Feed production control method and system based on deep learning

Abstract

The invention relates to the technical field of feed production, and discloses a feed production control method and system based on deep learning. Aiming at the problems of low production efficiency and unstable product quality caused by uncertain factors such as raw material quality fluctuation, equipment state change and the like which are difficult to deal with by a traditional feed production method, an improved CrossFormer deep learning model is adopted, and feed production process data are acquired and preprocessed in real time; and inputting the preprocessed data into a model to carry out validity evaluation on the current production strategy, and giving an adjustment instruction. And automatically adjusting a feed production strategy according to an output result of the model and an adjustment instruction, and recording all implemented adjustment measures, production data comparison before and after adjustment and equipment monitoring information to form a production optimization log. According to the method, the intelligent level of feed production is remarkably improved, the production cost is reduced, and the product quality and the market competitiveness are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of feed production, and in particular to a feed production control method and system based on deep learning. Background Art

[0002] Feed production is a complex and highly sophisticated process that traditionally relies on manual experience and fixed production processes. However, this approach often has difficulty coping with uncertainties such as fluctuations in raw material quality and changes in equipment status, resulting in low production efficiency and unstable product quality.

[0003] With the rapid development of big data and artificial intelligence technologies, more and more companies are trying to apply intelligent technologies to the feed production process. However, most of the existing intelligent feed production control methods are based on simple machine learning models, which have limited processing capabilities and prediction accuracy for complex and changeable feed production process data, making it difficult to effectively evaluate and optimize production strategies. Therefore, there is an urgent need for a feed production control method that can efficiently process feed production process data, accurately evaluate production strategies, and give optimization suggestions. Summary of the invention

[0004] In view of the above-mentioned technical deficiencies, the purpose of the present invention is to provide a feed production control method and system based on deep learning to solve the problem that it is difficult to efficiently process feed production process data, accurately evaluate production strategies and provide optimization suggestions in the prior art.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a feed production control method based on deep learning, comprising: Step S100: collecting feed production process data in real time and preprocessing the feed production process data; Step S200: constructing an improved CrossFormer deep learning model based on historical feed production process data; Step S300: input the preprocessed real-time feed production process data into the optimized CrossFormer deep learning model, and the deep learning model evaluates the effectiveness of the current production strategy and gives adjustment instructions; Step S400: adjusting the feed production strategy according to the output results of the deep learning model and the adjustment instructions; Step S500: After the adjustment is completed, the system automatically records all the implemented adjustment measures, the comparison of production data before and after the adjustment, and the equipment monitoring information to form a production optimization log.

[0006] Preferably, in a possible implementation of the first aspect, the feed production process data includes raw material quality data, formula ratio data, production temperature data, production humidity data, production stirring speed data and equipment status data; the historical feed production process data includes historical raw material quality data, historical formula ratio data, historical production temperature data, historical production humidity data, historical stirring speed data and historical equipment status data.

[0007] Preferably, in a possible implementation manner of the first aspect, the step of preprocessing the feed production process data in step S100 includes data cleaning, missing value processing, outlier detection and processing, and data normalization processing.

[0008] Preferably, in a possible implementation of the first aspect, the improved CrossFormer deep learning model structure includes: The input layer receives the pre-processed real-time feed production process data and converts the data into a format that the model can recognize; The dimension segmentation embedding layer divides the time series into patches, uses relative position encoding, and embeds them into the feature vector; Two-stage attention layer: the first stage performs attention mechanism calculation in the time dimension, and the second stage performs attention mechanism calculation between multiple variables to align the encodings of different variables at each time step; Hierarchical encoding and decoding layer, the encoder extracts information from different granularities, and the decoder uses different levels of coding to make predictions, adding the prediction values ​​of each layer to obtain the final prediction result; The output layer outputs the effectiveness evaluation results and adjustment instructions of the current production strategy based on the prediction results.

[0009] Preferably, in a possible implementation of the first aspect, the dimension segmentation embedding layer is specifically implemented as follows: the preprocessed real-time feed production process data is divided into multiple patches in a time series, and for each patch, a relative position encoding technique is used to encode the relative position information of each element in the patch into a vector form, and the relative position encoding vector is merged with the corresponding data feature vector to form a feature vector containing the position information.

[0010] Preferably, in a possible implementation manner of the first aspect, the adjustment instruction includes adjusting the raw material ratio, changing the production temperature, adjusting the production humidity and adjusting the stirring speed.

[0011] Preferably, in a possible implementation manner of the first aspect, step S400 is implemented by controlling automated equipment of the feed production line, including raw material ratio adjustment, production temperature adjustment, production humidity adjustment and stirring speed adjustment.

[0012] Preferably, in a possible implementation of the first aspect, the feed production control method further includes comparing the real-time data in the feed production process with the prediction results of the improved CrossFormer deep learning model, and iteratively optimizing the deep learning model according to the comparison results.

[0013] In a second aspect, the present invention provides a feed production control system based on deep learning, comprising: A data collection module collects feed production process data in real time and pre-processes the feed production process data; Model building module, which builds an improved CrossFormer deep learning model based on historical feed production process data; The model processing module is used to input the pre-processed real-time feed production process data into the optimized CrossFormer deep learning model. The deep learning model evaluates the effectiveness of the current production strategy and gives adjustment instructions; The production adjustment module adjusts the feed production strategy based on the output results and adjustment instructions of the deep learning model; The log module is used to automatically record all adjustment measures implemented, production data comparison before and after adjustment, and equipment monitoring information after the production adjustment module is completed, forming a production optimization log.

[0014] The beneficial effect of the present invention is that by introducing the improved CrossFormer deep learning model, historical feed production process data can be fully utilized to achieve accurate modeling and prediction of complex production processes. The method can collect feed production process data in real time and preprocess it, and input the preprocessed data into the deep learning model. The model can quickly evaluate the effectiveness of the current production strategy and give specific adjustment instructions. According to the prediction results and adjustment instructions of the model, the system can automatically adjust the feed production process to ensure production efficiency and product quality. In addition, the system can also automatically record all implemented adjustment measures, production data comparisons before and after adjustments, and equipment monitoring information to form a production optimization log, providing data support for subsequent production optimization. This method significantly improves the level of intelligence in feed production, reduces production costs, and improves product quality and market competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 A flow chart of a feed production control method based on deep learning is provided for this application.

[0017] Figure 2 A structural schematic diagram of a feed production control system based on deep learning is provided for this application.

[0018] Explanation of the accompanying drawings: 1-data acquisition module, 2-model building module, 3-model processing module, 4-production adjustment module, 5-log module. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] Example 1: Figure 1 As shown, the present invention provides a feed production control method based on deep learning, comprising: Step S100: collecting feed production process data in real time and preprocessing the feed production process data.

[0021] Specifically, feed production process data includes raw material quality data, formula ratio data, production temperature data, production humidity data, production stirring speed data and equipment status data.

[0022] The raw material quality data is measured in real time by raw material testing instruments; the formula ratio data is obtained by comparing the formula set before production with the actual measurement data of the raw materials input; the production temperature data and production humidity data are monitored in real time by temperature sensors and humidity sensors respectively; the production stirring speed data is obtained by the speed sensor of the stirring equipment; the equipment status data relies on the equipment's own monitoring system for real-time detection, including the equipment's operating status, fault alarm and other information.

[0023] In this embodiment, the preprocessing steps include data cleaning, missing value processing, outlier detection and processing, and data normalization processing.

[0024] First, use Python's Pandas library to read the real-time collected data and convert it into DataFrame format. At the same time, import necessary libraries, such as numpy for numerical calculations and MinMaxScaler in sklearn.preprocessing for data normalization.

[0025] For continuous variables such as raw material quality, formula ratio, production temperature, and production humidity, if there are few missing values, linear interpolation is used to fill in; if there are many missing values, the median is used to fill in. For categorical variables such as equipment status, if there are few missing values, they are filled in based on context inference; if there are many missing values, the most common status is filled in by default.

[0026] Outliers were detected using the box plot method. For continuous variables, the quartiles (Q1 and Q3) and interquartile range (IQR) were calculated for each variable. The upper and lower bounds were set as Q1 minus 1.5 times the IQR and Q3 plus 1.5 times the IQR. Any value outside these bounds was considered an outlier. For detected outliers, if the number was small, they were replaced or deleted according to the context; if the number was large, they were filled with the most common state by default.

[0027] In order to eliminate the influence of different dimensions on the data, the Min-Max normalization method is used to normalize the continuous variables. Specifically, the MinMaxScaler in sklearn.preprocessing is used to scale the value of each variable to between 0 and 1 and store it in the normalized_df variable.

[0028] Step S200: construct an improved CrossFormer deep learning model based on historical feed production process data.

[0029] Specifically, the historical feed production process data includes historical raw material quality data, historical formula ratio data, historical production temperature data, historical production humidity data, historical stirring speed data and historical equipment status data.

[0030] The improved CrossFormer deep learning model structure includes an input layer, a dimensionally segmented embedding layer, a two-stage attention layer, a hierarchical encoding and decoding layer, and an output layer.

[0031] The input layer receives the preprocessed real-time feed production process data and converts it into a format that the model can recognize.

[0032] The input layer uses the InputLayer class, which inherits from the nn.Module base class in the PyTorch framework and is used to process real-time feed production process data. In the constructor of the InputLayer class, a parameter called input_dim is defined. This parameter is used to specify the dimension of the input feature, that is, the total number of features of the historical data. In this embodiment, the number of input features is 6, corresponding to six feed production process features. The constructor first initializes the parent class nn.Module by calling super(InputLayer,self).__init__(), and then saves the value of the input_dim parameter to the instance variable of the class.

[0033] In addition to the constructor, the InputLayer class also defines a method called forward. This method implements the forward propagation process of the data. The forward method receives a parameter called x, which is a three-dimensional tensor with a shape of (batch_size, sequence_length, input_dim). Among them, batch_size represents the batch size, that is, the number of data samples processed simultaneously; sequence_length represents the sequence length, that is, the number of steps in the time series in each data sample; and input_dim corresponds to the input_dim parameter in the constructor, indicating the dimension of the input feature.

[0034] The dimension segmentation embedding layer divides the time series into patches, uses relative position encoding, and embeds them into the feature vector.

[0035] The dimension segmentation embedding layer divides the time series into patches and obtains patch embedding as the input of the subsequent model. The time series of each variable is divided into multiple blocks according to a certain window. Each block is mapped through full connection. The dimension segmentation embedding is expressed as:

[0036]

[0037] in, Indicates the length of each segment in each dimension, T is the total length of the sequence, and D is the number of dimensions. is the i-th segment of dimension d.

[0038] Each fragment is then embedded into a vector using linear projection adding the positional embedding:

[0039] Where E represents the learnable projection matrix, Indicates location The learnable relative position embedding of . After embedding, a two-dimensional vector array is obtained, and the two axes represent time dependency and dimension dependency respectively:

[0040] Two-stage attention layer. The first stage performs attention mechanism calculation in the time dimension. The second stage performs attention mechanism calculation between multiple variables to align the encodings of different variables at each time step.

[0041] The two-stage attention layer refers to the time dimension attention layer and the space dimension attention layer. The input first passes through a time dimension attention layer to independently model each sequence in time series; then it is input into a space dimension attention layer to align the encoding of each time step of different variables. After the time series of each variable in the time dimension attention layer is converted into a patch, it is input into the Transformer, and the representation of each patch of each variable after self-attention is output; the space dimension attention layer aligns each time step of different variables. The purpose is to find the corresponding relationship between time steps between variables.

[0042] Given an array As the input of the two-stage attention layer, where L and D are the number of segments and dimensions respectively. Here Z is the embedding of the dimension segment embedding layer or the output of the lower two-stage attention layer. represents a vector of all dimensions at time step i, Represents the vector of all time steps at dimension d. In the cross-time stage, multi-head self-attention is directly applied to each dimension:

[0043]

[0044] in , LayerNorm represents layer normalization, MLP represents a multi-layer feedforward network, and MSA(Q, K,V) represents a multi-head self-attention layer, where Q, K, and V are used as query, key, and value. All dimensions share the same MSA layer. , Represents the output of MSA and MLP.

[0045] Use large To reduce the number of L across time phases. In the cross - dimensional phase, a fixed small number (c << D) of learnable vectors are set as routers for each time step i. These routers first act as queries in the MSA and use vectors from all dimensions as keys and values to aggregate messages from all dimensions. Then, the routers use the dimensional vectors as queries and the aggregated messages as keys and values to distribute the received messages between dimensions. In this way, a full - to - full connection between D dimensions is established:

[0046]

[0047]

[0048]

[0049] where is an array of learnable vectors used as routers. is the aggregated message of all dimensions. represents the output of the router mechanism. The same , are shared across all time steps. , represent the outputs of the skip connection and the MLP respectively.

[0050] The final output of the two - stage attention layer is Y:

[0051] where represent the input and output vector arrays of the two - stage attention layer respectively.

[0052] Hierarchical encoder - decoder layer. The encoder extracts information from different granularities, and the decoder uses encodings at different levels for prediction. The predicted values of each layer are added together to obtain the final prediction result.

[0053] In this embodiment, a dimensional - segment embedding layer, a two - stage attention layer, and segment merging are used to construct the hierarchical encoder - decoder layer. In each layer of the encoder (except the first layer), every two adjacent vectors in the time domain are merged to obtain a coarser - level representation. Then, the two - stage attention layer is applied to capture the dependencies at this scale. This process is modeled as :

[0054]

[0055] where H represents a 2D array obtained by embedding through the dimensional - segment embedding layer; represents the The output of the encoder layer; M represents the learnable matrix for segment merging; Indicates a serial operation; Indicates The number of segments in each dimension of the layer, if not divisible by 2, we will Fill to the appropriate length; Represents the array after merging the i-th layer segment. Assuming there are N layers in the encoder, use , , ..., ( =H) represents the N + 1 outputs of the encoder.

[0056] Get the N + 1 feature arrays output by the encoder and use the N + 1 layers in the decoder for prediction. The lth layer takes the lth encoded array as input and then outputs the decoded 2D array of the lth layer. This process is :

[0057]

[0058]

[0059]

[0060] in represents the learnable position embedding of the decoder. is the output of the two-stage attention layer. MSA is As a query, As keys and values, the connection between the encoder and decoder is constructed. The output of MSA is represented as . Represent the output of the skip connection and MLP respectively. , , ..., Represents the decoder output.

[0061] Linear projection is applied to the output of each layer to produce the prediction for that layer, and the predictions of each layer are added together to obtain the final prediction:

[0062]

[0063]

[0064] in is a learnable matrix used to project vectors into time series segments. represents the i-th segment in the d-dimensional prediction.

[0065] The output layer outputs the effectiveness evaluation results and adjustment instructions of the current production strategy based on the prediction results.

[0066] In the output layer, multiple fully connected layers are first applied to transform the final output of the hierarchical encoder-decoder layer. The high-dimensional hidden state vector is mapped to a lower-dimensional space that is easier to interpret and process. Specifically, the linear layer obtains multiple output vectors by calculating the product of the input vector and the weight matrix and adding a bias term. Next, each output vector is normalized by applying a Softmax activation function to the output of the linear layer. The vector is converted into a probability distribution where each element has a value between 0 and 1 and the sum of all element values ​​is 1. This makes each output represent a discrete probability distribution corresponding to different feed production strategies and adjustment instructions.

[0067] Step S300: The preprocessed real-time feed production process data is input into the optimized CrossFormer deep learning model. The deep learning model evaluates the effectiveness of the current production strategy and gives adjustment instructions.

[0068] Specifically, after receiving the data, the model uses its feature extraction and sequence modeling capabilities to evaluate the production strategy, analyze whether the current raw material ratio is optimal, whether the production temperature and humidity are appropriate, whether the mixing time is sufficient, etc.

[0069] Based on the evaluation results, the CrossFormer deep learning model gives adjustment instructions for the current production strategy. The adjustment instructions include adjusting the raw material ratio, changing the production temperature, adjusting the production humidity, and adjusting the stirring speed.

[0070] In this embodiment, the output layer outputs a total of 4 probability values ​​( , , , ), corresponding to adjusting the raw material ratio, changing the production temperature, adjusting the production humidity, and adjusting the stirring speed. When P1>0.3, it means that the instruction adjusts the current raw material ratio, and optimizes the ratio according to the specific value to achieve a better production effect; if P2>0.4, it is considered that the current production temperature is not optimal and needs to be adjusted. According to the specific value of P2, it is decided whether to increase or decrease the production temperature; when P3>0.3, the production humidity needs to be adjusted. According to the value of P3 and the current humidity conditions, it is decided whether to increase or decrease the humidity; if P4>0.2, it is considered that the stirring speed needs to be adjusted. According to the value of P4, it is decided whether to increase or decrease the stirring speed to ensure that the raw materials are fully mixed.

[0071] Step S400: Adjust the feed production strategy according to the output results of the deep learning model and the adjustment instructions.

[0072] Specifically, according to the adjustment instructions output by the model, a specific adjustment plan is automatically generated. This plan includes the equipment to be adjusted, the adjustment parameters, the adjustment target value and the adjustment time point, and is sent to relevant personnel for confirmation.

[0073] After confirmation by relevant personnel, the control instructions in the plan are sent to the corresponding production equipment through the communication protocol. For raw material ratio adjustment, the new raw material ratio is automatically calculated, and the new ratio data is sent to the batching controller through the interface with the automated batching system. The batching controller adjusts the amount of each raw material according to the received data.

[0074] For production temperature adjustment, read the model output adjustment instructions, determine the temperature adjustment direction, and send the temperature setting value to the temperature control device through the communication interface of the temperature control system. The temperature control device adjusts the output of the heating or cooling system according to the instructions to make the production environment temperature reach the set value.

[0075] For production humidity adjustment, the output adjustment instructions of the model are used to determine the direction of humidity adjustment. The humidity set value is sent through the interface with the humidification or dehumidification equipment. The humidification or dehumidification equipment adjusts its working state according to the instructions to make the humidity of the production environment reach the target value.

[0076] For stirring speed adjustment, the output adjustment instruction of the model is read to determine the stirring speed adjustment direction, and the speed setting value is sent through the communication interface with the mixer. The mixer adjusts its stirring speed according to the received instruction to ensure that the raw materials are fully mixed.

[0077] In this embodiment, the Modbus industrial communication protocol is used as the communication protocol to ensure the stability and reliability of data during transmission. It not only supports multiple data transmission methods, but also has a powerful error detection and recovery mechanism.

[0078] For the interface with the automated batching system, a standardized API interface is used to achieve seamless connection with the batching controller. This interface can send new raw material ratio data to the batching controller in real time to ensure the accuracy and efficiency of the batching process.

[0079] In terms of sending and executing adjustment instructions, instruction format and content specifications have been developed. Each instruction contains key information such as device identification, adjustment parameters, target value, and execution time to ensure that the device can accurately understand and execute instructions. At the same time, an instruction verification mechanism is set up to verify the integrity and correctness of the instructions.

[0080] In addition, in order to verify whether the equipment has been adjusted according to the instructions and achieved the expected results, a real-time monitoring and feedback mechanism was introduced. Through sensors and monitoring equipment, various data in the production process, such as raw material ratio, production temperature, production humidity and stirring speed, are obtained in real time. These data are used to compare with the target values ​​set in the instructions to verify whether the adjustment effect meets the standards.

[0081] Finally, in terms of exception handling, emergency plans and fault handling procedures have been formulated. Once abnormal situations such as equipment failure and communication interruption occur during the adjustment process, the emergency plan will be immediately activated and corresponding measures will be taken to troubleshoot and repair the fault to ensure the continuity and stability of production.

[0082] Step S500: After the adjustment is completed, the system automatically records all the implemented adjustment measures, the comparison of production data before and after the adjustment, and the equipment monitoring information to form a production optimization log.

[0083] Specifically, all adjustment measures triggered by the deep learning model are automatically captured and recorded. These measures include changes in raw material ratios, fine-tuning of production temperature, increase or decrease in mixing time, etc. Each adjustment record is marked with detailed adjustment time, specific values ​​of adjustment parameters, and reasons for adjustment.

[0084] Then the production data before and after the adjustment are automatically collected and compared, including feed quality indicators, production efficiency, energy consumption, etc. Through comparison, the direct impact of the adjustment measures on production results can be intuitively displayed.

[0085] In addition to production data, the monitoring information of equipment operation during the adjustment period is integrated, including the equipment's operating status, operating efficiency, and maintenance records.

[0086] Based on the information recorded above, a detailed production optimization log is automatically generated. The log not only contains the specific content of the adjustment measures, but also includes data comparison charts before and after the adjustment, a summary report of equipment monitoring, and preliminary analysis conclusions based on the data.

[0087] In one embodiment, the method further includes comparing the real-time data in the feed production process with the prediction results of the improved CrossFormer deep learning model, and iteratively optimizing the deep learning model based on the comparison results.

[0088] Specifically, the difference between real-time data and predicted results is calculated, and the source of the difference is traced, such as fluctuations in raw material quality, changes in equipment status, or unstable production environment. Based on the comparative analysis results, the training data is updated. For example, to address the accuracy of production temperature prediction, the relevant parameters of time series analysis are adjusted or relevant historical data is added. In addition, the production optimization logs are used to verify the model optimization effect. These logs contain adjustment measures, production data comparisons, and equipment monitoring information to provide support for model optimization and production strategy adjustments.

[0089] Example 2: Figure 2 As shown, the present invention provides a feed production control system based on deep learning, comprising: Data collection module 1, collects feed production process data in real time and pre-processes the feed production process data; Model building module 2, building an improved CrossFormer deep learning model based on historical feed production process data; Model processing module 3, used to input the pre-processed real-time feed production process data into the optimized CrossFormer deep learning model, the deep learning model evaluates the effectiveness of the current production strategy and gives adjustment instructions; Production adjustment module 4 adjusts the feed production strategy according to the output results and adjustment instructions of the deep learning model; The log module 5 is used to automatically record all the adjustment measures implemented, the comparison of production data before and after the adjustment, and the equipment monitoring information after the adjustment of the production adjustment module is completed, so as to form a production optimization log.

[0090] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A feed production control method based on deep learning, characterized in that: include: Step S100: collecting feed production process data in real time and preprocessing the feed production process data; Step S200: constructing an improved CrossFormer deep learning model based on historical feed production process data; Step S300: input the preprocessed real-time feed production process data into the optimized CrossFormer deep learning model, and the deep learning model evaluates the effectiveness of the current production strategy and gives adjustment instructions; Step S400: adjusting the feed production strategy according to the output results of the deep learning model and the adjustment instructions; Step S500: After the adjustment is completed, the system automatically records all the implemented adjustment measures, the comparison of production data before and after the adjustment, and the equipment monitoring information to form a production optimization log.

2. The feed production control method according to claim 1, characterized in that: The feed production process data includes raw material quality data, formula ratio data, production temperature data, production humidity data, production stirring speed data and equipment status data; the historical feed production process data includes historical raw material quality data, historical formula ratio data, historical production temperature data, historical production humidity data, historical stirring speed data and historical equipment status data.

3. The feed production control method according to claim 1, characterized in that: The steps of preprocessing the feed production process data in step S100 include data cleaning, missing value processing, outlier detection and processing, and data normalization processing.

4. The feed production control method according to claim 1, characterized in that: The improved CrossFormer deep learning model structure includes: The input layer receives the pre-processed real-time feed production process data and converts the data into a format that the model can recognize; The dimension segmentation embedding layer divides the time series into patches, uses relative position encoding, and embeds them into the feature vector; Two-stage attention layer: the first stage performs attention mechanism calculation in the time dimension, and the second stage performs attention mechanism calculation between multiple variables to align the encodings of different variables at each time step; Hierarchical encoding and decoding layer, the encoder extracts information from different granularities, and the decoder uses different levels of coding to make predictions, adding the prediction values ​​of each layer to obtain the final prediction result; The output layer outputs the effectiveness evaluation results and adjustment instructions of the current production strategy based on the prediction results.

5. The feed production control method according to claim 3, characterized in that: The specific implementation method of the dimension segmentation embedding layer is: the pre-processed real-time feed production process data is divided into multiple patches in the time series, and for each patch, the relative position encoding technology is used to encode the relative position information of each element in the patch into a vector form, and the relative position encoding vector is merged with the corresponding data feature vector to form a feature vector containing position information.

6. The feed production control method according to claim 1, characterized in that: The adjustment instructions include adjusting the raw material ratio, changing the production temperature, adjusting the production humidity and adjusting the stirring speed.

7. The feed production control method according to claim 1, characterized in that: The step S400 is achieved by controlling the automated equipment of the feed production line, including raw material ratio adjustment, production temperature adjustment, production humidity adjustment and stirring speed adjustment.

8. The feed production control method according to claim 1, characterized in that: The feed production control method also includes comparing the real-time data in the feed production process with the prediction results of the improved CrossFormer deep learning model, and iteratively optimizing the deep learning model according to the comparison results.

9. A feed production control system based on deep learning, characterized in that: include: A data collection module collects feed production process data in real time and pre-processes the feed production process data; Model building module, which builds an improved CrossFormer deep learning model based on historical feed production process data; The model processing module is used to input the pre-processed real-time feed production process data into the optimized CrossFormer deep learning model. The deep learning model evaluates the effectiveness of the current production strategy and gives adjustment instructions; The production adjustment module adjusts the feed production strategy based on the output results and adjustment instructions of the deep learning model; The log module is used to automatically record all adjustment measures implemented, production data comparison before and after adjustment, and equipment monitoring information after the production adjustment module is completed, forming a production optimization log.

Images