Pet self-adaptive feeding recommendation method and system based on AI image analysis

Abstract

The invention provides a pet self-adaptive feeding recommendation method and system based on AI image analysis, and the method comprises the steps: firstly monitoring food images of a feeding disc of feeding equipment at the moment of food discharge stopping and pet feeding leaving, obtaining a food state image set of the feeding disc, carrying out the AI image analysis of the food state image set of the feeding disc, and obtaining a food state image set of the feeding disc; feeding time and food intake information are extracted, a pet food intake data set is formed, the pet food intake data set is input into a pre-trained feeding demand prediction model, and the next hunger time and food intake prediction result of the pet is obtained through fitting analysis. Self-adaptive feeding recommendation information containing recommended feeding time and recommended feeding amount is generated according to the next hunger time and food intake prediction result of the pet, and finally the self-adaptive feeding recommendation information is pushed to an associated mobile terminal for a user to check and decide whether to execute feeding operation or not. Personalized, intelligent and precise management of pet feeding is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method and system for adaptive feeding recommendations for pets based on AI image analysis. Background Technology

[0002] In the pet care industry, scientifically and rationally arranging pet feeding is crucial for the healthy growth of pets. Currently, common pet feeding methods include manual feeding at fixed times and in fixed quantities or using simple automatic feeders to feed according to preset programs. However, manual feeding requires users to spend a lot of time and energy observing the pet's eating behavior and manually adjusting the feeding plan, which is often difficult for busy modern people to achieve accurately and promptly. While existing automatic feeders can feed at fixed intervals and in fixed amounts, they lack the ability to dynamically sense and analyze the pet's actual eating situation. Different pets have different hunger times and food intake due to differences in breed, age, health status, activity level, etc. Fixed feeding patterns cannot meet the individual dietary needs of pets, easily leading to overeating or undereating, which in turn affects the pet's health. In addition, existing feeding methods cannot provide timely feedback on the pet's eating information to the user, making it difficult for the user to fully understand the pet's eating status and make scientific and reasonable feeding decisions. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a pet adaptive feeding recommendation method based on AI image analysis, the method comprising: The food images of the feeding tray of the feeding device are collected at the time when food dispensing stops and when the pet leaves after eating, thus obtaining a set of food status images of the feeding tray; AI image analysis processing is performed on the set of food status images in the feeding bowl to extract the feeding time information and the amount of food consumed in each feeding, forming a pet feeding data set containing multiple sets of corresponding feeding time information and food consumption information. The pet feeding dataset is input into a pre-trained feeding demand prediction model. The correlation between time and food intake in the pet feeding dataset is fitted and analyzed to obtain the prediction results of the pet's next hunger time and the pet's next food intake. Based on the prediction results of the pet's next hunger time and the pet's next food intake, generate adaptive feeding recommendation information for the pet, which includes recommended feeding time and recommended food amount; The adaptive feeding recommendation information for pets is pushed to the mobile terminal associated with the feeding device, so that users can view the adaptive feeding recommendation information for pets and decide whether to perform the feeding operation based on the adaptive feeding recommendation information for pets.

[0004] Furthermore, embodiments of the present invention also provide a pet adaptive feeding recommendation system based on AI image analysis, characterized in that it includes: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the aforementioned AI image analysis-based adaptive pet feeding recommendation method by executing the machine-executable instructions.

[0005] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, a processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the above-described AI image analysis-based pet adaptive feeding recommendation method.

[0006] Based on the above, by monitoring food images from the feeding tray at the moment the food dispensing stops and when the pet leaves after eating, a set of food state images from the feeding tray is constructed. This allows for a comprehensive and accurate record of the food state of the feeding tray before and after each feeding. AI image analysis is then performed on this set of images to extract feeding time and amount information for each feeding, forming a pet feeding data set containing multiple sets of corresponding relationships. This enables detailed analysis and quantitative assessment of pet feeding behavior. The pet feeding data set is then input into a pre-trained feeding demand prediction model for fitting analysis, accurately predicting the pet's next hunger time and food intake, fully considering the dynamic changes in pet feeding patterns. Based on the prediction results, adaptive feeding recommendations for pets, including recommended feeding times and amounts, are generated and pushed to mobile terminals associated with the feeding device. This allows users to obtain timely and scientifically sound feeding advice, achieving personalized, intelligent, and precise management of pet feeding, effectively improving pets' dietary health, and providing users with a convenient and efficient pet care experience. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the execution flow of the AI ​​image analysis-based adaptive feeding recommendation method for pets provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of exemplary hardware and software components of the AI ​​image analysis-based adaptive feeding recommendation system for pets provided in an embodiment of the present invention. Detailed Implementation

[0009] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1This is a flowchart illustrating a pet adaptive feeding recommendation method based on AI image analysis, provided in one embodiment of the present invention. The following is a detailed description of this pet adaptive feeding recommendation method based on AI image analysis.

[0010] Step S110: Monitor the food images of the feeding tray of the feeding device at the time when food dispensing stops and at the time when the pet leaves after eating, and obtain a set of food status images of the feeding tray.

[0011] In this embodiment, the use of a smart feeding device for a pet dog named "Wangcai" in a home environment is used as an example. A wide-angle camera is fixedly installed above the feeding tray of the feeding device. The camera's field of view covers the entire feeding tray area, and the camera is connected to the control unit of the feeding device via a data cable. The control unit has preset image acquisition trigger conditions. When the food dispensing device stops rotating and no food falls into the dispensing channel, the control unit determines that food dispensing has stopped and then sends a shooting command to the camera. After receiving the command, the camera takes a picture of the food in the feeding tray at a preset resolution and frame rate, which serves as the image at the moment when food dispensing stops.

[0012] During the feeding of the pet dog "Wangcai," the pressure sensor of the feeding device continuously monitors the weight change of the feeding bowl. When the weight change value remains below a set threshold for a preset time period, the control unit determines that the pet has finished eating and left, and sends a shooting command to the camera again. The camera then captures an image of the food on the feeding bowl at this moment, which serves as the image of the moment the pet left after eating. Each captured image is stored with corresponding acquisition time information, which is obtained through the clock module of the control unit. The camera transmits the captured image data to the control unit, which creates a separate folder for each feeding in its local storage module. The images of the moment the food was dispensed and the moment the pet left after eating are stored in the same folder. Multiple folders together constitute a collection of images showing the food status of the feeding bowl.

[0013] Step S120: Perform AI image analysis processing on the set of food status images in the feeding bowl to extract the feeding time information and the amount of food consumed in each feeding, forming a pet feeding data set containing multiple sets of corresponding feeding time information and amount of food information.

[0014] Step S121: Separate the food image of the feeding tray at the moment the food dispensing stops and the food image of the feeding tray at the moment the pet leaves after eating from the set of food state images of the feeding tray, so that the food image of the feeding tray at the moment the food dispensing stops and the food image of the feeding tray at the moment the pet leaves after eating correspond to the initial state and the end state of the same feeding, respectively.

[0015] The control unit invokes the file parsing module to read the names of all folders and the acquisition time information of the images within them in the food status image set of the feeding tray. Folder names are formatted using the date and sequence number of the feeding, such as "year-month-day-sequence number". Parsing the folder names determines the feeding order. For each folder, the acquisition time information of two images is extracted. The image acquired earlier is marked as the food image from the feeding tray at the moment the feeding stopped, and the image acquired later is marked as the food image from the feeding tray at the moment the pet left after eating. After marking, the file parsing module establishes the association between the two images by adding an association identifier to the storage path, ensuring that the initial and final status images of the same feeding can be accurately identified and retrieved.

[0016] Step S122: Call the pre-trained image feature extraction model to extract features from the food image of the feeding tray at the moment when the feeding stops, and obtain the morphological and distribution features of the food in the feeding tray at the moment when the feeding stops.

[0017] Step S1221: Input the food image of the feeding tray at the moment of stopping feeding into the image preprocessing layer of the pre-trained image feature extraction model for noise removal. Input the processed food image of the feeding tray into the feature detection layer of the image feature extraction model to detect the food contour and extract the contour features of the food. The contour features are used as a component of the morphological features.

[0018] The image preprocessing layer first normalizes the size of the input image at the moment the food was dispensed, adjusting it to the fixed size required by the pre-trained model. Next, a Gaussian filtering algorithm is used to remove noise from the normalized image. During the filtering process, the parameters of the filtering kernel are dynamically adjusted according to the brightness distribution of the image, smoothing out noise points while preserving food edge information. The processed image then enters the feature detection layer. This layer scans the image using edge detection operators, identifying the boundary lines between food and non-food areas. The pixel coordinates on the boundary lines are connected sequentially to form the food's contour curve. Contour features include parameters such as the perimeter, area, and curvature changes at points on the contour. These parameters are integrated into a contour feature vector, serving as part of the morphological features.

[0019] Step S1222: The height distribution of food in the processed feeding plate food image is detected by the feature detection layer, the accumulation height of food is determined by the grayscale difference in the processed feeding plate food image, and the height feature of food is extracted. The height feature is used as a component of the morphological feature.

[0020] After contour detection, the feature detection layer performs grayscale analysis on the food regions in the image. Different stacking heights of food result in varying light reflection in corresponding areas of the image, manifesting as different grayscale values. Regions with higher stacking heights typically have lower grayscale values, while regions with lower stacking heights have higher grayscale values. The feature detection layer divides the food region into multiple equal-area sub-regions, calculates the average grayscale value of each sub-region, and converts the average grayscale value into a corresponding height value based on the grayscale-height mapping relationship stored in the pre-trained model. The height values ​​of all sub-regions form a height matrix, where each element represents the height of the corresponding sub-region. This height matrix represents the height feature of the food, which, together with the contour feature vector, constitutes the morphological feature.

[0021] Step S1223: The feature detection layer detects the distribution range of food in the processed feeding plate food image within the feeding plate, determines the proportion of the area covered by food in the feeding plate, and extracts the coverage range feature of the food, which is used as a component of the distribution feature.

[0022] The feature detection layer uses an image segmentation algorithm to divide the processed image into food regions and background regions (feeding dish regions). The background region is the non-food area within the pre-defined feeding dish outline. The proportion of the feeding dish covered by food is obtained by calculating the ratio of the number of pixels in the food region to the total number of pixels within the feeding dish outline. The coverage feature also includes the centroid coordinates of the food region, which are obtained by calculating the mean coordinates of all pixels in the food region, representing the approximate distribution of the food within the feeding dish. The region proportion and centroid coordinates together constitute the coverage feature, as part of the distribution feature.

[0023] Step S1224: The density distribution of food in the processed feeding plate food image is detected by the feature detection layer. The density difference of food is judged by the density of pixels in the processed feeding plate food image. The density features of food are extracted and used as a component of the distribution features.

[0024] The feature detection layer randomly selects multiple detection windows within the food region, with the window size set based on the estimated size of the food particles. For each detection window, the number of food pixels within the window is calculated, and the ratio of this number to the total area of ​​the window is used as the pixel density of that window. Based on the pixel density, the detection windows are divided into different density levels, such as high, medium, and low. The distribution number and location of detection windows of different density levels within the food region are statistically analyzed to form a density distribution histogram. The horizontal axis of the density distribution histogram represents the density level, and the vertical axis represents the number of windows at the corresponding level. The density distribution histogram and the position coordinates of the windows at each density level together constitute the density feature, which, together with the coverage feature, forms the distribution feature.

[0025] Step S1225: Integrate the extracted contour features and height features to form the morphological features of the food in the feeding tray at the moment when feeding stops; integrate the extracted coverage features and density features to form the distribution features of the food in the feeding tray at the moment when feeding stops, and output the morphological features and distribution features of the food in the feeding tray at the moment when feeding stops.

[0026] The feature integration module receives the contour feature vector, height matrix, coverage features (region proportion and centroid coordinates), and density features (density distribution histogram and window position coordinates) output by the feature detection layer. For morphological features, the parameters in the contour feature vector are concatenated column-wise with the height matrix to form a multi-dimensional vector, the dimension of which is equal to the sum of the number of contour feature parameters and the number of rows in the height matrix. For distribution features, the region proportion and centroid coordinates of the coverage features are converted into numerical vectors and concatenated sequentially with the bin values ​​of the density distribution histogram to form a multi-dimensional vector. After integration, the morphological feature vector and distribution feature vector are transmitted to the subsequent food state recognition model.

[0027] Step S123: Call the pre-trained image feature extraction model to extract features from the food image of the feeding dish at the moment the pet leaves the feeding dish, and obtain the morphological and distribution features of the food in the feeding dish at the moment the pet leaves the feeding dish.

[0028] The feature extraction process for the food image from the feeding bowl at the moment the pet leaves after eating is exactly the same as step S122. It also involves noise removal and size normalization in the image preprocessing layer, followed by feature detection layer extracting contour features, height features, coverage features, and density features sequentially. Finally, the feature integration module forms morphological and distribution features. Throughout this process, the parameter settings of the image feature extraction model remain unchanged to ensure standard consistency between the two feature extractions, facilitating subsequent comparisons.

[0029] Step S124: Input the morphological and distribution features of the food in the feeding tray at the moment the feeding stops into the pre-trained food state recognition model to obtain the initial food state of the feeding tray at the moment the feeding stops.

[0030] The food state recognition model consists of a fully connected layer and an output layer connected sequentially. The number of input nodes in the fully connected layer matches the total dimension of the concatenated morphological feature vector and distribution feature vector. The morphological and distribution feature vectors at the moment of food dispensing cessation are concatenated at the input to form a combined feature vector, which is then input to the fully connected layer. The fully connected layer performs a linear transformation on the combined feature vector using a weight matrix and outputs an intermediate feature vector after processing with an activation function. This intermediate feature vector is input to the output layer, which uses a softmax activation function to output probability values ​​for multiple categories. Each category corresponds to a food state, which includes information such as food volume, average height, and density level. The category with the highest probability value is selected as the recognition result, which represents the initial food state of the feeding tray at the moment of food dispensing cessation.

[0031] Step S125: Input the shape and distribution characteristics of the food in the feeding dish at the moment the pet leaves the feeding dish into the pre-trained food state recognition model to obtain the remaining food state of the feeding dish at the moment the pet leaves the feeding dish.

[0032] Similar to step S124, the morphological feature vector and distribution feature vector at the moment the pet leaves after eating are concatenated into a combined feature vector, which is then input into the food state recognition model. After linear transformation and activation function processing in the fully connected layer, and probability calculation in the output layer, the remaining food state in the feeding dish at the moment the pet leaves after eating is obtained. The remaining food state also includes information such as volume, average height, and density level.

[0033] Step S126: Extract the feed dispensing stop time information from the operation record of the feeding device, and determine the feed dispensing stop time information as the feeding time information for this feeding.

[0034] The control unit of the feeding equipment records the time of feed dispensing stop in an operation log file after each dispensing stop. The operation log file is stored in text format, and each record contains information such as the time of feed dispensing stop and the duration of dispensing. The control unit calls the data reading module to search for the corresponding feed dispensing stop time information in the operation log file based on the name of the folder containing the currently processed feeding image. For example, the sequence number in the folder name corresponds to the record number in the operation log. By matching the sequence number, the corresponding record line is found, and the feed dispensing stop time field is extracted. This feed dispensing stop time field is in the format of "hour:minute," and is determined as the feeding time information for this feeding.

[0035] Step S127: Based on the difference between the initial food state and the remaining food state, and combined with the pre-stored feeding tray capacity parameters and food density parameters, extract the food intake information corresponding to this feeding.

[0036] First, initial volume parameters are extracted from the initial food state, and remaining volume parameters are extracted from the remaining food state. The difference between the initial and remaining volume parameters is calculated to obtain the change in food volume. The control unit's storage module pre-stores the feeding dish's capacity parameters, including the bottom area and maximum height. However, in this step, the bottom area parameter is mainly used to assist in verifying the volume calculation results, ensuring that the volume change is within a reasonable range. Simultaneously, the storage module also pre-stores the density parameters of the currently used pet food, which are pre-input and stored based on information from the food packaging. Multiplying the food volume change by the density parameters yields the pet's food intake mass for this feeding, i.e., the amount of food consumed. The unit for the amount of food consumed is grams, and the calculation result is rounded to the nearest integer.

[0037] Step S128: Associate and store the feeding time information and the corresponding food intake information for this feeding. Repeat the above steps to process the food status image set of the feeding plate corresponding to multiple feedings to obtain multiple sets of correspondence between feeding time information and food intake information.

[0038] The control unit creates a pet feeding data table in the local database, which includes two fields: "feeding time" and "food amount". The feeding time information determined in step S126 and the food amount information calculated in step S127 are written into the corresponding fields of the data table, forming a complete record. For each folder in the food status image set of the feeding tray (i.e., each feeding), steps S121 to S127 are repeated to obtain the corresponding feeding time and food amount information, which are then added to the pet feeding data table, thus forming multiple sets of corresponding relationships.

[0039] Step S129: Arrange the correspondence between multiple sets of feeding time information and food intake information in the order of feeding time information to form a pet feeding data set containing the correspondence between multiple sets of feeding time information and food intake information.

[0040] The control unit calls the database sorting module to sort the records in the pet feeding data table in ascending order according to the "feeding time" field. The sorted records are in the same order as the feedings occurred. After sorting, the entire data table is exported as a structured file, such as a JSON file. This JSON file is the pet feeding data collection, and each record in the JSON file contains two key-value pairs: "feeding time" and "food amount".

[0041] Step S130: Input the pet feeding data set into the pre-trained feeding demand prediction model, perform fitting analysis on the correspondence between time and food intake in the pet feeding data set, and obtain the prediction results of the pet's next hunger time and the pet's next food intake.

[0042] Step S131: Organize the feeding time information in the pet feeding data set into a time series so that all feeding time information forms a continuous time series data in chronological order.

[0043] The input preprocessing module of the feeding demand prediction model reads a JSON file containing pet feeding data and extracts all "feeding time" field values. It converts the feeding times into timestamp format, with each timestamp starting from a fixed date and in hours, such as the number of hours counted from "year-month-day 00:00:00". The converted timestamps are then arranged in ascending order to form a one-dimensional array. This one-dimensional array represents the continuous time series data, with each element representing a feeding timestamp.

[0044] Step S132: Perform correlation matching on the food intake information in the pet food intake data set, so that each food intake information corresponds one-to-one with the feeding time information in the corresponding time series data, forming a time-food intake correlation sequence.

[0045] Extract the "food intake" field values ​​corresponding to each timestamp in the time series data from the pet feeding dataset. Arrange the food intake values ​​in the order of the timestamps to form another one-dimensional array. Combine the timestamp array and the food intake array into a two-dimensional array. Each row of the two-dimensional array contains a timestamp element and the corresponding food intake element, thus forming a time-food intake correlation sequence.

[0046] Step S133: Input the time-feed amount correlation sequence into the sequence processing layer of the pre-trained feeding demand prediction model, and perform preliminary analysis on the time interval and feeding amount change pattern in the time-feed amount correlation sequence through the sequence processing layer of the feeding demand prediction model.

[0047] The sequence processing layer consists of multiple recurrent neural network units, such as LSTM units. The time-feed amount correlation sequence is divided into multiple time windows, each containing a predetermined number of consecutive data rows. The data within each time window is input into the LSTM unit. The LSTM unit learns features from the time intervals (differences between adjacent timestamps) and feed amounts within the time window through a gating mechanism, capturing the trends in time interval changes and fluctuations in feed amounts. The preliminary analysis results are represented by the hidden state vector corresponding to each time window, which contains abstract features of the relationship between time and feed amount within that window.

[0048] Step S134: After the preliminary analysis results are output from the sequence processing layer of the feeding demand prediction model, the preliminary analysis results are input into the fitting analysis layer of the feeding demand prediction model. Based on the changing trend of the historical time-food intake correlation sequence, the fitting analysis layer of the feeding demand prediction model fits the feeding interval pattern and food intake fluctuation pattern of the pet.

[0049] Step S1341: After receiving the preliminary analysis results output by the sequence processing layer of the feeding demand prediction model through the fitting analysis layer of the feeding demand prediction model, extract the feeding interval data and feeding amount change data from the historical time-feeding amount correlation sequence from the preliminary analysis results.

[0050] The fitting analysis layer receives all hidden state vectors output by the sequence processing layer and inputs them into the feature parsing module. The feature parsing module uses a backpropagation algorithm to reconstruct the corresponding feeding interval data and food intake change data from the hidden state vectors. The feeding interval data is the set of differences between adjacent time stamps in the time-food intake correlation sequence, and the food intake change data is the set of differences between adjacent food intake values.

[0051] Step S1342: Perform trend analysis on the feeding interval data, calculate the change in the time interval between two adjacent feedings, and determine the stable trend or fluctuation trend of the feeding interval.

[0052] The fitting analysis layer calls the trend analysis module to perform a moving average on the feeding interval data. The size of the moving window is dynamically adjusted according to the amount of data. The average feeding interval and standard deviation within the moving window are calculated. By comparing the average feeding intervals of different windows, it is determined whether the feeding intervals show a gradually increasing trend, a gradually decreasing trend, or remain stable within a certain range. For example, if the absolute value of the difference between the average feeding intervals of multiple consecutive windows is less than a set threshold, it is determined to be a stable trend; if the average feeding interval continues to increase and the difference exceeds the threshold, it is determined to be an increasing trend within a fluctuating trend.

[0053] Step S1343: Perform trend analysis on the feed intake change data, calculate the difference in feed intake between two adjacent feedings, and determine the upward trend, downward trend, or stable trend of feed intake.

[0054] Using a method similar to step S1342, the data on changes in food intake are processed using a moving average to calculate the average difference in food intake and the standard deviation within each moving window. The trend is determined based on the sign and magnitude of the average difference in food intake: if the average difference is positive and continuously increasing, it is determined to be an upward trend; if the average difference is negative and continuously decreasing, it is determined to be a downward trend; if the average difference fluctuates slightly around zero, it is determined to be a stable trend.

[0055] Step S1344: Based on the trends of feeding intervals and food intake, construct a two-dimensional fitting model. The two-dimensional fitting model uses time as the horizontal axis and food intake as the vertical axis to map the data points in the historical time-food intake correlation sequence to a two-dimensional coordinate system.

[0056] The module for constructing the two-dimensional fitting model selects an appropriate curve type, such as a linear curve, exponential curve, or polynomial curve, based on the analysis results of the feeding interval trend and the feeding amount trend. Using the timestamps in the time series data as the horizontal axis and the corresponding feeding amount as the vertical axis, all data points in the historical time-feeding amount correlation sequence are plotted on a two-dimensional coordinate system to form a scatter plot.

[0057] Step S1345: The two-dimensional fitting model calls the curve fitting algorithm to draw curves for the data points in the two-dimensional coordinate system, so that the drawn curves can fit all data points to the maximum extent; during the curve fitting process, the two-dimensional fitting model calls the weight adjustment algorithm to adjust the weights of data points that deviate far from the curve.

[0058] The curve fitting algorithm employs the least squares method, adjusting curve parameters to minimize the sum of squared differences between the ordinates of the data points on the curve and their corresponding ordinates on the scatter plot. For data points in the scatter plot that deviate significantly from the initially plotted curve (deviation exceeding a set threshold), the weight adjustment algorithm reduces their weight in the fitting calculation, thus minimizing their influence on the curve parameters. Conversely, it increases the weight of data points with smaller deviations to improve the overall curve fitting effect.

[0059] Step S1346: After the fitted curve is generated, the variation pattern of the time interval is extracted from the fitted curve, and the variation pattern of the time interval is the feeding interval pattern of the pet; at the same time, the variation pattern of the amount of food is extracted from the fitted curve, and the variation pattern of the amount of food is the fluctuation pattern of the pet's amount of food.

[0060] After the fitted curve is generated, its derivative is calculated to obtain the slope variation. The variation pattern of the time interval is obtained by analyzing the time intervals corresponding to the curve's slope being zero or extreme points. For example, the time interval between adjacent extreme points can serve as a reference for the feeding interval pattern. The fluctuation pattern of food intake is extracted through the variation of the curve's amplitude, that is, the change in the difference between the maximum and minimum values ​​of the curve in different time periods.

[0061] Step S1347: Perform cross-validation on the extracted feeding interval pattern and food intake fluctuation pattern to check whether the feeding interval pattern and food intake fluctuation pattern match each other and whether they conform to the pet's feeding behavior logic. If they do not match, return to adjust the curve parameters of the two-dimensional fitting model until they match each other.

[0062] The cross-validation module compares the patterns of feeding intervals and food intake fluctuations with a pre-defined rule base for pet feeding behavior. For example, the rule base includes rules such as "food intake usually increases when feeding intervals shorten" and "food intake may decrease when feeding intervals lengthen." If the extracted patterns match most rules in the rule base, a match is determined; otherwise, the process returns to step S1345, adjusts the weight parameters or curve type in the curve fitting algorithm, and re-performs curve fitting until cross-validation is successful.

[0063] Step S1348: After cross-validation is passed, output the pet's feeding interval pattern and food intake fluctuation pattern.

[0064] After cross-validation is successful, the fitting analysis layer outputs the feeding interval pattern and the food intake fluctuation pattern as a feature vector. The feature vector contains various parameters describing the pattern, such as the average feeding interval, the standard deviation of the feeding interval, and the average food intake fluctuation range.

[0065] Step S135: During the fitting process, the historical fitting parameter library built into the feeding demand prediction model is called, and the parameters in the current fitting process are compared with the optimal parameters in the historical fitting parameter library to adjust the weight distribution in the fitting process.

[0066] The historical fitting parameter library stores parameter combinations that performed well in past fitting processes, such as curve fitting weight coefficients and sliding window sizes. During curve fitting using the two-dimensional fitting model (step S1345), the currently used parameters are compared in real-time with the optimal parameters in the historical fitting parameter library, and the parameter difference is calculated. Based on the magnitude of the difference, the current parameters are adjusted proportionally to bring them closer to the optimal parameters. For example, if the current weight coefficient differs significantly from the optimal weight coefficient, the adjustment range is increased; conversely, the adjustment range is decreased.

[0067] Step S136: After the fitting is completed, the prediction output layer of the feeding demand prediction model calculates the time interval from the most recent feeding time to the next time the pet may feel hungry, based on the feeding interval pattern obtained by the fitting, and obtains the prediction result of the pet's next hunger time.

[0068] The prediction output layer receives the feeding interval pattern feature vector from the fitting analysis layer and extracts the average feeding interval parameter. It then obtains the last timestamp from the time series data, corresponding to the most recent feeding time. Adding the average feeding interval parameter to the most recent feeding timestamp yields the timestamp corresponding to the next hunger time. This timestamp is then converted to an "hour:minute" format, which is the prediction result for the pet's next hunger time.

[0069] Step S137: Based on the fitted food intake fluctuation pattern and the most recent food intake information, the prediction output layer of the feeding demand prediction model calculates the amount of food the pet may need to eat next time, and obtains the prediction result of the pet's next food intake.

[0070] Retrieve the last element from the food intake array, representing the most recent food intake. Adjust the most recent food intake based on the average fluctuation amplitude and trend direction in the food intake fluctuation feature vector. For example, if the trend is upward and the average fluctuation amplitude is positive, add the average fluctuation amplitude to the most recent food intake; if the trend is downward and the average fluctuation amplitude is negative, subtract the average fluctuation amplitude. The adjusted result is the predicted food intake for the pet's next feeding.

[0071] Step S138: Perform a consistency check on the prediction results of the pet's next hunger time and the pet's next food intake, and determine whether they conform to the changing trend in the historical time-food intake correlation sequence. If they do not conform, return to the fitting analysis layer of the feeding demand prediction model to readjust the fitting parameters until a prediction result that conforms to the changing trend is obtained.

[0072] The consistency verification module compares the predicted next hunger time and the predicted next food intake with the trends of historical time-food intake correlation sequences. For example, it checks whether the predicted hunger time interval is within the fluctuation range of historical food intake interval patterns, and whether the predicted food intake is within the fluctuation range of historical food intake patterns. If both are within the range, it is determined that the trend is consistent; otherwise, it returns to the fitting analysis layer, readjusts the curve parameters of the two-dimensional fitting model (such as the weight adjustment parameters in step S1345), and executes the fitting and prediction process again until the prediction results conform to the trend.

[0073] Step S140: Based on the prediction results of the pet's next hunger time and the pet's next food intake, generate adaptive feeding recommendation information for the pet, which includes recommended feeding time and recommended food intake.

[0074] Step S141: Extract the time information from the prediction result of the pet's next hunger time, and use this time information as the basic recommended time.

[0075] From the next hunger time prediction results output by the feeding demand prediction model, the time description part is directly extracted. This time description part contains a combination of hour and minute information. After the complete extraction, it is assigned to the basic recommendation time variable as the initial basis for subsequent time adjustments.

[0076] Step S142: Analyze the deviation between the feeding time information corresponding to the most recent feedings and the pet's actual eating time, and adjust the basic recommended time according to the deviation to obtain the final recommended feeding time.

[0077] Step S1421: Extract feeding time information corresponding to the most recent feedings from the pet feeding data set, and extract the actual start time information of the pet for each feeding from the image monitoring records of the feeding device.

[0078] In the pet feeding dataset, the most recent consecutive feeding records are selected chronologically, and the feeding time information for each record is extracted. These feeding time records are clearly marked with the hour and minute. Simultaneously, in the image monitoring records of the feeding device, for each feeding, images are retrieved showing the pet first entering the feeding dish area and beginning to eat after the food dispensing stopped. The capture time of this image represents the actual time the pet began eating, also marked with the hour and minute. The extracted feeding time information and the corresponding actual start-of-eating time information are stored in a one-to-one correspondence, forming two parallel time data sequences.

[0079] Step S1422: Calculate the time difference between the feeding time information for each feeding and the corresponding time information when the pet actually starts eating, to obtain the time deviation data for multiple feedings.

[0080] For each set of feeding time information and actual start feeding time information, both are converted into cumulative minutes, calculated by multiplying the hours by 60 and then adding the minutes. The cumulative feeding time is then subtracted from the cumulative start feeding time; the difference is the feeding time deviation data. If the actual start feeding time is later than the feeding time, the time deviation data is positive; if the actual start feeding time is earlier than the feeding time, the time deviation data is negative; if both are exactly the same, the time deviation data is zero. All calculated time deviation data are stored sequentially to form a time deviation data sequence.

[0081] Step S1423: Perform statistical analysis on the time deviation data to determine the common deviation range and deviation trend in the time deviation data, and determine whether the pet has a habit of eating earlier or later.

[0082] First, the arithmetic mean of the time deviation data series is calculated, reflecting the central tendency of multiple time deviations. Then, the standard deviation of the time deviation data series is calculated, reflecting the dispersion of the time deviations. The common deviation range is defined as the interval formed by adding or subtracting one standard deviation from the arithmetic mean. The proportion of data points falling within this common deviation range is counted; a higher proportion indicates a more concentrated time deviation. Next, the deviation trend is analyzed by plotting the time deviation data as a function of feeding frequency. With feeding frequency on the horizontal axis and time deviation data on the vertical axis, each data point is plotted on a coordinate system, and the overall trend of the data points is observed. If the data points show a clear upward trend, it indicates that the pet's delayed eating is gradually worsening; if they show a clear downward trend, it indicates that the pet's early eating is gradually worsening; if the data points show no clear overall trend, it indicates that the deviation trend is stable. Based on the concentration of common deviation ranges and the trend of deviation, when the deviation data is positive for most of the time and the deviation trend is upward, it is judged that the pet has a habit of delaying eating; when the deviation data is negative for most of the time and the deviation trend is downward, it is judged that the pet has a habit of eating early; otherwise, it is judged that there is no obvious habitual deviation.

[0083] Step S1424: If the statistical analysis results show that the pet has a habit of eating early, then the time adjustment module is called to adjust the basic recommended time in advance according to the early deviation value in the common deviation range.

[0084] Within the common deviation range, the smallest deviation value is selected as the advance deviation value. This advance deviation value is negative, and its absolute value reflects the maximum extent to which pets typically eat ahead of schedule. The base recommended time is converted into cumulative minutes, and then the absolute value of the advance deviation value is subtracted to obtain the adjusted cumulative minutes. The adjusted cumulative minutes are then converted back into a combination of hours and minutes to obtain the advance-adjusted time.

[0085] Step S1425: If the statistical analysis results show that the pet has a habitual tendency to delay eating, then the time adjustment module is invoked to adjust the basic recommended time according to the delay deviation value in the common deviation range.

[0086] Within the common deviation range, the largest deviation value is selected as the delay deviation value. This positive value reflects the maximum extent to which a pet typically delays eating. The base recommended time is converted to cumulative minutes, and then the delay deviation value is added to obtain the adjusted cumulative minutes. The adjusted cumulative minutes are then converted back to a combination of hours and minutes to obtain the time adjusted for delay.

[0087] Step S1426: If the statistical analysis results show that the pet has no obvious habitual deviation, then the basic recommended time remains unchanged.

[0088] The base recommended time will be used directly as the adjusted time, without any additional time adjustment operations.

[0089] Step S1427: During the adjustment process, refer to the time deviation data corresponding to the most recent feeding. If the most recent time deviation exceeds the boundary value of the common deviation range, call the weight adjustment module to reduce the weight ratio of the time deviation data in the adjustment calculation.

[0090] The most recent time deviation data is compared with the upper and lower boundaries of the common deviation range. If the deviation is less than the lower limit or greater than the upper limit, it is considered an abnormal deviation. In this case, the weight adjustment module is invoked to reduce the weight of this abnormal deviation data when calculating statistics such as the arithmetic mean. For example, its weight is reduced from one part to half of the normal weight to minimize the impact of the abnormal deviation on the overall statistical results, making the statistical analysis more reflective of the pet's normal eating deviation patterns.

[0091] Step S1428: After the adjustment is completed, the adjusted time is determined as the candidate recommended feeding time, and the interval verification module is called to check the rationality of the candidate recommended feeding time, calculate the interval between the candidate recommended feeding time and the most recent feeding time, and determine whether the interval conforms to the feeding interval pattern in the historical time-feeding volume association sequence.

[0092] The adjusted times are selected as candidate recommended feeding times. Then, the feeding time of the most recent feeding is extracted from the pet's feeding data set. The time interval between the candidate recommended feeding time and this most recent feeding time is calculated and converted to minutes. Next, the time intervals between all adjacent feeding times in the historical time-food intake correlation sequence are analyzed. The arithmetic mean and standard deviation of these historical feeding intervals are calculated to determine a reasonable range for historical feeding intervals. The interval between the candidate recommended feeding time and the most recent feeding time is compared with the reasonable range of historical feeding intervals. If it falls within the reasonable range, the reasonableness check is passed; otherwise, it is deemed unreasonable.

[0093] Step S1429: If the feeding interval does not conform to the feeding interval pattern, return to the weight adjustment module to readjust the deviation weight, and perform the time adjustment operation again until a candidate recommended feeding time that conforms to the feeding interval pattern is obtained.

[0094] When the interval between the candidate recommended feeding time and the most recent feeding time does not conform to the historical feeding interval pattern, return to step S1427 to further reduce the weight of the most recent abnormal deviation data, or re-analyze the common deviation range and deviation trend, and then readjust the time according to the process of steps S1424 to S1426 to generate a new candidate recommended feeding time, and perform interval verification again. This cycle continues until the candidate recommended feeding time passes the interval verification.

[0095] Step S14210: Determine the candidate recommended feeding times that conform to the feeding interval pattern as the final recommended feeding times.

[0096] After the interval verification confirms that the candidate recommended feeding time is reasonable, it is officially determined as the final recommended feeding time and used for the generation of subsequent feeding recommendation information.

[0097] Step S143: Extract the food intake information from the predicted food intake of the pet for the next time, and use this food intake information as the basic recommended feeding amount.

[0098] The feed intake prediction results from the feeding demand prediction model are directly extracted, and the feed intake value information is marked with a clear weight unit. After the complete extraction, it is assigned to the basic recommended feed intake variable as the initial basis for subsequent feed intake adjustments.

[0099] Step S144: Analyze the relationship between the food intake information corresponding to the most recent feedings and the actual amount of food left in the pet's food. Adjust the basic recommended feeding amount according to this relationship so that the adjusted feeding amount is within a stable range of the pet's actual food intake, and obtain the final recommended feeding amount.

[0100] Step S1441: Extract the food intake information corresponding to the most recent feedings from the pet feeding data set, and at the same time extract the remaining food amount information at the time the pet leaves after each feeding from the image monitoring record of the feeding device.

[0101] In the pet feeding dataset, the most recent consecutive feeding records with the same number of feedings as in step S1421 are selected, and the amount of food consumed in each feeding record is extracted. This amount of food consumption information includes specific units of weight. Simultaneously, in the image monitoring records of the feeding device, for each feeding, the image of the food in the feeding dish at the moment the pet leaves is retrieved. This image is analyzed using a food state recognition model to obtain the amount of food remaining, which also includes units of weight. The extracted amount of food consumed and the corresponding amount of food remaining are stored in a one-to-one correspondence, forming two parallel weight data sequences.

[0102] Step S1442: Calculate the difference between the amount of food consumed each time and the corresponding amount of food remaining, to obtain the actual amount of food consumed by the pet each time.

[0103] For each set of corresponding food intake and remaining food information, subtract the remaining food information from the food intake information; the difference is the actual amount of food consumed by the pet during that feeding. Store all calculated actual food intake data sequentially to form an actual food intake data sequence.

[0104] Step S1443: Perform statistical analysis on the actual food intake data to determine the stable range and fluctuation of the pet's actual food intake, and determine whether the pet has a fixed food intake preference.

[0105] First, calculate the arithmetic mean of the actual food intake data series. This arithmetic mean reflects the average actual intake of pets. Then, calculate the standard deviation of the actual food intake data series. The standard deviation reflects the degree of fluctuation in actual intake. The stable range of actual food intake is defined as the interval formed by adding or subtracting twice the standard deviation, centered on the arithmetic mean. Count the proportion of data points in the actual food intake data series that fall within this stable range. If this proportion is high, for example, more than three-quarters, it indicates that the pet's actual intake is relatively stable, with a relatively fixed food preference; if the proportion is low, it indicates that the pet's food intake fluctuates greatly, without a clear fixed food preference.

[0106] Step S1444: Based on the stable range obtained from statistical analysis, calculate the average amount of food actually ingested by the pet, and use this average amount as a reference benchmark for food intake.

[0107] The arithmetic mean of the actual food intake data sequence calculated in step S1443 is directly determined as the food intake reference benchmark, which reflects the average food intake of pets under normal conditions.

[0108] Step S1445: Compare the basic recommended feeding amount with the food intake reference benchmark. If the basic recommended feeding amount is higher than the upper limit of the food intake reference benchmark, then call the feeding amount adjustment module to reduce the basic recommended feeding amount to the range of the food intake reference benchmark.

[0109] The basic recommended feeding amount is compared with the upper limit of the stable range of the feed intake reference. If the basic recommended feeding amount is greater than the upper limit, the feeding amount adjustment module is invoked to reduce the basic recommended feeding amount by a certain percentage, such as reducing the difference between the basic recommended feeding amount and the upper limit by a certain percentage, so that the adjusted feeding amount is equal to or slightly lower than the upper limit, ensuring that it falls within the feed intake reference range.

[0110] Step S1446: If the basic recommended feeding amount is lower than the lower limit of the food intake reference benchmark, then the feeding amount adjustment module is invoked to increase the basic recommended feeding amount to the range of the food intake reference benchmark.

[0111] The basic recommended feeding amount is compared with the lower limit of the stable range of the feed intake reference. If the basic recommended feeding amount is less than the lower limit, the feeding amount adjustment module is invoked to increase the basic recommended feeding amount by a certain percentage, such as increasing the difference between the basic recommended feeding amount and the lower limit by a certain percentage, so that the adjusted feeding amount is equal to or slightly higher than the lower limit, ensuring that it falls within the feed intake reference range.

[0112] Step S1447: If the basic recommended feeding amount is within the range of the food intake reference benchmark, then further analyze the difference between the most recent actual food intake data and the previous K times. If the most recent actual intake exceeds the average fluctuation range of the actual food intake of the previous K times, then call the feeding amount adjustment module to appropriately reduce the basic recommended feeding amount; if the most recent actual intake is lower than the average fluctuation range of the actual food intake of the previous K times, then call the feeding amount adjustment module to appropriately increase the basic recommended feeding amount, where K times is a set positive integer.

[0113] When the basal recommended feeding amount is within the reference range, select the actual food intake data from the previous K times (e.g., the previous three times), calculate the arithmetic mean and standard deviation of these data, and determine the average fluctuation range of the actual food intake from the previous K times. Compare the most recent actual food intake data with this average fluctuation range. If the most recent actual intake is greater than the upper limit of the average fluctuation range, appropriately reduce the basal recommended feeding amount, and the reduction amount is determined based on how much it exceeds the upper limit. If the most recent actual intake is less than the lower limit of the average fluctuation range, appropriately increase the basal recommended feeding amount, and the increase amount is determined based on how much it is below the lower limit. If the most recent actual intake is within the average fluctuation range, keep the basal recommended feeding amount unchanged.

[0114] Step S1448: During the adjustment process, refer to the feed dispensing accuracy information of the feeding equipment. If there is a deviation in the feed dispensing accuracy, call the accuracy correction module to make a second correction to the adjusted feeding amount according to the deviation. After the adjustment is completed, determine the adjusted feeding amount as the candidate recommended feeding amount.

[0115] The system acquires the deviation data between the actual feed dispensing amount and the target feed dispensing amount from the feeding equipment in recent feeding cycles, and calculates the arithmetic mean of these deviations as the current feed dispensing accuracy deviation. If the absolute value of this average deviation is large, exceeding the preset allowable range for feed dispensing accuracy, then the feed dispensing accuracy is considered to be inaccurate. In this case, the accuracy correction module is invoked to perform a secondary correction on the adjusted feed amount based on the sign of the feed dispensing accuracy deviation. If the feed dispensing accuracy deviation is positive, meaning the actual feed dispensing amount is generally higher than the target feed dispensing amount, the average deviation value is subtracted from the adjusted feed amount; if the feed dispensing accuracy deviation is negative, meaning the actual feed dispensing amount is generally lower than the target feed dispensing amount, the absolute value of the average deviation value is added to the adjusted feed amount. The feed amount after this secondary correction is determined as the candidate recommended feed amount.

[0116] Step S1449: Call the food quantity verification module to calculate the difference between the candidate recommended feeding amount and the most recent actual food intake data, and determine whether the difference is within a reasonable fluctuation range. If it does not meet the reasonable fluctuation range, return to the food quantity reference benchmark calculation step to readjust the weight of the food quantity reference benchmark, and perform the feeding amount adjustment operation again until a candidate recommended feeding amount that meets the reasonable fluctuation range is obtained. The candidate recommended feeding amount that meets the reasonable fluctuation range is determined as the final recommended feeding amount.

[0117] Calculate the absolute value of the difference between the candidate recommended feeding amount and the most recent actual food intake data. Compare this absolute value with a certain percentage (e.g., 10%) of the most recent actual food intake data. If the absolute value is less than or equal to this percentage, the difference is considered to be within a reasonable fluctuation range; otherwise, it is determined to be non-compliant. If non-compliant, return to step S1444. When calculating the food intake reference baseline, adjust the weight of each actual food intake data, for example, increase the weight of the most recent actual intake, recalculate the food intake reference baseline, and then readjust the feeding amount according to steps S1445 to S1448 to generate a new candidate recommended feeding amount. Perform the food intake verification again. Repeat this process until the candidate recommended feeding amount meets the reasonable fluctuation range, and then determine it as the final recommended feeding amount.

[0118] Step S145: Collect the feed dispensing status information of the most recent feeding device and use the feed dispensing status information of the most recent feeding device as auxiliary information. The feed dispensing status information of the most recent feeding device includes whether the feed dispensing is smooth and whether the feed dispensing accuracy meets the preset requirements.

[0119] The operation record of the feeding equipment's most recent feeding process is queried to extract the operating parameters of the feed dispensing motor, such as operating current and speed stability, to determine if there were any jams or blockages during the feed dispensing process, thus confirming the smoothness of the dispensing. Simultaneously, the actual feed dispensing amount from the most recent dispensing is compared with the target dispensing amount to calculate the dispensing accuracy deviation and determine if this deviation is within the preset dispensing accuracy requirements. The results of the smoothness of dispensing and the accuracy of dispensing are combined to form the feed dispensing status information of the most recent feeding equipment, serving as supplementary information.

[0120] Step S146: Collect the pet's behavior status information after the most recent feeding, and use the pet's behavior status information after the most recent feeding as auxiliary information. The pet's behavior status information after the most recent feeding includes the activity frequency and rest duration after feeding.

[0121] A pet activity monitoring device connected to the feeding equipment acquires activity data within a preset time period (e.g., two hours) following the pet's most recent meal. The number of times the pet is active within this time period is counted to calculate the activity frequency; simultaneously, the duration of continuous rest is recorded, including the longest rest time and the total rest time. This data, including activity frequency and rest duration, is combined to create behavioral status information about the pet after its most recent meal, serving as supplementary information.

[0122] Step S147: Integrate the recommended feeding time, the recommended feeding amount, and the two types of auxiliary information to make the information logically coherent and form a preliminary feeding recommendation information framework.

[0123] With recommended feeding times and amounts as the core content, two types of auxiliary information are used as supplementary explanations, organized in a logical order. For example, the recommended feeding times and amounts are explained first, followed by information on food dispensing status and pet behavior status, and the relationship between this auxiliary information and the core recommendations is explained. For instance, smooth food dispensing helps ensure the accuracy of the recommended feeding amount, and high activity levels after eating may indicate that the amount of food is appropriate. This allows all types of information to be combined into a logically coherent whole, forming a preliminary framework for feeding recommendations.

[0124] Step S148: Call the information labeling module to add core identification tags to the recommended feeding time and recommended feeding amount in the preliminary feeding recommendation information framework, and add auxiliary identification tags to the food dispensing status information of the feeding device in the most recent time and the behavior status information of the pet after the most recent feeding.

[0125] The information annotation module is invoked to scan the text content within the initial feeding recommendation framework, identifying text paragraphs corresponding to recommended feeding times and amounts. These paragraphs are then labeled with core identifiers, such as specific symbols or prefixes, to clearly define their central recommendation status. Simultaneously, text paragraphs corresponding to the most recent feeding status and the pet's behavior status after the most recent meal are identified, and these paragraphs are labeled with auxiliary identifiers to distinguish them from the core identifiers and indicate their supplementary informational role.

[0126] Step S149: The recommended feeding time, recommended feeding amount and two types of auxiliary information with identification tags are arranged in the order of core information first and auxiliary information last through the information sorting module to form a feeding recommendation information structure.

[0127] The information sorting module reads the information from each tagged section and organizes it according to the order of priority for information with core tags, followed by information with auxiliary tags. Within the core information, it is arranged in the order of recommended feeding time first, followed by recommended feeding amount; within the auxiliary information, it is arranged in the order of feeding status information first, followed by pet behavior status information, thus forming a clearly structured feeding recommendation information structure.

[0128] Step S1410: Call the format verification module to verify the format of the feeding recommendation information structure, check whether the labels are correct and whether the information arrangement order conforms to the preset rules. If the verification fails, return to the information labeling module to re-add labels or return to the information sorting module to readjust the arrangement order.

[0129] The format verification module performs a comprehensive check on the feeding recommendation information structure according to preset format rules. First, it checks whether the core and auxiliary identification tags are added correctly, and whether there are any missing, incorrect, or duplicate tags. Then, it checks whether the arrangement of each part of the information conforms to the preset rule of prioritizing core information and placing auxiliary information later, and whether the order of sub-information within the core and auxiliary information is correct. If any non-compliance is found, the verification fails. Depending on the specific problem, the module returns to the information labeling module to re-add identification tags, or returns to the information sorting module to readjust the information's order, until the format verification passes.

[0130] Step S1411: After the verification is passed, generate pet adaptive feeding recommendation information containing recommended feeding time and recommended feeding amount.

[0131] After the feeding recommendation information structure passes format validation, it is converted into a preset text format, such as paragraphs in natural language, to ensure that the information is clear and easy to understand. The final generated pet adaptive feeding recommendation information includes recommended feeding time, recommended feeding amount, and two types of auxiliary information, presenting complete feeding suggestions to the user.

[0132] Step S150: Push the pet adaptive feeding recommendation information to the mobile terminal associated with the feeding device, so that the user can view the pet adaptive feeding recommendation information and decide whether to perform the feeding operation based on the pet adaptive feeding recommendation information.

[0133] Step S151: Call the communication establishment module to establish a communication connection between the feeding device and the associated mobile terminal. After converting the data format of the pet adaptive feeding recommendation information, transmit the converted pet adaptive feeding recommendation information to the mobile terminal through the established communication connection. This allows the mobile terminal to display the recommended feeding time, recommended feeding amount, and auxiliary information through the terminal display interface. During the information display process, operation options are provided for the user to select. The operation options include the option to perform the feeding operation and the option not to perform the feeding operation.

[0134] The feeding device's communication module supports Wi-Fi connectivity. The communication establishment module scans for preset Wi-Fi networks, enters a password, and establishes a connection with the home router. The mobile terminal also connects to the same Wi-Fi network. The communication establishment module searches for and identifies associated mobile terminals within the local area network using the mobile terminal's device identifier (such as its Bluetooth MAC address) and establishes a TCP / IP connection. The text string containing the pet's adaptive feeding recommendations is converted to JSON format data, and a message type identifier (such as "Recommendation Information") is added. The JSON data is sent to the mobile terminal via the established TCP / IP connection. The mobile terminal's receiving module parses the JSON data, extracts the recommended feeding time, recommended feeding amount, and auxiliary information, and displays it in a designated area of ​​the display interface. Two buttons are located at the bottom of the display interface: "Execute Feeding" and "Do Not Feed," corresponding to the options of executing or not executing the feeding operation.

[0135] Step S152: If the user selects to perform the feeding operation option, the user's operation command is transmitted to the command receiving module of the feeding device, so that after receiving the operation command, the command receiving module of the feeding device calls the control module to prepare the feeding operation according to the recommended feeding time and recommended feeding amount.

[0136] For example, in step S1521: the instruction selected by the user to perform the feeding operation is encoded, and the encoded operation instruction is transmitted to the feeding device through the established communication connection between the feeding device and the mobile terminal. The feeding device then calls the instruction decoding algorithm to decode the encoded operation instruction, obtains the instruction content to perform the feeding operation, reads the recommended feeding time and recommended feeding amount, compares the recommended feeding time with the current time, and determines whether the feeding operation needs to be performed immediately or at the recommended feeding time.

[0137] After the user clicks the "Execute Feeding" button, the mobile terminal's instruction generation module encodes the operation instruction into binary data, adding an instruction header (containing instruction length and checksum) and an instruction tail identifier. The encoded instruction is then sent to the feeding device via an established TCP / IP connection. Upon receiving the instruction, the feeding device's instruction receiving module first verifies the checksum. If verification is successful, it calls an instruction decoding algorithm (such as Base64 decoding) to convert the binary data into text instruction content, which includes the "Execute Feeding" identifier, recommended feeding time, and recommended feeding amount. The control module reads the current time (obtained through the clock module) and compares it with the recommended feeding time. If the difference is within a preset range (e.g., within five minutes), it determines that the feeding operation needs to be executed immediately; otherwise, it determines that the feeding operation should be executed at the recommended feeding time.

[0138] Step S1522: If the current time is consistent with the recommended feeding time, the control module of the feeding device sends a feeding instruction to the feeding module of the feeding device. The feeding instruction contains the feeding parameters corresponding to the recommended feeding amount.

[0139] The feed dispensing parameters include the number of revolutions and the speed of the feed dispensing motor. The control module calculates the required number of revolutions (recommended feed amount divided by the feed dispensing rate) based on the recommended feed amount and the feed dispensing rate (feed amount per revolution) of the feed dispensing module. The rotation speed is preset to a fixed value. The number of revolutions and the rotation speed are packaged into a feed dispensing command and sent to the feed dispensing module via the internal bus.

[0140] Step S1523: If the current time is earlier than the recommended feeding time, the control module of the feeding device stores the recommended feeding time and the recommended feeding amount, and monitors the current time in real time. When the recommended feeding time is reached, it sends a feeding instruction containing the feeding parameters corresponding to the recommended feeding amount to the feeding module, so that the feeding module can perform feeding operation according to the feeding parameters in the feeding instruction after receiving the feeding instruction. During the feeding process, the image monitoring module of the feeding device takes pictures of the food in the feeding tray at preset intervals to check whether the feeding amount meets the recommended feeding amount.

[0141] The control module stores the recommended feeding time and amount in a temporary cache, starts a timer, and retrieves the current time every minute, comparing it with the recommended feeding time. When the current time reaches the recommended feeding time, the control module calculates the feed dispensing parameters according to step S1522 and generates a feed dispensing command, which is then sent to the feed dispensing module. The motor of the feed dispensing module starts rotating at the set speed and number of revolutions, driving the feed dispensing screw to push the food to the feeding tray. During the feed dispensing process, the image monitoring module takes an image of the feeding tray at preset intervals (e.g., ten seconds), calculates the current feed dispensing amount in real time using a food status recognition model, and compares it with the recommended feeding amount. If the current feed dispensing amount reaches more than 95% of the recommended feeding amount, the control module issues a stop feed dispensing command.

[0142] Step S1524: After the feed is dispensed, the feed dispensing module sends a feed dispensing completion signal to the control module. The control module of the feeding equipment calls the recording module to record the execution time and actual feed dispensing amount of this feeding, and calls the data addition module to add them to the operation record of the feeding equipment.

[0143] After the motor of the feed dispensing module stops rotating, it sends a feed dispensing completion signal containing the actual number of rotations to the control module. The control module calculates the actual feed dispensing amount (actual number of rotations multiplied by the feed dispensing rate) based on the actual number of rotations and the feed dispensing rate. It then calls the recording module to write the execution time (current time) and the actual feed dispensing amount to the operation log file (text format). Each record is formatted as "Execution Time - Actual Feed Dispensing Amount - Recommended Feeding Amount". The data appending module appends this record to the end of the operation log file.

[0144] Step S1525: At the same time, the control module of the feeding device transmits the execution time and actual feed output of this feeding to the mobile terminal, so that the information display module of the mobile terminal can display it to the user through the display interface.

[0145] The control module encapsulates the execution time and actual feed output into JSON format data, adds a "feeding result" message type identifier, and sends it to the mobile terminal via the communication module. Upon receiving the data, the mobile terminal displays a pop-up message box showing "Feeding executed, execution time: XX:XX, actual feed output: XX grams".

[0146] Step S1526: The control module of the feeding device associates the execution time of this feeding with the actual amount of food dispensed with the pet feeding data set, and updates the pet feeding data set.

[0147] The control module converts the execution time into a timestamp, forms a new time-feed amount data pair with the actual food output, adds it to the JSON file of the pet feeding data set, and reorders it according to the timestamp order to complete the update of the data set.

[0148] Step S153: If the user chooses not to perform the feeding operation, record the user's operation selection and feed it back to the feeding demand prediction model as a reference factor for subsequent fitting analysis.

[0149] After the user clicks the "Do not feed" button, the mobile terminal encodes the operation command and sends it to the feeding device. The control module records the operation selection as "do not execute" and stores this information in the user feedback record. When the feeding demand prediction model is fitted and analyzed next time (step S135), the user feedback record is called. If there is a "do not execute" record, the weight of the corresponding historical fitting parameter is appropriately reduced, and the fitting process is adjusted.

[0150] Step S154: If the user does not make a selection immediately, control the mobile terminal to start timing. After the preset time is reached, generate an information viewing reminder command and trigger the reminder operation through the terminal display interface or sound prompt.

[0151] The mobile device starts a timer while displaying feeding recommendations; the preset time is ten minutes. If the user does not click any action option within ten minutes, the timer triggers a reminder, the mobile device's display screen begins to flash, and a preset prompt sound plays until the user performs an action or manually turns off the reminder.

[0152] Step S160: During the above push interaction process, record the information transmission time, user viewing time, and user operation time, and add the time information to the pet eating data set.

[0153] During communication, the feeding device and the mobile terminal record the time of information transmission and reception respectively; the mobile terminal records the user's viewing time when the recommended information is displayed (triggered by the display interface rendering completion event); the user's operation time is recorded when the user clicks the operation option. The above time information is compiled into a time log. The control module adds the key time points (such as user operation time) in the time log to the corresponding records in the pet feeding data set as reference data for subsequent analysis.

[0154] Step S210: Pre-train the image feature extraction model.

[0155] Step S211: Collect a large number of pet food images of different types, shapes and distribution states as training samples, and annotate the training samples. The annotation content includes the outline feature points, height regions, coverage boundaries and density level regions of the food.

[0156] The training sample images were captured by multiple cameras under different lighting conditions, covering different types of food, such as dry and wet food, different forms such as piled up and laid flat, and different distribution states such as centrally clustered and peripherally dispersed. Each image was annotated using professional annotation tools, and contour feature points (such as the coordinates of inflection points on the contour) were manually marked. Different colors were used to divide height regions (high, medium, and low), the boundary lines of the coverage area were drawn, and different gray values ​​were used to annotate density level regions (high density, medium density, and low density).

[0157] Step S212: Perform data augmentation processing on the training sample images, including operations such as rotation, scaling, cropping, and brightness adjustment, to expand the training sample set.

[0158] The data augmentation module performs multiple transformations on each original image: random rotation (-15 degrees to 15 degrees), random scaling (0.8 to 1.2 times the original size), random cropping (center cropping to preserve the food region), and random brightness adjustment (0.7 to 1.3 times the original brightness). Each transformation generates multiple new images, which, together with the original images, form the augmented training sample set.

[0159] Step S213: Construct the network structure of the image feature extraction model, including an image preprocessing layer, a feature detection layer, and a feature integration module.

[0160] The image preprocessing layer includes convolutional layers (for noise removal, with a preset kernel size) and pooling layers (for size normalization, with a preset kernel size). The feature detection layer consists of multiple convolutional blocks, each containing a convolutional layer, a batch normalization layer, and an activation function layer to extract features from different levels. The feature integration module is a fully connected layer that concatenates and integrates the features output from each feature detection layer.

[0161] Step S214: Input the labeled training samples into the image feature extraction model, use the backpropagation algorithm for training, calculate the difference between the predicted features and the labeled features through the loss function, adjust the model parameters until the loss function value converges to the preset threshold.

[0162] During training, the training sample set is divided into a training set and a validation set (e.g., 8:2). Images from the training set are input into the model, and the model outputs predicted contour features, height features, coverage features, and density features. Mean squared error loss is used as the loss function to calculate the error between the predicted features and the labeled features. Through backpropagation, the weights and bias parameters of each layer are adjusted layer by layer from the output layer to the input layer. After each training round, the loss value is calculated on the validation set. When the validation set loss value no longer decreases for several consecutive rounds and is less than a preset threshold, training is stopped, and the model parameters at this point are saved.

[0163] Step S220: Pre-train the food state recognition model.

[0164] Step S221: Collect a large number of food state samples. Each sample contains the food's morphological feature vector, distribution feature vector, and corresponding real food state (volume, average height, density level, etc.).

[0165] Food state samples are obtained by extracting features from food images in known states using an image feature extraction model. The true food state parameters are obtained by measuring the volume with a graduated cylinder, the height with a laser rangefinder, and the density with a densitometer.

[0166] Step S222: Construct a food state recognition model. The model consists of an input layer, multiple fully connected layers, and an output layer. The input layer receives a concatenated vector of morphological feature vectors and distribution feature vectors, and the output layer outputs food state parameters.

[0167] The number of nodes in the input layer is equal to the total dimension of the morphological feature vector and the distribution feature vector. The number of nodes in the fully connected layer decreases layer by layer. The number of nodes in the output layer is consistent with the number of food state parameters (e.g., if there are three parameters such as volume, average height, and density level, then the output layer has three nodes).

[0168] Step S223: Train the food state recognition model using labeled food state samples, adopt the mean squared error loss function, and optimize the model parameters through gradient descent algorithm until the model's prediction accuracy on the test set reaches the preset requirements.

[0169] The food state samples are divided into training, validation, and test sets (e.g., a ratio of 7:1:2). After the training set samples are input into the model, the predicted food state parameters are output, and the mean squared error loss is calculated by comparing them with the true parameters. The weight matrix and bias vector of the fully connected layers are adjusted using the gradient descent algorithm. After each training round, the accuracy is evaluated on the validation set. When the accuracy on the test set reaches a preset threshold (e.g., 90%), training is stopped, and the model parameters are saved.

[0170] Step S230: Pre-train the feeding demand prediction model.

[0171] Step S231: Collect a large amount of historical feeding data of pets as training data. The historical feeding data includes multiple sets of time-feed amount correlation sequences.

[0172] Historical feeding data comes from pet feeding records of multiple households, covering pets of different breeds and ages. Each time-feeding sequence contains at least a preset number of consecutive feeding records (e.g., 30 times).

[0173] Step S232: Construct a feeding demand prediction model, including a sequence processing layer, a fitting analysis layer, and a prediction output layer. The sequence processing layer adopts an LSTM network structure, the fitting analysis layer includes a two-dimensional fitting module and a parameter adjustment module, and the prediction output layer is a fully connected layer.

[0174] The number of LSTM units in the sequence processing layer is set according to the length of the input sequence. The two-dimensional fitting module of the fitting analysis layer has multiple built-in curve fitting algorithms. The parameter adjustment module is connected to the historical fitting parameter library. The number of nodes in the fully connected layer of the prediction output layer is 2 (to output the next hunger time and food intake respectively).

[0175] Step S233: Input the training data into the feeding demand prediction model, use the time series prediction loss function (such as MAE loss), train the model through the backpropagation algorithm, adjust the weight parameters of each layer, and continuously optimize the historical fitting parameter library until the prediction error of the model is within the preset range.

[0176] Training data is divided into input and target sequences in chronological order. The input sequence consists of the time-food intake correlation data from the first N iterations, while the target sequence contains the time and food intake data from the (N+1)th iteration. The Model-Based Error (MAE) loss is calculated between the model's prediction results and the target sequence. Backpropagation is used to update the gating parameters of the LSTM units, the curve parameters of the fitting analysis layer, and the weights of the prediction output layer. During training, the optimal parameters from each training iteration (e.g., the parameters that minimize the loss) are stored in a historical fitting parameter database for reference in subsequent predictions. Training is complete when the MAE loss on the validation set is less than a preset threshold.

[0177] Regarding privacy protection: Throughout the entire data collection and processing process, all pet feeding data, image data, etc., are stored locally on devices (feeding devices and mobile terminals) and are not uploaded to cloud servers. Data transmission uses encrypted communication methods (such as AES encryption), and communication between the mobile terminal and the feeding device requires device identification authentication to prevent unauthorized devices from accessing the network. After image data is processed locally, only feature vectors are retained, and original images are automatically deleted periodically (e.g., retaining original images from the most recent month) to reduce the risk of privacy data leakage.

[0178] Based on the same inventive concept, please refer to Figure 2 This paper shows a schematic block diagram of a pet adaptive feeding recommendation system 100 based on AI image analysis, which is provided in an embodiment of this application for executing the above-described pet adaptive feeding recommendation method based on AI image analysis. The pet adaptive feeding recommendation system 100 based on AI image analysis may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0179] In this embodiment, both the machine-readable storage medium 120 and the processor 130 are located within the AI ​​image analysis-based pet adaptive feeding recommendation system 100 and are separately configured. However, it should be understood that the machine-readable storage medium 120 may also be independent of the AI ​​image analysis-based pet adaptive feeding recommendation system 100 and may be accessed by the processor 130 via a bus interface. Alternatively, the machine-readable storage medium 120 may be integrated into the processor 130 and may communicate with external systems via the communication unit 110.

[0180] The processor 130 is the control center of the AI ​​image analysis-based pet adaptive feeding recommendation system 100. It connects various parts of the system via various interfaces and lines, and executes software programs and / or modules stored in the machine-readable storage medium 120, as well as accessing data stored in the machine-readable storage medium 120. This allows for the execution of various functions and data processing by the AI ​​image analysis-based pet adaptive feeding recommendation system 100, thereby providing overall monitoring of the system. Optionally, the processor 130 may include one or more processing cores; for example, it may integrate an application processor and a modem processor. The application processor primarily handles the operating system, user interface, and applications, while the modem processor primarily handles wireless communication. It is understood that the modem processor may not be integrated into the processor. The machine-readable storage medium 120 is used to store machine-executable instructions for executing the scheme of this application, and the processor 130 is used to execute the machine-executable instructions stored in the machine-readable storage medium 120 to implement the AI ​​image analysis-based pet adaptive feeding recommendation method provided in the aforementioned method embodiments.

[0181] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A pet adaptive feeding recommendation method based on AI image analysis, characterized in that, The method includes: The food images of the feeding tray of the feeding device are collected at the time when food dispensing stops and when the pet leaves after eating, thus obtaining a set of food status images of the feeding tray; AI image analysis processing is performed on the set of food status images in the feeding bowl to extract the feeding time information and the amount of food consumed in each feeding, forming a pet feeding data set containing multiple sets of corresponding feeding time information and food consumption information. The pet feeding dataset is input into a pre-trained feeding demand prediction model. The correlation between time and food intake in the pet feeding dataset is fitted and analyzed to obtain the prediction results of the pet's next hunger time and the pet's next food intake. Based on the prediction results of the pet's next hunger time and the pet's next food intake, generate adaptive feeding recommendation information for the pet, which includes recommended feeding time and recommended food amount; The adaptive feeding recommendation information for pets is pushed to the mobile terminal associated with the feeding device, so that users can view the adaptive feeding recommendation information for pets and decide whether to perform the feeding operation based on the adaptive feeding recommendation information for pets.

2. The pet adaptive feeding recommendation method based on AI image analysis according to claim 1, characterized in that, The AI ​​image analysis processing is performed on the set of food status images in the feeding bowl to extract the feeding time information and the amount of food consumed in each feeding, forming a pet feeding data set containing multiple sets of correspondences between feeding time information and food consumption information, including: Separate the food image of the feeding tray at the moment when food dispensing stops and the food image of the feeding tray at the moment when the pet eats and leaves from the set of food state images of the feeding tray, so that the food image of the feeding tray at the moment when food dispensing stops and the food image of the feeding tray at the moment when the pet eats and leaves correspond to the initial state and the end state of the same feeding, respectively. The pre-trained image feature extraction model is invoked to extract features from the food image of the feeding tray at the moment when the feeding stops, so as to obtain the morphological and distribution features of the food in the feeding tray at the moment when the feeding stops. The pre-trained image feature extraction model is invoked to extract features from the food image of the feeding dish at the moment the pet leaves the feeding dish, thereby obtaining the morphological and distribution features of the food in the feeding dish at the moment the pet leaves the feeding dish. The morphological and distribution characteristics of the food in the feeding tray at the moment when feeding stops are input into the pre-trained food state recognition model to obtain the initial food state of the feeding tray at the moment when feeding stops. The shape and distribution characteristics of the food in the feeding dish at the moment the pet leaves the feeding dish are input into the pre-trained food state recognition model to obtain the remaining food state in the feeding dish at the moment the pet leaves the feeding dish. Extract the feed dispensing stop time information from the operation record of the feeding equipment, and determine the feed dispensing stop time information as the feeding time information for this feeding; Based on the difference between the initial food state and the remaining food state, and combined with the pre-stored feeding tray capacity parameters and food density parameters, the food intake information corresponding to this feeding is extracted; The feeding time information and the corresponding food intake information are associated and stored. The above steps are repeated to process the food status image set of the feeding plate corresponding to multiple feedings, so as to obtain multiple sets of correspondence between feeding time information and food intake information. The correspondence between multiple sets of feeding time information and food intake information is arranged in chronological order of the feeding time information to form a pet feeding data set containing the correspondence between multiple sets of feeding time information and food intake information.

3. The pet adaptive feeding recommendation method based on AI image analysis according to claim 2, characterized in that, The pre-trained image feature extraction model is invoked to extract features from the food image in the feeding tray at the moment the feeding stops, obtaining the morphological and distribution features of the food in the feeding tray at the moment the feeding stops, including: The food image of the feeding tray at the moment when feeding stops is input into the image preprocessing layer of the pre-trained image feature extraction model for noise removal. The processed food image of the feeding tray is then input into the feature detection layer of the image feature extraction model to detect the food contours and extract the contour features of the food. These contour features are used as a component of the morphological features. The feature detection layer detects the height distribution of food in the processed feeding plate food image, determines the stacking height of food by the grayscale difference in the processed feeding plate food image, and extracts the height features of food, which are used as a component of morphological features. The feature detection layer detects the distribution range of food within the feeding plate in the processed feeding plate food image, determines the proportion of the area covered by food in the feeding plate, and extracts the coverage range feature of the food, which is used as a component of the distribution feature. The feature detection layer detects the density distribution of food in the processed feeding plate food image, judges the density difference of food by the density of pixels in the processed feeding plate food image, and extracts the density features of food, which are used as a component of the distribution features. The extracted contour and height features are integrated to form the morphological features of the food in the feeding tray at the moment when feeding stops; the extracted coverage and density features are integrated to form the distribution features of the food in the feeding tray at the moment when feeding stops, and the morphological and distribution features of the food in the feeding tray at the moment when feeding stops are output.

4. The pet adaptive feeding recommendation method based on AI image analysis according to claim 1, characterized in that, The step involves inputting the pet's feeding data set into a pre-trained feeding demand prediction model, performing a fitting analysis on the correlation between time and food intake in the pet's feeding data set, and obtaining predictions for the pet's next hunger time and next food intake, including: The feeding time information in the pet feeding data set is organized into a time series so that all feeding time information is formed into a continuous time series data in chronological order. The pet feeding data set is associated and matched to ensure that each feeding amount corresponds one-to-one with the feeding time information in the corresponding time series data, forming a time-feeding amount association sequence. The time-feed amount correlation sequence is input into the sequence processing layer of the pre-trained feeding demand prediction model. The sequence processing layer of the feeding demand prediction model performs a preliminary analysis on the time interval and feeding amount variation pattern in the time-feed amount correlation sequence. After the sequence processing layer of the feeding demand prediction model outputs the preliminary analysis results, the preliminary analysis results are input into the fitting analysis layer of the feeding demand prediction model. Based on the changing trend of the historical time-food intake correlation sequence, the fitting analysis layer of the feeding demand prediction model fits the feeding interval pattern and food intake fluctuation pattern of the pet. During the fitting process, the historical fitting parameter library built into the feeding demand prediction model is called, and the parameters in the current fitting process are compared with the optimal parameters in the historical fitting parameter library to adjust the weight distribution in the fitting process. After the fitting is completed, the prediction output layer of the feeding demand prediction model calculates the time interval from the most recent feeding time to the next time the pet may feel hungry, based on the feeding interval pattern obtained from the fitting, and obtains the prediction result of the pet's next hunger time. The prediction output layer of the feeding demand prediction model calculates the amount of food the pet may need to eat next time based on the fitted pattern of food intake fluctuations and the most recent food intake information, thus obtaining the prediction result of the pet's next food intake. The consistency of the prediction results of the pet's next hunger time and the pet's next food intake is checked to determine whether they conform to the changing trend in the historical time-food intake correlation sequence. If they do not conform, the model returns to the fitting analysis layer of the feeding demand prediction model to readjust the fitting parameters until a prediction result that conforms to the changing trend is obtained.

5. The pet adaptive feeding recommendation method based on AI image analysis according to claim 4, characterized in that, The fitting analysis layer of the feeding demand prediction model fits the pet's feeding interval pattern and feeding amount fluctuation pattern based on the changing trend of historical time-feeding quantity correlation sequence, including: After receiving the preliminary analysis results output by the sequence processing layer of the feeding demand prediction model through the fitting analysis layer of the feeding demand prediction model, the feeding interval data and feeding amount change data in the historical time-feeding amount correlation sequence are extracted from the preliminary analysis results. Trend analysis is performed on the feeding interval data to calculate the change in the time interval between two adjacent feedings and to determine the stable trend or fluctuation trend of the feeding interval. Perform trend analysis on the data on changes in feed intake, calculate the difference in feed intake between two consecutive feedings, and determine whether the feed intake is increasing, decreasing, or stable. Based on the trends of feeding intervals and food intake, a two-dimensional fitting model is constructed. The two-dimensional fitting model uses time as the horizontal axis and food intake as the vertical axis to map the data points in the historical time-food intake correlation sequence to a two-dimensional coordinate system. The two-dimensional fitting model calls a curve fitting algorithm to draw curves for the data points in the two-dimensional coordinate system, so that the drawn curves can fit all data points as closely as possible. During the curve fitting process, the two-dimensional fitting model calls a weight adjustment algorithm to adjust the weights of data points that deviate far from the curve. After the fitted curve is generated, the variation pattern of the time interval is extracted from the fitted curve, and the variation pattern of the time interval is the feeding interval pattern of the pet; at the same time, the variation pattern of the food intake is extracted from the fitted curve, and the variation pattern of the food intake is the food intake fluctuation pattern of the pet. Cross-validate the extracted feeding interval pattern and food intake fluctuation pattern to check whether the feeding interval pattern and food intake fluctuation pattern match each other and whether they conform to the pet's feeding behavior logic. If they do not match, return to adjust the curve parameters of the two-dimensional fitting model until they match each other. After successful cross-validation, the pet's feeding interval pattern and food intake fluctuation pattern are output.

6. The pet adaptive feeding recommendation method based on AI image analysis according to claim 1, characterized in that, The step of generating adaptive feeding recommendation information for the pet, including recommended feeding time and recommended feeding amount, based on the predicted time of the pet's next hunger and the predicted amount of food the pet will eat next, includes: Extract the time information from the prediction results of the pet's next hunger time, and use this time information as the base recommendation time; The deviation between the feeding time information corresponding to the most recent feedings and the pet's actual eating time is analyzed, and the basic recommended time is adjusted according to the deviation to obtain the final recommended feeding time. Extract the food intake information from the predicted food intake of the pet for the next time, and use this food intake information as the basic recommended feeding amount; The relationship between the food intake information corresponding to the most recent feedings and the actual amount of food left in the pet is analyzed. Based on this relationship, the basic recommended feeding amount is adjusted so that the adjusted feeding amount is within a stable range of the pet's actual food intake, thus obtaining the final recommended feeding amount. Collect the most recent feed dispensing status information of the feeding device, and use the most recent feed dispensing status information of the feeding device as auxiliary information. The most recent feed dispensing status information of the feeding device includes whether the feed dispensing is smooth and whether the feed dispensing accuracy meets the preset requirements. Collect the pet's behavior status information after the most recent feeding, and use the pet's behavior status information after the most recent feeding as auxiliary information. The pet's behavior status information after the most recent feeding includes the pet's activity frequency and rest duration after feeding. The recommended feeding time, the recommended feeding amount, and the two types of auxiliary information are integrated to make the various types of information logically coherent and form a preliminary feeding recommendation information framework. The information labeling module is called to add core identification tags to the recommended feeding time and recommended feeding amount in the initial feeding recommendation information framework, and to add auxiliary identification tags to the food dispensing status information of the feeding device in the most recent time and the behavior status information of the pet after the most recent feeding. The information sorting module arranges the recommended feeding time, recommended feeding amount and two types of auxiliary information with labels in the order of core information first and auxiliary information last, forming a feeding recommendation information structure. The format verification module is called to verify the format of the feeding recommendation information structure, check whether the labels are correct and whether the information arrangement order conforms to the preset rules. If the verification fails, the information labeling module is returned to re-add labels or the information sorting module is returned to readjust the arrangement order. After verification, adaptive feeding recommendations for pets, including recommended feeding times and amounts, are generated.

7. The pet adaptive feeding recommendation method based on AI image analysis according to claim 6, characterized in that, The analysis examines the discrepancies between the feeding times of the most recent feedings and the pet's actual eating times. Based on these discrepancies, the baseline recommended feeding times are adjusted to obtain the final recommended feeding times, including: Extract feeding time information corresponding to the most recent feedings from the pet feeding data set, and extract the actual start time information of the pet for each feeding from the image monitoring records of the feeding device. Calculate the time difference between the feeding time information for each feeding and the corresponding time information when the pet actually starts eating, to obtain the time deviation data for multiple feedings; Statistical analysis of the time deviation data is performed to determine the common deviation ranges and trends in the time deviation data, and to determine whether the pet has a habit of eating earlier or later than usual. If the statistical analysis results show that the pet has a habit of eating early, then the time adjustment module is invoked to adjust the basic recommended time in advance according to the early deviation value in the common deviation range; If the statistical analysis results show that the pet has a habit of delaying its eating, then the time adjustment module is invoked to adjust the basic recommended time according to the delay deviation value in the common deviation range. If the statistical analysis results show that the pet has no obvious habitual deviation, then the basic recommended time remains unchanged; During the adjustment process, the time deviation data corresponding to the most recent feeding is referenced. If the most recent time deviation exceeds the boundary value of the common deviation range, the weight adjustment module is called to reduce the weight ratio of the time deviation data in the adjustment calculation. After the adjustment is completed, the adjusted time is determined as the candidate recommended feeding time, and the interval verification module is called to check the rationality of the candidate recommended feeding time, calculate the interval between the candidate recommended feeding time and the most recent feeding time, and determine whether the interval conforms to the feeding interval pattern in the historical time-feeding volume correlation sequence. If the feeding interval does not conform to the feeding interval pattern, return to the weight adjustment module to readjust the deviation weight, and perform the time adjustment operation again until a candidate recommended feeding time that conforms to the feeding interval pattern is obtained; The candidate recommended feeding times that conform to the feeding interval pattern were determined as the final recommended feeding times.

8. The pet adaptive feeding recommendation method based on AI image analysis according to claim 6, characterized in that, The analysis establishes a relationship between the food intake information from the most recent feedings and the actual amount of food remaining in the pet's diet. Based on this relationship, the basic recommended feeding amount is adjusted to ensure that the adjusted feeding amount falls within a stable range of the pet's actual food intake, resulting in the final recommended feeding amount, including: Extract the food intake information corresponding to the most recent feedings from the pet feeding data set, and extract the remaining food information at the time the pet leaves after each feeding from the image monitoring records of the feeding device; Calculate the difference between the amount of food consumed and the corresponding amount of food left over from each feeding to obtain the actual amount of food consumed by the pet at each feeding. Statistical analysis was performed on the actual food intake data to determine the stable range and fluctuation of the pet's actual food intake, and to determine whether the pet has a fixed food intake preference. Based on the stable range obtained from statistical analysis, calculate the average amount of food actually ingested by the pet, and use this average as a reference benchmark for food intake. If the basic recommended feeding amount is compared with the food intake reference benchmark, and the basic recommended feeding amount is higher than the upper limit of the food intake reference benchmark, then the feeding amount adjustment module is invoked to reduce the basic recommended feeding amount to the range of the food intake reference benchmark. If the basic recommended feeding amount is lower than the lower limit of the feeding reference benchmark, the feeding amount adjustment module is invoked to increase the basic recommended feeding amount to the range of the feeding reference benchmark. If the basic recommended feeding amount is within the range of the food intake reference benchmark, the difference between the most recent actual food intake data and the previous K times is further analyzed. If the most recent actual intake exceeds the average fluctuation range of the actual food intake of the previous K times, the feeding amount adjustment module is invoked to appropriately reduce the basic recommended feeding amount; if the most recent actual intake is lower than the average fluctuation range of the actual food intake of the previous K times, the feeding amount adjustment module is invoked to appropriately increase the basic recommended feeding amount, where K is a set positive integer. During the adjustment process, the feed dispensing accuracy information of the feeding equipment is referenced. If there is a deviation in the feed dispensing accuracy, the accuracy correction module is called to make a second correction to the adjusted feeding amount based on the deviation. After the adjustment is completed, the adjusted feeding amount is determined as the candidate recommended feeding amount. The feed intake verification module is invoked to calculate the difference between the candidate recommended feed intake and the most recent actual food intake data. It is then determined whether the difference is within a reasonable fluctuation range. If it is not within a reasonable fluctuation range, the feed intake reference benchmark calculation step is returned to readjust the weight of the feed intake reference benchmark, and the feed intake adjustment operation is performed again until a candidate recommended feed intake that is within a reasonable fluctuation range is obtained. The candidate recommended feed intake that is within a reasonable fluctuation range is determined as the final recommended feed intake.

9. The pet adaptive feeding recommendation method based on AI image analysis according to claim 1, characterized in that, The step of pushing the pet adaptive feeding recommendation information to a mobile terminal associated with the feeding device, allowing the user to view the pet adaptive feeding recommendation information and decide whether to perform the feeding operation based on the pet adaptive feeding recommendation information, includes: The communication establishment module is invoked to establish a communication connection between the feeding device and the associated mobile terminal. After converting the data format of the pet adaptive feeding recommendation information, the converted pet adaptive feeding recommendation information is transmitted to the mobile terminal through the established communication connection. The mobile terminal then displays the recommended feeding time, recommended feeding amount and auxiliary information through the terminal display interface, and provides operation options for the user to select during the information display process. The operation options include the option to perform the feeding operation and the option not to perform the feeding operation. If the user selects to perform the feeding operation, the user's operation command is transmitted to the command receiving module of the feeding device, so that after receiving the operation command, the command receiving module of the feeding device calls the control module to prepare the feeding operation according to the recommended feeding time and recommended feeding amount; If the user chooses not to perform the feeding operation, the user's operation choice is recorded and fed back to the feeding demand prediction model as a reference factor for subsequent fitting analysis; If the user does not make a selection immediately, the mobile terminal is controlled to start timing. After the preset time is reached, an information viewing reminder instruction is generated and the reminder operation is triggered through the terminal display interface or sound prompt. During the above push interaction process, the information transmission time, user viewing time, and user operation time are recorded, and the time information is added to the pet eating data set.

10. A pet adaptive feeding recommendation system based on AI image analysis, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the AI ​​image analysis-based adaptive feeding recommendation method for pets as described in any one of claims 1 to 9 by executing the machine-executable instructions.

Images