Pet positioner electric quantity optimization system based on deep learning

Abstract

The invention relates to the technical field of pet locator electric quantity optimization, in particular to a pet locator electric quantity optimization system based on deep learning, which performs preprocessing on acceleration, direction and locating coordinate data, constructs a time sequence feature vector, and performs time sequence coding on the feature vector based on a deep learning model to optimize the electric quantity of a pet locator. And generating a motion trend prediction result containing the speed variation and the direction estimator so as to realize dynamic inference of the motion state of the pet. And further realizing self-adaptive distribution of the sampling frequency by distinguishing a static state from a moving state and combining the direction change amplitude. And performing position extrapolation calculation on the current position based on the speed variation and the direction estimator, generating a positioning precision adjustment strategy according to the distance between the predicted position and the geo-fence boundary, and adaptively adjusting the positioning precision from the angle of space risk. And finally, integrating the sampling frequency and the positioning precision strategy to uniformly update the parameters of the positioner, thereby realizing the energy consumption optimization of the electric quantity and the dynamic balance of the positioning quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pet locator power optimization, and in particular to a pet locator power optimization system based on deep learning. BACKGROUND

[0002] The existing pet locator usually relies on multi-source sensors such as acceleration sensors, electronic compasses, GNSS modules for data collection, and its positioning strategy is usually run in a fixed sampling frequency and fixed accuracy mode. However, pets have the characteristics of frequent changes in motion state and uneven spatial movement in daily activities, and the demand for positioning frequency and positioning accuracy is significantly different in different states. For example, frequent sampling in a stationary state not only cannot improve the positioning accuracy, but also causes a large amount of invalid energy consumption; in the high-speed motion state or near the boundary of the geographic fence, the fixed accuracy mode is difficult to reflect the position change in time, thereby reducing the monitoring effectiveness.

[0003] In order to alleviate the problem of power consumption, the existing technology uses low-power chips, prolongs the sampling interval, or reduces the number of high-precision positioning, but these methods often rely on static rules or fixed thresholds, and lack the ability to dynamically adapt to changes in pet behavior. When the pet's direction changes rapidly, the speed increases suddenly, or the trajectory is complex, the static rule is difficult to accurately identify the current behavior mode, resulting in lag or inaccuracy in sampling frequency adjustment, so that energy consumption control and positioning reliability are difficult to be considered at the same time. SUMMARY

[0004] To solve the above problems, the present application provides a pet locator power optimization system based on deep learning.

[0005] To achieve the above purpose, the technical scheme adopted by the present application is: A pet locator power optimization system based on deep learning, comprising: A data acquisition module for collecting acceleration sequences, direction sequences and positioning coordinates through a pet locator and preprocessing, constructing a time series feature vector and inputting a deep learning prediction model; A prediction module for performing time series encoding on the time series feature vector according to the deep learning prediction model to generate a motion trend prediction result, the motion trend prediction result including a speed change and a direction estimate; A frequency adjustment module for generating a sampling frequency adjustment strategy according to the motion trend prediction result; An accuracy adjustment module for calculating a predicted position based on the motion trend prediction result, and generating a positioning accuracy adjustment strategy according to the distance between the predicted position and the boundary of the geographic fence; An optimization control module for performing power optimization control on the pet locator based on the sampling frequency and accuracy adjustment results.

[0006] Furthermore, the process of collecting acceleration sequences, direction sequences, and location coordinates through a pet locator and performing preprocessing includes the following steps: The acceleration sequence is subjected to noise reduction filtering, and the acceleration change is normalized based on the time window to obtain acceleration preprocessed data; Angle smoothing and outlier angle correction are performed on the direction sequence to obtain direction preprocessed data; The positioning coordinates are subjected to trajectory interpolation and drift correction processing to obtain preprocessed positioning coordinate data; Based on the acceleration preprocessing data, orientation preprocessing data, and positioning coordinate preprocessing data, time alignment and feature concatenation processing are performed to obtain the temporal feature vector.

[0007] Furthermore, the deep learning prediction model is a long short-term memory network.

[0008] Furthermore, the prediction module is used to perform the following steps: Based on the temporal feature vector, a long short-term memory network is used to perform a gated update process on the feature input at each time step to obtain the first temporal hidden feature containing the forward state. Based on the temporal feature vector, the feature input at each time step is subjected to gated update processing of the inverse long short-term memory network to obtain the second temporal hidden feature containing the backward state; Perform feature concatenation processing on the first and second temporal hidden features to obtain bidirectional temporal hidden features; Based on the bidirectional temporal hidden features, nonlinear mapping processing is performed to generate motion trend prediction results.

[0009] Furthermore, the frequency adjustment module is used to perform the following steps: Based on the velocity change in the motion trend prediction results, threshold discrimination processing is performed on the motion state to obtain a stationary label or a motion label. Based on the motion state, the current sampling frequency is adjusted by either increasing or decreasing the sampling frequency, including decreasing the sampling frequency when the sampling frequency is at rest and increasing the sampling frequency when the sampling frequency is at motion. The change amplitude is calculated based on the direction estimate in the motion trend prediction results to obtain the direction change amplitude. Based on strategy selection processing, the sampling frequency value is determined according to the directional amplitude, thus obtaining the sampling frequency adjustment strategy.

[0010] Furthermore, the precision adjustment module is used to perform the following steps: Based on the velocity change and direction estimate in the motion trend prediction results, the current positioning coordinates are extrapolated to obtain the predicted position at the next preset time. Based on the predicted location, the nearest distance calculation is performed on the geofence boundary to obtain the distance data between the predicted location and the geofence. Based on the distance data, a threshold discrimination process is performed on the positioning accuracy requirement to obtain an accuracy level identifier; The accuracy adjustment range is calculated based on the accuracy level identifier to obtain the positioning accuracy adjustment strategy.

[0011] Furthermore, the process of performing position extrapolation calculation on the current positioning coordinates based on the velocity change and direction estimate in the motion trend prediction results includes: Calculate the displacement direction vector per unit time based on the current positioning coordinates and direction estimate; Based on the velocity change and the displacement direction vector per unit time, the displacement increment is calculated at the current position to obtain the predicted displacement increment. Based on the predicted displacement increment, the current positioning coordinates are updated to obtain the predicted position at the next preset time.

[0012] Furthermore, the nearest distance calculation includes the following steps: Based on the predicted location, perform point-by-point distance calculation on the polygon vertices of the geofence boundary to obtain vertex distance data from the predicted location to each vertex; Based on the predicted location, the vertical projection distance is calculated for each edge segment of the geofence boundary to obtain the line segment distance data from the predicted location to each edge segment. The vertex distance data and line segment distance data are subjected to minimum value determination processing to obtain the distance data between the predicted location and the geofence.

[0013] Furthermore, the optimization control module is used to perform the following steps: Based on the sampling frequency adjustment strategy, the current sampling frequency is numerically updated to obtain the target sampling frequency; and based on the positioning accuracy adjustment strategy, the current positioning accuracy parameter is adjusted by calculating the adjustment range to obtain the target positioning accuracy parameter. Based on the target sampling frequency and target positioning accuracy parameters, control parameters are sent to the pet locator.

[0014] Furthermore, the system also includes a power monitoring module, which is used to obtain the real-time power of the pet locator and issue an alarm when the real-time power is lower than a preset power threshold.

[0015] The beneficial effects of this invention are as follows: By preprocessing the collected acceleration, direction, and positioning coordinate data and constructing a consistent temporal feature vector, the input features maintain continuity and modelability in the time dimension. Temporal encoding is performed on the temporal feature vector based on a deep learning model, generating motion trend prediction results that include velocity changes and direction estimates. This prediction mechanism enables the system to infer the pet's motion trend, thus avoiding state switching lag and insufficient recognition accuracy caused by relying on static thresholds. The sampling frequency is strategically adjusted based on the motion trend prediction results. By distinguishing between static and dynamic scenarios and combining the magnitude of direction changes, the sampling frequency is dynamically allocated. This allows the locator to reduce the sampling frequency in low-activity states to reduce ineffective energy consumption, and to increase the sampling frequency in motion or sudden direction changes to maintain real-time trajectory updates. Furthermore, based on the velocity changes and direction estimates, position extrapolation calculations are performed on the current positioning coordinates, and a positioning accuracy adjustment strategy is generated based on the distance between the predicted position and the geofence boundary. This strategy dynamically adjusts positioning accuracy based on the risk level of spatial location, shifting the positioning accuracy requirement from a fixed mode to an adaptive configuration based on behavioral trends and spatial location, thereby reducing unnecessary calls to high-precision mode. Finally, by combining the sampling frequency adjustment strategy and the positioning accuracy adjustment strategy, the operating parameters of the pet locator are updated uniformly, effectively compensating for the shortcomings of existing technologies that rely on fixed rules and cannot adapt to changes in pet movement behavior, and significantly improving the locator's power efficiency. Attached Figure Description

[0016] Fig. 1 This is a schematic diagram of a deep learning-based pet locator power optimization system according to the present invention.

[0017] Fig. 2 This is a flowchart of the execution steps of the frequency adjustment module in this invention. Detailed Implementation

[0018] Please see Figs. 1-2 As shown, this invention relates to a deep learning-based pet locator power optimization system, comprising: The data acquisition module is used to collect acceleration sequences, orientation sequences, and positioning coordinates through the pet locator, perform preprocessing, construct time-series feature vectors, and input them into a deep learning prediction model; The prediction module is used to perform temporal encoding on the temporal feature vector according to the deep learning prediction model to generate motion trend prediction results, the motion trend prediction results including velocity change and direction estimate; A frequency adjustment module is used to generate a sampling frequency adjustment strategy based on the motion trend prediction results. The accuracy adjustment module is used to calculate the predicted location based on the motion trend prediction results, and generate a positioning accuracy adjustment strategy based on the distance between the predicted location and the geofence boundary. The optimized control module performs power optimization control on the pet locator based on the sampling frequency and accuracy adjustment results.

[0019] In some embodiments, the acceleration sequence, direction sequence, and positioning coordinates from the pet locator are first acquired through a data acquisition module, and preprocessing is performed on various types of raw data. Taking acceleration data as an example, the system uses a bandpass filtering model to remove high-frequency mechanical noise and low-frequency drift components, and then normalizes the acceleration changes based on a fixed-length window to maintain comparability under different activity intensities. The direction sequence is processed through angle smoothing and abnormal angle correction to solve the jitter and jumps caused by the electronic compass in rapidly rotating scenarios. The positioning coordinates are corrected by trajectory interpolation and drift correction strategies to correct the trajectory offset caused by GNSS in obstructed or weak signal areas. The above preprocessed data are aligned based on a unified time reference axis and spliced ​​into a structurally stable temporal feature vector according to the time sequence. This construction method not only preserves the dynamic change attributes during the pet's movement, but also avoids the state recognition bias caused by information gaps in traditional feature selection. In order to perform deep dynamic modeling of the temporal feature vector, this embodiment uses a Long Short-Term Memory (LSTM) network to construct a prediction module, and extracts short-term transient features and long-term movement trends through its gating units. The system performs bidirectional temporal modeling on the temporal feature vector, enabling the model to simultaneously capture the correlation between consecutive moments. This is particularly suitable for predicting pet behavior when movement direction changes frequently or trajectories are irregular. For example, when a pet suddenly changes from a slow walk to a fast run or makes a sharp turn, the bidirectional LSTM can output the velocity change and direction estimate through coupled analysis of historical and current states, forming an accurate prediction of movement trends. Compared to existing solutions that rely on acceleration thresholds or step detection algorithms, this embodiment can extract implicit dynamic patterns from complex temporal data, making state recognition more sensitive and prediction capabilities stronger. In the frequency adjustment module, the system determines whether the pet is stationary, moving slowly, or moving rapidly based on the velocity change, and dynamically determines the sampling frequency strategy based on the magnitude of the change in direction estimate. For example, when the pet is stationary for a long time with minimal directional change, the system will automatically reduce the sampling frequency to reduce unnecessary energy consumption; when a sudden increase in velocity or a sharp change in direction is detected, the system quickly increases the sampling frequency to improve the trackability of the spatial trajectory. Unlike traditional methods that rely solely on peak acceleration to determine state, this embodiment utilizes trend quantification indicators obtained through a deep learning model. This allows the frequency adjustment strategy to respond proactively to motion changes, significantly reducing energy waste or positioning delays caused by sampling lag. The accuracy adjustment module extrapolates the current positioning coordinates based on velocity changes and direction estimates, assessing proximity to the geofence boundary by predicting possible future locations. For example, if the extrapolated location indicates the pet may approach the fence boundary in seconds, the system immediately increases positioning accuracy to ensure timely detection of boundary breaches. Conversely, when the pet remains within the fence for an extended period and the extrapolated path deviates from the boundary, the system reduces positioning accuracy to minimize energy consumption in high-accuracy mode.This predictive accuracy allocation mechanism based on motion trends differs from existing methods that only improve accuracy after boundary triggering, significantly enhancing energy utilization efficiency.

[0020] Ultimately, the optimization control module jointly executes the sampling frequency adjustment strategy and the positioning accuracy adjustment strategy, achieving overall energy consumption control through unified updates of the sampling frequency and positioning accuracy parameters. This module directly acts on the locator hardware through the parameter distribution interface, enabling the deep learning inference results to be reflected in the device's operating mode in real time, achieving closed-loop optimization from data perception and trend prediction to resource scheduling.

[0021] Furthermore, the process of collecting acceleration sequences, direction sequences, and location coordinates through a pet locator and performing preprocessing includes the following steps: The acceleration sequence is subjected to noise reduction filtering, and the acceleration change is normalized based on the time window to obtain acceleration preprocessed data; Angle smoothing and outlier angle correction are performed on the direction sequence to obtain direction preprocessed data; The positioning coordinates are subjected to trajectory interpolation and drift correction processing to obtain preprocessed positioning coordinate data; Based on the acceleration preprocessing data, orientation preprocessing data, and positioning coordinate preprocessing data, time alignment and feature concatenation processing are performed to obtain the temporal feature vector.

[0022] In some embodiments, denoising is first performed on the signal using a band-limited filter to suppress high-frequency noise caused by factors such as device vibration, electromagnetic interference, and loose pet clothing. In one specific implementation, a second-order low-pass filter based on the Butterworth structure is used, with the cutoff frequency set to 3–5 Hz according to the pet's activity characteristics, preserving mainly the acceleration variation components related to the pet's gait cycle. Subsequently, the system normalizes the acceleration variation based on a fixed-length sliding time window. Taking a 1-second window (approximately 20–50 sampling points) as an example, the system calculates the minimum and maximum values ​​of the acceleration magnitude within the window and maps each sampling point to the [0,1] space, thereby eliminating the amplitude inconsistency caused by different activity intensities and individual differences, making the acceleration variation patterns more comparable in the feature space. During the direction sequence processing, a weighted moving average model is used to smooth the angle signal to eliminate the jitter and jumps common in electronic compasses. To address the instantaneous anomalies in directional data caused by magnetic interference, this embodiment introduces an anomaly correction algorithm based on angular continuity. This algorithm compares the rate of angular change between adjacent time slices; if the rate of change exceeds a preset threshold (e.g., above 100° / s and inconsistent with acceleration), the sampling point is considered an anomaly and replaced with the interpolation result of two reliable points before and after it. This correction mechanism ensures the continuity and stability of the directional sequence even under rapid rotation and obstruction environments. For positioning coordinates, a trajectory interpolation algorithm is first used to recover positioning discontinuities caused by satellite obstruction or signal loss. For example, when the time interval between two consecutive positioning samples exceeds twice the normal sampling period, piecewise linear interpolation or interpolation based on velocity changes is used to fill in the intermediate nodes on the time axis. Furthermore, GNSS coordinates often drift near buildings. This embodiment uses a historical trajectory smoothing model (such as the Kalman filter algorithm) to correct this drift. By jointly considering position, velocity, and direction data, the trajectory becomes more consistent globally, thereby reducing error accumulation during the position extrapolation stage. Finally, the acceleration preprocessed data, orientation preprocessed data, and positioning coordinate preprocessed data are aligned on the same time axis. Specifically, the system resamples the data sequence according to their respective sampling frequencies, ensuring that all three types of data have corresponding timestamps on a unified set of sampling points. Finally, the system concatenates the three types of preprocessed data according to feature dimensions, sorted by time, to form a consistent temporal feature vector. Each time slice contains multi-dimensional features including changes in acceleration magnitude, orientation angle, and corrected coordinates.

[0023] Furthermore, the deep learning prediction model is a long short-term memory network.

[0024] Furthermore, the prediction module is used to perform the following steps: Based on the temporal feature vector, a long short-term memory network is used to perform a gated update process on the feature input at each time step to obtain the first temporal hidden feature containing the forward state. Based on the temporal feature vector, the feature input at each time step is subjected to gated update processing of the inverse long short-term memory network to obtain the second temporal hidden feature containing the backward state; Perform feature concatenation processing on the first and second temporal hidden features to obtain bidirectional temporal hidden features; Based on the bidirectional temporal hidden features, nonlinear mapping processing is performed to generate motion trend prediction results.

[0025] It should be noted that the prediction module uses a Long Short-Term Memory (LSTM) network as the core prediction model. It performs temporal encoding on the constructed temporal feature vector to extract long-term dependency features and short-term dynamic change features of pet movement behavior over time. In the specific implementation, the system first inputs the temporal feature vector into the feedforward LSTM network in chronological order. For each input feature vector xt at time t, the input gate, forget gate, and output gate within the LSTM dynamically adjust the current input, historical memory, and output states, respectively. Specifically, the gating weights are calculated using sigmoid and tanh functions to determine how much historical state and current information to retain. This mechanism enables the network to automatically adjust the memory span according to changes in the pet's state, thereby generating a first temporal hidden feature hf(t) containing the forward movement trend. This feature can effectively reflect the continuity of movement changes from the past to the present, such as the trend of a smooth transition from low speed to high speed. To capture potential inverse dependencies in the pet's movement patterns, such as subtle decelerations or behavioral precursors before a sudden change in direction, a reverse LSTM network is further introduced. The reverse network traverses the same temporal feature vectors in a time-reverse manner, generating a second temporal hidden feature hb(t) that includes the backward state through a gating update mechanism consistent with the forward network structure. This feature represents the correlation between the current moment and future motion trends; for example, slight fluctuations in the direction sequence are more pronounced in the reverse state when a pet is about to turn. Subsequently, vector-level concatenation is performed on the forward hidden feature hf(t) and the backward hidden feature hb(t) to construct the bidirectional temporal hidden feature hbi(t). This feature integrates temporal dependencies into the same representation, enabling the model to have a comprehensive perception of the global temporal sequence when inferring velocity changes and direction estimates. Compared to unidirectional LSTM, the bidirectional structure is more robust to handling complex motion patterns such as abrupt changes, short-term acceleration, or continuous deceleration. Finally, the bidirectional temporal hidden feature is input into a nonlinear mapping layer, which performs feature compression and mapping based on a multilayer perceptron structure, generating velocity changes and direction estimates through ReLU or GELU activation functions. During the training phase, network parameters are optimized using real speed and direction labels in a supervised manner, enabling the model to learn stable prediction patterns in various motion scenarios (such as normal walking, fast running, sharp turns, and pauses).

[0026] Furthermore, the frequency adjustment module is used to perform the following steps: Based on the velocity change in the motion trend prediction results, threshold discrimination processing is performed on the motion state to obtain a stationary label or a motion label. Based on the motion state, the current sampling frequency is adjusted by either increasing or decreasing the sampling frequency, including decreasing the sampling frequency when the sampling frequency is at rest and increasing the sampling frequency when the sampling frequency is at motion. The change amplitude is calculated based on the direction estimate in the motion trend prediction results to obtain the direction change amplitude. Based on strategy selection processing, the sampling frequency value is determined according to the directional amplitude, thus obtaining the sampling frequency adjustment strategy.

[0027] In some embodiments, the speed change output by the prediction module is first processed by threshold discrimination. By comparing it with a preset speed change threshold, it is determined whether the pet is currently stationary or in motion. This discrimination is not a simple judgment of speed magnitude, but rather an analysis of the rate of change of speed trend, which can identify the transition process of the start or end of movement in advance. For example, when the speed change continuously decreases and remains at an extremely low level for multiple time windows, even if the speed itself has not dropped to zero, the system can determine in advance that the pet has entered a stationary state. Through this trend-based discrimination mechanism, the problem that fixed thresholds cannot identify micro-movements, slow movements, or noise interference is effectively solved. After obtaining the stationary or moving indicator, the direction of sampling frequency adjustment is selected according to the movement state. If the pet is stationary, the system reduces the current sampling frequency proportionally to reduce the energy consumption of invalid positioning behavior; if the pet is in motion, the sampling frequency is increased according to the predicted speed change trend to maintain the spatiotemporal continuity of trajectory recording. In a specific implementation, an exponential adjustment function is used to smoothly map the frequency change. For example, a larger adjustment coefficient is used for high-speed movement, while a smaller adjustment coefficient is used for low-speed stable movement, thereby achieving dynamic elasticity of frequency adjustment. Further, the directional amplitude is calculated based on the directional estimates to characterize the degree of change in the pet's movement direction per unit time. The calculation of directional amplitude is based on the normalization of angle differences, for example, by calculating the angle difference between two consecutive directional estimates and combining it with a standardization function to obtain the intensity of the directional change. In practical application scenarios, such as when a pet suddenly turns to chase a target or frequently circles around in an outdoor environment, the directional amplitude often increases significantly. Such behaviors are difficult to respond to in a timely manner in traditional rule systems, but this embodiment can detect them in advance before the movement has fully produced a significant speed change by modeling continuous directional amplitude.

[0028] Furthermore, the precision adjustment module is used to perform the following steps: Based on the velocity change and direction estimate in the motion trend prediction results, the current positioning coordinates are extrapolated to obtain the predicted position at the next preset time. Based on the predicted location, the nearest distance calculation is performed on the geofence boundary to obtain the distance data between the predicted location and the geofence. Based on the distance data, a threshold discrimination process is performed on the positioning accuracy requirement to obtain an accuracy level identifier; The accuracy adjustment range is calculated based on the accuracy level identifier to obtain the positioning accuracy adjustment strategy.

[0029] In some embodiments, the velocity change is considered a dynamic parameter describing the change in motion rate, and the direction estimate is considered a vector basis characterizing the displacement direction. By substituting both into the position extrapolation function and combining them with the current positioning coordinates, the predicted position is solved, allowing the prediction process to maintain a continuous simulation of the displacement trend even with a finite time step. This extrapolation calculation can be solved using a linear extrapolation model or a state transition equation based on a Kalman filter framework, ensuring that the predicted position reflects the combined influence of velocity and direction at future moments. A geofence-based boundary model performs nearest-distance calculations, obtaining distance data describing the degree of potential boundary crossing risk by constructing the shortest path distance function between the predicted position and the fence boundary. To ensure calculation accuracy, distance solutions are typically based on vector projection or spatial geometric shortest distance algorithms, ensuring accurate acquisition of the minimum distance from the predicted position to the fence boundary even with complex boundary shapes. Based on this distance data, a threshold discrimination process is performed on the positioning accuracy requirements, constructing an accuracy level discrimination function that maps different distance intervals to corresponding accuracy level identifiers. For example, when the predicted location is within a high-risk zone from the fence boundary, a high-accuracy level identifier will be output; when the predicted location is far from the boundary and the movement trend is stable, a low-accuracy level identifier will be output. This discrimination result is used to guide the generation of subsequent accuracy adjustment strategies, giving the triggering of accuracy control a clear boundary condition basis. Finally, based on the accuracy level identifier, the accuracy adjustment range is calculated. By mapping the accuracy level to a specific upper limit of measurement error or positioning parameter configuration value, a positioning accuracy adjustment strategy is obtained, enabling the positioning system to automatically improve positioning accuracy in high-risk areas and reduce accuracy requirements in low-risk areas to reduce energy consumption.

[0030] Furthermore, the process of performing position extrapolation calculation on the current positioning coordinates based on the velocity change and direction estimate in the motion trend prediction results includes: Calculate the displacement direction vector per unit time based on the current positioning coordinates and direction estimate; Based on the velocity change and the displacement direction vector per unit time, the displacement increment is calculated at the current position to obtain the predicted displacement increment. Based on the predicted displacement increment, the current positioning coordinates are updated to obtain the predicted position at the next preset time.

[0031] In some embodiments, firstly, a displacement direction vector per unit time is calculated based on the current positioning coordinates and the direction estimate. The direction estimate, output by a deep learning prediction model, has higher stability and semantic interpretation capabilities than simple sensor direction readings. It not only reflects the current direction of travel but also implies the short-term trend of direction change. Therefore, the determination of the direction vector not only uses a simple angle conversion formula but also combines the time smoothing result of the direction estimate to ensure the continuity and predictability of the inferred direction in noisy environments. Subsequently, displacement increment calculation is performed based on the velocity change and the direction vector per unit time. The velocity change describes the dynamic trend of velocity increase or decrease at the current moment, rather than simply the magnitude of velocity. Therefore, in calculating the displacement increment, this embodiment uses a dynamic velocity estimation model to map the velocity change to the instantaneous velocity at the prediction moment. For example, when detecting a pet's acceleration trend, the system uses the velocity change to correct the base velocity estimate, enabling the prediction model to simulate the short-term motion trajectory under the influence of real acceleration. The displacement increment is achieved through vector multiplication, that is, scaling the unit direction vector according to the time step corrected by the velocity change to obtain a vector result containing both direction and displacement magnitude. After obtaining the predicted displacement increment, the displacement increment is compared with the current positioning coordinates to calculate the predicted position at the next preset time. The update process is typically based on standard two-dimensional coordinate translation formulas; however, to enhance the model's spatial adaptability, this embodiment can choose to perform latitude and longitude transformation in the geographic coordinate system, employing projection algorithms or spherical distance approximation methods to ensure high accuracy in predicted positions across large-scale movement scenarios. Through this extrapolation mechanism, the system can estimate the pet's future position within the actual positioning sampling interval, providing lead time for subsequent geofence distance assessment and accuracy adjustment strategies.

[0032] Furthermore, the nearest distance calculation includes the following steps: Based on the predicted location, perform point-by-point distance calculation on the polygon vertices of the geofence boundary to obtain vertex distance data from the predicted location to each vertex; Based on the predicted location, the vertical projection distance is calculated for each edge segment of the geofence boundary to obtain the line segment distance data from the predicted location to each edge segment. The vertex distance data and line segment distance data are subjected to minimum value determination processing to obtain the distance data between the predicted location and the geofence.

[0033] In some embodiments, the predicted location is first used as a reference point to perform point-by-point distance calculations on all polygon vertices of the geofence boundary. The distance calculation is based on the Euclidean distance formula, obtaining a set of vertex distance data by calculating the straight-line distance between the predicted location and each vertex. This type of distance is advantageous when handling near-corner scenarios because, near a vertex, the shortest path from the predicted location to the corner usually does not fall on the edge segment but points directly to the vertex. Traditional algorithms that only consider line segment distances tend to underestimate the risk in this case. Further, a vertical projection distance calculation is performed on each edge segment based on the predicted location. Specifically, by constructing a vector projection model from the predicted location to the line segment, it is determined whether the projection point falls within the line segment's range. If the projection point is inside the line segment, the vertical distance from the predicted location to the projection point is used as the effective distance; if the projection point is not inside the line segment, it indicates that the line segment direction does not contribute to the distance, and the projection result of that line segment is marked as invalid and the vertex distance is used as a substitute. Through this projection mechanism, the system can accurately capture the shortest distance characteristics of the predicted location near the edge segment region, exhibiting higher accuracy, especially in continuous boundary regions. After obtaining all vertex distance and line segment distance data, a minimum value determination process is performed on both types of data to obtain the final minimum distance between the predicted location and the geofence. By forming a unified candidate distance set from vertex distances to line segment distances, the process of selecting the minimum value is equivalent to solving for the precise shortest distance from the predicted location to the polygon boundary. This shortest distance not only serves as a core input parameter for the positioning accuracy adjustment strategy but is also used to construct a quantitative model of boundary crossing risk, enabling the system to improve accuracy or issue early warnings when the geofence is approaching.

[0034] Furthermore, the optimization control module is used to perform the following steps: Based on the sampling frequency adjustment strategy, the current sampling frequency is numerically updated to obtain the target sampling frequency; and based on the positioning accuracy adjustment strategy, the current positioning accuracy parameter is adjusted by calculating the adjustment range to obtain the target positioning accuracy parameter. Based on the target sampling frequency and target positioning accuracy parameters, control parameters are sent to the pet locator.

[0035] In some embodiments, firstly, the sampling frequency adjustment strategy generated by the receiving frequency adjustment module is used as the base parameter to obtain the target sampling frequency through numerical update processing. This update process can be implemented through linear mapping, exponential smoothing, or a rate-limited adjustment model. For example, when the movement trend prediction shows that the pet is moving at high speed and with a large change in direction, the sampling frequency is increased according to the strategy output to maintain a smooth change while meeting real-time requirements; conversely, when the pet is stationary, the sampling frequency is decreased according to the strategy to reduce energy consumption. Through this adaptive update algorithm, the target sampling frequency can accurately reflect the needs of the pet's current movement state, unlike the coarse-grained adjustment method based on fixed value switching in the prior art. Simultaneously, the adjustment range of the current positioning accuracy parameter is calculated based on the positioning accuracy adjustment strategy. This process maps the distance between the predicted location and the geofence to a priority level of positioning accuracy, and then calculates the corresponding accuracy adjustment range according to the mapping function. For example, when the predicted location is detected to be close to the geofence boundary, the positioning accuracy parameter will be increased according to the accuracy adjustment strategy, allowing the GNSS or fusion positioning module to operate within a narrower error range. Conversely, when the distance to the geofence is greater and the movement trend is stable, the system will decrease the accuracy parameter to reduce the additional power consumption of the high-accuracy mode. This adjustment mechanism is implemented using a differentiable or piecewise function, ensuring a continuous response between accuracy adjustment and spatial location risk, thus avoiding the jump behavior caused by hard threshold switching in traditional systems. After obtaining the target sampling frequency and target positioning accuracy parameters, the optimization control module sends control parameters to the pet locator through the device control interface. During the sending process, the system verifies the feasibility of the target parameters, including whether the sampling frequency is within the device's supported range and whether the positioning accuracy parameters are compatible with the sensor resource configuration, thereby ensuring that the update strategy can be stably executed on the device. This closed-loop control method, from strategy generation to device regulation, enables the locator to dynamically adjust resource usage based on real-time prediction results, significantly improving power utilization efficiency while maintaining the reliability of geofence detection and trajectory recording.

[0036] Furthermore, the system also includes a power monitoring module, which is used to obtain the real-time power of the pet locator and issue an alarm when the real-time power is lower than a preset power threshold.

[0037] Specifically, the power monitoring module reads the real-time voltage, current, and remaining capacity estimates output by the locator's power management unit, and combines this with a power estimation model based on the discharge characteristic curve to obtain the accurate percentage of battery charge. To avoid misjudgments caused by instantaneous voltage fluctuations, a sliding window averaging mechanism is used during monitoring, and a temperature correction factor is incorporated when necessary to improve the stability and robustness of power estimation. Based on real-time power assessment, when the detected power value is lower than a preset power threshold (e.g., 20% or a threshold dynamically set according to the device's operating strategy), the system will trigger an alarm process. The alarm signal can be output in various ways, including voice prompts on the device, push notifications from mobile applications, or uploading a low-power event to the server, enabling users to take timely measures to charge or recycle the device.

[0038] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A deep learning based pet locator power optimization system, comprising: The method comprises the following steps: a data acquisition module is used to collect acceleration sequences, direction sequences and positioning coordinates through a pet locator, and to preprocess the collected data, construct a time sequence feature vector and input the time sequence feature vector into a deep learning prediction model; a prediction module is used to perform time sequence coding on the time sequence feature vector according to the deep learning prediction model, and to generate a motion trend prediction result, wherein the motion trend prediction result comprises a speed change amount and a direction estimation amount; a frequency adjustment module is used to generate a sampling frequency adjustment strategy according to the motion trend prediction result; an accuracy adjustment module is used to calculate a predicted position based on the motion trend prediction result, and to generate a positioning accuracy adjustment strategy according to the distance between the predicted position and the boundary of a geofence; an optimization control module is used to perform power optimization control on the pet locator based on the sampling frequency and the accuracy adjustment result.

2. The deep learning based pet locator power optimization system of claim 1, wherein, The preprocessing of the acceleration sequences, direction sequences and positioning coordinates collected by the pet locator comprises the following steps: denoising filter processing is performed on the acceleration sequences, and normalization processing is performed on the acceleration change amount based on a time window to obtain acceleration pretreatment data; angle smoothing and abnormal angle correction processing are performed on the direction sequences to obtain direction pretreatment data; trajectory interpolation and drift correction processing are performed on the positioning coordinates to obtain positioning coordinate pretreatment data; time alignment and feature splicing processing are performed based on the acceleration pretreatment data, direction pretreatment data and positioning coordinate pretreatment data to obtain a time sequence feature vector.

3. The deep learning based pet locator power optimization system of claim 2, wherein, The deep learning prediction model is a long short-term memory network.

4. The deep learning based pet locator power optimization system of claim 3, wherein, The prediction module is used to perform the following steps: based on the time sequence feature vector, a long short-term memory network gating update process is performed on the feature input of each time point to obtain a first time sequence hidden feature containing a forward state; based on the time sequence feature vector, a reverse long short-term memory network gating update process is performed on the feature input of each time point to obtain a second time sequence hidden feature containing a backward state; feature splicing processing is performed on the first time sequence hidden feature and the second time sequence hidden feature to obtain a bidirectional time sequence hidden feature; nonlinear mapping processing is performed based on the bidirectional time sequence hidden feature to generate a motion trend prediction result.

5. The deep learning based pet locator power optimization system of claim 1, wherein, The frequency adjustment module is used to perform the following steps: based on the speed change amount in the motion trend prediction result, threshold discrimination processing is performed on the motion state to obtain a static identifier or a motion identifier; based on the motion state, strategy selection processing is performed on the current sampling frequency, including down-regulating the sampling frequency under the static identifier and up-regulating the sampling frequency under the motion identifier; based on the direction estimation amount in the motion trend prediction result, a direction amplitude is calculated to obtain a direction amplitude; based on the strategy selection processing, the sampling frequency value is determined according to the direction amplitude to obtain a sampling frequency adjustment strategy.

6. The deep learning based pet locator power optimization system of claim 1, wherein, The accuracy adjustment module is used to perform the following steps: based on the speed change amount and the direction estimation amount in the motion trend prediction result, position extrapolation calculation processing is performed on the current positioning coordinates to obtain a predicted position at a next preset time; the closest distance calculation is performed on the geofence boundary according to the predicted position to obtain distance data between the predicted position and the geofence. Based on the distance data, a threshold discrimination process is performed on the positioning accuracy requirement to obtain an accuracy level identifier; According to the accuracy level identifier, an accuracy adjustment range is calculated to obtain a positioning accuracy adjustment strategy.

7. The deep learning based pet locator power optimization system of claim 6, wherein, The position extrapolation calculation process performed on the current positioning coordinates based on the speed change amount and the direction estimation amount in the motion trend prediction result includes: According to the current positioning coordinates and the direction estimation amount, a unit time displacement direction vector is calculated; Based on the speed change amount and the unit time displacement direction vector, a displacement increment calculation process is performed on the current position to obtain a predicted displacement increment; According to the predicted displacement increment, a coordinate update process is performed on the current positioning coordinates to obtain a predicted position at the next preset time.

8. The deep learning based pet locator power optimization system of claim 6, wherein, The nearest distance calculation includes the following steps: According to the predicted position, a point-by-point distance calculation process is performed on the polygon vertices of the geofence boundary to obtain vertex distance data from the predicted position to each vertex; Based on the predicted position, a perpendicular projection distance calculation process is performed on each edge segment of the geofence boundary to obtain line segment distance data from the predicted position to each edge segment; The minimum value determination process is performed on the vertex distance data and the line segment distance data to obtain distance data between the predicted position and the geofence.

9. The deep learning based pet locator power optimization system of claim 1, wherein, The optimization control module is used to perform the following steps: Based on the sampling frequency adjustment strategy, a numerical update process is performed on the current sampling frequency to obtain a target sampling frequency; and based on the positioning accuracy adjustment strategy, an adjustment range calculation process is performed on the current positioning accuracy parameter to obtain a target positioning accuracy parameter; According to the target sampling frequency and the target positioning accuracy parameter, control parameters are issued to the pet locator.

10. The deep learning based pet locator power optimization system of claim 1, wherein, The system further includes a power monitoring module for obtaining the real-time power of the pet locator and issuing an alarm when the real-time power is lower than the preset power threshold.

Images