Intelligent breeding feed dynamic management method, device and equipment and storage medium
By constructing a heat stress strategy switching action index and a three-state switching state space model, the problem of rapid identification and correction of feed demand forecasting under heat stress scenarios in livestock farms was solved, and reliable feed demand forecasting and management were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SICHUAN ANIMAL SCI ACAD
- Filing Date
- 2026-02-02
- Publication Date
- 2026-05-01
Smart Images
Figure CN121615959B_ABST
Abstract
Description
Intelligent feeding dynamic management methods, devices, equipment and storage media Technical Field
[0001] This invention relates to the field of intelligent aquaculture technology, and in particular to a method, apparatus, equipment and storage medium for dynamic feed management in intelligent aquaculture. Background Technology
[0002] In large-scale livestock farming, the feeding process is highly time-sensitive and strategic. Typically, farms develop feeding plans based on factors such as animal breed, age, and weight gain curves. Automated feeding equipment executes feeding at preset times, resulting in a relatively stable pattern of average feed intake. Under this stable pattern, using historical feeding data for regression extrapolation, moving averages, or estimation methods based on fixed growth curves can, to some extent, predict feed demand.
[0003] However, in actual production operations, there are some special stress scenarios. For example, when the combination of temperature and humidity in the barn causes a sustained increase in the heat load on animals, their appetite decreases, and their diurnal feeding rhythm changes significantly. To reduce the risk of heat stress, a clear management action is usually implemented on-site: triggering a strategy switch of stress-induced feed reduction + nighttime feeding shift, that is, reducing the amount of feed given during the day and shifting some of the feed to the nighttime period. This action has significant mechanistic mutation characteristics: it not only changes the total daily feed level, but also significantly changes the intraday feeding structure, causing the feed demand generation mechanism of the same breeding unit to show segmented differences in the normal, stress, and recovery stages, and there is a recovery lag and uncertainty of rebound.
[0004] Existing technologies typically face the following shortcomings when dealing with the aforementioned special stress scenarios: First, traditional prediction methods often rely on daily totals or smooth trends, making it difficult to identify the occurrence and intensity of nighttime feeding movements from intraday structural changes; second, traditional methods often assume smooth evolution of feeding behavior, making it difficult to quickly switch prediction mechanisms on the day the movement occurs or the following day, thus resulting in significant prediction bias during the stress phase; third, when data conditions at the breeding site are limited, only a small amount of key data such as feeding logs, temperature and humidity, number of animals, and body weight can be obtained. If there is a lack of modeling methods that can couple the intensity of the movement with the state switching, it is difficult to achieve reliable prediction of sudden changes under limited data conditions; fourth, the rate of recovery in feeding during the recovery phase has significant uncertainty. If there is a lack of interval prediction and online correction mechanisms, it can easily lead to insufficient feed supply or inventory backlog, thereby affecting production continuity and economic efficiency.
[0005] In summary, under the special application scenario of heat stress triggering strategy switching, there is an urgent need for a technical solution that can construct a characterization of action intensity based on a small amount of key data and deeply integrate action intensity into the state switching and time-series prediction process, thereby realizing dynamic prediction and rapid correction of feed demand. Summary of the Invention
[0006] This invention provides a method, apparatus, equipment, and storage medium for dynamic feed management in intelligent aquaculture, aiming to solve at least one of the above-mentioned technical problems.
[0007] To achieve the above objectives, the present invention provides an intelligent feed dynamic management method for aquaculture, comprising the following steps:
[0008] S1: Obtain feeding log data, environmental temperature and humidity data, and basic production data of the target breeding unit, and construct key time-series data under the same date index;
[0009] S2: Perform preprocessing on the key time series data to obtain cleaned time series data;
[0010] S3: Calculate the thermal stress strategy switching action index based on the cleaning time series data, and output the action index as an action feature sequence characterizing the intensity of stress reduction and nighttime feeding;
[0011] S4: Construct an input-driven three-state switching state space model based on the action feature sequence and the average feeding residual sequence, and determine the output hidden state estimate and state probability;
[0012] S5: Based on the latent state estimation, state probability and inventory data, recursively generate the future feed demand forecast value and forecast range, and perform online bias update on the forecast error based on the action index to output the corrected feed demand forecast.
[0013] Optionally, step S1, which involves acquiring feeding log data, environmental temperature and humidity data, and basic production data of the target breeding unit, and constructing key time-series data under the same date index, specifically includes:
[0014] S11: Obtain feeding log data including single feeding amount and timestamp, environmental data including daily statistical temperature and daily statistical relative humidity, and production data including daily closing number of animals and average weight;
[0015] S12: Aggregate the feeding amount into an hourly sequence based on the timestamp of the feeding log data, and summarize to generate the total daily feeding amount. Generate temperature and humidity sequences based on the date field of environmental data, and generate stock and weight sequences based on the date field of production basic data.
[0016] S13: Using date as the key, perform key joins on hourly series, total daily feed amount, temperature series, humidity series, livestock stock series, and weight series, and output key time series data under the same date index.
[0017] Optionally, step S2: perform preprocessing on the key time series data, specifically including: non-negative constraints on the total daily feeding amount and quantile pruning of the sliding window, synchronous pruning and scaling of the hourly feeding amount according to the daily ratio, and linear interpolation for missing weight.
[0018] Optionally, in step S3, the heat stress strategy switching action index is calculated, specifically including:
[0019] S31: Calculate the nighttime feeding ratio based on the total daily feeding amount and hourly feeding amount in the key time series data; wherein, the nighttime feeding ratio is configured as the ratio of the sum of the hourly feeding amounts corresponding to the preset nighttime hourly set intervals to the total daily feeding amount;
[0020] S32: Calculate the average feeding amount based on the total daily feeding amount and the daily closing number of animals in the key time series data, and calculate the feed reduction intensity based on the median of the average feeding amount within the historical window; wherein, the average feeding amount is configured as the ratio of the total daily feeding amount to the daily closing number of animals, and the feed reduction intensity is configured as the relative decrease determined by the average feeding amount and the median.
[0021] S33: Calculate the temperature and humidity index based on the temperature and humidity sequences in the key time series data, and convert the temperature and humidity index into heat stress intensity; wherein, the temperature and humidity index is configured as a temperature and humidity index parameter determined by the temperature value and the humidity value, and the heat stress intensity is configured to map the degree of heat stress into a continuous intensity value using Sigmoid.
[0022] S34: Perform weighted fusion processing on the nighttime feeding ratio, the feed reduction intensity, and the heat stress intensity to generate a heat stress strategy switching action index.
[0023] Optionally, in step S4, an input-driven three-state switching state-space model is constructed based on the action feature sequence and the average feeding residual sequence to determine the output hidden state estimate and state probability, specifically including:
[0024] S41: Construct three sets of state space models including normal state, stress state and recovery state, and input the action feature sequence into the state transition equation of each state to perform hidden state evolution;
[0025] The expression for the state transition equation is as follows: ;
[0026] ;
[0027] In the formula, This represents the average feeding residual sequence. This represents the hidden state, defined as appetite shift. , Indicates the noise term. Indicates state inertia, This indicates the strength of the effect of the action index on appetite shift. Represents a sequence of action features;
[0028] S42: Perform discrete state transition probability calculation on the action feature sequence, generate a transition probability matrix, and perform three sets of Kalman recursion to output the hidden state estimate of each state respectively;
[0029] The expression for calculating the state transition probability is as follows:
[0030] ;
[0031] In the formula, This represents the discrete state variable on day t, with values corresponding to the normal state N, the stress state S, and the recovery state R. This represents the probability of transitioning from the previous state i to the current state j. This represents the basic transfer tendency parameter, used to express the natural transfer tendency in the absence of action. This represents the action-driven coefficient, used to express the modulation strength of the action index on the transfer tendency. This represents a function that normalizes the scores of all candidate states to form a probability distribution;
[0032] S43: Calculate the likelihood of each state based on the average feeding residual sequence and update the state probability. Then, weight and fuse the hidden state estimates according to the state probabilities to output a unified hidden state estimate.
[0033] Optionally, in step S4, the determination of the average feeding residual sequence specifically includes:
[0034] S411: Fit the normal average feeding baseline on a non-stressed sample set based on the weight sequence and average feeding amount in the cleaning time series data;
[0035] S412: Calculate the average feeding residual based on the average feeding amount and the normal average feeding baseline; wherein the average feeding residual is configured as the difference between the average feeding amount and the normal average feeding baseline.
[0036] Optionally, step S5: Based on the latent state estimation, state probability, and inventory data, recursively generate future feed demand forecasts and forecast intervals, and perform online bias updates on the forecast error based on the action index to output the corrected feed demand forecast, specifically including:
[0037] S51: Based on the hidden state estimation and the state probability, the future hidden states are recursively calculated to generate a future offset prediction sequence;
[0038] S52: Calculate the weight gain rate based on the weight sequence in the cleaned time series data and recursively extrapolate the future weight sequence. Calculate the future normal baseline sequence based on the future weight sequence and add the future offset prediction sequence to the future normal baseline sequence to generate the future average demand prediction sequence.
[0039] S53: Based on the inventory sequence in key time series data, convert the future average demand forecast sequence into the future total demand forecast sequence to generate future feed demand forecast values.
[0040] S54: Generate future demand forecast intervals based on the uncertainty quantification results of future offset predictions and output the upper and lower bounds of demand.
[0041] S55: After obtaining the actual average feeding amount for the day, calculate the prediction error, and update the prediction bias by generating an accelerated learning rate based on the action feature sequence. Then, superimpose the prediction bias onto the future average demand prediction sequence to output the corrected future feed demand prediction value.
[0042] Furthermore, to achieve the above objectives, the present invention also provides an intelligent feed dynamic management device for aquaculture, comprising:
[0043] The module is used to acquire feeding log data, environmental temperature and humidity data, and basic production data of the target breeding unit, and to build key time-series data under the same date index.
[0044] The preprocessing module is used to perform preprocessing on the key time series data to obtain cleaned time series data;
[0045] The calculation module is used to calculate the thermal stress strategy switching action index based on the cleaning time series data, and output the action index as an action feature sequence characterizing the intensity of stress reduction and nighttime feeding.
[0046] The determination module is used to construct an input-driven three-state switching state-space model based on the action feature sequence and the average feeding residual sequence, and to determine the output hidden state estimate and state probability.
[0047] The prediction module is used to recursively generate future feed demand forecasts and forecast intervals based on latent state estimation, state probabilities, and inventory data, and to perform online bias updates on the prediction error based on the action index to output the corrected feed demand forecast.
[0048] In addition, to achieve the above objectives, the present invention also provides a dynamic feed management device for intelligent aquaculture, the dynamic feed management device for intelligent aquaculture comprising: a memory, a processor, and a dynamic feed management program for intelligent aquaculture stored in the memory and executable on the processor, wherein when the dynamic feed management program for intelligent aquaculture is executed by the processor, the steps of the dynamic feed management method for intelligent aquaculture as described above are implemented.
[0049] In addition, to achieve the above objectives, the present invention also provides a storage medium storing a dynamic feed management program for intelligent aquaculture, wherein the dynamic feed management program for intelligent aquaculture is executed by a processor to implement the steps of the above-described dynamic feed management method for intelligent aquaculture.
[0050] The beneficial effects of this invention are as follows: It proposes a method, device, equipment, and storage medium for dynamic feed management in intelligent aquaculture. By constructing key time-series data and performing preprocessing, cleaning time-series data is obtained. Based on the cleaning time-series data, a heat stress strategy switching action index is calculated. The action index is output as an action feature sequence characterizing the intensity of stress-induced feed reduction and nighttime feeding. Based on the action feature sequence and the average feeding residual sequence, an input-driven three-state switching state-space model is constructed to determine the output latent state estimate and state probability. Combined with the stock data, future feed demand prediction values and prediction intervals are recursively generated. Based on the action index, online bias updates are performed on the prediction error to output the corrected feed demand prediction. Therefore, this invention, by introducing the proportion of nighttime feeding and the intensity of feed reduction and fusing them with heat stress intensity to construct a strategy switching action index, and simultaneously injecting the action index into continuous latent state evolution and discrete state transition probability calculation, enables the system to effectively identify action intensity from a small amount of key data and quickly complete the prediction mechanism switch on the day the action occurs or the next day. Simultaneously, it outputs a reliable interval for uncertainties in the recovery phase and performs rapid correction, realizing intelligent dynamic feed management in special stress scenarios. Attached Figure Description
[0051] Figure 1 is a schematic diagram of the hardware operating environment device structure involved in the embodiment of the present invention;
[0052] Figure 2 is a flowchart illustrating an embodiment of the intelligent aquaculture feed dynamic management method of the present invention;
[0053] Figure 3 is a structural block diagram of a dynamic feed management device for intelligent aquaculture in an embodiment of the present invention. Detailed Implementation
[0054] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0056] As shown in Figure 1, Figure 1 is a schematic diagram of the device structure of the hardware operating environment involved in the embodiment of the present invention.
[0057] As shown in Figure 1, the device may include: a processor 1001, such as a CPU; a communication bus 1002; a user interface 1003; a network interface 1004; and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.
[0058] Those skilled in the art will understand that the structure of the device shown in FIG1 does not constitute a limitation on the device, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.
[0059] As shown in Figure 1, the memory 1005, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a dynamic feed management program for intelligent aquaculture.
[0060] In the terminal shown in Figure 1, the network interface 1004 is mainly used to connect to the backend server and communicate with it; the user interface 1003 is mainly used to connect to the client (user terminal) and communicate with it; while the processor 1001 can be used to call the intelligent aquaculture feed dynamic management program stored in the memory 1005 and perform the following operations:
[0061] Obtain feeding log data, environmental temperature and humidity data, and basic production data of the target breeding unit, and construct key time-series data under the same date index;
[0062] Preprocessing is performed on the key time-series data to obtain cleaned time-series data;
[0063] The thermal stress strategy switching action index is calculated based on the cleaning time series data, and the action index is output as an action feature sequence characterizing the intensity of stress-induced feed reduction and nighttime feeding.
[0064] Based on the action feature sequence and the average feeding residual sequence, an input-driven three-state switching state space model is constructed to determine the output hidden state estimate and state probability.
[0065] Based on latent state estimation, state probability, and inventory data, the predicted value and prediction range of future feed demand are recursively generated. Based on the action index, the prediction error is updated online to output the corrected feed demand prediction.
[0066] The specific embodiments of the present invention applied to the device are basically the same as the embodiments of the intelligent aquaculture feed dynamic management method described below, and will not be repeated here.
[0067] This invention provides a method for dynamic feed management in intelligent aquaculture. Referring to Figure 2, which is a flowchart of an embodiment of the method for dynamic feed management in intelligent aquaculture according to this invention.
[0068] In this embodiment, a method for dynamic feed management in intelligent aquaculture includes the following steps:
[0069] S1: Obtain feeding log data, environmental temperature and humidity data, and basic production data of the target breeding unit, and construct key time-series data under the same date index;
[0070] S2: Perform preprocessing on the key time series data to obtain cleaned time series data;
[0071] S3: Calculate the thermal stress strategy switching action index based on the cleaning time series data, and output the action index as an action feature sequence characterizing the intensity of stress reduction and nighttime feeding;
[0072] S4: Construct an input-driven three-state switching state space model based on the action feature sequence and the average feeding residual sequence, and determine the output hidden state estimate and state probability;
[0073] S5: Based on the latent state estimation, state probability and inventory data, recursively generate the future feed demand forecast value and forecast range, and perform online bias update on the forecast error based on the action index to output the corrected feed demand forecast.
[0074] It should be noted that existing technologies typically face the following shortcomings when dealing with special stress scenarios: First, traditional prediction methods often rely on daily totals or smooth trends, making it difficult to identify the occurrence and intensity of nighttime feeding movements from intraday structural changes; second, traditional methods often assume smooth evolution of feeding behavior, making it difficult to quickly switch prediction mechanisms on the day the movement occurs or the following day, thus resulting in significant prediction bias during the stress phase; third, when data conditions at the breeding site are limited, only a small amount of key data such as feeding logs, temperature and humidity, number of animals, and body weight can be obtained. If there is a lack of modeling methods that can couple the intensity of the movement with the state switching, it is difficult to achieve reliable prediction of sudden changes under limited data conditions; fourth, the rate of recovery of feeding during the recovery phase has significant uncertainty. If there is a lack of interval prediction and online correction mechanisms, it can easily lead to insufficient feed supply or inventory backlog, thereby affecting production continuity and economic efficiency.
[0075] To address the aforementioned issues, this embodiment constructs key time-series data and performs preprocessing to obtain cleaning time-series data. Based on this data, a heat stress strategy switching action index is calculated. This action index is output as an action feature sequence characterizing the intensity of stress-induced feed reduction and nighttime feeding. An input-driven three-state switching state-space model is constructed based on this action feature sequence and the average feeding residual sequence to determine the output hidden state estimate and state probability. Combined with livestock inventory data, future feed demand predictions and prediction intervals are recursively generated. Online bias updates are performed on the prediction error based on the action index to output a corrected feed demand prediction. Therefore, this invention, by introducing the nighttime feeding ratio and feed reduction intensity and fusing them with heat stress intensity to construct a strategy switching action index, simultaneously injects the action index into continuous hidden state evolution and discrete state transition probability calculation. This enables the system to effectively identify action intensity from a limited amount of key data and quickly complete the prediction mechanism switch on the day the action occurs or the following day. Simultaneously, it outputs a reliable interval for uncertainties during the recovery phase and performs rapid correction, achieving intelligent dynamic feed management in special stress scenarios.
[0076] In a preferred embodiment, step S1, which involves acquiring feeding log data, environmental temperature and humidity data, and basic production data of the target breeding unit, and constructing key time-series data under the same date index, specifically includes:
[0077] S11: Obtain feeding log data including single feeding amount and timestamp, environmental data including daily statistical temperature and daily statistical relative humidity, and production data including daily closing number of animals and average weight;
[0078] S12: Aggregate the feeding amount into an hourly sequence based on the timestamp of the feeding log data, and summarize to generate the total daily feeding amount. Generate temperature and humidity sequences based on the date field of environmental data, and generate stock and weight sequences based on the date field of production basic data.
[0079] S13: Using date as the key, perform key joins on hourly series, total daily feed amount, temperature series, humidity series, livestock stock series, and weight series, and output key time series data under the same date index.
[0080] Specifically, feeding log data is typically a flow record output by automatic feeding equipment, including the timestamp of each feeding and the amount fed. To enable the feeding data to be used for subsequent "nighttime feeding" action identification, the flow record needs to be aggregated into an hourly sequence. Using date 't' as the index and hour 'h' as the time bucket, the feeding amounts within the same hour are summed to obtain an hourly feeding amount sequence. After obtaining the hourly feeding amount sequence, the feeding amounts over the 24 hours on the same date are summed to obtain the total daily feeding amount. This allows for consistency verification of the hourly aggregation results during the data processing stage, preventing distortion of subsequent nighttime proportions due to missing hourly aggregation data.
[0081] Furthermore, environmental temperature and humidity data are typically collected by in-house sensors at fixed sampling intervals. To align these data with the daily index, daily statistical operations can be performed on the original temperature and humidity sampling sequences. For example, the daily average or maximum temperature value can be calculated, and the daily average humidity value can be calculated to obtain the daily statistical temperature and humidity for day t, generating temperature and humidity sequences. Basic production data includes the daily inventory count and average weight. Inventory counts are usually calculated daily and are complete, while weight may be collected weekly or by period, resulting in gaps. To preserve the source characteristics of weight data collection, missing weight data points can be marked.
[0082] Finally, using date t as the key, key join alignment is performed on the hourly series, total daily feed amount, temperature series, humidity series, livestock inventory series, and weight series to output key time series data. This alignment process ensures that the construction and prediction of the action index for each subsequent date are based on the same time series benchmark.
[0083] In a preferred embodiment, step S2: preprocessing is performed on the key time-series data, specifically including: non-negative constraints on the total daily feeding amount and quantile pruning of the sliding window, synchronous pruning and scaling of the hourly feeding amount by day, and linear interpolation for missing weight.
[0084] Preprocessing for the total daily feed amount: Considering that the total daily feed amount may have extreme outliers due to instantaneous false alarms, duplicate recordings, or communication jitter in the feeding equipment, in order to reduce the impact of outliers on the calculation of motion intensity, non-negative constraint processing can be performed on the total daily feed amount first, and quantile clipping can be performed within the sliding window (for example, clipping the total daily feed amount with the 1% and 99th quantiles within the window as the upper and lower bounds).
[0085] Preprocessing for hourly feeding amounts: After pruning to match the daily total, to ensure the conservation consistency between the hourly series and the daily total, the pruned hourly series is scaled according to the daily ratio. This avoids the total deviation caused by pruning the hourly series being further amplified in the calculation of the nighttime proportion.
[0086] Preprocessing for missing weight data: Linear interpolation is performed between two consecutive valid weight records to obtain a continuous weight sequence. It should be noted that the purpose of weight interpolation is to establish a continuous baseline trend, rather than pursuing high-frequency accuracy of weight data; therefore, linear interpolation is sufficient for subsequent power-law baseline fitting.
[0087] In a preferred embodiment, step S3, calculating the heat stress strategy switching action index, specifically includes:
[0088] S31: Calculate the nighttime feeding ratio based on the total daily feeding amount and hourly feeding amount in the key time series data; wherein, the nighttime feeding ratio is configured as the ratio of the sum of the hourly feeding amounts corresponding to the preset nighttime hourly set intervals to the total daily feeding amount;
[0089] S32: Calculate the average feeding amount based on the total daily feeding amount and the daily closing number of animals in the key time series data, and calculate the feed reduction intensity based on the median of the average feeding amount within the historical window; wherein, the average feeding amount is configured as the ratio of the total daily feeding amount to the daily closing number of animals, and the feed reduction intensity is configured as the relative decrease determined by the average feeding amount and the median.
[0090] S33: Calculate the temperature and humidity index based on the temperature and humidity sequences in the key time series data, and convert the temperature and humidity index into heat stress intensity; wherein, the temperature and humidity index is configured as a temperature and humidity index parameter determined by the temperature value and the humidity value, and the heat stress intensity is configured to map the degree of heat stress into a continuous intensity value using Sigmoid.
[0091] S34: Perform weighted fusion processing on the nighttime feeding ratio, the feed reduction intensity, and the heat stress intensity to generate a heat stress strategy switching action index.
[0092] Regarding the determination of the nighttime feeding ratio: The preset nighttime hourly set interval can be configured as {20,21,22,23,0,1,2,3,4,5}. Based on this, the nighttime feeding ratio is obtained by dividing the sum of the hourly feeding amounts in the nighttime hourly set interval by the total daily feeding amount.
[0093] Regarding the determination of feed reduction intensity: After calculating the average feed amount using the ratio of total daily feed amount to the daily closing number of animals, to characterize the intensity of stress-induced feed reduction, a baseline value (e.g., the median) can be constructed based on the average feed amount over the past K days. Then, the relative decrease is calculated to obtain the feed reduction intensity, expressed as: ;
[0094] In the formula, Indicates the strength of reduced material. Indicates the baseline value. This indicates the average amount of feed given. This indicates the constant for preventing division by zero. The larger the value, the more significant the reduction in material compared to the normal level.
[0095] Regarding the determination of heat stress intensity: Based on the temperature and humidity series, the temperature and humidity index (THI) is obtained using the conventional formula. Then, to map the degree of heat stress to a continuous intensity value, a sigmoid mapping is performed on the THI to obtain the heat stress intensity, expressed as:
[0096] ;
[0097] In the formula, The thermal stress threshold, For the mapping slope parameter, Let represent the temperature and humidity index on day t. This mapping allows for greater sensitivity to heat stress intensity near a threshold, while it tends to saturate further away from the threshold, making it more suitable for subsequent state transition probability-driven applications.
[0098] Finally, the intensity of feed reduction, the intensity of nighttime shift, and the intensity of thermal stress are weighted and fused to generate a thermal stress strategy switching action index. Through this fusion method, the action index increases significantly only when the thermal stress is strong and the characteristics of feed reduction and shifting are obvious, thus avoiding mistaking structural changes caused by non-stress reasons as thermal stress strategy switching.
[0099] In a preferred embodiment, step S4 involves constructing an input-driven three-state switching state-space model based on the action feature sequence and the average feeding residual sequence, and determining the output hidden state estimate and state probability. Specifically, this includes:
[0100] S41: Construct three sets of state space models including normal state, stress state and recovery state, and input the action feature sequence into the state transition equation of each state to perform hidden state evolution.
[0101] In this embodiment of the invention, a hidden state is first defined to represent appetite shift, and the average feeding residual sequence is used as the observation to construct an observation equation, the expression of which is:
[0102] ;
[0103] In the formula, This represents the average feeding residual sequence. This represents the hidden state, defined as appetite shift. This indicates the noise term.
[0104] After this, the discrete state Defined into three categories: normal state N, stress state S, and recovery state R. Within each category, the latent state... Evolving according to different dynamic parameters, and injecting the action index as a control input into the state equation, the expression is:
[0105] ;
[0106] In the formula, Indicates the noise term. Indicates state inertia, This indicates the strength of the effect of the action index on appetite shift. This represents a sequence of action features.
[0107] Under normal circumstances, under stress When the value is relatively large, the motion index increases. Rapidly becomes negative; in the recovery state The parameters are used to characterize the regression rate; the parameters under normal conditions are used to characterize small, stable fluctuations.
[0108] S42: Perform discrete state transition probability calculation on the action feature sequence to generate a transition probability matrix, and perform three sets of Kalman recursion to output the hidden state estimates of each state respectively.
[0109] S43: Calculate the likelihood of each state based on the average feeding residual sequence and update the state probability. Then, weight and fuse the hidden state estimates according to the state probabilities to output a unified hidden state estimate.
[0110] In this embodiment of the invention, to enable the model to quickly switch from a normal state to a stressed state on the day the action occurs, the invention injects an action index into the calculation of discrete state transition probabilities. For any previous state i and next state j, a score is constructed and the transition probability is obtained through softmax normalization, expressed as:
[0111] ;
[0112] In the formula, This represents the discrete state variable on day t, with values corresponding to the normal state N, the stress state S, and the recovery state R. This represents the probability of transitioning from the previous state i to the current state j. This represents the basic transfer tendency parameter, used to express the natural transfer tendency in the absence of action. This represents the action-driven coefficient, used to express the modulation strength of the action index on the transfer tendency. This represents a function that normalizes the scores of all candidate states to form a probability distribution.
[0113] It should be noted that in the calculation of this transition probability, it is achieved through setting or learning. This can significantly increase the probability of N transitioning to S when the action index increases, through setting or learning. This can increase the probability of S shifting to R when the action index weakens.
[0114] After obtaining the transition probabilities, interactive multiple model filtering (IMM) can be used for online inference. Specifically, a Kalman filter is constructed for each state class, and the three models are predicted and updated in parallel at the same time. The likelihood of each model is normalized and updated based on the observation residuals to obtain the state probability weights. Then, the hidden state estimates of the three models are fused by probability weighting to obtain a unified hidden state estimate. For ease of understanding, the prediction form and observation update of each model m can be written as:
[0115] ;
[0116] ;
[0117] In the formula, This represents the prior estimate of the hidden state. This represents the delayed estimation of the hidden state. This represents the Kalman gain.
[0118] Based on this, the hidden state estimates of the three models are fused using probability weighting to obtain a unified hidden state estimate, which can be expressed as:
[0119] ;
[0120] In the formula, This represents the estimation of hidden states. Represents the state probability weight.
[0121] As is easily understood, the key challenge in the heat stress strategy switching scenario of this invention lies in the fact that the feed demand mechanism is not singular, but rather switches between three mechanisms: normal, stress, and recovery. During the normal period, feed intake fluctuates little and is significantly influenced by weight gain trends; during the stress period, feed intake declines rapidly and is driven by activity intensity; during the recovery period, feed intake gradually recovers, but the speed is uncertain and there is a lag. If a single model is used, two types of problems easily arise: first, the model cannot quickly switch when stress occurs, leading to prediction lag; second, the model cannot express uncertainty during the recovery period, resulting in missing prediction intervals or unstable bias.
[0122] To improve this problem, this embodiment of the invention introduces a three-state switching prediction framework and uses the action index as the input driver, which acts on two paths simultaneously: at the continuous level, the action index drives the change of the latent state (appetite shift), causing the latent state to immediately shift negative when the action occurs; at the discrete level, the action index drives the state transition tendency, increasing the probability of normal transition to stress when the action occurs, enabling the model to quickly switch from the normal mechanism to the stress mechanism.
[0123] Furthermore, an interactive multi-model inference approach is adopted to maintain three sets of state models in parallel. The model probabilities are updated by observing residuals, and the hidden state estimates are then weighted and fused to obtain a unified estimation result that can both respond quickly and express the uncertainty during the recovery period.
[0124] In a preferred embodiment, step S4, determining the average feeding residual sequence, specifically includes:
[0125] S411: Fit a normal average feeding baseline on a non-stressed sample set based on the weight sequence and average feeding amount in the cleaned time series data.
[0126] In this embodiment of the invention, to isolate the natural upward trend in feed intake resulting from weight gain, a normal average baseline needs to be fitted on a set of non-stress days. The baseline model can be in power-law form:
[0127] ;
[0128] In the formula, This represents the weight sequence in the time-series data being cleaned. This represents the fitted parameters.
[0129] To facilitate fitting, a logarithmic transformation is performed on the above relationships:
[0130] ;
[0131] In the formula, This is the average feeding amount. This is the fitting residual. After fitting and obtaining a and b, calculations can be performed over the entire time period. This serves as the baseline for average feeding during the normal period.
[0132] S412: Calculate the average feeding residual based on the average feeding amount and the normal average feeding baseline; wherein the average feeding residual is configured as the difference between the average feeding amount and the normal average feeding baseline.
[0133] In this embodiment of the invention, a difference operation is performed on each date to generate the average feeding residual:
[0134] ;
[0135] In the formula, The average feeding residual sequence is used to characterize the degree of deviation from the normal baseline. When stress-induced feed reduction and nighttime feeding strategies are switched, the residuals typically become negative rapidly and then gradually return to near zero during the recovery phase.
[0136] In a preferred embodiment, step S5: Based on the latent state estimation, state probability, and inventory data, a future feed demand forecast and forecast interval are recursively generated, and an online bias update is performed on the forecast error based on the action index to output a corrected feed demand forecast, specifically including:
[0137] S51: Based on the hidden state estimation and the state probability, the future hidden states are recursively calculated to generate a future offset prediction sequence.
[0138] Specifically, to achieve multi-day forecasting, it is necessary to recursively extrapolate the intensity of future actions. Considering that strategy switching is usually continuous and gradual, this invention uses exponential decay recursion for the intensity of material reduction and nighttime shift, and employs autoregression based on historical THI for the intensity of thermal stress, thereby forming a future action index. The attenuation form of the reduction intensity and nighttime offset can be written as:
[0139] ;
[0140] ;
[0141] In the formula, The attenuation coefficient is used to characterize the gradual release of the action. This represents the offset value for the proportion of feeding at night. This indicates the intensity of material reduction.
[0142] For the intensity of thermal stress, a first-order autoregression can be performed on the THI:
[0143] ;
[0144] In the formula, For the fitting parameters, To predict the temperature and humidity index.
[0145] Recursively obtained Then mapped to The final result obtained by fusion is:
[0146] ;
[0147] In the formula, This indicates the predicted intensity of heat stress. This indicates an index of future actions.
[0148] Following this, based on the current state probability and the parameters of each state model, the future hidden states are recursively deduced. The recursion is performed for each state model m as follows: ;
[0149] In the formula, Indicates future action index, This represents the hidden state prediction under model m on the (t+k)th day in the future;
[0150] And the current state probability is used as the weight for fusion:
[0151] ;
[0152] In the formula, This represents the prediction of the future hidden states after fusion;
[0153] Simultaneously, to predict the output interval, the uncertainty of future offsets needs to be quantified. This uncertainty can be decomposed into within-model variance and between-model differences, and approximated as follows:
[0154] ;
[0155] In the formula, The recursive variance within the state model. This represents the variance of the future hidden state prediction. In this way, the uncertainty of whether the recovery phase is in stress or recovery will be reflected as a difference term between models, thus allowing the interval width to be reasonably increased during the recovery phase.
[0156] S52: Calculate the weight gain rate based on the weight sequence in the cleaned time series data and recursively extrapolate the future weight sequence. Calculate the future normal baseline sequence based on the future weight sequence and add the future offset prediction sequence to the future normal baseline sequence to generate the future average demand prediction sequence.
[0157] S53: Based on the inventory sequence in key time-series data, the future average demand forecast sequence is converted into the future total demand forecast sequence to generate future feed demand forecasts. Specifically, the weight gain rate over the past G days is calculated based on the continuous body weight sequence, and future body weight is extrapolated.
[0158] ;
[0159] ;
[0160] In the formula, Represents a continuous weight sequence. This represents the weight sequence over the past G days. This represents the average weight gain rate over the past G days. Indicates future weight prediction. This indicates the prediction step size.
[0161] Then, the future normal baseline sequence is calculated using the power-law parameters obtained from the fitting:
[0162] ;
[0163] In the formula, This indicates the future normal baseline sequence.
[0164] It should be noted that this baseline extrapolation is used to preserve growth trends and prevent excessive downward revision of demand estimates during stress phases.
[0165] Based on this, the future normal baseline is added to the future latent state offset to obtain the future average demand forecast, and then the average demand is converted into the total demand forecast by combining the number of animals in stock:
[0166] ;
[0167] ;
[0168] In the formula, This indicates a forecast of average future demand. This indicates a forecast of future aggregate demand. This represents the number of animals in stock on day t.
[0169] S54: Generate future demand forecast intervals based on the uncertainty quantification results of future offset predictions and output the upper and lower bounds of demand.
[0170] Specifically, the future average demand variance is synthesized based on the future offset variance and observation noise, and the confidence interval is output. This can be written as: ;
[0171] ;
[0172] ;
[0173] After that, convert to the total demand range:
[0174] ;
[0175] ;
[0176] In the formula, This is a quantile parameter; for example, 1.96 can be used for the 95% interval. The standard deviation of the observed noise, The standard deviation of the average future forecast. These are the upper and lower bounds of the average prediction, respectively. These represent the upper and lower bounds of the total demand forecast. By outputting this range, the uncertainty during the recovery period can be explicitly provided to the feed supply decision-making module, allowing adjustments to replenishment and inventory strategies while keeping risks under control.
[0177] S55: After obtaining the actual average feeding amount for the day, calculate the prediction error, and update the prediction bias by generating an accelerated learning rate based on the action feature sequence. Then, superimpose the prediction bias onto the future average demand prediction sequence to output the corrected future feed demand prediction value.
[0178] Specifically, after the actual feeding data arrives, to enable the model to quickly correct its course on the day the action occurs or the following day, this invention introduces a prediction bias and uses an action exponential adaptive learning rate for updating. First, the prediction error of the average feeding for the day is calculated:
[0179] ;
[0180] In the formula, This represents the average predicted value given by the model on that day. This represents the true average feeding amount. Indicates prediction error;
[0181] Then, construct the learning rate and update the bias based on the action index:
[0182] ;
[0183] ;
[0184] In the formula, represents the learning rate function. The parameters represent the learning rate function. Indicates online bias, This represents the strategy switching action index, specifically on day t. ;
[0185] Finally, the bias superposition value is used to predict the future, resulting in a corrected prediction:
[0186] ;
[0187] In the formula, This represents the corrected average forecast after adding bias. This represents the revised aggregate demand forecast.
[0188] Furthermore, to ensure consistency between the output correction and the state inference, the bias-corrected residuals can be constructed and fed back into subsequent inferences:
[0189] ;
[0190] The residual after bias correction at the next time step Alternative As an input to the IMM observation, the state probability update is made consistent with the correction result, avoiding the inconsistency of short-term output correction but state still lagging.
[0191] In this embodiment of the invention, based on the hidden state estimation and state probability output by S4, combined with the body weight sequence and inventory data, it is possible to recursively generate the predicted value and prediction range of feed demand for multiple days in the future; and after obtaining the actual feeding on the day, online bias update is performed based on the action index to output the corrected demand prediction result, and the corrected residual is fed back for subsequent inference.
[0192] Referring to Figure 3, which is a schematic diagram of an embodiment of the intelligent aquaculture feed dynamic management device of the present invention.
[0193] As shown in Figure 3, the intelligent aquaculture feed dynamic management device proposed in this embodiment of the invention includes:
[0194] Module 10 is used to acquire feeding log data, environmental temperature and humidity data, and production basic data of the target breeding unit, and to construct key time-series data under the same date index.
[0195] Preprocessing module 20 is used to perform preprocessing on the key time series data to obtain cleaned time series data;
[0196] The calculation module 30 is used to calculate the thermal stress strategy switching action index based on the cleaning time series data, and output the action index as an action feature sequence characterizing the intensity of stress reduction and nighttime feeding.
[0197] The determination module 40 is used to construct an input-driven three-state switching state space model based on the action feature sequence and the average feeding residual sequence, and to determine the output hidden state estimate and state probability.
[0198] The prediction module 50 is used to recursively generate future feed demand forecasts and forecast intervals based on latent state estimation, state probabilities, and inventory data, and to perform online bias updates on the prediction error based on the action index to output a corrected feed demand forecast. Other embodiments or specific implementations of the intelligent livestock feed dynamic management device of the present invention can refer to the above-described method embodiments, and will not be repeated here.
[0199] Furthermore, the present invention also proposes an intelligent aquaculture feed dynamic management device, which includes: a memory, a processor, and an intelligent aquaculture feed dynamic management program stored in the memory and executable on the processor. When the intelligent aquaculture feed dynamic management program is executed by the processor, it implements the steps of the intelligent aquaculture feed dynamic management method as described above.
[0200] The specific implementation method of the feed dynamic management equipment for intelligent aquaculture in this application is basically the same as the embodiments of the feed dynamic management method for intelligent aquaculture described above, and will not be repeated here.
[0201] Furthermore, this invention also proposes a readable storage medium, which includes a computer-readable storage medium storing a dynamic feed management program for intelligent aquaculture. The readable storage medium can be the memory 1005 in the terminal of Figure 1, or at least one of ROM (Read-Only Memory) / RAM (Random Access Memory), magnetic disk, and optical disk. The readable storage medium includes several instructions to cause a processor-equipped intelligent aquaculture feed management device to execute the dynamic feed management method for intelligent aquaculture described in various embodiments of this invention.
[0202] The specific implementation methods in the readable storage medium of this application are basically the same as the embodiments of the above-described intelligent aquaculture feed dynamic management method, and will not be repeated here.
[0203] It is understood that in the description of this specification, references to terms such as "one embodiment," "another embodiment," "other embodiments," or "first embodiment to Nth embodiment," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0204] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.
[0205] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0206] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.
[0207] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.
Claims
1. A method for dynamic feed management in intelligent aquaculture, characterized in that, Includes the following steps: S1: Obtain feeding log data, environmental temperature and humidity data, and basic production data of the target breeding unit, and construct key time-series data under the same date index; S2: Perform preprocessing on the key time series data to obtain cleaned time series data; S3: Calculate the heat stress strategy switching action index based on the cleaning time series data, and output the action index as an action feature sequence characterizing the intensity of stress-induced feed reduction and nighttime feeding. Specifically, calculating the heat stress strategy switching action index includes: S31: Calculating the nighttime feeding ratio based on the total daily feeding amount and hourly feeding amount in the key time series data; wherein, the nighttime feeding ratio is configured as the ratio of the sum of hourly feeding amounts corresponding to a preset nighttime hourly set interval to the total daily feeding amount; S32: Calculating the average feeding amount based on the total daily feeding amount and the daily closing number of animals in the key time series data, and calculating the feed reduction intensity based on the median of the average feeding amount within the historical window; wherein, the average feeding amount is configured as the ratio of the total daily feeding amount to the daily closing number of animals, and the feed reduction intensity is configured as the relative decrease determined by the average feeding amount and the median; S33: Based on the key time series data... Temperature and humidity sequences are used to calculate a temperature-humidity index, which is then converted into heat stress intensity. The temperature-humidity index is configured as a temperature and humidity parameter determined by temperature and humidity values, and the heat stress intensity is configured to map the degree of heat stress to a continuous intensity value using a sigmoid function. S34: The nighttime feeding ratio, the feed reduction intensity, and the heat stress intensity are weighted and fused to generate a heat stress strategy switching action index. S4: An input-driven three-state switching state space model is constructed based on the action feature sequence and the average feeding residual sequence to determine the output hidden state estimate and state probability. Specifically, this includes: S41: Constructing three sets of state space models containing normal, stressed, and recovered states, and inputting the action feature sequence into the state transition equations of each state to perform hidden state evolution. The expression for the state transition equation is: ； In the formula, This represents the average feeding residual sequence. This represents the hidden state, defined as appetite shift. 、 Indicates the noise term. Indicates state inertia, This indicates the strength of the effect of the action index on appetite shift. S42: Perform discrete state transition probability calculation on the action feature sequence to generate a transition probability matrix, and execute three sets of Kalman recursions to output the hidden state estimates for each state; wherein, the expression for calculating the state transition probability is: In the formula, This represents the discrete state variable on day t, with values corresponding to the normal state N, the stress state S, and the recovery state R. This represents the probability of transitioning from the previous state i to the current state j. This represents the basic transfer tendency parameter, used to express the natural transfer tendency in the absence of action. This represents the action-driven coefficient, used to express the modulation strength of the action index on the transfer tendency. S43: Calculate the likelihood of each state based on the average feeding residual sequence and update the state probability. The hidden state estimate is weighted and fused according to the state probability to output a unified hidden state estimate. S5: Based on the hidden state estimate, state probability and inventory data, the future feed demand forecast value and forecast interval are recursively generated. Based on the action feature sequence, the prediction error is updated online to output the corrected feed demand forecast.
2. The intelligent aquaculture feed dynamic management method as described in claim 1, characterized in that, Step S1 involves acquiring feeding log data, environmental temperature and humidity data, and production data for the target breeding unit, and constructing key time-series data under the same date index. Specifically, this includes: S11: acquiring feeding log data containing single feeding amounts and timestamps, environmental data containing daily statistical temperature and daily statistical relative humidity, and production data containing daily inventory count and average weight; S12: aggregating feeding amounts into hourly sequences based on the timestamps of the feeding log data, and summarizing them to generate the total daily feeding amount; generating temperature and humidity sequences based on the date fields of the environmental data; and generating inventory and weight sequences based on the date fields of the production data; S13: performing key joins on the hourly sequences, total daily feeding amounts, temperature sequences, humidity sequences, inventory sequences, and weight sequences using the date as the key, and outputting key time-series data under the same date index.
3. The intelligent aquaculture feed dynamic management method as described in claim 2, characterized in that, Step S2: Perform preprocessing on the key time series data, specifically including: non-negative constraints on the total daily feeding amount and quantile pruning of the sliding window, synchronous pruning and scaling of the hourly feeding amount according to the daily ratio, and linear interpolation for missing weight.
4. The intelligent aquaculture feed dynamic management method as described in claim 1, characterized in that, In step S4, the determination of the average feeding residual sequence specifically includes: S411: fitting a normal average feeding baseline on a non-stressed sample set based on the weight sequence and average feeding amount in the cleaned time series data; S412: calculating the average feeding residual based on the average feeding amount and the normal average feeding baseline; wherein, the average feeding residual is configured as the difference between the average feeding amount and the normal average feeding baseline.
5. The intelligent aquaculture feed dynamic management method as described in claim 4, characterized in that, Step S5: Based on the latent state estimation, state probability, and inventory data, recursively generate future feed demand forecasts and forecast intervals, and perform online bias updates on the forecast error based on the action index to output the corrected feed demand forecast. Specifically, this includes: S51: Recursively extrapolating future latent states based on the latent state estimation and state probability to generate a future offset prediction sequence; S52: Calculating the weight gain rate based on the weight sequence in the cleaned time series data and recursively extrapolating the future weight sequence, calculating the future normal baseline sequence based on the future weight sequence, and adding the future offset prediction sequence to the future normal baseline sequence to generate... S53: Based on the inventory sequence in the key time series data, the future average demand forecast sequence is converted into the future total demand forecast sequence to generate the future feed demand forecast value; S54: Based on the uncertainty quantification result of the future offset forecast, the future demand forecast interval is generated and the upper and lower bounds of demand are output; S55: After obtaining the actual average feeding amount for the day, the prediction error is calculated, and the prediction bias is updated by generating an accelerated learning rate based on the action feature sequence. The prediction bias is superimposed on the future average demand forecast sequence to output the corrected future feed demand forecast value.
6. A feed dynamic management device for intelligent aquaculture, used in the feed dynamic management method for intelligent aquaculture as described in any one of claims 1-5, characterized in that, include: The module is used to acquire feeding log data, environmental temperature and humidity data, and basic production data of the target breeding unit, and to build key time-series data under the same date index. The preprocessing module is used to perform preprocessing on the key time series data to obtain cleaned time series data; The calculation module is used to calculate the thermal stress strategy switching action index based on the cleaning time series data, and output the action index as an action feature sequence characterizing the intensity of stress reduction and nighttime feeding. The determination module is used to construct an input-driven three-state switching state space model based on the action feature sequence and the average feeding residual sequence, and determine the output hidden state estimate and state probability; the prediction module is used to recursively generate the future feed demand prediction value and prediction interval based on the hidden state estimate, state probability and inventory data, and perform online bias update on the prediction error based on the action index to output the corrected feed demand prediction.
7. A dynamic feed management device for intelligent aquaculture, characterized in that, The intelligent aquaculture feed dynamic management device includes: a memory, a processor, and an intelligent aquaculture feed dynamic management program stored in the memory and executable on the processor. When the intelligent aquaculture feed dynamic management program is executed by the processor, it implements the steps of the intelligent aquaculture feed dynamic management method as described in any one of claims 1 to 5.
8. A storage medium, characterized in that, The storage medium stores a dynamic feed management program for intelligent aquaculture. When the processor executes the dynamic feed management program for intelligent aquaculture, it implements the steps of the dynamic feed management method for intelligent aquaculture as described in any one of claims 1 to 5.
Citation Information
Patent Citations
Acoustic signal monitoring method for stress behavior of facility-breeding sheep
CN106847293A
In-situ stress-free full-life-cycle production performance measuring system for pasturing area
CN120087763A