Feed processing prediction and real-time control method based on deep learning

Abstract

The present invention belongs to the field of feed processing technology and relates to a feed processing prediction and real-time control method based on deep learning, comprising: performing feature processing on material preparation process data to obtain multiple material preparation process features and screening features; performing feature recognition on the screening features to predict the probability of processing abnormality and abnormal results; if the processing is abnormal, classifying and splicing the multiple material preparation process features to obtain processing operation features and processing equipment features; processing the processing operation features through a long short-term memory network to obtain operation-related features; processing the processing equipment features through a convolutional neural network to obtain equipment-related features; performing attention allocation operations on the operation-related features and the equipment-related features, and determining the equipment abnormality probability of the processing equipment based on the attention contribution of each material preparation process; achieving full-process quality prediction to avoid shutdowns and inspections due to unqualified feed processing.

Description

Technical Field

[0001] The present invention relates to the technical field of feed processing, and specifically discloses a feed processing prediction and real-time control method based on deep learning. Background Art

[0002] Feed processing involves various steps, including pulverization, raw material extraction, raw material mixing, and pelletizing. Traditional feed pulverization equipment often utilizes a single sieve plate design, which results in inadequate screening and uneven raw material particle size. This not only increases the difficulty of subsequent processing steps but also impacts the quality of feed additives and the efficiency of animal nutrient absorption. The screening process relies on a static sieve plate and a single vibration mode, which cannot dynamically adapt to the characteristics of different raw materials, resulting in incomplete separation of crushed materials. Traditional batching systems rely on fixed storage tanks and batching scales, which are subject to high structural costs and poor mobility. Manual handling and weighing not only consume significant manpower but also make it difficult to precisely control ingredient dosage. Ingredient extraction lacks a real-time dynamic control model, making it impossible to adaptively adjust the motor control amount based on parameters such as raw material fluidity and pipeline resistance. This results in low batching accuracy and insufficient production flexibility. Traditional mixing equipment often utilizes a single stirring method, which is difficult to effectively break up agglomerated raw materials and results in poor mixing uniformity of materials of different particle sizes. The mixing mechanism lacks a multi-dimensional motion mode and cannot simultaneously achieve vertical vibration and horizontal diffusion of the material, resulting in low mixing efficiency, especially insufficient adaptability to high-viscosity or easily agglomerated raw materials. Fault detection in traditional ring die pellet feed machines relies on manual inspections and single-parameter monitoring. Manual judgment is highly subjective, and traditional sensors can only capture single-dimensional data, making it impossible to comprehensively assess the overall status of the equipment. The lack of coordinated analysis of multiple parameters such as roller deformation, ring die gap changes, and motor vibration frequency deviations leads to delayed fault detection. For example, equipment anomalies may occur but no warning is given, leading to production interruptions and increased maintenance costs.

[0003] In view of this, this application provides a feed processing prediction and real-time control method based on deep learning. By constructing a spatiotemporal feature fusion model across processes, it realizes full-process quality prediction and avoids shutdowns and inspections due to unqualified feed processing; and by establishing a multi-dimensional anomaly index model, it improves the accuracy and timeliness of fault detection. Summary of the Invention

[0004] The purpose of the present invention is to provide a feed processing prediction method based on deep learning. The technical problems to be solved are: to achieve full-process quality prediction and avoid the situation where shutdown and investigation are caused by unqualified feed processing; and to improve the accuracy and timeliness of fault detection by establishing a multi-dimensional abnormality index model. The specific scheme includes: performing feature processing on the material preparation process data of multiple material preparation stages in the feed processing process to obtain multiple material preparation process features and screening features; performing feature recognition on the screening features to predict the processing abnormality probability and abnormal results of the feed processing; if the processing abnormality probability is greater than the preset processing abnormality probability threshold, then classifying the multiple material preparation process features to obtain multiple process operation features and process equipment features, and performing cross-process feature splicing to obtain processing operation features and processing equipment features; processing the processing operation features through a long short-term memory network to obtain operation-related features; processing the processing equipment features through a convolutional neural network to obtain equipment-related features; performing attention allocation operations on the operation-related features and equipment-related features, and determining the equipment abnormality probability of the processing equipment based on the attention contribution of each material preparation process.

[0005] Furthermore, for the crushing process, the process operation characteristics include the crushing blade speed, the discharge valve opening frequency and the vibration mechanism rotating rod speed; the process equipment characteristics include the inclination angle of the No. 1 screen plate, the aperture of the No. 2 screen plate and the proportion of the screen material; for the batching process, the process operation characteristics include the raw material extraction speed, the motor control amount and the pipeline pressure; the process equipment characteristics include the batching accuracy deviation and the correlation between the extraction speeds of different raw materials; for the mixing process, the process operation characteristics include the stirring rod speed, the vibration frequency of the vibration mechanism and the mixing time; the process equipment characteristics include the lifting amplitude of the receiving plate, the vibration displacement of the fixed block and the mixing uniformity CV value; the mixing uniformity CV value is used to quantify the uniformity of the distribution of each component after the feed is mixed; for the granulation process, the process operation characteristics include the pressure roller speed, the ring die temperature and the steam pressure; the process equipment characteristics include the pressure roller eccentricity, the ring die gap and the motor vibration frequency.

[0006] Furthermore, attention allocation operations are performed on the operation-related features and the equipment-related features to obtain attention contribution, including: performing self-attention operations on the operation-related features to obtain operation self-attention; performing self-attention operations on the equipment-related features to obtain equipment self-attention; performing cross-attention operations on the operation self-attention and the equipment self-attention to obtain cross-attention; fusing the operation self-attention, equipment self-attention and cross-attention to obtain the attention contribution of each material preparation process.

[0007] Furthermore, a self-attention operation is performed on the operation-related features to obtain operation self-attention, including: performing a self-attention operation on the operation-related features between processing procedures to extract the process operation self-attentions of multiple processing procedures; splicing and fusing the process operation self-attentions to obtain a fused operation self-attention; performing a residual connection between the fused operation self-attention and the operation-related features to obtain an initial operation self-attention; performing multi-layer perception on the initial operation self-attention and performing a residual connection between the multi-layer perception results and the initial operation self-attention to obtain the operation self-attention.

[0008] Furthermore, a self-attention operation is performed on the device-associated features to obtain device self-attention, including: performing a self-attention operation on the device-associated features between processing procedures to extract the process device self-attention of multiple processing procedures; splicing and fusing the process device self-attentions to obtain fused device self-attention; performing a residual connection on the fused device self-attention and the device-associated features to obtain the initial device self-attention; performing feature decoupling enhancement on the initial device self-attention to obtain the spatial feature part and the structural feature part in the device-associated features; performing global average pooling on the spatial feature part, and performing channel attention calculation on the pooling result to obtain the channel attention weight vector; calculating the spatial feature after channel attention enhancement through the spatial feature part and the channel attention weight vector; and calculating the device self-attention based on the spatial feature and structural feature part after channel attention enhancement.

[0009] Furthermore, a cross-attention operation is performed on the operation self-attention and the device self-attention to obtain the cross-attention, including: calculating the cross-attention of the operation self-attention and the device self-attention to obtain the cross-attention weight; fusing the cross-attention weight with the operation self-attention to obtain the fused operation feature; fusing the cross-attention weight with the device self-attention to obtain the fused device feature; performing a residual connection on the fused operation feature and the fused device feature to obtain the processing flow feature.

[0010] Furthermore, the calculation formula for attention contribution is:

[0011] ;

[0012] in, Indicates the attention contribution of the i-th material preparation stage; represents the cross attention weight of the i-th process; represents the self-attention feature of the i-th process; represents the self-attention feature of the equipment in the i-th process; represents the L2 general number; represents the cross attention weight of the j-th process; represents the running self-attention feature of the j-th process; Represents the equipment self-attention features of the j-th process.

[0013] Furthermore, determining the probability of equipment abnormality of the processing equipment includes: jointly training the initial attention model and the initial classification model to obtain a trained attention model and classification model; the loss function of the joint training is:

[0014] ;

[0015] in, represents the value of the loss function; k represents the training sample variable; K represents the total number of training samples; Indicates the true abnormal label of the k-th sample. When it is 1, it means the device is truly abnormal, and when it is 0, it means the device is truly normal. represents the abnormal probability prediction value of the kth sample by the classification model; ln represents the logarithmic function; the operation-related features and the equipment-related features are input into the attention model to obtain the attention contribution and the processing flow characteristics; the attention contribution and the processing flow characteristics are input into the classification model, and the classification model outputs the abnormal probability of the processing equipment.

[0016] The present invention also provides a real-time control method for feed processing of the feed processing prediction method as described in any of the above items, including: taking the processing equipment whose equipment abnormality probability is greater than the preset abnormality threshold as the target processing equipment; based on the abnormal result, adjusting the parameters of the target processing equipment in order of abnormality probability from large to small; based on the adjusted material preparation process data, updating the abnormality probability of the target processing equipment until the abnormality probability is less than the preset abnormality threshold; based on the adjustment parameters when the abnormality probability is less than the preset abnormality threshold, adjusting the operating parameters of the processing equipment, and performing feed processing with the new operating parameters.

[0017] Furthermore, it also includes identifying the adjustment parameters of the adjusted processing equipment, and issuing an alarm when it is identified that the adjustment parameters are not within the parameter threshold range; when it is identified that the adjustment parameters are within the parameter threshold range, feed processing is carried out with the adjusted processing equipment parameters, and the content of this modification is recorded.

[0018] The present invention has the following advantages and beneficial effects:

[0019] The present invention uses a long short-term memory network (LSTM) to process operation-related features and a convolutional neural network (CNN) to process device-related features, and combines self-attention and cross-attention mechanisms to achieve synergistic enhancement of features, which can effectively capture the nonlinear dependencies in the processing process and improve prediction accuracy.

[0020] The present invention uses self-attention operation to assign weights to the characteristics of each process, highlighting abnormal sensitive features and reducing noise interference; it also uses cross-attention to calculate the interaction weights of operation and equipment characteristics, identify cross-process anomalies, and improve the accuracy of processing anomaly probability prediction.

[0021] The present invention performs nonlinear transformation on the initial self-attention features through MLP to capture hidden feature interactions and improve the generalization ability of abnormal probability prediction.

[0022] The present invention can prioritize the processes with the highest risks by adjusting equipment parameters from high to low according to the abnormality probability, thus greatly shortening the abnormality response time. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 An exemplary flow chart of feed processing prediction based on deep learning provided by the present invention;

[0024] Figure 2 An exemplary flow chart for determining attention contribution provided by the present invention;

[0025] Figure 3 An operational flow chart of performing self-attention operations on associated features provided by the present invention;

[0026] Figure 4 This is an exemplary flow chart of the real-time control method for feed processing provided by the present invention. DETAILED DESCRIPTION

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of 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. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0028] Figure 1 This is an exemplary flow chart of feed processing prediction based on deep learning provided by the present invention. Figure 1 As shown, feed processing prediction based on deep learning includes the following:

[0029] Step 110 , performing feature processing on the obtained feed preparation process data of multiple feed preparation stages in the feed processing process to obtain multiple feed preparation process features and screening features.

[0030] The material preparation stage includes the crushing, batching, mixing, granulation, and screening processes. Material preparation process data includes crushing data, batching data, mixing data, granulation data, and screening data. Material preparation process characteristics include crushing characteristics, batching characteristics, mixing characteristics, and granulation characteristics.

[0031] Step 120 , performing feature recognition on the screening features to predict the probability of processing abnormalities and abnormal results of feed processing.

[0032] Screening characteristics include particle pulverization rate, particle distribution, and particle hardness. The probability of a processing abnormality can refer to the probability of a feed pellet pulverization rate exceeding 10%, the proportion of feed pellets outside the 2.0-8.0 mm range, the proportion of feed pellets larger than 2 mm, the proportion of pellets with a hardness less than 2 N, and the proportion of pellets with a hardness greater than 7.36 N. Abnormal results can include an excessively high pulverization rate, pellets that are too large, pellets that are too small, uneven particle distribution, excessively high hardness, and / or excessively low hardness. An excessively high pulverization rate indicates that the pellets have experienced significant wear during transportation or processing, potentially leading to nutritional loss and difficulty in feeding livestock and poultry. Causes of an excessively high pulverization rate can include insufficient conditioning temperature, low steam pressure, and / or overly coarse pulverization. Reasons for feed pellets not falling within the 2.0-8.0 mm range can include die wear and insufficient pelletizing pressure. Uneven particle distribution can be caused by screen blockage, abnormal screening equipment amplitude, or uneven pulverization. Low pellet hardness can lead to easy breakage and increased pulverization. Reasons for low pellet hardness include insufficient conditioning time and inadequate starch gelatinization. Excessive feed hardness can impair digestion and absorption in livestock and poultry. Reasons for high pellet hardness include excessively high conditioning temperatures (>90°C) and a large die aspect ratio. When the pulverization rate exceeds 10% and the hardness is less than 2N, it may be due to substandard conditioning parameters (such as temperature, time, and steam pressure). When the particle size distribution is uneven and the hardness is excessively high, it may be due to die wear or improper roller gap adjustment. When identifying screening features, the screening features from multiple time periods can be input into an anomaly training model to obtain predicted pulverization rate, predicted particle distribution, and predicted particle hardness. When the predicted pulverization rate, predicted particle distribution, and predicted particle hardness all meet the preset pellet standards, the feed processing is considered normal. Otherwise, an anomaly is identified and processing equipment anomaly detection is performed. By extracting historical screening data and processing the screening data in chronological order, a screening feature sequence can be obtained. By inputting the screening feature sequence into the long short-term memory network, the long short-term memory network extracts the change pattern of the screening features over time, and predicts the screening data in the future based on the time change pattern, and obtains a trained anomaly prediction model for processing anomaly prediction.

[0033] Step 130: If the processing abnormality probability is greater than the preset processing abnormality probability threshold, multiple material preparation process characteristics are classified to obtain multiple process operation characteristics and process equipment characteristics, and cross-process characteristics are spliced ​​to obtain processing operation characteristics and processing equipment characteristics; if the overall abnormality probability is less than or equal to the preset overall abnormality probability threshold, the feed processing flow of the processing process continues to be monitored.

[0034] The preset processing abnormality probability threshold is the maximum pre-set probability of no abnormalities in the feed processing process. When the threshold exceeds the preset processing abnormality probability, it is determined that an abnormality is likely to occur in the feed processing process within a certain period of time (e.g., the next 24 hours). When the threshold is equal to or less than the preset processing abnormality probability, it is determined that the feed processing process will not experience abnormalities or that the probability of an abnormality is extremely low within a certain period of time. Process operation characteristics refer to operational characteristics related to the feed processing process. Process equipment characteristics refer to equipment characteristics related to the feed processing process. For the crushing process, process operation characteristics include the crushing blade speed, discharge valve opening frequency, and vibration mechanism rotating rod speed; process equipment characteristics include the inclination angle of the first sieve plate, the aperture of the second sieve plate, and the percentage of oversize. For the batching process, process operation characteristics include the raw material extraction speed, motor control amount, and pipeline pressure; process equipment characteristics include batching accuracy deviation and the correlation between extraction speeds of different raw materials. For the mixing process, process operation characteristics include the stirring rod speed, vibration frequency of the vibrating mechanism, and mixing time; process equipment characteristics include the lifting amplitude of the receiving plate, the vibration displacement of the fixed block, and the CV value of the mixing uniformity. For the pelletizing process, process operation characteristics include roller speed, die temperature, and steam pressure; process equipment characteristics include roller eccentricity, die gap, and motor vibration frequency. Processing operation characteristics refer to the collection of equipment operation-related characteristics throughout the feed processing process. Processing equipment characteristics refer to the collection of equipment-related characteristics throughout the feed processing process. Cross-process feature concatenation combines equipment operation-related characteristics from multiple processes to obtain processing operation characteristics, and concatenates equipment-related characteristics from multiple processes to obtain processing equipment characteristics.

[0035] Step 140: Process the processing operation features using a long short-term memory network to obtain operation correlation features. Operation correlation features are used to represent the temporal dependencies of the feed processing process. The processing operation features are processed using an LSTM to extract temporal correlations between the processing processes.

[0036] Step 150: Process the processing equipment features using a convolutional neural network to obtain equipment association features. Equipment association features are used to represent the spatial dependencies of the feed processing process. The processing equipment features are processed using a CNN to extract local spatial associations between processing equipment.

[0037] Step 160 : performing an attention allocation operation on the operation-related features and the equipment-related features, and determining the equipment abnormality probability of the processing equipment based on the attention contribution of each material preparation process.

[0038] Attention allocation refers to assigning weights to each material preparation process. The weight represents the contribution of the material preparation process to feed processing anomalies. Attention contribution refers to the quantitative value of the importance of a single material preparation process in determining process anomalies. The probability of equipment anomaly is used to represent the probability that the current state of the equipment deviates from the normal working condition. Figure 2 As shown, the attention allocation operation is performed on the operation-related features and the equipment-related features to obtain the attention contribution, including: performing a self-attention operation on the operation-related features to obtain the operation self-attention. The operation self-attention is used to extract abnormal signals between processes in the time series dimension. Figure 3 As shown, a self-attention operation is performed on the associated features to obtain process self-attention; the process self-attentions are concatenated and fused to obtain fused self-attention; the fused self-attention is residually connected with the associated features to obtain initial self-attention; and self-attention is obtained by performing multi-layer perception on the initial self-attention and residually connecting the multi-layer perception result with the initial self-attention. Among them, the associated features include operation-related features and equipment-related features; the process self-attention includes process operation self-attention and process equipment self-attention; the fused self-attention includes fused operation self-attention and fused equipment self-attention; the initial self-attention includes initial operation self-attention and initial equipment self-attention; and the self-attention includes operation self-attention and equipment self-attention.

[0039] Performing self-attention operations on the operation-related features to obtain operation self-attention includes: performing self-attention operations on the operation-related features between the processing steps, and extracting the process operation self-attention of multiple processing steps. The calculation formula of the process operation self-attention is:

[0040] ;

[0041] Where i represents the process variable; represents the running self-attention of the i-th process; represents the activation function; represents the operation-related characteristics of the i-th process; Represents the operation-related characteristics of the j-th process.

[0042] The process operation self-attention is spliced ​​and fused to obtain the fused operation self-attention; the fused operation self-attention is residually connected with the operation-related features to obtain the initial operation self-attention. The calculation formula for the initial operation self-attention is:

[0043] ;

[0044] in, represents the initial run of self-attention; Indicates the running association feature; Concat indicates the concatenation function; Indicates the running fusion matrix, obtained through model training.

[0045] The running self-attention is obtained by performing multi-layer perception on the initial running self-attention and performing a residual connection between the multi-layer perception result and the initial running self-attention. The calculation formula of the running self-attention is:

[0046] ;

[0047] in, Indicates running self-attention; LayerNorm represents the layer normalization operation, which standardizes the features and stabilizes the training process; Indicates feature processing through a multi-layer perceptron; Represents the dimension of the i-th running process; Represents the dimension of the j-th running process.

[0048] Performing self-attention operations on device-related features to obtain device self-attention. Device self-attention is used to extract abnormal signals in the spatial dimension between processes. Performing self-attention operations on device-related features to obtain device self-attention includes: performing self-attention operations on device-related features between processing steps to extract process device self-attention of multiple processing steps. By performing single-head attention calculation on the process, the process device self-attention is obtained:

[0049] ;

[0050] in, represents the self-attention of the equipment in the i-th process; represents the activation function; Represents the equipment-related characteristics of the i-th process; Represents the equipment-related characteristics of the j-th process; Represents the dimension of the i-th equipment process; Represents the dimension of the j-th equipment process.

[0051] The process equipment self-attention is spliced ​​and fused to obtain the fused equipment self-attention. The process equipment self-attention of multiple processes is multi-head fused; and the fused equipment self-attention is residually connected with the equipment-related features to obtain the initial equipment self-attention. The calculation formula of the initial equipment self-attention is:

[0052] ;

[0053] in, represents the initial device self-attention; Indicates device association features; Concat indicates the concatenation function; Represents the device fusion matrix, which can be obtained through model training.

[0054] The initial device self-attention is subjected to feature decoupling enhancement to obtain the spatial feature part and the structural feature part of the device-related feature. The calculation formula for feature decoupling enhancement is:

[0055] ;

[0056] in, Represents the spatial feature part of the device-related features; Represents the structural feature part of the device-related features; represents the eigendecomposition function, which decouples device features according to the spatial-structural dimension.

[0057] Perform global average pooling on the spatial feature part and perform channel attention calculation on the pooling result to obtain the channel attention weight vector. The calculation formula of the channel attention weight vector is as follows:

[0058] ;

[0059] ;

[0060] Among them, z represents the feature vector after global average pooling, which is used for channel attention calculation; Represents the global average pooling operation, which averages the spatial feature matrix and compresses the spatial dimension; a represents the channel attention weight vector, which is generated by the Sigmoid function and has a value range of [0,1] and is used to weight spatial features; Represents the first layer weight matrix in channel attention, which is used to map the global pooling features; Represents the activation function, here is the ReLU function; Represents the second layer weight matrix in channel attention, which is used to implement feature weighting.

[0061] The spatial feature after channel attention enhancement is calculated by using the spatial feature part and the channel attention weight vector. The calculation formula of the spatial feature after channel attention enhancement is:

[0062] ;

[0063] in, Represents the spatial features after channel attention enhancement, by element-wise multiplication Fuse channel attention weight vector a.

[0064] Based on the spatial and structural features after channel attention enhancement, the device self-attention is calculated. The calculation formula of device self-attention is:

[0065] ;

[0066] in, Represents device self-attention; LayerNorm represents the layer normalization operation, which standardizes the features and stabilizes the training process; represents the modal mixing module, which realizes cross-modal feature interaction through MLP.

[0067] A cross-attention operation is performed on the running self-attention and the device self-attention to obtain a cross-attention. The cross-attention is used to establish a temporal-spatial coupling relationship between processes. In some embodiments, the cross-attention operation is performed on the running self-attention and the device self-attention to obtain a cross-attention, including: calculating the cross-attention of the running self-attention and the device self-attention to obtain a cross-attention weight. The calculation formula of the cross-attention weight is:

[0068] ;

[0069] in, Represents the cross attention weight, which quantifies the relative importance of temporal features and spatial features when they are fused. The greater the impact of temporal changes on the results, the higher the temporal weight. represents the dimension of running self-attention; Dimension representing the device's self-attention.

[0070] The cross attention weights are fused with the operation self-attention to obtain the fused operation features; the cross attention weights are fused with the device self-attention to obtain the fused device features; the fused operation features and the fused device features are residually connected to obtain the processing flow features. The calculation formula for the processing flow features is:

[0071] ;

[0072] in, It represents the processing flow characteristics, which are the fused operation characteristics and equipment characteristics. The weight represents the impact of the operation status and equipment configuration on the result, and is used to predict production anomalies.

[0073] By integrating the running self-attention, equipment self-attention and cross-attention, the attention contribution of each material processing process is obtained. The calculation formula of attention contribution is:

[0074] ;

[0075] in, represents the attention contribution of the i-th process; represents the cross attention weight of the i-th process; represents the self-attention feature of the i-th process; represents the self-attention feature of the equipment in the i-th process; represents the L2 general number; represents the cross attention weight of the j-th process; represents the running self-attention feature of the j-th process; Represents the equipment self-attention features of the j-th process.

[0076] In some embodiments, determining the probability of equipment abnormality of a processing device includes: jointly training an initial attention model and an initial classification model to obtain a trained attention model and classification model; the initial attention model and the initial classification model can be various feasible existing attention models and classification models. The loss function of the joint training is:

[0077] ;

[0078] in, represents the value of the loss function; k represents the training sample variable; K represents the total number of training samples; Indicates the true abnormal label of the k-th sample. When it is 1, it means the device is truly abnormal, and when it is 0, it means the device is truly normal. It represents the abnormal probability prediction value of the kth sample by the classification model; ln represents the logarithmic function.

[0079] The operation-related features and equipment-related features are input into the attention model to obtain the attention contribution and processing flow features. The attention contribution and processing flow features are then input into the classification model, which outputs the probability of processing equipment anomaly. The probability of processing equipment anomaly can be defined as the probability of an anomaly occurring in each process within the processing sequence.

[0080] Figure 4 This is an exemplary flow chart of the real-time control method for feed processing provided by the present invention, as shown in FIG. Figure 4 As shown, the real-time control method for feed processing includes the following contents:

[0081] Processing equipment with a probability of failure greater than a preset abnormality threshold is designated as a target processing equipment. The preset abnormality threshold is the maximum failure probability of the equipment under normal operation. When the probability of failure exceeds the preset abnormality threshold, it indicates a high probability of failure in the processing equipment and requires intervention to avoid more serious consequences (such as downtime for troubleshooting). Target processing equipment refers to processing equipment with a probability of failure greater than the preset abnormality threshold.

[0082] Based on the abnormal results, parameter adjustments are made to the target processing equipment in descending order of abnormality probability. Parameter adjustments refer to adjustments to the operating parameters of the feed processing equipment. For example, if the feed pellet pulverization rate is predicted to exceed 10% in the next time period, the corresponding processing equipment can be adjusted by increasing the conditioning temperature, steam pressure, and / or grinding degree.

[0083] Based on the adjusted material preparation process data, the abnormality probability of the target processing equipment is updated until the abnormality probability is less than the preset abnormality threshold.

[0084] Based on the adjustment parameters when the abnormal probability is less than the preset abnormal threshold, the operating parameters of the processing equipment are adjusted, and the feed processing is carried out with the new operating parameters.

[0085] In some embodiments, the adjustment parameters of the adjusted processing equipment are also identified. When it is identified that the adjustment parameters are not within the parameter threshold range, an alarm is issued to prompt an inspection of the equipment to determine whether the equipment needs to be replaced; when it is identified that the adjustment parameters are within the parameter threshold range, feed processing is performed with the adjusted processing equipment parameters, and the modification content is recorded to facilitate subsequent query and adjustment.

[0086] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A feed processing prediction method based on deep learning, characterized in that: include: Performing feature processing on the obtained feed preparation process data of multiple feed preparation stages in the feed processing process to obtain multiple feed preparation process features and screening features; Perform feature recognition on screening characteristics to predict the probability and results of abnormal processing in feed processing; If the probability of processing anomaly is greater than the preset processing anomaly probability threshold, multiple material preparation process features are classified to obtain multiple process operation features and process equipment features, and cross-process features are spliced ​​to obtain processing operation features and processing equipment features; Processing the processing operation features through a long short-term memory network to obtain operation correlation features; Processing the processing equipment features through a convolutional neural network to obtain equipment-related features; Perform attention allocation operations on operation-related features and equipment-related features, and determine the equipment abnormality probability of the processing equipment based on the attention contribution of each material preparation process; Get attention contribution, including: Perform self-attention operation on the running correlation features to obtain running self-attention; Perform self-attention operation on device-related features to obtain device self-attention; Perform cross attention operation on running self-attention and device self-attention to obtain cross attention; The attention contribution of each material preparation process is obtained by integrating the operation self-attention, equipment self-attention and cross attention. The calculation formula of the attention contribution is: ； in, Indicates the attention contribution of the i-th material preparation stage; represents the cross attention weight of the i-th process; represents the self-attention feature of the i-th process; represents the self-attention feature of the equipment in the i-th process; represents the L2 general number; represents the cross attention weight of the j-th process; represents the running self-attention feature of the j-th process; Represents the equipment self-attention features of the j-th process.

2. The feed processing prediction method based on deep learning according to claim 1, characterized in that For the crushing process, the process operation characteristics include the crushing blade speed, the discharge valve opening frequency and the vibration mechanism rotating rod speed; the process equipment characteristics include the inclination angle of the No. 1 screen plate, the aperture of the No. 2 screen plate and the proportion of oversize; For the batching process, the process operation characteristics include the raw material extraction speed, motor control amount and pipeline pressure; the process equipment characteristics include the batching accuracy deviation and the correlation between the extraction speeds of different raw materials; For the mixing process, the process operation characteristics include the stirring rod speed, the vibration frequency of the vibrating mechanism and the mixing time; the process equipment characteristics include the lifting amplitude of the receiving plate, the vibration displacement of the fixed block and the CV value of the mixing uniformity; For the pelletizing process, the process operation characteristics include roller speed, die temperature and steam pressure; the process equipment characteristics include roller eccentricity, die gap and motor vibration frequency.

3. The feed processing prediction method based on deep learning according to claim 1, characterized in that Perform self-attention operations on the running correlation features to obtain running self-attention, including: Perform self-attention operation on the operation correlation features between processing steps and extract the process operation self-attention of multiple processing steps; Concatenate and fuse the process operation self-attention to obtain the fused operation self-attention; Perform residual connection on the fused running self-attention and running correlation features to obtain the initial running self-attention; The running self-attention is obtained by performing multi-layer perception on the initial running self-attention and performing residual connection between the multi-layer perception results and the initial running self-attention.

4. The feed processing prediction method based on deep learning according to claim 1, characterized in that Perform self-attention operations on device-related features to obtain device self-attention, including: Perform self-attention operation on the equipment association features between processing steps to extract the process equipment self-attention of multiple processing steps; The process equipment self-attention is spliced ​​and fused to obtain the fused equipment self-attention; Perform residual connection on the fused device self-attention and device-related features to obtain the initial device self-attention; Perform feature decoupling and enhancement on the initial device self-attention to obtain the spatial feature part and structural feature part of the device-related features; Perform global average pooling on the spatial feature part and perform channel attention calculation on the pooling result to obtain the channel attention weight vector; The spatial feature after channel attention enhancement is calculated through the spatial feature part and the channel attention weight vector; Based on the spatial and structural features after channel attention enhancement, the device self-attention is calculated.

5. The feed processing prediction method based on deep learning according to claim 1, characterized in that Perform cross-attention operations on running self-attention and device self-attention to obtain cross-attention, including: Calculate the cross attention of running self-attention and device self-attention to obtain the cross attention weight; Fuse the cross attention weights with the running self-attention to obtain the fused running features; Fuse the cross attention weights with the device self-attention to obtain the fused device features; The fused operation features and fused equipment features are residually connected to obtain the processing flow features.

6. The feed processing prediction method based on deep learning according to claim 1, characterized in that Determine the probability of equipment failure for processing equipment, including: The initial attention model and the initial classification model are jointly trained to obtain the trained attention model and classification model; the loss function of the joint training is: ； in, represents the value of the loss function; k represents the training sample variable; K represents the total number of training samples; Indicates the true abnormal label of the k-th sample. When it is 1, it means the device is truly abnormal, and when it is 0, it means the device is truly normal. It represents the abnormal probability prediction value of the classification model for the kth sample; ln represents the logarithmic function; Input the operation-related features and device-related features into the attention model to obtain the attention contribution and processing flow features; The attention contribution and processing flow characteristics are input into the classification model, and the classification model outputs the abnormality probability of the processing equipment.

7. A method for real-time control of feed processing based on a deep learning-based feed processing prediction method according to any one of claims 1 to 6, characterized in that: include: The processing equipment with a probability of equipment abnormality greater than a preset abnormality threshold is regarded as the target processing equipment; Based on the abnormal results, adjust the parameters of the target processing equipment in descending order of abnormal probability; Based on the adjusted material preparation process data, the abnormal probability of the target processing equipment is updated until the abnormal probability is less than the preset abnormal threshold; Based on the adjustment parameters when the abnormal probability is less than the preset abnormal threshold, the operating parameters of the processing equipment are adjusted, and the feed processing is carried out with the new operating parameters.

8. The method for real-time control of feed processing according to claim 7, characterized in that: It also includes identifying the adjustment parameters of the adjusted processing equipment, When it is identified that the adjustment parameter is not within the parameter threshold range, an alarm is issued; When it is identified that the adjustment parameters are within the parameter threshold range, feed processing is performed with the adjusted processing equipment parameters, and the modification content is recorded.

Images