Pig feed consumption monitoring management method based on artificial intelligence

Abstract

This application relates to the fields of smart animal husbandry and the Internet of Things (IoT), and discloses an artificial intelligence-based method for monitoring and managing pig feed consumption. The method includes: collecting and synchronizing the operating parameters of the feeder motor with the video stream; extracting the acute feed condition variation variance based on the motor parameters after deducting the long-term mechanical drift baseline; extracting frequency domain features from the video stream to calculate the effective chewing time duty cycle; calculating the ineffective feed loss volume through motion pixel flux integration of a virtual suspended section; reconstructing the net effective consumption by combining the theoretical feed amount and the residual amount; performing cross-logic discrimination based on the aforementioned extracted parameters and outputting a multimodal diagnostic status; issuing negative feedback control parameters to the feeder according to the diagnostic status; synchronizing the effective consumption to the cloud to update the consumption rate model; and triggering a supply chain abnormality interruption command when a feed condition variation is confirmed. This invention achieves the differentiation and identification of pig spillage and waste behavior and physical deterioration of feed, as well as graded closed-loop control.

Description

Technical Field

[0001] This invention relates to the fields of smart animal husbandry and the Internet of Things, specifically to a method for monitoring and managing pig feed consumption based on artificial intelligence. Background Technology

[0002] Automated feeding systems are widely used in large-scale pig farming. Accurate feed consumption is fundamental for assessing pig growth and controlling farming costs. Existing feed consumption monitoring technologies typically rely on feed trough weighing modules or single-view vision sensors, primarily estimating consumption by statistically analyzing the theoretical amount thrown out by the feeding equipment. This one-dimensional monitoring method cannot capture the true conversion of feed intake to feed consumption. When pigs exhibit behaviors such as rooting for feed, some feed is spilled outside the trough. Existing systems include this uningested feed in the normal consumption figure, leading to inaccuracies in the calculation of effective feed consumption.

[0003] When attempting to visually measure spilled feed, existing technologies often use the ground as the monitoring background. Gaps in the slatted floor of the pig activity area and dirt on the ground create background interference, making it difficult for the system to extract the true volume of spilled feed from a two-dimensional image. Besides animal behavior, the physical state of the feed itself directly affects the feeding process. When feed becomes damp, clumps, or deteriorates, the mechanical resistance inside the feed auger pipe changes abruptly. Existing monitoring systems lack multi-source data synchronization and feature decoupling capabilities, making it impossible to distinguish between spillage caused by animal behavior and abnormal feed resistance caused by physical deterioration. Due to the lack of diagnostic dimensions, existing technologies struggle to implement targeted physical interventions and scheduling measures when anomalies occur. When behavioral spillage occurs, the front-end feeder continues to supply feed according to a preset fixed pulse, causing feed to continuously overflow from the trough; when feed clumps and deteriorates in the feed tower, the back-end supply chain management system, unable to detect the physical changes in feed conditions, continues to extrapolate the feed preparation plan based on the original consumption rate and triggers routine replenishment schedules. When a new batch of feed is injected into a problematic feed tower, it will mix and cover with the deteriorated materials, leading to cross-contamination and increased material loss. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based method for monitoring and managing pig feed consumption. This method solves the problems of existing feed monitoring technologies being unable to distinguish between pig spillage and feed clumping and deterioration, leading to deviations in the calculation of effective consumption. Furthermore, it lacks a closed-loop intervention mechanism under abnormal operating conditions, which can easily cause continuous feed overflow in feed troughs or trigger incorrect feeding plans, resulting in cross-contamination.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a method for monitoring and managing pig feed consumption based on artificial intelligence, comprising: Collect multi-source heterogeneous data including the operating parameters of the automated auger feeder motor and the fixed-focus video stream, and perform time-series synchronization processing based on a unified time base; The theoretical material feeding volume for a single feeding cycle is calculated based on the motor operating parameters. The original current fluctuation variance for a single feeding cycle is extracted, and the acute material condition variation variance is extracted based on the long-term mechanical drift baseline decoupling. The effective chewing time duty cycle is calculated by extracting the frequency domain features of the animal chewing action from the fixed-focus video stream, calculating the volume of ineffective feeding loss by integrating the motion pixels of the virtual suspended section, estimating the amount of residual feed in the trough based on the surface coverage of the trough area at the end of the monitoring period, and reconstructing the net effective consumption by combining the theoretical feeding volume, the volume of ineffective feeding loss, and the amount of residual feed in the trough. Based on the numerical states of the acute material condition variation variance, the effective chewing time duty cycle, and the ineffective loss volume of food intake, cross-logic discrimination is performed to output a multimodal diagnostic status; Based on the multimodal diagnostic status, negative feedback control parameters are sent to the motor drive controller, the net effective consumption is synchronized with the cloud ERP management system to update the dynamic consumption rate model, and an abnormal interruption command is triggered in the backend supply chain when an abnormal change in material condition is confirmed.

[0006] This invention independently extracts electromechanical features, visual frequency domain features, and spatial flux features, and uses the synchronous abnormal performance of multidimensional parameters to cross-verify, distinguishing between behavioral spillage and physical deterioration of feed, and executing closed-loop scheduling at different levels.

[0007] Furthermore, the timing synchronization processing based on a unified time reference specifically includes: The global time corresponding to the material feeding action start command is extracted as the starting reference point, and a time reference interval is established by reading the preset time window span; timestamps of arrival time are added to the received high-frequency stator current sampling values ​​and video images; all current sampling values ​​that satisfy the time reference interval are extracted to construct a homogeneous current sequence, and all video frames that satisfy the time reference interval are extracted to construct a homogeneous video sequence.

[0008] The above steps define the time window covering the physical feeding process and the subsequent feeding process, aligning electromechanical data and video streams in the time dimension and establishing the basis for cross-domain feature verification.

[0009] Furthermore, the extraction of the original current fluctuation variance for a single feeding cycle, and the extraction of acute material condition variation variance based on the long-term mechanical drift baseline specifically includes: The original current fluctuation variance for a single feeding cycle is obtained by performing variance calculation on the stator current sample values ​​within the same current sequence; outlier filtering is performed on the variance records within the historical monitoring cycle, and a simple moving average is calculated to construct a long-term mechanical drift baseline; the long-term mechanical drift baseline is subtracted from the original current fluctuation variance and a lower limit constraint is applied to extract the acute material condition variation variance.

[0010] The above extraction process uses historical data to build a baseline, removes the slowly varying current error caused by mechanical aging, and obtains characteristic parameters that reflect sudden changes in the physical resistance of the feed.

[0011] Furthermore, the step of extracting the frequency domain features of the animal's chewing motion from the fixed-focus video stream and calculating the effective chewing time duty cycle specifically includes: Within the video spatial domain of a fixed-focus video stream, a two-dimensional planar polygonal region inside the feeding trough is defined as the first region of interest. Dense optical flow fields between adjacent frames are extracted to construct a one-dimensional temporal motion signal that varies with time. The one-dimensional temporal motion signal is divided into multiple short segments of fixed duration, and a fast Fourier transform is performed on each short segment to generate a power spectral density distribution. The power spectral density integral value within the target frequency band is extracted from the power spectral density distribution. The effective chewing time is obtained by accumulating the lengths of time segments where the power spectral density integral value is greater than the system noise floor threshold. The effective chewing time duty cycle is obtained by calculating the ratio of the effective chewing time to the set time window span.

[0012] The above frequency domain feature extraction process eliminates invalid occupancy actions of animals within the feeding trough area and identifies the length of time animals are in a true feeding state.

[0013] Furthermore, the calculation of the ineffective loss volume of food intake through motion pixel integration of the virtual suspended cross section specifically includes: In the fixed-focus video stream, a virtual cross-section located below the outer edge of the feeding trough and above the manure-suspension board is defined as the second region of interest. Background subtraction is performed on the second region of interest to obtain the foreground moving pixel set. The foreground moving pixel set is then constrained and filtered vertically downwards using optical flow field vector data, retaining pixel clusters that conform to the downward movement characteristics. The coverage area of ​​the pixel clusters at the corresponding time is calculated to obtain the instantaneous drop flux. The instantaneous drop flux is integrated along the time axis within the monitoring period and multiplied by the volume conversion calibration factor to calculate the ineffective feeding loss volume.

[0014] By using suspended sections combined with directional constraints to filter out lateral motion interference, the interference of ground background occlusion and dirt on loss volume calculation is avoided.

[0015] Furthermore, the reconstructing of net effective consumption by combining the theoretical single feeding volume, the ineffective feeding loss volume, and the residual feed amount in the feed trough specifically includes: Calculate the difference between the theoretical feed volume and the ineffective feed loss volume, and apply lower limit protection logic to the difference; When the calculated volume of ineffective consumption exceeds the theoretical volume of a single feeding, the net effective consumption is forcibly reduced to zero.

[0016] By using forced zeroing logic, negative values ​​are avoided in the output, ensuring that the data boundaries of subsequent reconciliation operations do not become abnormal.

[0017] Furthermore, the step of performing cross-logic discrimination based on the numerical states of the acute material condition variance, the effective chewing time duty cycle, and the ineffective loss volume of food intake, and outputting a multimodal diagnostic state specifically includes: When the effective chewing time duty cycle is less than a preset chewing duty cycle threshold, the ineffective food loss volume is greater than a preset loss volume threshold, and the acute food condition variance is less than or equal to a preset upper limit threshold for acute variation parameters, the multimodal diagnostic status is output as behavioral waste mode; when the effective chewing time duty cycle is less than a preset chewing duty cycle threshold, the ineffective food loss volume is greater than a preset loss volume threshold, and the acute food condition variance is greater than a preset upper limit threshold for acute variation parameters, the multimodal diagnostic status is output as food condition variation mode; when the effective chewing time duty cycle is greater than or equal to a preset chewing duty cycle threshold and the ineffective food loss volume is less than or equal to a preset loss volume threshold, the multimodal diagnostic status is output as normal consumption mode.

[0018] By combining the numerical range of cross-domain parameters, different working condition characteristics are mapped to corresponding behavioral performance and material state categories.

[0019] Furthermore, the step of issuing negative feedback control parameters to the material feeder based on the multimodal diagnostic state specifically includes: When the multimodal diagnostic state is in the behavioral waste mode, the front-end negative feedback adjustment logic for the physical equipment is triggered; the amount of wasted feed is converted into the corresponding number of feeding pulses and deducted proportionally in the next feeding; negative feedback control parameters for the next feeding cycle are sent to the motor drive controller of the automated auger feeder to reduce the number of feeding pulses per cycle.

[0020] When wasteful feeding is detected, the subsequent feed quota is reduced at the source of feeding to reduce the accumulation and spillage of feed inside the trough.

[0021] Furthermore, the step of synchronizing the net effective consumption to the cloud-based ERP management system to update the dynamic consumption rate model specifically includes: After a single material supply cycle ends and the multimodal diagnostic status is not in the material condition variation mode and the net effective consumption is greater than zero, the net effective consumption is synchronously reported to the cloud ERP management system; the net effective consumption is used to perform a write-off update on the current actual total inventory of the physical material tower; a dynamic consumption rate model is constructed by aggregating the daily average effective consumption volume of historical sliding statistical days, and the expected material shortage time is deduced based on the dynamic consumption rate model; when the expected material shortage time is less than the replenishment lead time threshold, a regular replenishment schedule is generated.

[0022] By aggregating historical effective consumption data, the timing of feed shortages can be estimated, adapting to the dynamic upward trend of average daily feed intake in pigs and reducing errors in feed preparation time calculations.

[0023] Furthermore, the specific instructions for triggering abnormal interruption of the backend supply chain upon confirmation of material condition anomalies include: When the multimodal diagnostic status is in material condition variation mode, the abnormal isolation scheduling mechanism is triggered; an abnormal interruption command containing the physical address number of the equipment is sent to the cloud ERP management system; the inventory status of the corresponding physical material tower and material batch is updated from available to frozen and downgraded in the system database; the automatically generated regular replenishment schedule is intercepted and suspended at the server process level; and an abnormal investigation and cleanup work order is dispatched.

[0024] When abnormalities in feed conditions are confirmed, the routine logistics scheduling process is intercepted to prevent a new batch of feed from being mixed into the problematic feed tower and causing cross-contamination.

[0025] This invention provides an artificial intelligence-based method for monitoring and managing pig feed consumption. It has the following beneficial effects: 1. This invention synchronizes motor operating parameters with video streams in a timely manner, separates the acute feed condition variation variance after deducting the mechanical aging baseline, and combines it with the effective chewing duty cycle extracted from the video and the volume of ineffective feeding loss for cross-logic discrimination. This multi-dimensional data verification method overcomes the limitation of a single sensor in distinguishing between animal play waste and feed physical deterioration, enabling the output diagnostic status to objectively reflect the correlation between feed pipeline resistance and pig feeding behavior, thus improving the objectivity of the system's judgment results on abnormal feeding conditions.

[0026] 2. This invention defines a suspended cross-section below the edge of the feeding trough within the video frame. Combining this with downward-moving optical flow vector constraints, it performs time-axis flux integration on the pixel area passing through the suspended cross-section to calculate the amount of ineffective feeding loss. This monitoring method based on a spatially hollow region eliminates visual interference caused by the lateral movement of animal limbs, while avoiding volumetric underestimation errors caused by physical obstruction from gaps in the slatted floor and the influence of fecal contamination when relying on ground background subtraction. This allows for the acquisition of a loss baseline that accurately reflects the actual physical scattering process.

[0027] 3. This invention establishes a hierarchical closed-loop control mechanism between local equipment and the cloud system based on diagnostic status output results. When a behavioral waste pattern is identified, the monitoring node sends negative feedback parameters to the motor driver to reduce the number of subsequent feeding pulses, thereby reducing feed overflow from the feed trough. When a feed condition variation pattern is confirmed, the monitoring node reports an interruption command to the cloud to freeze the feed tower inventory status and suspend the regular replenishment schedule. The aforementioned intervention mechanism physically reduces animal spillage behavior at the front-end feeding stage and blocks back-end logistics scheduling in the early stages of feed deterioration, preventing large-scale cross-contamination caused by mixing of newly added feed with deteriorated materials. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram of the multi-source data synchronization steps of the present invention; Figure 4 This is a schematic diagram of the electromechanical feature decoupling steps of the present invention; Figure 5 This is a schematic diagram of the visual flux integration steps of the present invention; Figure 6 This is a schematic diagram of the multidimensional collaborative diagnostic steps of the present invention; Figure 7 This is a schematic diagram of the closed-loop control scheduling steps of the present invention; Figure 8 This is a comparative chart showing the trend of feed spillage and waste in pigs under different monitoring and management modes of the present invention. Among them, 10 is the data synchronization module; 20 is the electromechanical decoupling module; 30 is the visual computing module; 40 is the collaborative diagnosis module; and 50 is the closed-loop control module. Detailed Implementation

[0029] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] See attached document Figure 1 The present invention provides a pig feed consumption monitoring and management system based on artificial intelligence, which may include: a data synchronization module 10, an electromechanical decoupling module 20, a visual computing module 30, a collaborative diagnosis module 40, and a closed-loop control module 50.

[0031] The system's physical hardware structure includes edge computing nodes and sensing devices and remote business systems that communicate with them. The sensing devices consist of automated auger feeders deployed in the pigsty, associated motor drive controllers, and fixed-focus monitoring cameras. Both the motor drive controllers and the fixed-focus monitoring cameras establish communication links with the edge computing nodes. The edge computing nodes establish a network connection with the cloud-based ERP management system to achieve command transmission and data aggregation.

[0032] The data synchronization module 10 runs on an edge computing node and receives motor operation pulse signals and high-frequency stator current sampling values ​​reported by the motor drive controller. It also receives video streams from a fixed-focus monitoring camera covering the trough area and the manure-slatted area. The data synchronization module 10 performs windowing alignment processing on the motor operation data and video stream data according to a set timestamp, generates time synchronization data packets for the same monitoring period, and distributes them to subsequent modules.

[0033] The electromechanical decoupling module 20 receives motor operation data from the time synchronization data packet. Based on the pulse count value in the motor operation pulse signal, the module calculates the theoretical feed rate and extracts the original current fluctuation variance for a single feeding cycle based on the high-frequency stator current sampling value. The module retrieves variance data from historical monitoring cycles to calculate the long-term mechanical drift baseline, subtracts this long-term mechanical drift baseline from the original current fluctuation variance, and calculates and outputs the acute material condition variation variance.

[0034] The visual computing module 30 receives the video stream from the time synchronization data packet. In one optional embodiment, the visual computing module 30 has a built-in trained object detection model or image segmentation model. The model takes fixed-focus video frames or a sequence of continuous video frames as input and outputs pixel masks of the animal head region, the feed trough surface region, and the area of ​​falling feed pellets. Based on the pixel masks, the visual computing module 30 further performs optical flow analysis, frequency domain analysis, and drop flux integration calculation. The visual computing module 30 delineates the interior of the feed trough as a first region of interest in the video spatial domain and delineates the suspended cross section from below the edge of the feed trough to above the manure slat as a second region of interest. The visual computing module 30 extracts the frequency domain peak value of the optical flow signal within the first region of interest and calculates the effective chewing time duty cycle.

[0035] The visual computing module 30 extracts the cluster area of ​​pixels with downward motion vectors that pass through the second region of interest, and calculates the ineffective feeding loss through instantaneous drop flux time integration. The visual computing module 30 also estimates the amount of residual feed in the trough based on the feed surface coverage area, feed texture area, or calibrated pixel area change within the first region of interest at the end of the monitoring period. Combining the theoretical feed supply, ineffective feeding loss, and residual feed in the trough, the visual computing module 30 calculates and outputs the net effective consumption.

[0036] The collaborative diagnostic module 40 receives the acute material condition variation variance output by the electromechanical decoupling module 20, and the effective chewing time duty cycle and ineffective feeding loss output by the visual computing module 30. The collaborative diagnostic module 40 retrieves preset system parameter benchmarks and, through the set conditional discrimination tree logic, outputs diagnostic status indicators including normal consumption patterns, behavioral waste patterns, or material condition variation patterns.

[0037] The closed-loop control module 50 receives the diagnostic status indicator output by the collaborative diagnostic module 40 and the net effective consumption calculated by the visual computing module 30. When the diagnostic status indicator indicates a behavioral waste mode, the closed-loop control module 50 sends negative feedback control parameters to the motor drive controller of the feeder to adjust the number of single feeding pulses. The closed-loop control module 50 synchronizes the net effective consumption to the cloud-based ERP management system to update the daily average inventory consumption rate model. When the diagnostic status indicator indicates a material condition variation mode, the closed-loop control module 50 writes an inventory degradation indicator to the corresponding material batch in the cloud-based ERP management system, triggers a reverse maintenance and cleaning work order, and intercepts the generation of a regular replenishment schedule.

[0038] See attached document Figure 2 The present invention provides an artificial intelligence-based method for monitoring and managing pig feed consumption, which may include: S100 collects multi-source heterogeneous data, including the operating parameters of the automatic auger feeder motor and the fixed-focus video stream of the pig house, and performs time-series synchronization processing based on a unified time reference. S200 calculates the theoretical volume of a single feeding based on motor operating parameters, extracts the original current fluctuation variance of a single feeding cycle, and extracts the acute material condition variation variance based on the historical long-term mechanical drift baseline. S300 extracts the frequency domain features of animal chewing actions from the video stream to calculate the effective chewing time duty cycle, calculates the ineffective feeding loss through motion pixel integration of the virtual suspended section, and estimates the amount of residual feed in the trough based on the feed surface coverage state of the trough area at the end of the monitoring period. Combines the theoretical volume, ineffective feeding loss and residual feed in the trough to reconstruct the net effective consumption. S400 performs cross-logic discrimination based on the numerical status of acute material condition variance, effective chewing time duty cycle, and ineffective feeding loss, and outputs a multimodal diagnostic status for material condition and animal consumption behavior. The S500 sends negative feedback parameters to the material controller to reduce the number of material supply pulses based on the diagnostic status, synchronizes the net effective consumption to the cloud business system to update the consumption rate model, and triggers a back-end supply chain maintenance circuit breaker command when a material condition anomaly is diagnosed.

[0039] This method establishes a global monitoring cycle based on the timestamps set by the underlying computing nodes, forming a multi-source data parallel analysis window.

[0040] After timing alignment, the system independently and in parallel performs feature extraction of electromechanical parameters and visual images. In the electromechanical data processing stage, the system uses moving average logic to eliminate slowly varying current errors caused by mechanical equipment aging and extracts acute physical parameters characterizing sudden changes in feed resistance within the pipeline. In the video processing stage, the system identifies animal swallowing actions in the frequency domain to eliminate invalid occupancy time. Simultaneously, the system calibrates a virtual monitoring section on the fixed-focus camera's imaging plane corresponding to the suspended area below the outer edge of the feed trough and above the slatted floor, using the time integral of the falling pixel clusters as the actual spillage loss, avoiding volume calculation errors caused by ground gaps. The system then subtracts the spillage loss from the theoretical total feed amount to obtain the net effective consumption data.

[0041] The system inputs independently extracted electromechanical features, visual frequency domain features, and spatial flux features into the decision logic tree. The system uses the synchronous or partial abnormal manifestations of the three-dimensional parameters for cross-verification to confirm whether the current cycle's behavior is due to normal feeding, behavioral spillage, or refusal to eat caused by deterioration of the physical properties of the feed.

[0042] Based on the specific diagnostic status, a hierarchical closed-loop scheduling mechanism is executed between local devices and the remote business platform. The local device directly modifies the driver's pulse output setting to limit the single feeding base. The remote resource management system receives the net effective consumption to extrapolate the actual inventory level, and suspends the regular logistics scheduling process and generates corresponding cleaning and inspection tasks upon receiving information about material deterioration.

[0043] See attached document Figure 3 The data synchronization module 10 runs within the edge computing node at the system front end, and the specific process for performing multi-source data synchronization steps is as follows: S101, in specific system operation scenarios, there is an objective time delay between the issuance of equipment control commands and the execution of physical actions, thus requiring the establishment of a unified monitoring benchmark. The data synchronization module 10 establishes communication links with both the motor drive controller and the fixed-focus monitoring camera to access multi-source heterogeneous data. The edge computing node periodically polls the address register inside the motor drive controller via an industrial field communication bus interface (e.g., RS485 bus or Ethernet industrial bus) to obtain the accumulated pulse count value of the automated auger feeder during operation. The edge computing node obtains the instantaneous stator current value sampled by the underlying hardware circuit of the motor drive controller within the control cycle, forming a high-frequency stator current sampling sequence. The edge computing node obtains a fixed-focus video stream containing continuous video frames pushed by the fixed-focus monitoring camera via a network interface. The field of view of the fixed-focus video stream covers the physical hardware feeding trough inside the pigsty and the surrounding slatted floor area. For the parsing of network streaming media communication protocols and the decapsulation and restoration of video frame data, those skilled in the art can use existing general network transmission protocols and video encoding and decoding standards. The underlying communication and decoding principles are well-known technologies in this field and will not be elaborated here.

[0044] S102, the data synchronization module 10 uses the motherboard system clock of the edge computing node as a unified time reference to perform hardware interrupt-based timestamp mapping and time window truncation processing on the received motor operating parameter sequence and fixed-focus video stream. Since there is a time lag between the physical resistance change caused by the feeding action and the visual action of the pigs eating, truncation of the synchronization data interval covering this causal chain is the data basis for subsequent cross-domain feature verification.

[0045] The edge computing node kernel responds to the material feeding control level signal issued by the control system, generates a system-level hardware interrupt request, and extracts the operating system global time t at the time of the interrupt. The data synchronization module 10 records the global time corresponding to the material feeding action start command as the starting reference point for this monitoring task. The data synchronization module 10 reads the preset time window span of a single feeding monitoring action. This establishes a time base interval. Time window span Covering both physical feeding time and the time spent on subsequent feeding by the animals, preferably, the time window spans [a certain range]. The value can be determined by historical data statistics or on-site calibration based on the preset total number of operating pulses of the feeding motor and the average feeding time corresponding to the standard quota for a single feeding. It can generally be set to several minutes to tens of minutes.

[0046] The data synchronization module 10 receives each high-frequency stator current sample value from the communication port. A marker variable indicating the arrival time of the system is appended to each frame of fixed-focus video image F. The marker timestamp is defined as follows: The current sample value is denoted as Mark timestamp as Video frame record To prevent audio-visual asynchrony caused by network transmission delays from leading to algorithmic logic deviations, timestamps are used to mark video frames. The system can prioritize extracting the display timestamp (PTS) provided by the video streaming media transmission protocol, and then compensate by calling the system arrival time when the transmission protocol does not provide it. The data synchronization module 10 filters the data queue in the edge computing node buffer area, and can remove or archive data elements whose timestamps are not within the current time base interval.

[0047] Data synchronization module 10 extracts data that meets time boundary conditions. All current sample values ​​are used to construct a sequence of currents originating from the same source in chronological order. Data synchronization module 10 extracts data that satisfy time boundary conditions. All video frames are sequentially used to construct a homogeneous video sequence. The data synchronization module 10 integrates the accumulated pulse counts, homogeneous current sequences, and homogeneous video sequences within the same time reference interval to construct a synchronization data packet for a single monitoring cycle. The homogeneous video sequence covers the entire time window span. In subsequent electromechanical decoupling processing, the same current sequence is further used to extract the effective operating range of the motor based on the motor enable signal, pulse output flag, or current amplitude threshold, in order to avoid the standby current or zero-value sampling after the motor stops participating in the calculation of the feeding resistance variance. The data synchronization module 10 writes the generated synchronization data packet to the operating system's shared memory area for concurrent reading by other processing modules.

[0048] See attached document Figure 4 The electromechanical decoupling module 20 reads the synchronization data packet from the shared memory area and performs the electromechanical feature decoupling steps in the following specific process: S201, the electromechanical decoupling module 20 receives the synchronized motor operation data and calculates the theoretical feeding volume corresponding to the current feeding command. For physical feeding systems using stepper or servo feedback mechanisms, the motor drives the auger to rotate and push feed; the number of auger rotations is positively correlated with the mechanical discharge volume. The electromechanical decoupling module 20 extracts the total number of valid motor execution pulses corresponding to the accumulated pulse count value in the synchronization data packet and converts it into the theoretical feeding volume. The theoretical feeding volume per cycle. The specific calculation logic is as follows: ; In the formula, N represents the theoretical feed volume per cycle; N is the total number of effective pulses executed by the motor. This is the calibrated feed volume coefficient corresponding to a single pulse. In practical application scenarios, it is related to the mechanical structure parameters of the equipment. As an feasible method, it can be obtained through on-site calibration. That is, during the equipment debugging phase, the actual total weight of the discharged material under a fixed number of pulses is collected and converted by combining it with the volume density of the current batch of feed. The physical calibration process is a routine equipment initialization operation in this field and will not be described in detail here.

[0049] S202, the electromechanical decoupling module 20, based on the same current sequence in the synchronization data packet, first determines the effective operating range of the motor according to the motor enable signal, pulse output flag, or continuous sampling interval where the current is greater than the standby threshold. It then extracts only the stator current sampling value within this effective operating range and calculates and extracts the original current fluctuation variance for a single feeding cycle. From the perspective of the motor's electrical operating principle, instantaneous changes in the motor output shaft load directly map to current amplitude fluctuations in the stator windings. In the working environment, when feed is pushed and conveyed within the auger pipe, if moisture and clumping occur, the friction coefficient and pushing resistance of the feed on the inner wall of the pipe will undergo high-frequency jumps. This change in mechanical resistance will cause high-frequency fluctuations in the motor stator current; therefore, variance statistics are used to characterize the physical characteristics of this resistance change. The process of variance calculation for the same current sequence by the electromechanical decoupling module 20 is as follows: ; ; In the formula, The index of the sequence sampling point is the independent variable in the formula. "To M" means taking the first to the Mth data points from the same source current sequence for calculation in sequence; The first in the same source current sequence One stator current sample value; M is the total number of sampling points contained in the same source current sequence, specifically the total number of stator current sampling points in the effective operating range of the motor. Its value depends on the hardware sampling frequency of the motor driver and the effective operating time intercepted. Usually, the hardware sampling frequency is set to the range of 50Hz to 200Hz to capture effective high-frequency physical resistance jumps. This is the average stator current for this feeding cycle; The corresponding original current fluctuation variance.

[0050] S203, the electromechanical decoupling module 20 establishes a time-scale historical sliding window to separate and calculate the long-term mechanical drift baseline and the variance of acute material condition variations. In actual operation, mechanical losses such as aging and wear of the auger bearing or lack of lubrication in the reducer can also cause a slow increase in the motor's operating baseline load. If the aforementioned original current fluctuation variance is directly used to assess the material state, the slow-changing interference from mechanical aging poses a risk of triggering logical misjudgments. Therefore, the electromechanical decoupling module 20 retrieves variance records containing the past K effective feeding actions from the system's historical database. Monitoring periods identified by the collaborative diagnosis module 40 as behavioral waste patterns or material condition variation patterns are not used as updated samples for the long-term mechanical drift baseline; instead, only variance records under normal consumption patterns are selected for the baseline sample set. To ensure the purity of the baseline data and avoid algorithm failure, the electromechanical decoupling module 20 performs outlier filtering on the retrieved variance records, removing extreme values ​​exceeding three standard deviations of the data distribution. Then, it constructs the long-term mechanical drift baseline by calculating its simple moving average. : ; In the formula, This represents the long-term mechanical drift baseline constructed by calculation; m represents the index variable of the historical sample sequence; K is the set number of historical valid feeding action samples, the value of which can be set according to the regular maintenance cycle of the farm's mechanical equipment; This represents the original current fluctuation variance after extreme value filtering for the m-th historical feeding cycle. To prevent the algorithm from malfunctioning due to a lack of historical data when the system is first launched, when the number of historical samples stored in the system is less than K, the electromechanical decoupling module 20 retrieves the calibration variance constant written by default at the factory to replace this long-term mechanical drift baseline.

[0051] Subsequently, the electromechanical decoupling module 20 subtracts the aforementioned long-term mechanical drift baseline from the original current fluctuation variance and calculates and extracts the variance of acute feed condition variation caused only by sudden changes in feed physical properties. : ; In the formula, Indicates the variance of acute material conditions; This represents the corresponding original current fluctuation variance; This represents the computationally constructed long-term mechanical drift baseline.

[0052] Considering that a temporary improvement in mechanical lubrication conditions may cause the resistance to fall below the historical baseline, to ensure the physical and operational significance of this characteristic parameter, the electromechanical decoupling module 20 applies a lower limit constraint to the above calculation results. When the calculated... When the value is less than 0, force to A value of 0 indicates that there is no abnormality in the current feeding action that increases resistance. Based on this time-scale decoupling and limiting logic, the electromechanical decoupling module 20 isolates low-frequency errors caused by physical aging factors of the equipment and outputs abrupt change characteristic quantity reflecting abnormal resistance of the feed.

[0053] See attached document Figure 5 The visual computing module 30 reads synchronization data packets containing the same video sequence from the shared memory area of ​​the edge computing node, and performs the visual flux integration in the following specific process: S301, the visual computing module 30 performs frequency domain transformation using a source video sequence to reconstruct the effective chewing time. Based on animal biological characteristics, their heads exhibit periodic undulating movements within a frequency range during feeding. The system uses this dynamic feature to characterize the actual feeding process, filtering out space-occupying behaviors that do not involve actual eating. Within the video spatial domain, the visual computing module 30 delineates a two-dimensional planar polygonal region inside the feeding trough as the first region of interest (ROI) based on preset coordinate parameters. These preset coordinate parameters can be generated manually by technicians during system initialization by selecting the trough edge coordinates from the monitoring screen, or automatically extracted using an image edge detection operator. For a continuous video frame sequence within the RIO, the visual computing module 30 extracts the dense optical flow field between adjacent frames, calculates the mean amplitude and sum of the optical flow motion vector amplitudes of all pixels within the RIO, or extracts local optical flow energy from candidate areas of the animal's head and then weights and sums it to construct a one-dimensional temporal motion signal that varies over time. For the specific extraction and calculation process of dense optical flow fields, those skilled in the art can use existing basic machine vision algorithms such as the Horn-Schunck algorithm or the Lucas-Kanade algorithm, which are well-known technologies in this field and will not be elaborated here.

[0054] To extract biological frequency band features, the visual computing module 30 introduces a short-time window segmentation mechanism, dividing the one-dimensional time-domain motion signal into multiple short-time segments of fixed duration (e.g., 2 seconds). A Fast Fourier Transform is applied to each short-time segment to map it to the frequency domain, generating a power spectral density distribution. The frequency of typical chewing or feeding head movements in pigs is usually concentrated in the 1.5Hz to 3.0Hz range. The visual computing module 30 extracts the power spectral density integral value within this target frequency band. The visual computing module 30 compares the power spectral density integral value of each short-time segment with a preset system noise threshold. This system noise threshold is determined by calibrating and confirming the average power spectral density integral value of background videos of the same duration collected in an empty pen, with an empirical bias constant added. When the integral value is greater than the noise threshold, it is determined that a real feeding action exists within the corresponding time segment. The visual computing module 30 accumulates the length of time segments that meet the noise threshold condition to obtain the effective chewing time within the monitoring period. The visual computing module 30 combines the time window span set by the system. Calculate the duty cycle of effective chewing time. : ; In the formula, Indicates the duty cycle of effective chewing time; Indicates the effective chewing time within the monitoring period; This indicates the time window span set by the system.

[0055] If the energy integral of the target frequency band does not trigger the noise floor threshold throughout the entire monitoring period, then A value of 0 indicates that no substantial feeding occurred during this period.

[0056] S302, the visual computing module 30 constructs a dynamic drop flux integral model of the hollow cross-section to calculate the ineffective loss from feeding. In the physical environment of the breeding site, the slatted floor has geometric gaps, and the ground is covered with dynamically increasing manure and urine. Relying on the conventional ground background difference method to extract the spillage area is easily affected by both physical masking and false positives from dirt. To address the above measurement bias, the visual computing module 30 defines a virtual cross-section in the video image located below the outer edge of the feed trough and above the slatted floor as the second region of interest, thereby capturing the spatial dynamic characteristics of the feed before it falls into the gaps in the ground. The second region of interest can be calibrated during the system initialization stage based on the camera installation height, installation angle, feed trough edge position, and the upper surface position of the slatted floor. When the camera has perspective distortion, the visual computing module 30 can first perform perspective correction on the video frame, and then determine the second region of interest in the corrected image coordinate system.

[0057] Time window span During the period, the visual computing module 30 performs background subtraction on the second region of interest to obtain a set of foreground moving pixels, and performs directional constraint filtering on the foreground moving pixel set in combination with optical flow field vector data. An angular tolerance range is set based on the vertically downward direction; preferably, the allowed directional deviation angle is limited to ±15 degrees. The vertically downward direction is the image direction corresponding to the direction of gravity after camera installation posture calibration or perspective correction, to ensure that the optical flow direction filtering result is consistent with the actual feed falling direction. The visual computing module 30 removes pixels that do not conform to the above downward motion vector constraint, retaining only pixel clusters that conform to the downward motion characteristics, thereby eliminating interference caused by the animal's limbs moving laterally across the area. If the foreground moving pixel set extracted in the current frame is empty or does not conform to the directional constraint, the area of ​​the pixel cluster retained at the corresponding time is recorded as 0. The visual computing module 30 calculates the coverage area of ​​the retained pixel clusters at time t, which is defined as the instantaneous fall flux. .

[0058] The instantaneous drop flux is calculated by integrating along the time axis over the entire monitoring period, and a volume conversion calibration factor is introduced. The visual computing module 30 calculates the volume of ineffective physical spillage loss due to food not being blocked by the ground. : ; In the formula, This represents the calculated volume of ineffective physical spillage during feeding; This represents the volume conversion calibration conversion factor; This indicates the time window span set by the system; Indicates instantaneous drop in flux; The time differential element represents the integral performed within an extremely short time infinitesimal; the calibration conversion coefficients are used. This is the mapping weight for converting pixel area integral values ​​to physical volume. As a specific implementation, this coefficient is determined by linear fitting calculations using a known volume of pre-shaped feed pellets scattered in a controlled environment and statistically analyzing the system's output flux integral results. The calibration process is preferably performed under conditions similar to or the same as those in an actual pigsty, including camera installation location, lighting conditions, feed pellet size, and feed color, to reduce scene deviation during the conversion of pixel area integral values ​​to physical volume.

[0059] S303, the visual computing module 30 reconstructs the net effective consumption based on the theoretical calculation data output from the previous steps. The system restores the actual intake of individual animals by eliminating spatial spillage losses and the amount of uneaten residual feed in the trough. The visual computing module 30 retrieves the theoretical single feeding volume calculated by the electromechanical decoupling module 20. Combined with the ineffective loss volume of food intake calculated in this step The amount of residual feed in the trough is estimated based on the feed surface coverage area, feed color and texture area, or calibrated pixel volume mapping relationship in the first region of interest at the end of the monitoring period, and the net effective consumption of a single feeding is calculated. Due to uncontrollable visual noise interference such as changes in ambient lighting or flying insects, the flux integral may shift upwards. To prevent negative output results from causing data anomalies in subsequent inventory accumulation models, the visual computing module 30 introduces a lower limit protection logic, the calculation expression of which is: ; In the formula, This represents the calculated net effective consumption per feeding. This represents the maximum value function in mathematics. This indicates that the material volume exceeds the theoretical feed volume; This indicates the volume lost due to ineffective food intake.

[0060] When the calculated ineffective loss volume of food intake Exceeding the theoretical feed volume At that time, the net effective consumption will be forced to be Zeroing out ensures the logical integrity of the underlying business data loop.

[0061] See attached document Figure 6 The collaborative diagnosis module 40 receives the acute material condition variation variance output by the electromechanical decoupling module 20, and the effective chewing time duty cycle and ineffective feeding loss output by the visual computing module 30. The specific process of performing multidimensional collaborative diagnosis is as follows: S401, the collaborative diagnostic module 40 acquires and loads preset system status discrimination benchmark parameters. Considering the physical phenomena of pig feeding and the electrical characteristics of equipment operation, data from a single sensor is prone to diagnostic bias. For example, simple feed waste might be due to playful behavior by individual animals rather than a physical problem with the feed itself. Therefore, it is necessary to introduce multi-dimensional judgment benchmarks to establish a discrimination tree. The collaborative diagnostic module 40 reads multi-dimensional benchmark thresholds, including chewing duty cycle thresholds, during the system initialization or equipment calibration phase. Loss volume threshold Upper limit threshold of acute variant parameters And the feed trough residue threshold. These thresholds are determined based on the physical properties of the aquaculture environment and historical data. Specifically, the chewing duty cycle threshold... Based on the statistical lower limit of the time required for the theoretical quota volume of feed intake in normal-aged pigs, due to the duty cycle of effective chewing time... for and The ratio, therefore The threshold value is dimensionless and typically ranges from 0.3 to 0.5; the loss volume threshold... The setting is based on the maximum reasonable redundancy allowed by the physical equipment on site, for example, set as the theoretical volume of material to be discharged in a single operation. 5%; the trough residue threshold is based on the theoretical single feeding volume. The allowable residue ratio in feed troughs and the upper limit for residue statistics after normal feed intake by pigs of the same age; upper limit threshold for acute variation parameters. The upper limit of the confidence interval for variance fluctuation was used to calibrate the equipment by collecting data from no-load and standard load tests on similar motor equipment using normal and dry batches of feed, in order to define the tolerance boundary for normal operation of the equipment.

[0062] S402, the collaborative diagnosis module 40 establishes a single-dimensional behavioral waste diagnosis logic based on cross-domain parameter states. During the daily feeding process of pigs, competition for space or playful rooting actions between individuals can cause physical spillage, but the physical state of the feed itself does not deteriorate in such cases. The collaborative diagnosis module 40 performs a logical AND operation on the input visual and electromechanical parameters. When the conditions are met, the effective chewing time duty cycle is determined. And the amount of food lost ineffectively And acute material condition variation At that time, the collaborative diagnosis module 40 determines that the current pipeline transport resistance is within the normal baseline fluctuation range, the physical transport of materials is smooth, and the spillage loss is mainly due to behavioral behaviors such as insufficient feeding time and excessive play by individual animals. The system outputs a corresponding diagnostic status indicator. =1, defined as a behavioral waste pattern.

[0063] S403, the collaborative diagnostic module 40 constructs a joint diagnostic logic for material condition variations based on cross-domain collaboration of mechanical and visual perspectives to diagnose multimodal physical concurrent events. Feed becoming damp and clumping or formula deterioration can simultaneously cause increased pipeline propulsion resistance and animal refusal to eat or scattering behavior due to decreased palatability. The collaborative diagnostic module 40 captures this multimodal concurrent event by jointly discriminating electromechanical and visual features. When the conditions meet the effective chewing time duty cycle... And the amount of food lost ineffectively Simultaneously satisfying the variance of acute material condition variation At that time, the collaborative diagnosis module 40 confirmed the existence of a cross-domain correlation anomaly. The system determined that the current material state had undergone physical deterioration, causing the feeding motor to encounter a sudden change in high-frequency resistance, and the animal exhibited behaviors such as refusing to swallow and overturning and throwing. The collaborative diagnosis module 40 output the corresponding diagnostic status indicator. =2, defined as the material condition variation mode.

[0064] To ensure the completeness of the discrimination logic tree and avoid algorithm dead zones, the collaborative diagnosis module 40 performs fallback judgment processing on other parameter combination boundary states that do not fall into the two types of abnormal branches mentioned above. When the conditions are met... and At this time, the system determines it to be normal healthy feeding. For other parameter combinations, such as a low effective chewing time duty cycle but no actual spillage loss, or a chewing time that meets the standard but there is a certain degree of overflow, since there is no excessive waste of physical materials or the equipment resistance does not show synchronous abnormalities, the front-end interception or back-end circuit breaker mechanism is not triggered. The collaborative diagnosis module 40 classifies such combinations of states that do not meet the triggering degradation and interception conditions into the normal consumption mode, and outputs the corresponding diagnostic status indicator. =0. Through this logic tree structure, the collaborative diagnosis module 40 realizes the mapping processing of different equipment operating conditions and animal behavior.

[0065] See attached document Figure 7 The closed-loop control module 50 receives the diagnostic status identifier output by the collaborative diagnostic module 40 and the net effective consumption calculated by the visual computing module 30. Combining the network link between the edge computing node and the cloud ERP management system, the specific process of executing closed-loop control scheduling is as follows: S501, the closed-loop control module 50 triggers the front-end negative feedback adjustment logic for the local physical device based on the diagnostic status indicator value. When the received diagnostic status indicator value... When the value is 1, the system determines that the current state is a behavioral waste mode, meaning that the feed has not physically deteriorated but the animals are exhibiting excessive spillage. By converting the wasted feed amount into a corresponding number of feeding pulses and deducting it proportionally in the next feeding cycle, the risk of subsequent feed accumulation and continuous spillage can be reduced, thereby reducing the continuous accumulation and spillage of feed at the physical source. In a specific embodiment of the present invention, the closed-loop control module 50 sends negative feedback control parameters for the next feeding cycle to the motor drive controller of the automated auger feeder, reducing the number of feeding pulses per cycle. The specific formula for calculating the reduction in the number of pulses is as follows: ; In the formula, The target total number of valid pulses set for the next cycle; This represents the maximum value function in mathematics; N is the total number of valid pulses executed in the current cycle; Indicates the volume of food lost due to ineffective feeding; This is the calibrated feed volume coefficient corresponding to a single pulse; This represents the rounding down operation; This establishes a minimum guaranteed pulse threshold to maintain the animals' basic physiological needs. As an optional configuration, this threshold is configured and confirmed by technicians based on the daily minimum standard for pig age and weight. Through this negative feedback logic, the system directly deducts the corresponding feed allocation when it detects wasteful behavior.

[0066] S502, the closed-loop control module 50 performs consumption rate prediction and business scheduling based on the inventory decline rule. At the end of each supply cycle, when the diagnostic status indicator is not in the material condition variation mode and the net effective consumption is... If the value is greater than 0, the closed-loop control module 50 will transmit the net effective consumption via network communication protocol. The data is simultaneously reported to the cloud-based ERP management system for routine inventory write-offs and consumption rate updates. When the diagnostic status indicator is in material condition variation mode, the closed-loop control module 50 can still report the net effective consumption. Abnormal data is synchronized to the cloud-based ERP management system, but this data is recorded as an abnormal periodic record and does not trigger regular replenishment scheduling. The cloud-based ERP management system uses this data to determine the current actual total inventory of the physical material silos. Perform write-off update: ; In the formula, This is the total inventory at the end of the previous calculation cycle. This value is taken from the full-load calibration weight of the silo during system initialization. The known physical bulk density of the current batch of feed; Indicates the total actual inventory; This represents the calculated net effective consumption per feeding.

[0067] Considering that the average daily feed intake of pigs dynamically increases with age, using a fixed-baseline consumption model can easily lead to deviations in feed preparation time. Therefore, the cloud-based ERP management system aggregates the average daily effective consumption volume over the past D days to construct a dynamic consumption rate model. Based on this rate, the expected material shortage time can be estimated. : ; In the formula, The known physical bulk density of the current batch of feed; This indicates the predicted time of material shortage. This represents the dynamic consumption rate model of the system construction; This represents the actual total inventory; D is the sliding statistical days, typically ranging from 3 to 7 days based on the fluctuation characteristics of the typical feeding cycle in farms. The cloud-based ERP management system will calculate the expected feed shortage time. The feed supply lead time is compared with a preset lead time threshold. Understandably, this preset lead time threshold can be determined by combining the feed mill's stock preparation response time with the time required for logistics and transportation, for example, set to 2 to 3 days. When the expected feed shortage time is less than the lead time threshold, the cloud-based ERP management system automatically generates a regular feed supply schedule based on standard business processes and coordinates backend logistics and transportation.

[0068] S503, the closed-loop control module 50 triggers an anomaly isolation scheduling mechanism based on multi-dimensional diagnostic results. When the received diagnostic status identifier... When the value is 2, the system determines that the current equipment has experienced a material condition change caused by feed becoming damp and clumping or physical deterioration. If the replenishment action is triggered according to the normal business logic at this time, the newly issued healthy feed will be added to and cover the feed towers or pipelines containing deteriorated feed, which can easily cause large-scale cross-contamination of materials and exacerbate the risk of mechanical blockage.

[0069] To disrupt this business chain, the closed-loop control module 50 sends an abnormal interruption command containing the physical address number of the equipment to the cloud-based ERP management system. This abnormal interruption command has a higher priority than regular inventory write-offs, consumption rate predictions, and replenishment schedule generation processes. Upon receiving the abnormal interruption command, the cloud-based ERP management system prioritizes freezing inventory, suspending schedules, and generating maintenance work orders. After receiving the command, the cloud-based ERP management system updates the inventory status of the corresponding physical material tower and material batch in its system database from available to frozen / downgraded. Simultaneously, at the server process level, the cloud-based ERP management system intercepts and suspends the regular replenishment schedule automatically generated in step S502, blocking front-end replenishment scheduling. Based on the bound local maintenance personnel information, the cloud-based ERP management system dispatches anomaly investigation and cleanup work orders containing spatial coordinates and equipment numbers. Through this cross-system collaborative action, the system achieves state isolation and manual intervention scheduling in the early stages of physical condition deterioration in the logistics supply chain, preventing the physical spread of deteriorated materials.

[0070] Specific application examples: The initial operating parameters for setting a specific pig pen in a breeding site include: a time window span of 600 seconds, a total number of effective pulses for a single standard feeding of 500, a single pulse calibration feed volume coefficient of 5 ml / pulse, a single theoretical feed volume of 2500 ml, a system minimum pulse threshold of 200, and preset system diagnostic thresholds including a chewing duty cycle threshold of 0.3, a loss volume threshold of 5% of the theoretical feed volume (125 ml), and an upper limit threshold for acute variation parameters of 0.5 ampere-squared.

[0071] During the first monitoring cycle, the visual module counted an effective chewing time of 270 seconds with an effective chewing time duty cycle of 0.45. The calculated ineffective loss volume was 45 ml. The electromechanical decoupling module extracted an acute material condition variance of 0.12 ampere squares. At this point, the duty cycle was greater than 0.3, the loss volume was less than 125 ml, and the variance was less than 0.5 ampere squares. The system output a diagnostic status indicator of 0. The visual computing module reconstructed the net effective consumption of 2450 ml and reported it to the cloud system to deduct inventory.

[0072] During the second monitoring cycle, the effective chewing time was 120 seconds, the effective chewing time duty cycle was 0.2, the volume of ineffective food loss was 380 ml, and the variance of acute food condition extracted by the electromechanical module was 0.18 ampere squares. At this time, the loss volume was greater than 125 ml and the variance was less than 0.5 ampere squares. The system output diagnostic status indicator was 1, the vision module calculated the net effective consumption to be 2120 ml, the closed-loop control module triggered negative feedback adjustment, the system calculated that the number of pulses to be deducted in the next cycle was 76, and the target number of execution pulses in the next cycle was reduced to 424.

[0073] During the third monitoring period, the effective chewing time was 45 seconds, the effective chewing time duty cycle was 0.075, the volume of ineffective feed loss was 210 ml, and the acute feed condition variance extracted by the electromechanical module was 1.45 ampere squares. At this time, the loss volume was greater than 125 ml and the acute feed condition variance was greater than 0.5 ampere squares. The system output a diagnostic status identifier of 2, the closed-loop control module sent an abnormal interruption command to the cloud, the cloud system marked the status of this batch of feed as frozen, the system suspended the feeding schedule and dispatched a pipeline cleaning and maintenance work order.

[0074] Experimental verification and effect comparison: See attached document Figure 8 ,Should Figure 8 The horizontal axis represents the experimental period, measured in days. Figure 8 The vertical axis represents the average daily amount of feed scattered and collected, measured in kilograms. Figure 8 The gray broken line with a circular mark represents the control group data, and the black solid line with a square mark represents the experimental group data.

[0075] Two independent pig houses with the same environmental layout and ventilation and lighting conditions were selected as the control group and the experimental group, respectively. Fifty Landrace pigs with an average age of 60 days and similar initial weight were introduced into each pig house. Both pig houses were equipped with the same model of automatic auger feeder and stainless steel feed trough.

[0076] The control group adopted a fixed timing and quantitative pulse strategy, while the experimental group deployed an artificial intelligence-based multidimensional collaborative diagnosis and closed-loop control scheduling system. The experiment lasted for 30 days. During the experiment, the actual total amount of feed and the actual amount of feed spilled were recorded daily. On the 15th day of the experiment, 5 kg of damp and clumped feed was mixed into the front end of the main feed pipeline of each of the two pens. After the experiment, the pigs in the two pens were weighed uniformly to obtain the average daily weight gain data.

[0077] Table 1: Data Record of Feed Consumption and Waste for Discrete Days Between Control and Experimental Groups

[0078] Table 2: Summary Table of Core Performance Indicators for 30 Days

[0079] From Table 1, Table 2 and appendix Figure 8 We can obtain: The total cumulative feeding amount in the control group was 4872.6 kg and the total cumulative spillage was 451.3 kg. The total cumulative feeding amount in the experimental group was 4415.3 kg and the total cumulative spillage was 68.4 kg. Compared with the control group, the total feeding amount in the experimental group was reduced by 9.38%, and the total spillage was reduced by 84.85%. After the introduction of damp and clumped feed on the 15th day, the response time in the control group was 14.5 hours, and the response time in the experimental group was 12 minutes. The final average daily weight gain of pigs in the two groups was 812.4 g and 808.7 g, respectively. The system of this invention reduced the amount of waste from physical spillage and located the abnormal feed conditions and blocked the pipeline operation.

Claims

1. A method for monitoring and managing pig feed consumption based on artificial intelligence, characterized in that, Includes the following steps: Collect multi-source heterogeneous data including the operating parameters of the automated auger feeder motor and the fixed-focus video stream, and perform time-series synchronization processing based on a unified time base; The theoretical material feeding volume for a single feeding cycle is calculated based on the motor operating parameters. The original current fluctuation variance for a single feeding cycle is extracted, and the acute material condition variation variance is extracted based on the long-term mechanical drift baseline decoupling. The effective chewing time duty cycle is calculated by extracting the frequency domain features of the animal chewing action from the fixed-focus video stream, calculating the volume of ineffective feeding loss by integrating the motion pixels of the virtual suspended section, estimating the amount of residual feed in the trough based on the surface coverage of the trough area at the end of the monitoring period, and reconstructing the net effective consumption by combining the theoretical feeding volume, the volume of ineffective feeding loss, and the amount of residual feed in the trough. Based on the numerical states of the acute material condition variation variance, the effective chewing time duty cycle, and the ineffective loss volume of food intake, cross-logic discrimination is performed to output a multimodal diagnostic status; Based on the multimodal diagnostic status, negative feedback control parameters are sent to the motor drive controller, the net effective consumption is synchronized with the cloud ERP management system to update the dynamic consumption rate model, and an abnormal interruption command is triggered in the backend supply chain when an abnormal change in material condition is confirmed.

2. The method for monitoring and managing pig feed consumption based on artificial intelligence according to claim 1, characterized in that, The timing synchronization processing based on a unified time reference specifically includes: The global time corresponding to the material feeding action start command is used as the starting reference point, and the preset time window span is read to establish the time reference interval. Add a timestamp to the received high-frequency stator current sample value and video image to indicate the time of arrival at the system; Extract all current sample values ​​that satisfy the time reference interval to construct a common current sequence, and extract all video frames that satisfy the time reference interval to construct a common video sequence.

3. The method for monitoring and managing pig feed consumption based on artificial intelligence according to claim 2, characterized in that, The extraction of the original current fluctuation variance for a single feeding cycle, and the extraction of acute material condition variation variance based on the long-term mechanical drift baseline, specifically includes: The variance of the original current fluctuation in a single feeding cycle is obtained by performing variance calculation on the stator current sample values ​​within the same current sequence. Retrieve variance records from historical monitoring periods, perform outlier filtering, and calculate a simple moving average to construct a long-term mechanical drift baseline; The acute material condition variation variance is extracted by subtracting the long-term mechanical drift baseline from the original current fluctuation variance and applying a lower limit constraint.

4. The method for monitoring and managing pig feed consumption based on artificial intelligence according to claim 1, characterized in that, The step of extracting the frequency domain features of the animal chewing action from the fixed-focus video stream and calculating the effective chewing time duty cycle specifically includes: Within the video spatial domain of a fixed-focus video stream, a two-dimensional planar polygonal region inside the trough is defined as the first region of interest. Dense optical flow fields between adjacent frames are extracted to construct a one-dimensional temporal motion signal that varies with time. The one-dimensional time-domain motion signal is divided into multiple short-time segments of fixed duration, and a power spectral density distribution is generated by performing a fast Fourier transform on each short-time segment. Extract the power spectral density integral value within the target frequency band from the power spectral density distribution, and accumulate the lengths of time segments where the power spectral density integral value is greater than the system noise floor threshold to obtain the effective chewing time. The effective chewing time duty cycle is obtained by calculating the ratio of the effective chewing time to the set time window span.

5. The method for monitoring and managing pig feed consumption based on artificial intelligence according to claim 1, characterized in that, The calculation of the ineffective loss volume of food intake through motion pixel integration of a virtual suspended cross section specifically includes: In the video frame of the fixed-focus video stream, a virtual cross-section located in the suspended area below the outer edge of the feeding trough and above the manure slat is defined as the second region of interest. Background subtraction is performed on the second region of interest to obtain the foreground moving pixel set. The foreground moving pixel set is then subjected to vertical downward constraint filtering in combination with optical flow field vector data, retaining pixel clusters that conform to the downward motion characteristics. The instantaneous drop flux is obtained by statistically analyzing the coverage area of ​​the pixel cluster at the corresponding time. The instantaneous drop flux during the monitoring period is integrated along the time axis and multiplied by the volume conversion calibration factor to calculate the ineffective loss volume of food intake.

6. The method for monitoring and managing pig feed consumption based on artificial intelligence according to claim 1, characterized in that, The specific steps for reconstructing the net effective consumption by combining the theoretical feed volume per feeding cycle, the ineffective feed loss volume, and the residual feed amount in the trough include: Calculate the difference between the theoretical feed volume and the ineffective feed loss volume, and apply lower limit protection logic to the difference; When the calculated volume of ineffective consumption exceeds the theoretical volume of a single feeding, the net effective consumption is forcibly reduced to zero.

7. The method for monitoring and managing pig feed consumption based on artificial intelligence according to claim 1, characterized in that, The step of performing cross-logic discrimination based on the numerical states of the acute material condition variance, the effective chewing time duty cycle, and the ineffective loss volume of food intake, and outputting the multimodal diagnostic status specifically includes: When the effective chewing time duty cycle is less than the preset chewing duty cycle threshold, the ineffective food loss volume is greater than the preset loss volume threshold, and the acute material condition variance is less than or equal to the preset upper limit threshold of the acute variation parameter, the multimodal diagnostic status is output as behavioral waste mode. When the effective chewing time duty cycle is less than the preset chewing duty cycle threshold, the ineffective food loss volume is greater than the preset loss volume threshold, and the acute material condition variation variance is greater than the preset upper limit threshold of the acute variation parameter, the multimodal diagnostic status is output as material condition variation mode. When the effective chewing time duty cycle is greater than or equal to a preset chewing duty cycle threshold and the ineffective food loss volume is less than or equal to a preset loss volume threshold, the multimodal diagnostic status is output as normal consumption mode.

8. The method for monitoring and managing pig feed consumption based on artificial intelligence according to claim 7, characterized in that, The specific steps of issuing negative feedback control parameters to the material feeder controller based on the multimodal diagnostic status include: When the multimodal diagnostic state is in the behavioral waste mode, the front-end negative feedback adjustment logic for physical devices is triggered. The wasted feed is converted into the corresponding number of feeding pulses and deducted proportionally in the next feeding cycle. The negative feedback control parameters for the next feeding cycle are sent to the motor drive controller of the automated auger feeder to reduce the number of feeding pulses per cycle.

9. The method for monitoring and managing pig feed consumption based on artificial intelligence according to claim 7, characterized in that, The process of synchronizing the net effective consumption to the cloud-based ERP management system to update the dynamic consumption rate model specifically includes: After a single material supply cycle ends and the multimodal diagnostic status is not in material condition variation mode and the net effective consumption is greater than zero, the net effective consumption will be synchronously reported to the cloud ERP management system. The current actual inventory of the physical silos is updated and reconciled using the net effective consumption. A dynamic consumption rate model is constructed by aggregating the daily average effective consumption volume of historical sliding statistical days, and the expected material shortage time is deduced based on the dynamic consumption rate model. When the expected material shortage time is less than the replenishment lead time threshold, a regular replenishment schedule is generated.

10. The method for monitoring and managing pig feed consumption based on artificial intelligence according to claim 7, characterized in that, The specific instructions for triggering abnormal supply chain interruption upon confirmation of material condition anomalies include: When the multimodal diagnostic status is in the material condition variation mode, the abnormal isolation scheduling mechanism is triggered; Send an abnormal interrupt command containing the device's physical address number to the cloud-based ERP management system; In the system database, the inventory status of the corresponding physical material tower and material batch is updated from available to frozen and downgraded. At the server process level, the automatically generated regular replenishment schedule is intercepted and suspended, and an anomaly investigation and cleanup work order is dispatched.

Images