Intelligent variable frequency feeding control method and system based on negative pressure air conveying

Abstract

The present application relates to the field of aquatic animal breeding, and more particularly, the present application relates to an intelligent variable frequency feeding control method and system based on negative pressure air feeding, the method comprising: inputting all historical variable frequency feeding data into a GBDT model to be trained, setting an initial learning rate during the training process, calculating the residual optimization degree after each decision tree partition, taking the product of the square root of the residual optimization degree and the initial learning rate as the adaptive learning rate of the next time, and optimizing the training process of the model according to the adaptive learning rate; inputting real-time variable frequency feeding data into the trained GBDT model, outputting a feeding frequency prediction result, and controlling the intelligent variable frequency feeding of the feeding equipment according to the prediction result. The present application closely links the feed quality and the model training effect, improves the scientificity and rationality of the feeding decision, and ensures that the breeding organisms obtain high-quality feed feeding.

Description

Technical Field

[0001] This invention relates to the field of aquatic animal farming. More specifically, this invention relates to an intelligent variable frequency feeding control method and system based on negative pressure air delivery. Background Technology

[0002] Against the backdrop of the current profound transformation of the aquaculture industry towards large-scale, intensive, and sustainable development, traditional manual feeding methods have revealed significant drawbacks—high feed loss rates, high labor input, and a growing risk of water pollution. By precisely controlling the feeding frequency, it is possible to ensure that feed is fully consumed by fish or other aquatic animals, reducing feed residue and waste in the water.

[0003] In the process of artificial variable frequency feeding of aquatic products, the timing of feeding mainly relies on human judgment. However, due to the uncertainty of human judgment, such as water quality changes and human subjectivity, the daily feeding judgment is inaccurate, which in turn affects the accuracy of the feeding frequency. Summary of the Invention

[0004] To address the aforementioned technical problem of how to achieve precise control over feeding frequency and thereby optimize feed utilization efficiency, this invention provides solutions in the following aspects.

[0005] In the first aspect, the intelligent variable frequency feeding control method based on negative pressure air delivery includes:

[0006] Collect and preprocess the variable frequency feeding data for each day of the preset historical number of days, input all the historical variable frequency feeding data into the GBDT model to be trained, calculate the adaptive learning rate based on the historical variable frequency feeding data during the training process, optimize the training process of the model based on the adaptive learning rate, and then complete the training of the GBDT model.

[0007] Real-time variable frequency feeding data is input into the trained GBDT model, and the feeding frequency prediction result is output. The feeding equipment is controlled to intelligently perform variable frequency feeding based on the prediction result.

[0008] The adaptive learning rate is calculated as follows:

[0009] Set the initial learning rate;

[0010] After each decision tree split, the residual optimization degree after the current split is calculated;

[0011] The product of the square root of the residual optimization degree and the initial learning rate is used as the adaptive learning rate for the next iteration.

[0012] Preferably, the variable frequency feeding data includes: daily water temperature, water yield, dissolved oxygen content, and the remaining feed amount before feeding the previous day.

[0013] Preferably, before calculating the degree of optimization of the residuals, the method further includes:

[0014] Calculate the feed remaining index before feeding on any past day; calculate the feed remaining quantity coordination degree on any past day.

[0015] The product of the feed surplus index and the synergy of feed surplus is used as the feeding quality index for any past day.

[0016] Preferably, the process of obtaining the optimality of the residual after the current partitioning includes:

[0017] After a split, the absolute value of the residuals of the data collected each day in the past is obtained. After at least four decision tree splits, the feeding quality index of each day in the past is normalized. The normalized feeding quality index is multiplied by the absolute value of the residual of the current split to obtain the weighted absolute value of the residual for the corresponding day. All weighted absolute values ​​of the residuals are summed to obtain the sum of the weighted residuals after the current split. The ratio of the preset acceptable residual to the sum of the weighted residuals is used as the quality residual ratio.

[0018] The calculation is based on the three consecutive partitions before the current partition. The average residual ratio between two adjacent partitions is calculated, and the average residual ratios are summed and averaged to obtain the average residual change rate.

[0019] The product of the quality residual ratio after the current partitioning and the average residual change rate is used as the degree of residual optimization after the current partitioning.

[0020] Preferably, the process of obtaining the feed residue index includes:

[0021] The feed surplus for each past day is converted using an exponential function to obtain the corresponding feed surplus index for that day.

[0022] Preferably, the process of obtaining the feed residue index includes:

[0023] Calculate the ratio of the remaining feed amount for each past day to the total feed amount for that day, and subtract the ratio from 1 to get the remaining feed amount for that day.

[0024] Preferably, the process of obtaining the feed surplus synergy includes:

[0025] Select any past day as the target day, and calculate the Euclidean distance between the target day and the feeding data collected from any other day; calculate the mean of the sum of the ratios of the difference in feed balance to the Euclidean distance between the target day and all other days, and obtain the feed balance coordination degree of the target day.

[0026] Secondly, an intelligent variable frequency feeding control system based on negative pressure air delivery includes: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the intelligent variable frequency feeding control method based on negative pressure air delivery described in any one of the claims is implemented.

[0027] The present invention has the following effects:

[0028] 1. By collecting data on various factors affecting feeding (such as water temperature, water yield, dissolved oxygen, and feed residue), and using the GBDT model for analysis and prediction, the feeding frequency can be accurately determined based on the actual aquaculture situation. This avoids the blindness and errors of manual experience-based feeding, ensuring that feed is added appropriately to meet the growth needs of aquaculture organisms while reducing feed waste.

[0029] 2. During GBDT model training, the training process is optimized by calculating an adaptive learning rate. This dynamic adjustment of the learning rate, compared to a fixed learning rate, allows the model to converge quickly in the early stages of training and adjust parameters more finely as it approaches the optimal solution. This avoids missing the optimal solution due to an excessively large learning rate or causing excessively long training times due to an excessively small learning rate, thereby improving the model's training efficiency and performance. Furthermore, by comprehensively considering the feed residue index and the feed surplus synergy degree to calculate the feeding quality index, the model not only focuses on the residuals themselves during training but also incorporates the actual feeding quality, further considering the quality and trends of historical data. This allows for a more accurate evaluation of the effect of each decision tree split, providing a more reliable basis for adjusting the adaptive learning rate and further improving the model's training effect. Attached Figure Description

[0030] Figure 1 This is a flowchart of steps S1-S2 in the intelligent variable frequency feeding control method based on negative pressure air delivery according to an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0032] Reference Figure 1 The intelligent variable frequency feeding control method based on negative pressure air delivery includes steps S1-S2, as follows:

[0033] S1: Collect and preprocess the variable frequency feeding data for each of the preset historical days, input all the historical variable frequency feeding data into the GBDT model to be trained, calculate the adaptive learning rate based on the historical variable frequency feeding data during the training process, optimize the training process of the model based on the adaptive learning rate, and then complete the training of the GBDT model.

[0034] In one embodiment, a fixed sampling time and location are determined each day to ensure data consistency and comparability. For example, data is collected from the same body of water every day at 8:00 AM.

[0035] The collected data includes: water temperature (measured using a thermometer), water production (recorded using a production metering device), dissolved oxygen (measured using a dissolved oxygen meter), and feed surplus from the previous day (obtained by weighing the remaining feed from the previous day before feeding).

[0036] Feeding data for each day of the past six months was collected using the methods described above. Further, the data underwent preprocessing, including data cleaning, integration, format conversion, and standardization.

[0037] In the process of artificial variable frequency feeding in aquaculture, feeding decisions mainly rely on human judgment. Due to the inherent uncertainty of human judgment, the daily feeding frequency may be inaccurate, resulting in inconsistent quality of the collected aquatic feeding data. By acquiring data quality indicators, the quality of daily feeding data can be quantitatively assessed, reducing the impact of inaccurate human judgment on feeding frequency and overall data quality.

[0038] If the amount of feed remaining on a particular day is excessive, it means that the feeding frequency may be inaccurate, possibly due to insufficient feeding or improper feeding timing. In addition, if the amount of feed remaining on a particular day differs significantly from the amount of feed remaining on other days with similar data, it indicates that the data collected on that day is inconsistent with other data and has low data quality.

[0039] In one embodiment, the feed surplus of each past day is transformed using an exponential function to obtain the feed surplus index for the corresponding day. For example, the feed surplus index for day i satisfies the following relationship:

[0040] S i =exp(-m i )

[0041] In the formula, S i Let m be the feed residue index for day i. i Let be the feed balance on day i, and exp() be an exponential function with base e. The higher the feed balance, the smaller the feed surplus index.

[0042] In one embodiment, any past day is selected as the target day, and the Euclidean distance between the target day and the feeding data collected from any other day is calculated. The mean of the sum of the ratios of the difference in feed balance between the target day and all other days to the Euclidean distance is calculated to obtain the feed balance synergy of the target day. For example, if day i is selected as the target day, the feed balance synergy of day i satisfies the following relationship:

[0043]

[0044] In the formula, X i Let m be the feed balance synergy on day i, N be the total number of days for which data was collected, and m be the feed balance synergy on day i. i Let m be the remaining feed on day i. k Let D(i,k) be the feed balance on day k, and D(i,k) be the Euclidean distance between the feeding data on day i and day k. exp() is an exponential function with base e. The greater the difference between the feed balance on day i and the feed balance on other days, the lower the consistency of the feed balance on day i, and the lower the data quality.

[0045] In another embodiment, the ratio of the remaining feed amount for each past day to the total feed amount given on that day is calculated, and the difference between 1 and this ratio is taken as the feed surplus index for that day. For example, the feed surplus index for day i satisfies the following relationship:

[0046]

[0047] In the formula, S i Let m be the feed residue index for day i. i M represents the remaining feed on day i. i This represents the total feed intake on day i. This operation directly reflects the proportion of feed surplus to the total feed intake for that day, and more intuitively shows the relative extent of feed surplus and the amount of feed given that day.

[0048] Furthermore, the product of the feed surplus index and the feed surplus synergy obtained from the above calculation is used as the feeding quality index for day i. Similarly, the feeding quality index corresponding to all days collected can be obtained.

[0049] Considering that the quality of data collected during aquaculture data acquisition may vary due to equipment malfunctions, environmental interference, and other factors, the GBDT (Gradient Boosting Decision Tree) algorithm can differentiate historically collected feeding data by incorporating data quality indicators. During model construction, it gives more consideration to data with higher quality indicators and reduces the impact of data with lower quality indicators on the model, thereby improving the model's accuracy and reliability.

[0050] In the GBDT algorithm, the learning rate is a crucial parameter during iteration, affecting the model's convergence speed and accuracy. Both excessively high and low learning rates can cause problems; too high a learning rate can lead to oscillations and inaccurate model construction, while too low a rate increases computational cost. This control method dynamically adjusts the learning rate by calculating the residual optimization degree, allowing the learning rate to adapt to the model's training progress, thereby obtaining a more accurate prediction model and improving the precision of feeding frequency prediction.

[0051] In one embodiment, firstly, an initial learning rate and a maximum number of iterations are set, and a base value (such as 0.1) is selected as the starting point for the learning rate through cross-entropy verification or experience.

[0052] Then, in each iteration, the following process is executed:

[0053] 1. Calculate the current residual, that is, for all samples, calculate the difference (residual) between the current model prediction value and the true value, and use it as the fitting target for the next tree.

[0054] 2. Fitting a decision tree involves training a classification and regression tree based on the residual data, and determining the partitioning of the data space through feature selection and split points.

[0055] 3. Evaluate the partitioning effect. Calculate the residual optimization degree to measure the contribution of the current partitioning to reducing error.

[0056] In one embodiment, after each partition, the absolute value of the residuals of all feeding data is calculated. After at least four decision tree partitions (because the model has undergone at least three partitions, and the trend of residual change begins to stabilize), the feeding quality index for each past day is normalized. The normalized feeding quality index is multiplied by the absolute value of the current partition's residual to obtain the weighted absolute value of the residual for that day. All weighted absolute values ​​of residuals are summed to obtain the weighted residual sum after the current partition. The ratio of the preset acceptable residual to the sum of the weighted residuals is used as the quality residual ratio. Based on the three consecutive partitions before the current partition, the average residual ratio of the two adjacent partitions is calculated. All the calculated average residual ratios are summed and averaged to obtain the average residual change rate. The product of the quality residual ratio after the current partition and the average residual change rate is used as the residual optimization degree after the current partition.

[0057] For example, the optimality of the residuals after the current partitioning satisfies the following relationship:

[0058] C h =Z h ×P h

[0059]

[0060] Ch Z represents the optimality of the residual after the h-th partition; h Let Q be the quality residual ratio after the h-th partition, and E0 be the acceptable residual (it can be set to 0.3 in the early stages of model training, and E0 can be gradually reduced in the later stages of training). i E represents the feeding quality index for day i. h,i P represents the absolute value of the residual of the feeding data on the i-th day after the h-th partition, N is the total number of days collected, and norm() represents the normalization process; h Let h be the average residual change rate after the h-th partition. Let be the average residual after the ha-th partitioning. Let be the average residual after the (h-a+1)th partition, where h ≥ 4.

[0061] The above Z h By weighting residuals using data quality metrics, the impact of low-quality data can be reduced. h By using the residual ratios of the three most recent partitions, we can assess whether the residuals are gradually decreasing. Combining these two parts yields the residual optimization degree, which is used to determine whether the current partition is better than the previous partitions.

[0062] Finally, the initial learning rate is adjusted based on the degree of residual optimization to obtain an adaptive learning rate.

[0063] The higher the residual optimization level, the better the current learning rate; therefore, the smaller the adjustment range should be. The product of the square root of the residual optimization level (it should be considered that directly using the residual optimization level might lead to excessive fluctuations in the learning rate; the square root can suppress the influence of extreme values, making the adjustment more stable) and the initial learning rate is used as the adaptive learning rate for the next iteration. This new learning rate is used in the next decision tree split.

[0064] In subsequent iterations, the residual optimization degree is calculated according to the above steps, and the learning rate is adjusted according to the residual optimization degree to gradually optimize the model's learning rate and improve the model's predictive performance.

[0065] Training stops when the maximum number of iterations set above is reached, and the output model is the final trained model.

[0066] S2: Input the real-time variable frequency feeding data into the trained GBDT model, output the feeding frequency prediction result, and control the feeding equipment to intelligently perform variable frequency feeding based on the prediction result.

[0067] In one embodiment, the same process of collecting and preprocessing historical frequency conversion feeding data for each day according to S1 is used to collect real-time frequency conversion feeding data. This real-time data is then input into the GBDT model trained through S1. The model dynamically adjusts its internal parameters based on an adaptive learning rate and outputs a predicted feeding frequency for the current moment. The adaptive learning rate allows the model to automatically adjust its weights based on changes in real-time data, thereby predicting the feeding frequency more accurately.

[0068] Based on the predicted output feeding frequency and the hardware control requirements of the negative pressure air delivery feeding system, corresponding control commands are generated, such as setting the output frequency of the frequency converter (corresponding to the operating frequency of the feeding device), and sending it to the control unit of the feeding system to drive the feeding device to perform precise feeding operations according to the calculated frequency. This realizes intelligent frequency conversion feeding control based on negative pressure air delivery, ensuring that feed can be supplied to aquatic organisms on demand and in a timely manner.

[0069] The system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the intelligent variable frequency feeding control method based on negative pressure air delivery according to the first aspect of the present invention.

[0070] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0071] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A smart variable frequency feeding control method based on negative pressure air delivery, characterized in that, include: Collect and preprocess the variable frequency feeding data for each day of the preset historical number of days, input all the historical variable frequency feeding data into the GBDT model to be trained, calculate the adaptive learning rate based on the historical variable frequency feeding data during the training process, optimize the training process of the model based on the adaptive learning rate, and then complete the training of the GBDT model. Real-time variable frequency feeding data is input into the trained GBDT model, and the feeding frequency prediction result is output. The feeding equipment is controlled to intelligently perform variable frequency feeding based on the prediction result. The adaptive learning rate is calculated as follows: Set the initial learning rate; After each decision tree split, the residual optimization degree after the current split is calculated. The process of obtaining the residual optimization degree after the current split includes: after a split, obtaining the absolute value of the residual of the data collected each day in the past; after at least four decision tree splits, normalizing the feeding quality index of each day in the past, multiplying the normalized feeding quality index by the absolute value of the residual of the current split to obtain the weighted absolute value of the residual for the corresponding day, summing all the weighted absolute values ​​of the residual to obtain the sum of the weighted residuals after the current split; and using the ratio of the preset acceptable residual to the sum of the weighted residuals as the quality residual ratio. The calculation is based on the three consecutive partitions before the current partition. The average residual ratio between two adjacent partitions is calculated, and the average residual ratios are summed and averaged to obtain the average residual change rate. The product of the quality residual ratio after the current partitioning and the average residual change rate is taken as the degree of residual optimization after the current partitioning; The product of the square root of the residual optimization degree and the initial learning rate is used as the adaptive learning rate for the next iteration.

2. The intelligent variable frequency feeding control method based on negative pressure air delivery according to claim 1, characterized in that, The variable frequency feeding data includes: daily water temperature, water yield, dissolved oxygen level, and the amount of feed remaining before feeding the previous day.

3. The intelligent variable frequency feeding control method based on negative pressure air delivery according to claim 2, characterized in that, The acquisition of the feeding quality indicators includes: calculating the feed remaining indicators before feeding on any past day; and calculating the feed remaining quantity coordination degree on any past day. The product of the feed surplus index and the synergy of feed surplus is used as the feeding quality index for any past day.

4. The intelligent variable frequency feeding control method based on negative pressure air delivery according to claim 3, characterized in that, The process of obtaining the feed surplus index includes: converting the feed surplus of each past day into the corresponding feed surplus index by using an exponential function.

5. The intelligent variable frequency feeding control method based on negative pressure air delivery according to claim 3, characterized in that, The process of obtaining the feed surplus index includes: calculating the ratio of the feed surplus of each past day to the total feed amount fed on the corresponding day, and subtracting the ratio from 1 as the feed surplus index for the corresponding day.

6. The intelligent variable frequency feeding control method based on negative pressure air delivery according to claim 3, characterized in that, The process of obtaining the feed balance synergy includes: selecting any past day as the target day, calculating the Euclidean distance between the target day and the feeding data collected from any other day; calculating the mean of the sum of the ratios of the feed balance difference to the Euclidean distance between the target day and all other days, and obtaining the feed balance synergy of the target day.

7. An intelligent variable frequency feeding control system based on negative pressure air delivery, characterized in that, include: The processor and memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the intelligent variable frequency feeding control method based on negative pressure air delivery according to any one of claims 1-6.

Images