sequence, sequentially restoring the clock supply to the central processing unit core, resetting key registers, enabling necessary peripheral modules, and finally switching the microcontroller from deep sleep mode to operating mode. The completion time of the above wake-up sequence is typically on the order of microseconds, ensuring that the timing data of the vibration event is not lost due to wake-up delay. After the microcontroller enters the working mode, it completes all operations in sequence, including three-axis acceleration timing data acquisition, multi-dimensional structural vibration feature extraction, optimal feature subset selection, weighted normalized Euclidean distance calculation, and door closing event and environmental interference discrimination. The identification process for the current vibration event then ends.
[0084] This event-driven architecture forms the final link in the control loop. After determining whether a single vibration event is a door-closing event or an environmental disturbance event, the microcontroller is controlled to return to the deep sleep mode. This return operation is executed by the power management unit within the microcontroller, specifically: first, key state data of the current operating mode is saved to non-volatile memory to ensure correct context recovery upon wake-up; then, the clocks of enabled peripheral modules are sequentially disabled; finally, a sleep command is sent to the central processing unit core to stop the core clock supply, causing the microcontroller to return to deep sleep. Through this coordinated operation, the microcontroller remains in working state during the brief time window of a vibration event and in deep sleep mode throughout the long time interval between two vibration events. This event-driven power management mechanism reduces the average power consumption of the terminal sensing node to an extremely low level. In actual deployment scenarios, based on approximately twenty vibration events triggered daily, a 500 mAh lithium battery can support continuous operation for more than eighteen months, fully meeting the application requirements of long-term maintenance-free deployment in home monitoring scenarios.
[0085] This application provides a non-invasive vibration behavior recognition device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the non-invasive vibration behavior recognition method in Embodiment 1 above.
[0086] The following is for reference. Figure 7 The diagram illustrates a structural schematic suitable for implementing a non-invasive vibration behavior recognition device according to embodiments of this application. The non-invasive vibration behavior recognition device in embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 7 The non-invasive vibration behavior recognition device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.
[0087] like Figure 7As shown, the non-invasive vibration behavior recognition device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the non-invasive vibration behavior recognition device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the non-invasive vibration behavior recognition device to communicate wirelessly or wiredly with other devices to exchange data. Although a non-invasive vibration behavior recognition device with various systems is shown in the figure, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.
[0088] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.
[0089] The non-invasive vibration behavior recognition device provided in this application, employing the non-invasive vibration behavior recognition method described in the above embodiments, solves the technical problem of high false alarm rates caused by existing vibration sensors using fixed amplitude thresholds for refrigerator door opening and closing recognition. This is due to the inability to adapt to the vibration characteristics differences of different installation locations and device models, and the inability to distinguish between door closing actions and daily environmental interference. Compared with the prior art, the beneficial effects of the non-invasive vibration behavior recognition device provided in this application are the same as those of the non-invasive vibration behavior recognition method provided in the above embodiments, and other technical features of this non-invasive vibration behavior recognition device are the same as those disclosed in the previous embodiment method, and will not be repeated here.
[0090] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.
[0091] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
[0092] This application provides a storage medium, which is a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the non-intrusive vibration behavior recognition method in the above embodiments.
[0093] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.
[0094] The aforementioned computer-readable storage medium may be included in the non-invasive vibration behavior recognition device; or it may exist independently and not assembled into the non-invasive vibration behavior recognition device.
[0095] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the non-invasive vibration behavior recognition device, enable the non-invasive vibration behavior recognition device to implement the technical content of the non-invasive vibration behavior recognition method embodiment shown above.
[0096] Computer program code for performing the operations of this application can be written in one or more programming languages or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, and C++, and conventional procedural programming languages such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).
[0097] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.
[0098] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.
[0099] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned non-invasive vibration behavior recognition method. This solves the technical problem of high false alarm rates caused by existing vibration sensors using fixed amplitude thresholds for refrigerator door opening and closing recognition, which cannot adapt to differences in vibration characteristics across different installation locations and device models, and cannot distinguish between door closing actions and everyday environmental interference. Compared to the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the non-invasive vibration behavior recognition method provided in the above embodiments, and will not be elaborated upon here.
Claims
1. A non-invasive vibration behavior recognition method, characterized in that, Includes the following steps: Vibration signals are collected by a triaxial accelerometer attached to a certain installation position on the outer wall of the refrigerator. When the vibration amplitude exceeds a preset trigger threshold, the microcontroller is awakened and the triaxial acceleration timing data of a single vibration event is recorded. Multi-dimensional physically interpretable structural vibration features are extracted from the triaxial acceleration time series data. These structural vibration features include total energy features, peak values for each axis, power attenuation ratio features, power proportion features for each axis, zero-crossing rate features for each axis, and normalized peak time features for each axis. According to predetermined screening criteria, an optimal feature subset is selected from the multi-dimensional structural vibration features. A feature vector is constructed based on the optimal feature subset. The screening criteria for the optimal feature subset are predetermined based on the installation location, and the optimal feature subsets corresponding to different installation locations are different. The optimal feature subset is determined as follows: the discrimination index of the structural vibration features in each dimension is calculated. The discrimination index is used to distinguish between door closing events and interference events. The features are sorted from high to low according to the discrimination index. Multiple candidate feature subsets are constructed based on the sorting results. The Pareto optimal feature subset is selected from the candidate feature subset with the false alarm rate and inference delay as optimization objectives. Calculate the weighted normalized Euclidean distance between the feature vector and the door closing event feature template; If the weighted normalized Euclidean distance is less than the preset classification threshold, it is determined to be a refrigerator door closing event; otherwise, it is determined to be an environmental interference event.
2. The non-invasive vibration behavior recognition method as described in claim 1, characterized in that, The steps for calculating the discrimination index of structural vibration characteristics in each dimension include: Obtain the values of each sample in the door closing event sample set for the structural vibration characteristics in each dimension, and calculate the first mean and first variance of the structural vibration characteristics in each dimension in the door closing event sample set; Obtain the values of each sample in the disturbance event sample set for the structural vibration characteristics in each dimension, and calculate the second mean and second variance of the structural vibration characteristics in each dimension in the disturbance event sample set; For each dimension of structural vibration characteristics, the square of the difference between the first mean and the second mean is calculated. The square is then divided by the sum of the first variance and the second variance. The resulting ratio is used as the discrimination index of the structural vibration characteristics of that dimension. The larger the value of the discrimination index, the stronger the ability of the structural vibration characteristics of that dimension to distinguish between door closing events and disturbance events.
3. The non-invasive vibration behavior recognition method as described in claim 1, characterized in that, The step of constructing multiple candidate feature subsets based on the ranking results, and selecting the Pareto optimal feature subset from the candidate feature subsets with false positive rate and inference latency as optimization objectives, includes: Based on the sorting results of the structural vibration features of each dimension from high to low according to the discrimination index, the structural vibration features of the first number of dimensions ranked first are selected to form the first candidate subset, the structural vibration features of the second number of dimensions ranked second are selected to form the second candidate subset, and so on, until all structural vibration features of all dimensions are selected to form the final candidate subset, thus obtaining multiple candidate feature subsets. For each candidate feature subset, the false alarm rate for distinguishing between door closing events and interference events based on the candidate feature subset is tested, and the inference latency required for performing feature extraction and distance calculation on the microcontroller based on the candidate feature subset is tested. Among all candidate feature subsets, a subset of candidate features whose inference latency does not exceed a preset latency budget is selected, and the subset of candidate features with the lowest false alarm rate is selected as the optimal feature subset.
4. The non-invasive vibration behavior recognition method as described in claim 1, characterized in that, The step of pre-determining the selection criteria for the optimal feature subset based on the installation location, and ensuring that the optimal feature subsets corresponding to different installation locations are different, includes: When the installation location is the side wall opposite the refrigerator hinge, the optimal feature subset is the top ten structural vibration features in the order of the discrimination index from high to low. When the installation location is on the side wall near the refrigerator hinge, the optimal feature subset is the top eleven dimensions of structural vibration features after sorting according to the discrimination index from high to low. When the installation location is the top surface of the refrigerator, the optimal feature subset is the top seven dimensions of structural vibration features after sorting according to the discrimination index from high to low.
5. The non-invasive vibration behavior recognition method as described in claim 1, characterized in that, The step of calculating the weighted normalized Euclidean distance between the feature vector and the door closing event feature template includes: For each dimension of the structural vibration feature in the feature vector, the feature value of the structural vibration feature in that dimension is subtracted from the template value of the corresponding dimension of the structural vibration feature in the door closing event feature template, and the difference is taken as the difference of the structural vibration feature in that dimension. The squared value of the difference is then multiplied by the component value of the weight vector corresponding to the structural vibration feature of the dimension to obtain the weighted squared distance of the structural vibration feature of the dimension. The weighted squared distances of the structural vibration features across all dimensions are summed, and the square root of the summation is taken. The resulting square root value is the weighted normalized Euclidean distance between the feature vector and the door closing event feature template.
6. The non-invasive vibration behavior recognition method as described in claim 1, characterized in that, Before the step of calculating the weighted normalized Euclidean distance between the feature vector and the door closing event feature template, the method further includes: In response to a user-triggered learning mode command, the microcontroller is controlled to enter a self-learning calibration mode; In the self-learning calibration mode, door closing vibration event samples corresponding to multiple door closing actions performed by the user and interference vibration event samples corresponding to multiple typical interference actions performed by the user are collected sequentially. Based on each vibration event sample collected, the feature vector corresponding to the optimal feature subset in the vibration event sample is extracted, and the mean and standard deviation of the structural vibration features of each dimension in the door closing event sample set, as well as the mean and standard deviation of the structural vibration features of each dimension in the interference event sample set, are calculated online in a preset recursive manner. After collecting all vibration event samples, the discrimination index of structural vibration features in each dimension is calculated based on the mean and standard deviation of the structural vibration features of the door closing event sample set and the interference event sample set in each dimension. The discrimination index is then normalized to generate the weight vector required for the weighted normalized Euclidean distance calculation. At the same time, the door closing event feature template is generated based on the mean of the structural vibration features of the door closing event sample set in each dimension.
7. The non-invasive vibration behavior recognition method as described in claim 6, characterized in that, The typical interference actions include vibrations caused by colliding with the refrigerator, vibrations caused by jumping and landing, and vibrations caused by accidentally bumping the refrigerator; the door closing vibration event samples were collected five times, and the interference vibration event samples were collected three times.
8. The non-invasive vibration behavior recognition method according to any one of claims 1 to 7, characterized in that, The microcontroller is a low-power microcontroller that supports deep sleep mode, and the method further includes: During periods when no vibration signal is detected, the microcontroller is controlled to maintain a deep sleep mode, in which the current consumption is less than microamperes; The microcontroller is activated to perform vibration signal acquisition, feature extraction, feature filtering, and weighted normalized Euclidean distance calculation operations only when the vibration amplitude detected by the triaxial accelerometer exceeds the preset trigger threshold. After determining whether a single vibration event is a door closing event or an environmental disturbance event, the microcontroller is controlled to return to the deep sleep mode.
9. A non-invasive vibration behavior recognition device, characterized in that, The non-invasive vibration behavior recognition device stores a computer program, which, when executed by a processor, implements the non-invasive vibration behavior recognition method according to any one of claims 1-8.
10. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the non-invasive vibration behavior recognition method according to any one of claims 1-8.
