LSTM-based feed feeding process dynamic adjustment method and system and storage medium

Abstract

The embodiment of the application provides a kind of based on the dynamic adjustment method, system and storage medium of feed feeding process of LSTM, belong to process control technical field.The method comprises: obtaining the running data of target pen in historical management period;Effective forage event sequence for representing real foraging behavior is constructed;Based on effective forage event sequence and trough remaining material weight data, forage interval length and remaining material decay slope are calculated, and pen demand feature sequence is constructed;Pen demand feature sequence is input into preset LSTM feed demand prediction model, and the feed demand prediction value of target pen in future management period is output, and feeding management parameter for subsequent feeding process is generated based on feed demand prediction value.The application scheme is combined by effective forage event screening and LSTM feed demand prediction, so that feeding management parameter can be dynamically generated based on real foraging rhythm and future demand change, and invalid triggering interference and supply deviation are reduced.

Description

Technical Field

[0001] This invention relates to the field of LSTM prediction technology, and more specifically to a method, system and storage medium for dynamic adjustment of feed feeding process based on LSTM. Background Technology

[0002] In large-scale farms, automated feeding equipment and feed management platforms have been gradually applied to daily production management. Current methods mostly generate feeding plans based on fixed age curves, historical experience-based feeding amounts, or leftover feed in the troughs. Some solutions estimate feed demand based on daily feed consumption, animal numbers, and environmental data. While these methods can complete basic feeding management in typical production scenarios, their data foundation remains relatively coarse. They often directly use feeding records or leftover feed results as the basis for demand prediction, making it difficult to reflect the actual changes in animal feeding behavior.

[0003] In actual operation, the amount of feed remaining in the trough is only a result at a certain moment. Its variation may originate from normal animal feeding, or it may be affected by rooting, shaking, falling feed impact, or disturbances caused by the operation of the feeding device. Although trigger signals or video recognition results can reflect animal behavior when approaching the trough, during continuous feeding, the sound of falling feed and mechanical vibration generated when the feeding device starts can easily induce animals to gather briefly, causing trigger signals to occur frequently, while actual feed consumption does not increase synchronously. If this type of trigger data is directly used as input for demand forecasting, the model may learn false feeding feedback, leading to an overestimation of feed demand forecasts for future management cycles, which in turn affects the generation of planned feeding amounts, supplementary feed demand, and the allocation ratio of feeding periods.

[0004] Furthermore, the size of animals in different pens, their feeding habits, and their environmental responses vary significantly. Animal feeding behavior also exhibits clear group dynamics and rhythmicity, often manifesting as concentrated feeding, intermittent intake, and feeding decline. Existing prediction methods based on fixed thresholds, moving averages, or single historical feed consumption are insufficient to continuously correlate effective feeding events, feeding interval lengths, feed decay slopes, pen inventory data, and environmental data. They also struggle to correct prediction inputs in the presence of invalid triggers or rhythmic shifts, easily leading to prolonged overfeeding or underfeeding in some pens.

[0005] Therefore, how to accurately identify real feeding behavior within the historical management cycle, eliminate invalid triggers caused by feeding disturbances, and transform the effective feeding event sequence, feeding rhythm characteristics, pen inventory data, and environmental data into a pen demand characteristic sequence that can be used for time series prediction, and then construct a model suitable for feed demand prediction and feeding management parameter generation in future management cycles, is an urgent technical problem to be solved in the dynamic adjustment of feed feeding process in large-scale farms. Summary of the Invention

[0006] The purpose of this invention is to provide a dynamic adjustment method, system and storage medium for feed feeding process based on LSTM, so as to at least solve the problems in the existing feed demand prediction process, which make it difficult to distinguish between real feeding behavior and invalid triggers caused by feeding disturbances, resulting in distortion of demand characteristics, and the difficulty in reflecting real feeding rhythm changes when generating feeding plans based on a single historical feed consumption or instantaneous feed surplus.

[0007] To achieve the above objectives, the first aspect of the present invention provides a method for dynamic adjustment of feed feeding process based on LSTM. The method includes: acquiring operational data of a target pen during a historical management period, the operational data including at least the weight of leftover feed in the feed trough, feeding flow rate data, pre-feed trigger data, pen inventory data, and environmental data; constructing an effective feeding event sequence to characterize actual feeding behavior based on the weight of leftover feed in the feed trough, the feeding flow rate data, and the pre-feed trigger data; calculating the feeding interval length and the leftover feed decay slope based on the effective feeding event sequence and the weight of leftover feed in the feed trough, and constructing a pen demand feature sequence based on the feeding interval length, the leftover feed decay slope, the pen inventory data, and the environmental data; inputting the pen demand feature sequence into a preset LSTM feed demand prediction model, outputting the predicted feed demand value of the target pen in a future management period, and generating feeding management parameters for subsequent feeding processes based on the predicted feed demand value.

[0008] Optionally, the operation data of the target pen within the historical management period is obtained, including: within the historical management period, synchronously collecting the remaining feed weight data of the feed trough corresponding to the target pen, the feeding flow rate data of the feeding device, the trigger data of animals in the feed trough area, the pen inventory data of the target pen, and the environmental data of the breeding environment corresponding to the target pen according to a preset sampling period; performing time alignment processing on the remaining feed weight data of the feed trough, the feeding flow rate data, the trigger data of animals in the feed trough area, the pen inventory data, and the environmental data based on the preset sampling period to form operation data corresponding to the same time index.

[0009] Optionally, based on the remaining feed weight data in the feed trough, the feeding flow rate data, and the trigger data before the feed trough, an effective feeding event sequence for characterizing actual feeding behavior is constructed, including: calculating the net remaining feed change within each sampling interval based on the change in remaining feed weight and the feeding input at each sampling time in the operational data, and using the net remaining feed change as a criterion for characterizing feeding consumption; combining the trigger data before the feed trough, identifying the trigger event interval corresponding to each trigger time, extracting the corresponding net remaining feed change for each trigger event interval, and determining whether the net remaining feed change meets a preset consumption judgment condition, so as to perform filtering processing on the trigger event interval; The trigger event interval that meets the preset consumption judgment condition is determined as a valid feeding event, and each valid feeding event is associated and organized in chronological order to generate a valid feeding event sequence. The valid feeding event sequence is used to construct the column demand feature sequence.

[0010] Optionally, for each trigger event interval, the corresponding net remaining material change is extracted, and it is determined whether the net remaining material change meets a preset consumption judgment condition to perform a filtering process on the trigger event interval. This includes: calculating the net remaining material change corresponding to the trigger event interval based on the remaining material weight value at the start time, the remaining material weight value at the end time, and the total feeding input within the corresponding interval; constructing a net consumption per unit time based on the net remaining material change and the duration of the trigger event interval, and comparing it with a preset consumption judgment threshold to determine whether the corresponding trigger event interval meets the preset consumption judgment condition; determining the trigger event intervals that meet the preset consumption judgment condition as valid feeding event intervals and using them to construct the valid feeding event sequence; determining the trigger event intervals that do not meet the preset consumption judgment condition as invalid trigger event intervals, and generating an invalid trigger ratio based on the proportion of invalid trigger event intervals in the corresponding time index. The invalid trigger ratio is used to construct the column demand feature sequence.

[0011] Optionally, based on the effective feeding event sequence and the remaining food weight data in the feed trough, the feeding interval length and the remaining food decay slope are calculated, including: traversing each effective feeding event in the effective feeding event sequence in chronological order, extracting the occurrence time of two adjacent effective feeding events, and calculating the time interval between the two adjacent effective feeding events as the feeding interval length under the corresponding time index; constructing a remaining food weight sequence corresponding to the sampling time based on the remaining food weight data in the feed trough, and calculating the ratio of the difference in remaining food weight to the corresponding time interval for each adjacent sampling time to obtain the remaining food decay slope for each sampling interval; performing correlation processing on the feeding interval length and the remaining food decay slope based on the same time index to form basic data of feeding rhythm for constructing the stall demand feature sequence.

[0012] Optionally, a stall demand feature sequence is constructed based on the feeding interval length, the residual feed decay slope, the stall inventory data, and the environmental data. This includes: normalizing the feeding interval length and the corresponding residual feed decay slope after time alignment to obtain the interval representation value and slope representation value for each time index; constructing a feeding rhythm deviation value to represent the degree of change in the actual feeding rhythm based on the deviation relationship between the interval representation value and the slope representation value; and concatenating the feeding rhythm deviation value, the number of effective feeding events, the feeding interval length, the residual feed decay slope, the stall inventory data, and the environmental data according to the time index to generate a stall demand feature sequence corresponding to the target stall.

[0013] Optionally, the process of inputting the stall demand feature sequence into a preset LSTM feed demand prediction model and outputting the feed demand prediction value for the target stall in the future management cycle includes: inputting the stall demand feature sequence into the feature input layer of the preset LSTM feed demand prediction model, and generating rhythm correction weights for corresponding time indices based on the feeding rhythm deviation and invalid trigger ratio; inputting the rhythm correction weights into the rhythm correction layer of the preset LSTM feed demand prediction model to correct the feature weights for corresponding time indices in the stall demand feature sequence, thereby obtaining a corrected stall demand feature sequence; inputting the corrected stall demand feature sequence into the LSTM time series prediction layer of the preset LSTM feed demand prediction model, and outputting the latent state prediction result for the future management cycle; and inputting the latent state prediction result into the demand output layer of the preset LSTM feed demand prediction model, and outputting the feed demand prediction value for the target stall in the future management cycle.

[0014] Optionally, feeding management parameters for subsequent feeding processes are generated based on the predicted feed demand, including: determining the demand level of the target pen in the future management cycle based on the deviation between the predicted feed demand and the historical feed consumption benchmark value corresponding to the target pen; generating the planned feeding amount, feeding priority, supplementary feed demand, and feeding time allocation ratio corresponding to the target pen based on the predicted feed demand, the demand level, and the feed inventory data of the target farm; and using the planned feeding amount, the feeding priority, the supplementary feed demand, and the feeding time allocation ratio as feeding management parameters for subsequent feeding processes.

[0015] A second aspect of the present invention provides a dynamic adjustment system for feed feeding based on LSTM. The system includes: a data acquisition unit for acquiring operational data of a target pen during a historical management period, the operational data including at least the weight of leftover feed in the feed trough, feeding flow rate data, pre-feed trigger data, pen inventory data, and environmental data; a time series determination unit for constructing an effective feeding event sequence characterizing actual feeding behavior based on the weight of leftover feed in the feed trough, the feeding flow rate data, and the pre-feed trigger data; a demand sequence determination unit for calculating the feeding interval length and the leftover feed decay slope based on the effective feeding event sequence and the weight of leftover feed in the feed trough, and constructing a pen demand feature sequence based on the feeding interval length, the leftover feed decay slope, the pen inventory data, and the environmental data; and a prediction unit for inputting the pen demand feature sequence into a preset LSTM feed demand prediction model, outputting a predicted feed demand value for the target pen in a future management period, and generating feeding management parameters for subsequent feeding processes based on the predicted feed demand value.

[0016] On the other hand, the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described LSTM-based dynamic adjustment method for feed feeding.

[0017] Through the above technical solution, this invention synchronously acquires data on the weight of leftover feed in the feed trough, feeding flow rate, trigger data before the feed trough, pen inventory data, and environmental data within a historical management cycle. Based on these data, an effective feeding event sequence is constructed, enabling the extraction of genuine feeding behavior from historical operational data and reducing the impact of feeding disturbances and invalid triggers on demand assessment. Furthermore, by calculating the feeding interval length and the slope of leftover feed decay, and combining pen inventory data and environmental data, a pen demand characteristic sequence is constructed. This allows feed demand prediction to no longer solely rely on historical feed consumption or instantaneous leftover feed results, but rather to reflect the actual feeding rhythm changes in the target pen. Further, the pen demand characteristic sequence is input into a preset LSTM feed demand prediction model, which outputs predicted feed demand values ​​for future management cycles. Feeding management parameters are then generated based on these values, ensuring that subsequent feeding plans match actual pen demand changes. This reduces problems such as feed supply deviation, leftover feed accumulation, and delayed replenishment, improving the stability and adaptability of feed management in large-scale farms.

[0018] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the steps of a dynamic adjustment method for feed feeding based on LSTM provided in one embodiment of the present invention; Figure 2 This is a structural diagram of an LSTM feed demand prediction model based on rhythm-corrected weights provided in one embodiment of the present invention; Figure 3 This is a graph showing the change in residual material in the target compartment feed trough and the feeding event, provided by one embodiment of the present invention. Figure 4 This is a curve showing the net consumption rate of the target field versus the determination of effective feeding events, provided by one embodiment of the present invention. Figure 5 This is a graph showing the feed demand forecast results for the target pen during the future management cycle, provided by one embodiment of the present invention. Figure 6 This is a system structure diagram of a dynamic adjustment system for feed feeding based on LSTM provided in one embodiment of the present invention; Figure 7 This is an internal structural diagram of a computer device provided in one embodiment of the present invention. Detailed Implementation

[0020] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] like Figure 1 As shown, embodiments of the present invention provide a method for dynamic adjustment of feed feeding process based on LSTM, the method comprising: Step S1: Obtain the operating data of the target field within the historical management period. The operating data includes at least the weight of the remaining material in the trough, the feeding flow rate, the trigger data before the trough, the field inventory data, and the environmental data.

[0022] Specifically, acquiring the operational data of the target enclosure within the historical management period includes: synchronously collecting, according to a preset sampling period, the remaining feed weight data of the feed trough corresponding to the target enclosure, the feeding flow rate data of the feeding device, the trigger data of animals in the feed trough area, the enclosure stock data of the target enclosure, and the environmental data of the breeding environment corresponding to the target enclosure; performing time alignment processing on the remaining feed weight data of the feed trough, the feeding flow rate data, the trigger data of animals in the feed trough area, the enclosure stock data, and the environmental data based on the preset sampling period to form operational data corresponding to the same time index.

[0023] In this embodiment of the invention, in practical applications, the target pen is typically equipped with a feed trough weighing sensor, a feeding device flow meter, and a trigger sensing component for detecting animals approaching or contacting the feed trough. During operation, these devices output signals for the weight of remaining feed in the trough, the feeding flow rate, and a trigger signal at the trough, respectively. Meanwhile, the pen inventory data can originate from daily inventory records, ear tag statistics, or automatic inventory devices in the farm management system. Environmental data can be obtained from temperature sensors, humidity sensors, or environmental monitoring terminals installed within the corresponding area of ​​the pen, reflecting changes in the farming environment over a given time period.

[0024] Considering the differences in sampling frequency, update time, and response characteristics among data from different sources, a preset sampling period set by the controller or data management platform is typically used as a unified time benchmark during the data acquisition phase. For example, data on the weight of remaining material in the trough, feeding flow rate, and pre-trough trigger data are synchronously collected at fixed time intervals, and the inventory data and environmental data at the corresponding time granularity are synchronously mapped or resampled, ensuring that various types of data have a corresponding relationship in the time dimension. In this way, data from different sources within the historical management period can form a corresponding relationship under the same time index, avoiding time misalignment during the subsequent construction of requirement features.

[0025] During data organization, the weight of leftover feed in the trough, feeding flow rate, trigger status at the trough, number of animals in each pen, and environmental parameters corresponding to each sampling time are arranged in chronological order to construct the original operational data sequence. For cases with sampling time offsets or missing data, data from different sources can be mapped to the same time index based on timestamp alignment or interpolation, thus forming a unified time series structure. This unified time series not only preserves the dynamic relationship between changes in leftover feed in the trough and feeding input, but also records the animal behavior triggers in the trough area, changes in the number of animals in the corresponding pens, and changes in the breeding environment, providing a foundation for subsequently constructing effective feeding event sequences and pen demand characteristic sequences.

[0026] Furthermore, in some implementation scenarios, stall inventory data can be updated at a daily time granularity and mapped to multiple sampling periods within the corresponding time range; environmental data can be generated by performing regional averaging on multiple environmental sampling points to produce environmental data for the corresponding stall. Through the above processing, stall inventory data and environmental data can be kept consistent with feed trough weight data, feeding flow data, and pre-trough trigger data in the time dimension, which facilitates subsequent stall demand feature construction and LSTM feed demand prediction based on a unified time index.

[0027] It should be noted that the specific method of acquiring the above-mentioned operational data can be adjusted according to the site conditions. For example, the trigger data at the feed trough can come from infrared sensors, pressure sensing devices, visual recognition results, or ultrasonic detection results, as long as it can reflect the interaction behavior between the animal and the feed trough, it can be included in the scope of protection of this invention. Similarly, environmental data can also include at least one of temperature, humidity, ammonia concentration, or light intensity. The sampling period can be set according to the feeding rhythm, animal feeding characteristics, and farm data management cycle, and all of these are implementation methods of this invention as long as they do not affect the temporal correlation of the data.

[0028] Step S2: Based on the remaining weight data of the feed trough, the feeding flow rate data, and the trigger data in front of the trough, construct an effective feeding event sequence to characterize the actual feeding behavior.

[0029] Specifically, based on the changes in the weight of leftover food and the amount of feed input at each sampling time in the operational data, the net change in leftover food within each sampling interval is calculated, and the net change in leftover food is used as the criterion for characterizing feed consumption. Combined with the pre-fence trigger data, the trigger event intervals corresponding to each trigger time are identified, and the corresponding net change in leftover food is extracted for each trigger event interval. It is then determined whether the net change in leftover food meets the preset consumption judgment conditions, and the trigger event intervals are filtered accordingly. Trigger event intervals that meet the preset consumption judgment conditions are identified as valid feed events, and each valid feed event is associated and organized in chronological order to generate a valid feed event sequence. This valid feed event sequence is used to construct the stall demand feature sequence.

[0030] Furthermore, for each trigger event interval, the corresponding net remaining material change is extracted, and it is determined whether the net remaining material change meets the preset consumption judgment condition to perform a filtering process on the trigger event interval. This includes: calculating the net remaining material change corresponding to the trigger event interval based on the remaining material weight value at the start time, the remaining material weight value at the end time, and the total feeding input within the corresponding interval; constructing the net consumption per unit time based on the net remaining material change and the duration of the trigger event interval, and comparing it with the preset consumption judgment threshold to determine whether the corresponding trigger event interval meets the preset consumption judgment condition; determining the trigger event intervals that meet the preset consumption judgment condition as valid feeding event intervals and using them to construct the valid feeding event sequence; determining the trigger event intervals that do not meet the preset consumption judgment condition as invalid trigger event intervals, and generating an invalid trigger ratio based on the proportion of invalid trigger event intervals in the corresponding time index. The invalid trigger ratio is used to construct the column demand feature sequence.

[0031] In this embodiment of the invention, the change in the weight of remaining feed in the trough between adjacent sampling times is calculated based on the unified time index running data formed in step S1. Specifically, for any sampling interval, the weight of remaining feed in the trough at the beginning of the sampling interval, the weight of remaining feed in the trough at the end of the sampling interval, and the feeding input amount input to the trough by the feeding device within the sampling interval are first obtained; then, based on the feeding input amount, the net consumption amount of feed due to animal feeding within the sampling interval is determined.

[0032] In other words, the net feed consumption is used to characterize the reduction in the weight of the feed remaining in the trough at the end of the sampling period compared to the weight of the feed already in the trough at the beginning and the weight of the newly added feed during the sampling period after the animal has consumed it. This process incorporates the increase in trough weight caused by feeding behavior into the correction process, avoiding the misinterpretation of weight changes caused by feeding input as changes in feed consumption. This allows subsequent analysis to focus primarily on actual feed consumption rather than directly relying on the instantaneous change in the weight of the remaining feed in the trough.

[0033] Based on this, animal behavior segments within the feed trough area are identified by combining pre-trough trigger data. Pre-trough trigger signals typically appear as discrete pulses or state changes. Continuous trigger signals can be grouped into trigger event intervals by setting trigger duration or interval thresholds. Let the start time of a certain trigger event interval be... The end time is Then the time range corresponding to the triggering event interval is This range reflects a continuous behavioral process of the animal near the feed trough. It may include actual feeding behavior, or it may only correspond to the animal's brief stay, approach behavior after the feeding device is activated, or disturbance behavior that does not result in effective feeding.

[0034] For each triggered event interval, the weight of remaining feed in the trough at the start and end of the interval, as well as the total feeding input within the interval, are extracted from the unified time index running data. Based on the total feeding input, the net feed consumption caused by animal feeding within the triggered event interval is determined. The net feed consumption is used to characterize the reduction in the weight of remaining feed in the trough at the end of the interval compared to the weight of the original feed in the trough at the start of the interval and the newly added feed during the interval after animal feeding.

[0035] Furthermore, to avoid biases in the determination process due to different durations of trigger event intervals, a normalization process using net time consumption per unit time is introduced. Let the duration of the trigger event interval be... The corresponding net consumption per unit time is expressed as: ; This quantity is used to reflect the actual food consumption intensity per unit time. Compared with a single total consumption, it is more suitable for comparing between trigger event intervals of different lengths.

[0036] After obtaining the net consumption per unit time, it is compared with a preset consumption threshold. This preset consumption threshold can be determined based on different breeding stages, historical operating data of the pen, or historical feed consumption benchmarks, for example, based on the average consumption level of the target pen under stable feeding conditions. When the consumption threshold is exceeded, it is considered that there is continuous and effective feeding behavior within the triggering event interval, and it is marked as a valid feeding event interval; when If the value is less than or equal to the preset consumption threshold, the triggering event interval is considered to have not generated effective food consumption, and is thus determined as an invalid triggering event interval and removed in subsequent processing.

[0037] Furthermore, the proportion of invalid triggering event intervals within the corresponding time index is determined as the invalid triggering ratio. Let the total number of triggering event intervals within the corresponding time index be... The number of invalid triggering event intervals is The invalid trigger ratio is then expressed as: ; in, This represents the proportion of invalid triggers under the corresponding time index. This proportion characterizes the degree of triggering caused by feeding disturbances, short-term proximity, or invalid behavior within the current time range, and is used as a correction feature in the subsequent construction of the demand feature sequence for each column. This approach reduces the impact of frequent invalid triggers on subsequent LSTM feed demand predictions, preventing the model from mislearning feeding disturbances as changes in actual feed intake.

[0038] Through the above screening process, the original trigger event intervals are divided into a set of valid feeding event intervals and a set of invalid trigger event intervals. Valid feeding event intervals are typically characterized by a continuous decrease in the weight of remaining feed in the trough, with the decrease rate matching the time range. Invalid trigger event intervals, on the other hand, often occur shortly after the feeding device is activated, or when animals briefly remain without sustained feeding. This screening process allows subsequent analysis to move beyond relying on a single trigger signal and instead make a combined judgment based on changes in the weight of remaining feed in the trough and the feeding input information.

[0039] After obtaining the effective feeding event intervals, the effective feeding events are linked and organized in chronological order to form an effective feeding event sequence. The effective feeding event sequence is based on time, with each event node corresponding to an effective feeding event interval that meets the consumption judgment conditions. The time interval relationship will be directly used to calculate the subsequent feeding interval length, residual feed decay slope, and feeding rhythm deviation.

[0040] In another possible implementation, field observation revealed that animals typically experience short pauses during continuous feeding. These pauses may be identified by the feeder trigger signal as multiple independent trigger event intervals, leading to over-segmentation of effective feeding events. Based on this, the interval between adjacent trigger event intervals... Make a judgment when And the difference in net consumption per unit time between adjacent triggering event intervals satisfies At that time, the two trigger event intervals are merged to form a new candidate feeding event interval. This indicates the preset merging threshold. This indicates the preset net consumption continuity threshold.

[0041] The merged candidate feeding event intervals are recalculated for changes in net feed surplus and net consumption per unit time, and then used in subsequent validity assessments. This process avoids incorrectly segmenting continuous feeding behavior into multiple discrete events, making the resulting valid feeding event sequence more closely resemble the actual feeding process and improving the stability of subsequent pen demand feature sequence construction and LSTM feed demand prediction.

[0042] It should be noted that the method for identifying the trigger event interval is not limited to a specific implementation. For example, in an infrared triggering scheme, the interval range can be determined based on the duration of the trigger signal; in a visual recognition scheme, the trigger event interval can be determined by the time period during which the animal's head enters the feed trough area in the target detection results. As long as the time range corresponding to the feed trough interaction behavior can be obtained, it can be used for the implementation of this invention. Similarly, the calculation of the change in net remaining feed can also be performed at different sampling accuracies, as long as the weight data of the remaining feed in the feed trough and the feeding flow rate data have a corresponding relationship in the time dimension.

[0043] Step S3: Based on the effective feeding event sequence and the remaining food weight data in the feed trough, calculate the feeding interval length and the remaining food decay slope, and construct a stall demand feature sequence based on the feeding interval length, the remaining food decay slope, the stall inventory data, and the environmental data.

[0044] Specifically, based on the effective feeding event sequence and the remaining food weight data in the feed trough, the feeding interval length and the remaining food decay slope are calculated, including: traversing each effective feeding event in the effective feeding event sequence in chronological order, extracting the occurrence time of two adjacent effective feeding events, and calculating the time interval between the two adjacent effective feeding events as the feeding interval length under the corresponding time index; constructing a remaining food weight sequence corresponding to the sampling time based on the remaining food weight data in the feed trough, and calculating the ratio of the remaining food weight difference to the corresponding time interval for each adjacent sampling time to obtain the remaining food decay slope for each sampling interval; performing correlation processing on the feeding interval length and the remaining food decay slope based on the same time index to form the basic data of feeding rhythm for constructing the demand feature sequence of the feeding area.

[0045] Furthermore, a stall demand feature sequence is constructed based on the feeding interval length, the residual feed decay slope, the stall inventory data, and the environmental data. This includes: normalizing the feeding interval length and the corresponding residual feed decay slope after time alignment to obtain the interval representation value and slope representation value for each time index; constructing a feeding rhythm deviation value to represent the degree of change in the actual feeding rhythm based on the deviation relationship between the interval representation value and the slope representation value; and concatenating the feeding rhythm deviation value, the number of effective feeding events, the feeding interval length, the residual feed decay slope, the stall inventory data, and the environmental data according to the time index to generate a stall demand feature sequence corresponding to the target stall.

[0046] In this embodiment of the invention, step S3 quantifies the actual feeding rhythm of the target pen based on the effective feeding event sequence, and converts this feeding rhythm, along with pen inventory data and environmental data, into a pen demand feature sequence that can be input into a preset LSTM feed demand prediction model. Unlike directly using historical feeding amounts or feed trough residue for prediction, the pen demand feature sequence constructed in this step is derived from actual feeding behavior after being filtered by effective feeding events. Therefore, it can reduce the impact of invalid triggers, feeding disturbances, and short-term weight fluctuations on the subsequent prediction process.

[0047] Specifically, the sequence of valid feeding events is arranged in chronological order. Let the first... The time of occurrence of a valid foraging event is , No. The time of occurrence of a valid foraging event is The length of the feeding interval between two adjacent valid feeding events can be expressed as: ; in, This indicates the length of the feeding interval under the corresponding time index. This feeding interval length reflects the pause time between consecutive actual feeding behaviors in the target enclosure, and its numerical variation characterizes the tightness of the feeding rhythm. When When the length continues to shorten, it usually indicates that feeding behavior in the target area is becoming more concentrated; when If the number continues to increase, it may indicate a decrease in the willingness to feed or that the feeding process has entered an intermittent state.

[0048] Meanwhile, a sequence of remaining material weights in the trough is constructed based on the time-aligned remaining material weight data in step S1. The slope of the remaining material decay is calculated based on the change in the weight of the remaining material in the trough between adjacent sampling times. Let any adjacent sampling time be... and The corresponding weights of the remaining material in the trough are respectively and Then the slope of the residual material decay within this sampling interval can be expressed as:

[0049] in, Indicates the first The rate of decrease of remaining feed in the trough within each sampling interval. Since the previous steps have already isolated the influence of feeding input on the change in remaining feed through the net change in remaining feed, the feed decay slope here is mainly used to characterize the intensity of feed consumption. If there are instantaneous fluctuations in the feed weight data caused by animal touching the feed trough, short-term shaking, or sensor noise, the feed weight sequence can be analyzed before calculating the feed decay slope. Perform moving average or median filtering to reduce the impact of high-frequency interference on slope calculation.

[0050] Considering that the feeding interval length is derived from the sequence of effective feeding events, while the feed decay slope is derived from a continuous sampling sequence, and that the two are not entirely consistent in terms of temporal granularity, it is necessary to perform correlation processing on them based on the same time index. Specifically, this can be done within each feeding interval. Within this time range, the slope of residual food decay is statistically analyzed to obtain the average slope of residual food decay for the corresponding feeding interval. Let the number of sampling points within this feeding interval be... Then the average residual material decay slope can be expressed as: ; in, Indicates the length of the feeding interval The corresponding average residual material decay slope, This indicates the number of sampling points included in the statistics within the feeding interval. Through the above processing, the length of each feeding interval can be... The average residual material decay slope of the corresponding interval Establish a one-to-one correspondence to form the basic data of feeding rhythm used to construct the sequence of stall demand characteristics.

[0051] Furthermore, to ensure comparability between the feeding interval length and the residual feed decay slope, normalization is applied to both. Based on historical operational data or statistical results within a sliding time window for the target field, reference ranges for the feeding interval length and the average residual feed decay slope are determined. The minimum reference value for the feeding interval length is set as follows: The maximum reference value is Then the first The intermittent representation value under each time index can be expressed as: ; Similarly, let the minimum reference value for the average residual material decay slope be... The maximum reference value is Then the first The slope representation value under each time index can be expressed as: ; in, Indicates the first Intermittent representation values ​​under each time index Indicates the first The slope representation value under each time index. Through the above normalization process, the feeding interval length and residual feed decay slope of different dimensions can be mapped to a unified numerical range, so that they can be used for subsequent deviation relationship calculations. The above reference range can be derived from fixed settings or from real-time updated statistical results. As long as it can reflect the historical operating characteristics of the current target column, it can be used as the basis for normalization.

[0052] Obtaining intermittent characterization values and slope characterization value Next, a feeding rhythm deviation measure is constructed to describe the deviation relationship between the two. In actual feeding, shorter feeding intervals are usually accompanied by higher feeding intensity, while longer feeding intervals usually correspond to lower feeding intensity. When a significant deviation occurs between the interval representation value and the slope representation value, it usually indicates a change in the feeding rhythm of the target field. Based on this, a feeding rhythm deviation measure can be constructed using a deviation function, for example: ; in, Indicates the first Deviation in feeding rhythm under each time index. When the interval is small, it indicates a good match between the feeding interval length and the feed decay slope, and the target pen is in a relatively stable feeding state; when When the value increases, it indicates a shift between the length of the feeding interval and the slope of the remaining food decay, which may correspond to changes in the feeding rhythm caused by concentrated feeding, feeding decay, short pauses, or environmental disturbances.

[0053] In practical applications, the method for constructing the deviation from the feeding rhythm is not limited to the absolute difference form mentioned above; weighted relationships can also be introduced. For example: ; in, and The preset weights are used to adjust the contribution ratios of feeding interval length and feed decay slope to the deviation in feeding rhythm. The weight values ​​can be set based on the historical operating characteristics of different rearing stages, animal breeds, or target pens. Using a weighted approach can make the deviation in feeding rhythm more closely match actual feeding behavior, especially in scenarios where the target pen is more sensitive to changes in feeding interval or consumption intensity.

[0054] After obtaining the deviation from the feeding rhythm, it is concatenated with the number of effective feeding events, feeding interval length, feed decay slope, stall inventory data, and environmental data according to the same time index to form a stall demand feature sequence for the corresponding target stall. Let the first... The number of valid feeding events corresponding to each time index is The column data is Environmental data is Then, the feature vector of the column requirements under the corresponding time index can be represented as: ; If the environmental data includes multiple environmental parameters such as temperature, humidity, ammonia concentration, or light intensity, then This can be represented as a sub-vector of environmental features under the corresponding time index. Furthermore, by arranging the column demand feature vectors corresponding to each time index within the historical management period in chronological order, a column demand feature sequence can be obtained: ; in, This indicates the number of time indices involved in modeling within the historical management period. The column requirement feature sequence. As input to the pre-defined LSTM feed demand forecasting model, it is used to express the actual feeding rhythm changes, inventory changes, and environmental changes of the target pen during the historical management cycle, so that subsequent feed demand forecasting no longer relies solely on historical feed consumption or instantaneous feed surplus results, but is based on the continuous expression of effective feeding behavior and pen operating status.

[0055] Step S4: Input the demand feature sequence of the pen into the preset LSTM feed demand prediction model, output the feed demand prediction value of the target pen in the future management cycle, and generate feeding management parameters for the subsequent feeding process based on the feed demand prediction value.

[0056] Specifically, the process involves inputting the stall demand feature sequence into a preset LSTM feed demand prediction model and outputting the feed demand prediction value for the target stall in the future management cycle. This includes: inputting the stall demand feature sequence into the feature input layer of the preset LSTM feed demand prediction model and generating rhythm correction weights for corresponding time indices based on the feeding rhythm deviation and invalid trigger ratio; inputting the rhythm correction weights into the rhythm correction layer of the preset LSTM feed demand prediction model to correct the feature weights for the corresponding time indices in the stall demand feature sequence, resulting in a corrected stall demand feature sequence; inputting the corrected stall demand feature sequence into the LSTM time series prediction layer of the preset LSTM feed demand prediction model and outputting the latent state prediction result for the future management cycle; and inputting the latent state prediction result into the demand output layer of the preset LSTM feed demand prediction model and outputting the feed demand prediction value for the target stall in the future management cycle.

[0057] Furthermore, feeding management parameters for subsequent feeding processes are generated based on the predicted feed demand, including: determining the demand level of the target pen in the future management cycle based on the deviation between the predicted feed demand and the historical feed consumption benchmark value corresponding to the target pen; generating the planned feeding amount, feeding priority, supplementary feed demand, and feeding time allocation ratio corresponding to the target pen based on the predicted feed demand, the demand level, and the feed inventory data of the target farm; and using the planned feeding amount, the feeding priority, the supplementary feed demand, and the feeding time allocation ratio as feeding management parameters for subsequent feeding processes.

[0058] In this embodiment of the invention, the pen demand feature sequence generated in step S3 is input into the feature input layer of a preset LSTM feed demand prediction model. Let the pen demand feature sequence constructed within the historical management cycle be represented as: ; in, Indicates the first The feature vector corresponding to each time index. This indicates the number of time indices within the historical management period. The field demand feature vector includes at least the feeding rhythm deviation, the number of effective feeding events, the feeding interval length, the residual feed decay slope, field inventory data, and environmental data.

[0059] Considering that the stability of feeding behavior varies across different time indices, before inputting into the LSTM time series prediction layer, rhythm correction weights are further generated based on the feeding rhythm deviation and the proportion of invalid triggers for the corresponding time index. Let the... The feeding rhythm deviation corresponding to each time index is: The invalid trigger ratio is Then the rhythm correction weight under the corresponding time index can be expressed as: ; in, Indicates the first The rhythm correction weights corresponding to each time index and This represents the preset weighting coefficient, used to adjust the influence of feeding rhythm deviation and invalid trigger ratio in the weighting calculation. When the feeding rhythm deviation increases or the invalid trigger ratio rises, the rhythm correction weight under the corresponding time index will decrease, thereby reducing the impact of abnormal feeding behavior or false triggering behavior on subsequent prediction results.

[0060] After obtaining the rhythm correction weights, they are input into the rhythm correction layer of the preset LSTM feed demand prediction model to perform correction processing on the feature weights corresponding to the time indices in the pen demand feature sequence. Let the first... The original column requirement feature vector corresponding to each time index is: The corrected column requirement feature vector can then be expressed as: ; in, Indicates the first The corrected column demand feature vectors correspond to each time index. Through the above processing, the feature information formed under stable feeding conditions has higher weight in the model training and prediction process, while the feature information formed during abnormal fluctuations is appropriately suppressed, thereby improving the stability of feed demand prediction.

[0061] Furthermore, the corrected column demand feature sequence is input into the LSTM time series prediction layer of the preset LSTM feed demand prediction model. Because the LSTM structure has temporal memory capabilities, it can predict feed demand for future management cycles based on the trend of feeding rhythm changes within historical time windows. Let the first... The hidden state corresponding to each time index is: Cell state is The LSTM timing update process can then be represented as: ; in, Used to characterize the The time-series state features corresponding to each time index It is used to preserve information on historical feeding rhythm changes. As the time series input progresses, the model can gradually learn the changes in feed demand of the target column under different breeding stages, different environmental conditions, and different feeding rhythm states.

[0062] After completing the LSTM time series prediction, the final hidden state prediction result is input into the demand output layer to generate the feed demand forecast value for the target column in the future management cycle. Let the output function corresponding to the demand output layer be... Then, the forecast value of feed demand during the future management cycle can be expressed as: ; in, This indicates the projected feed demand for the target column during the next management cycle. This indicates the hidden state prediction result corresponding to the last time index of the historical management cycle.

[0063] Furthermore, after obtaining the feed demand forecast, feeding management parameters for subsequent feeding processes are generated based on this forecast. Specifically, firstly, the demand level corresponding to the target field is determined based on the deviation between the feed demand forecast and the historical feed consumption baseline. Let the historical feed consumption baseline corresponding to the target field be... The degree of demand deviation can then be expressed as: ; in, This indicates the degree of deviation of the current predicted demand in the target field from the historical material consumption baseline. When the demand continues to increase, it indicates that the current food consumption demand in the target area is higher than the historical average; when A continuous decrease indicates that current food demand is below the historical average.

[0064] Based on the degree of demand deviation and the preset demand level classification rules, the target field is divided into different levels, such as high demand level, normal demand level, or low demand level. Different demand levels correspond to different feeding control strategies. For example, under the high demand level, the planned feeding amount can be increased and the feeding priority can be raised; under the low demand level, the feeding frequency can be reduced and the amount of feeding per batch can be decreased to reduce the risk of leftover material accumulation.

[0065] Furthermore, by combining the current feed inventory data of the target farm, the planned feeding amount, feeding priority, supplementary feed requirement, and feeding time allocation ratio for the corresponding target column are generated. Let the current available feed inventory of the target farm be... The planned feeding amount corresponding to the target column can be expressed as: ; in, This indicates the planned feeding amount corresponding to the target field. This represents the allocation coefficient, which is dynamically adjusted based on inventory status and demand levels. When the target farm has sufficient inventory, the allocation coefficient can be appropriately increased; when inventory is tight, differentiated allocation can be implemented for different pens based on feeding priorities.

[0066] Furthermore, based on demand levels and historical feeding rhythm distribution, the feeding periods within the future management cycle are divided. Let the future management cycle be divided into... During the feeding period, the first feeding period... The allocation ratio of each feeding period can be expressed as: ; in, Indicates the first The feeding time allocation ratio corresponding to each feeding time period This indicates the model's prediction of the phased feed demand within the corresponding time period. Using this method, subsequent feeding processes can move away from fixed-time, uniform feeding and instead dynamically adjust the feeding rhythm based on the actual feed intake of the target pen at different time periods.

[0067] Ultimately, the planned feeding amount, feeding priority, supplementary feed demand, and feeding time allocation ratio are collectively output as feeding management parameters for subsequent feeding processes. These feeding management parameters can be directly sent to automatic feeding equipment for feeding control, or used as a scheduling basis in the aquaculture management platform for optimizing feed distribution among multiple pens. This approach transforms the feeding control process from a traditional passive adjustment based on instantaneous feed surplus feedback to an active adjustment based on actual feeding rhythms and temporal demand predictions, thereby improving the stability and adaptability of the feeding process.

[0068] like Figure 2 As shown, the preset LSTM feed demand prediction model includes a feature input layer, a rhythm correction layer, an LSTM time series prediction layer, a demand output layer, and a rhythm correction weight generation module. The feature input layer receives the demand feature sequence from the field. Each column requires a feature vector. Including deviation from feeding rhythm Number of effective feeding events Feeding interval length Residual material decay slope Column storage data and environmental data .

[0069] The rhythm correction weight generation module is based on the deviation of the feeding rhythm. and invalid trigger ratio Calculate rhythm correction weights and will Input rhythm correction layer. Rhythm correction layer according to... The feature vectors of field demand under the corresponding time index are weighted and corrected to obtain the corrected feature sequence of field demand. Subsequently, the LSTM time series prediction layer receives the corrected feature sequence of field demand. And pass the hidden state between adjacent time indices. and cell state This preserves the impact of historical feeding rhythms on subsequent demand changes. The demand output layer performs fully connected mapping and activation processing on the hidden states output by each time index to obtain the predicted feed demand for the future management cycle. This predicted value is used to generate the feeding management parameters for the target field.

[0070] In another possible implementation, when the target field switches from the first feed batch to the second feed batch, the switching time is recorded, and the changes in the number of effective feeding events, feeding interval length, and residual feed decay slope are continuously calculated within a preset observation window after the switch. If the number of effective feeding events decreases and the feeding interval length increases, while the residual feed decay slope is lower than the historical baseline before the switch, a feed palatability offset is generated; if the above parameters gradually return to the historical baseline range before the switch within multiple consecutive time indices, a feed palatability stabilization amount is generated.

[0071] The feed palatability offset and palatability recovery amount are added to the pen demand feature sequence and input into a preset LSTM feed demand prediction model. This allows the model to distinguish between a decrease in actual demand and short-term feed intake fluctuations caused by feed adaptation periods after batch changes. Based on this processing, when generating subsequent feeding management parameters, it is possible to avoid prematurely reducing the planned feeding amount due to a short-term decrease in palatability, and to promptly adjust the supplementary feeding demand and feeding time allocation ratio after feed intake recovers.

[0072] Example: The method of this invention is described using a fattening pen in a large-scale pig farm as an application scenario. This farm has multiple independent pens, each equipped with an automatic feeding device, a feed trough weighing sensor, a feeding flow meter, and a feed trough trigger sensor. Taking target pen A as an example, the average number of pigs in pen A is 38. The feed trough weighing sensor collects the weight of remaining feed in the trough, the feeding flow meter records the feeding input for each feeding event, and the feed trough trigger sensor records the triggering of animals entering the feed trough area. The system uses a preset sampling period of 30 seconds, performs time alignment processing on the feed trough weight data, feeding flow data, and feed trough trigger data, and simultaneously records pen inventory data and environmental data.

[0073] like Figure 3 As shown, within the historical management period from 0 min to 360 min, the overall weight of the remaining feed in the feed trough of target column A decreased from approximately 50.2 kg to approximately 40.8 kg, with four significant feeding input events occurring during this period. Feeding event F1 occurred around 30 min, with a feeding input of 1.8 kg; feeding event F2 occurred around 120 min, with a feeding input of 2.4 kg; feeding event F3 occurred around 210 min, with a feeding input of 1.6 kg; and feeding event F4 occurred around 295 min, with a feeding input of 2.0 kg. Since feeding inputs can cause a short-term rebound in the weight of the remaining feed in the feed trough, this embodiment does not directly use the decrease in the weight of the remaining feed in the feed trough as the basis for judging feeding demand. Instead, it first deducts the feeding input to obtain the net change in remaining feed. The operational data for some time points are shown in Table 1.

[0074] Table 1. Example of partial operational data within the historical management period for target field A.

[0075] After obtaining unified time index operation data, the system identifies trigger event intervals by combining them with pre-slot trigger data and calculates the net consumption per unit time for each interval. The system sets a preset consumption judgment threshold of 0.06 kg / min. When the net consumption per unit time corresponding to a trigger event interval reaches or exceeds this threshold, it is determined as a valid feeding event; when the net consumption per unit time is lower than this threshold, it is determined as an invalid trigger event. Figure 4 As shown, the net consumption rate in intervals E1, E3, and E5 significantly exceeded the judgment threshold and were thus determined as valid feeding events. Although there were triggering events at the feeder in intervals E2 and E4, the net consumption rate did not reach the judgment threshold, and therefore they were determined as invalid triggering events. This result indicates that E2 and E4 are more likely to correspond to short-term animal approach, feeding disturbance, or behavior that did not result in actual feeding. The corresponding triggering event screening results are shown in Table 2.

[0076] Table 2 Example of Trigger Event Range Filtering Results

[0077] As shown in Table 2, the net consumption per unit time for E1, E3, and E5 is all higher than 0.06 kg / min; the net consumption per unit time for E2 and E4 is lower than 0.06 kg / min. This forms the valid feeding event sequence for target field A, namely {E1, E3, E5}. Simultaneously, the system calculates the invalid trigger ratio based on the number of invalid trigger events and the total number of trigger events. In this embodiment, there are 2 invalid trigger events and 5 total trigger events; therefore, the invalid trigger ratio is 0.40.

[0078] After obtaining the valid feeding event sequence, the system further calculates the feeding interval length and the feed decay slope. Taking E1 and E3 as examples, their center times are approximately 30 min and 150 min, respectively, corresponding to a feeding interval length of approximately 120 min; the center times of E3 and E5 are approximately 150 min and 270 min, respectively, also corresponding to a feeding interval length of approximately 120 min. Combined with... Figure 3 The system extracts the residual weight curve of the feed trough, the slope of residual weight decay within the corresponding interval, and constructs the feeding rhythm deviation after normalization. The feeding rhythm deviation, the number of effective feeding events, the length of feeding interval, the residual weight decay slope, the stock data of the target stock, the environmental data, and the proportion of invalid triggers are concatenated according to the time index to form the stock demand feature sequence of target stock A.

[0079] In this embodiment, the aforementioned pen demand feature sequence is input into a preset LSTM feed demand prediction model. The model's rhythm correction layer generates rhythm correction weights based on the deviation of the feeding rhythm and the proportion of invalid triggers, reducing the feature contributions of time periods corresponding to E2 and E4, while retaining the true feeding features corresponding to E1, E3, and E5. Subsequently, the LSTM time series prediction layer outputs feed demand prediction results for future management cycles based on the corrected pen demand feature sequence.

[0080] like Figure 5 As shown, the model outputs the cumulative predicted demand curve for the next 5 hours. The predicted demand at 30min, 60min, 90min, 120min, 150min, 180min, 210min, and 240min are 38.7kg, 57.6kg, 77.8kg, 96.5kg, 108.8kg, 118.6kg, 124.3kg, and 128.4kg, respectively; the corresponding historical cumulative consumption baselines are 35.6kg, 52.8kg, 69.1kg, 83.2kg, 94.6kg, 104.8kg, 112.5kg, and 119.6kg, respectively. Therefore, the predicted demand for the next 4 hours is 128.4kg, the historical baseline is 119.6kg, the predicted increase is 8.8kg, and the growth rate is approximately 7.4%.

[0081] Based on the above prediction results, target stall A is identified as having a high demand state, and feeding management parameters are generated. Specifically, the planned feeding amount for the next 4 hours is set to be between approximately 128.4 kg and 131 kg, the supplemental feed requirement is set to 18 kg, the feeding priority is set to level one, and based on the increased feeding trend of target stall A in the later period, the feeding time allocation ratio during nighttime or low-temperature periods is set to 32%. This embodiment demonstrates that the method of the present invention can first eliminate invalid triggering events and then input the actual feeding rhythm into the LSTM feed demand prediction model, thereby making the feeding management parameters more closely match the actual feeding needs of the target stall.

[0082] like Figure 6As shown, this invention provides an LSTM-based dynamic adjustment system for feed feeding. The system includes: a data acquisition unit for acquiring operational data of a target pen during a historical management period, the operational data including at least the weight of leftover feed in the feed trough, feeding flow rate data, pre-feed trigger data, pen inventory data, and environmental data; a time series determination unit for constructing an effective feeding event sequence characterizing actual feeding behavior based on the leftover feed weight data, feeding flow rate data, and pre-feed trigger data; a demand sequence determination unit for calculating the feeding interval length and the leftover feed decay slope based on the effective feeding event sequence and the leftover feed weight data, and constructing a pen demand feature sequence based on the feeding interval length, the leftover feed decay slope, the pen inventory data, and the environmental data; and a prediction unit for inputting the pen demand feature sequence into a preset LSTM feed demand prediction model, outputting a predicted feed demand value for the target pen in a future management period, and generating feeding management parameters for subsequent feeding processes based on the predicted feed demand value.

[0083] The present invention also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described LSTM-based dynamic adjustment method for feed feeding.

[0084] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor A01, a network interface A02, memory (not shown), and a database (not shown) connected via a system bus. The processor A01 provides computing and control capabilities. The memory includes internal memory A03 and a non-volatile storage medium A04. The non-volatile storage medium A04 stores an operating system B01, a computer program B02, and a database (not shown). The internal memory A03 provides an environment for the operation of the operating system B01 and the computer program B02 stored in the non-volatile storage medium A04. The network interface A02 is used for communication with external terminals via a network connection. When the computer program B02 is executed by the processor A01, it implements a dynamic adjustment method for the feed feeding process based on LSTM.

[0085] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0086] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details described above. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe the various possible combinations.

[0087] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the embodiments of the present invention, they should also be regarded as the content disclosed by the embodiments of the present invention.

Claims

1. A method for dynamic adjustment of feed feeding process based on LSTM, characterized in that, The method includes: Obtain the operating data of the target field within the historical management period. The operating data includes at least the weight of the remaining material in the trough, the feeding flow rate, the trigger data in front of the trough, the field inventory data, and the environmental data. Based on the remaining material weight data in the feed trough, the feeding flow rate data, and the trigger data in front of the feed trough, an effective feeding event sequence is constructed to characterize the actual feeding behavior. Based on the effective feeding event sequence and the remaining food weight data in the feed trough, the feeding interval length and the remaining food decay slope are calculated, and a stall demand feature sequence is constructed based on the feeding interval length, the remaining food decay slope, the stall inventory data, and the environmental data. The demand feature sequence of the pen is input into a preset LSTM feed demand prediction model, which outputs the feed demand prediction value of the target pen in the future management cycle, and generates feeding management parameters for subsequent feeding processes based on the feed demand prediction value.

2. The method for dynamic adjustment of feed feeding process based on LSTM according to claim 1, characterized in that, Retrieve the operational data of the target field within the historical management period, including: During the historical management cycle, the remaining feed weight data of the feed trough corresponding to the target pen, the feeding flow rate data of the feeding device, the trigger data of animals in the feed trough area, the pen inventory data of the target pen, and the environmental data of the breeding environment corresponding to the target pen are collected synchronously according to the preset sampling cycle. Based on the preset sampling period, time alignment processing is performed on the remaining material weight data in the trough, the feeding flow rate data, the trough trigger data, the column storage data, and the environmental data to form running data corresponding to the same time index.

3. The method for dynamic adjustment of feed feeding process based on LSTM according to claim 1, characterized in that, Based on the remaining feed weight data in the feed trough, the feeding flow rate data, and the trigger data in front of the feed trough, an effective feeding event sequence is constructed to characterize actual feeding behavior, including: Based on the changes in the weight of leftover food and the amount of feed input at each sampling time in the operational data, the net change in leftover food within each sampling interval is calculated, and the net change in leftover food is used as the basis for judging the consumption of feed. Based on the pre-tank trigger data, the trigger event intervals corresponding to each trigger time are identified, and the corresponding net remaining material change is extracted for each trigger event interval. It is then determined whether the net remaining material change meets the preset consumption judgment condition in order to perform filtering processing on the trigger event intervals. The trigger event interval that meets the preset consumption judgment condition is determined as a valid feeding event, and each valid feeding event is associated and organized in chronological order to generate a valid feeding event sequence. The valid feeding event sequence is used to construct the column demand feature sequence.

4. The method for dynamic adjustment of feed feeding process based on LSTM according to claim 3, characterized in that, For each trigger event interval, extract the corresponding net remaining material change amount, determine whether the net remaining material change amount meets the preset consumption judgment condition, and perform filtering processing on the trigger event interval, including: Based on the weight of remaining material at the start time and the weight of remaining material at the end time of each trigger event interval, and the total amount of feeding input within the corresponding interval, calculate the net change in remaining material for the corresponding trigger event interval; The net consumption per unit time is constructed based on the change in net remaining material and the duration of the triggering event interval, and compared with a preset consumption judgment threshold to determine whether the corresponding triggering event interval meets the preset consumption judgment condition. Triggering event intervals that meet the preset consumption judgment conditions are determined as valid feeding event intervals and used to construct the valid feeding event sequence. Triggering event intervals that do not meet the preset consumption judgment conditions are determined as invalid triggering event intervals. An invalid triggering ratio is generated based on the proportion of invalid triggering event intervals in the corresponding time index. The invalid triggering ratio is used to construct the column demand feature sequence.

5. The method for dynamic adjustment of feed feeding process based on LSTM according to claim 1, characterized in that, Based on the effective feeding event sequence and the remaining food weight data in the feed trough, the feeding interval length and the remaining food decay slope are calculated, including: The effective feeding events in the sequence of effective feeding events are traversed in chronological order. The occurrence times of two adjacent effective feeding events are extracted, and the time interval between the two adjacent effective feeding events is calculated as the feeding interval length under the corresponding time index. Based on the remaining material weight data of the trough, a sequence of remaining material weights corresponding to the sampling time is constructed, and the ratio of the difference in remaining material weight to the corresponding time interval is calculated for each adjacent sampling time to obtain the remaining material attenuation slope of each sampling interval. Based on the same time index, the length of the feeding interval and the slope of the remaining feed decay are correlated to form the basic data of the feeding rhythm for constructing the demand feature sequence of the stall.

6. The method for dynamic adjustment of feed feeding process based on LSTM according to claim 5, characterized in that, Based on the feeding interval length, the residual feed decay slope, the stall inventory data, and the environmental data, a stall demand feature sequence is constructed, including: Normalization is performed on the feeding interval length and the corresponding residual food decay slope after time alignment to obtain the interval characterization value and slope characterization value for each time index. Based on the deviation relationship between the intermittent characterization value and the slope characterization value, a feeding rhythm deviation value is constructed to characterize the degree of change in the actual feeding rhythm. The feeding rhythm deviation, number of effective feeding events, feeding interval length, residual feed decay slope, stall inventory data, and environmental data are concatenated according to time index to generate a stall demand feature sequence corresponding to the target stall.

7. The method for dynamic adjustment of feed feeding process based on LSTM according to claim 6, characterized in that, The feed demand feature sequence of the target pen is input into a preset LSTM feed demand prediction model, and the predicted feed demand value of the target pen in the future management cycle is output, including: The column demand feature sequence is input into the feature input layer of the preset LSTM feed demand prediction model, and rhythm correction weights corresponding to the time index are generated based on the feeding rhythm deviation and invalid trigger ratio. The rhythm correction weights are input into the rhythm correction layer of the preset LSTM feed demand prediction model to correct the feature weights of the corresponding time index in the stall demand feature sequence, so as to obtain the corrected stall demand feature sequence. The modified column demand feature sequence is input into the LSTM time series prediction layer of the preset LSTM feed demand prediction model, and the hidden state prediction results in the future management cycle are output. The hidden state prediction results are input into the demand output layer of the preset LSTM feed demand prediction model, and the predicted feed demand value of the target column in the future management cycle is output.

8. The method for dynamic adjustment of feed feeding process based on LSTM according to claim 1, characterized in that, Feeding management parameters for subsequent feeding processes are generated based on the predicted feed demand values, including: Based on the deviation between the predicted feed demand and the historical feed consumption baseline value corresponding to the target pen, the demand level of the target pen in the future management cycle is determined. Based on the feed demand forecast, the demand level, and the feed inventory data of the target farm, the planned feeding amount, feeding priority, supplementary feed demand, and feeding time allocation ratio for the corresponding target column are generated. The planned feeding amount, the feeding priority, the supplementary feeding requirement, and the feeding time period allocation ratio are used as feeding management parameters for subsequent feeding processes.

9. A dynamic adjustment system for feed feeding process based on LSTM, characterized in that, The system includes: The data acquisition unit is used to acquire the operating data of the target field within the historical management period. The operating data includes at least the weight data of the remaining material in the trough, the feeding flow rate data, the trigger data in front of the trough, the field storage data, and the environmental data. The time series determination unit is used to construct an effective feeding event sequence to characterize the actual feeding behavior based on the weight data of the remaining material in the feed trough, the feeding flow rate data, and the trigger data in front of the feed trough. The demand sequence determination unit is used to calculate the feeding interval length and the residual material decay slope based on the effective feeding event sequence and the residual material weight data in the feed trough, and to construct a stall demand feature sequence based on the feeding interval length, the residual material decay slope, the stall inventory data and the environmental data. The prediction unit is used to input the demand feature sequence of the pen into a preset LSTM feed demand prediction model, output the feed demand prediction value of the target pen in the future management cycle, and generate feeding management parameters for subsequent feeding processes based on the feed demand prediction value.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the LSTM-based dynamic adjustment method for feed feeding as described in any one of claims 1-8.

Images