Abnormality analysis method and system for feed production control system

Abstract

According to the anomaly analysis method and system of the feed production control system provided by the embodiment of the invention, efficient analysis and anomaly identification of system control parameter path data are realized by introducing the target deep learning network. The method comprises the following steps: firstly, performing graph self-coding representation on system control parameter path data by utilizing a first graph generation module to generate a graph self-coding representation vector sequence with representation capability; then, through a feature focusing unit in a second graph generation module, multi-round feature focusing weight distribution is performed on the graph self-encoding representation vector sequence, key features are highlighted, and a first feature focusing vector sequence with higher distinction degree is generated. After the system control parameter path data and the system control parameter path data are fused, anomaly classification is carried out through the full-connection mapping module, anomaly classification data in the system control parameter path data can be accurately recognized, the accuracy and efficiency of anomaly detection are improved, and the stability and reliability of the feed production control system are enhanced.

Description

Technical Field

[0001] This application relates to the technical field of machine learning, and more specifically, to an abnormal analysis method and system for a feed production control system. Background Art

[0002] During the feed production process, the stability and reliability of the feed production control system are crucial for ensuring production quality and efficiency. However, due to the complexity and diversity of the feed production process, system control parameters are often affected by various factors and are prone to abnormal fluctuations or failures, thus affecting the normal operation of the entire production line. Therefore, timely and accurate detection and analysis of abnormalities in the feed production control system are of great significance for ensuring production safety and improving production efficiency.

[0003] Traditional abnormal analysis methods often rely on manual experience or simple threshold settings and are difficult to cope with complex and changing abnormal situations of system control parameters. With the continuous development of artificial intelligence and deep learning technologies, using deep learning networks to intelligently analyze system control parameters has become a new solution. However, existing deep learning networks still have some deficiencies in abnormal analysis, such as inaccurate feature extraction and poor abnormal classification effects, and are difficult to meet the high requirements of the feed production control system for abnormal analysis. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide an abnormal analysis method and system for a feed production control system.

[0005] According to the first aspect of this application, an abnormal analysis method for a feed production control system is provided. The method includes: Obtain the system control parameter path data corresponding to the feed production control system, and load the system control parameter path data into a target deep learning network, where the target deep learning network includes a first graph generation module, a second graph generation module, and a fully connected mapping module; Use the first graph generation module to perform graph auto-encoding representation to generate a graph auto-encoding representation vector sequence corresponding to the feed production control system; Use the second graph generation module to perform multi-round feature focusing weight allocation on the graph auto-encoding representation vector sequence according to a feature focusing unit to generate a first feature focusing vector sequence, where the second graph generation module is obtained by cascading an encoder and the feature focusing unit; Fuse the graph auto-encoding representation vector sequence and the first feature focusing vector sequence to generate a first fused vector sequence; Load the first fused vector sequence into the fully connected mapping module for abnormal classification to generate abnormal classification data of the system control parameter path data.

[0006] According to a second aspect of the present application, an abnormal analysis system for a feed production control system is provided. The abnormal analysis system for the feed production control system includes a processor and a readable storage medium. The readable storage medium stores a program, and when the program is executed by the processor, the foregoing abnormal analysis method for the feed production control system is implemented.

[0007] According to a third aspect of the present application, a computer-readable storage medium is provided. Computer-executable instructions are stored in the computer-readable storage medium, and when it is monitored that the computer-executable instructions are executed, the foregoing abnormal analysis method for the feed production control system is implemented.

[0008] According to any of the above aspects, through the introduction of a target deep learning network, the embodiments of the present application achieve efficient analysis and anomaly recognition of system control parameter path data. First, the first graph generation module is used to perform graph auto-encoding representation on the system control parameter path data, generating a graph auto-encoding representation vector sequence with characterization ability. Subsequently, through the feature focusing unit in the second graph generation module, multi-round feature focusing weight distribution is performed on the graph auto-encoding representation vector sequence, highlighting key features and generating a more discriminative first feature focusing vector sequence. After fusing the two, anomaly classification is performed through the fully connected mapping module, which can accurately identify the anomaly classification data in the system control parameter path data, not only improving the accuracy and efficiency of anomaly detection, but also enhancing the stability and reliability of the feed production control system. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present application and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0010] Figure 1 Shows a schematic flowchart of the abnormal analysis method for the feed production control system provided by the embodiments of the present application; Figure 2 Shows a schematic component structure diagram of the abnormal analysis system for the feed production control system for implementing the foregoing abnormal analysis method for the feed production control system. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0011] To enable those skilled in the art to better understand the solution of this application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application without creative efforts shall fall within the scope of protection of this application.

[0012] It should be noted that the terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of this application described here can be implemented in an order other than those illustrated or described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0013] Figure 1 The flow diagram of the abnormal analysis method of the feed production control system provided by the embodiment of this application is shown. It should be understood that in other embodiments, the order of some steps of the abnormal analysis method of the feed production control system can be interchanged according to actual needs, or some of the steps can also be omitted or deleted. The detailed steps of the abnormal analysis method of the feed production control system are introduced as follows.

[0014] Step S110, obtain the system control parameter path data corresponding to the feed production control system, and load the system control parameter path data into the target deep learning network, where the target deep learning network includes a first graph generation module, a second graph generation module, and a fully connected mapping module.

[0015] In this embodiment, in a feed production enterprise, the server is responsible for the data management and analysis work of the entire feed production control system. The feed production control system covers multiple links, such as raw material feeding, operating parameters of processing equipment, mixing and stirring time, and many other control parameters. The system control parameter path data reflects the change trajectories of these parameters during the production process. For example, in the raw material feeding link, there may be parameters such as the feeding amount and feeding time interval of different raw materials, and the operating parameters of the processing equipment include the rotation speed, temperature, pressure, etc. of the equipment. These parameters form a complete path data over time in the entire production process.

[0016] The server obtains these system control parameter path data from the database of the feed production control system. These data may be stored in a relational database or a dedicated industrial data storage system. Then, the server loads the obtained system control parameter path data into the target deep learning network. The target deep learning network is a network structure specifically constructed for analyzing the data of the feed production control system. The first graph generation module in it is responsible for converting the system control parameter path data into a graph structure-related representation form, the second graph generation module collaborates with the feature focusing unit for feature processing, and the fully connected mapping module is used for the final anomaly classification task.

[0017] When loading data into the target deep learning network, the server adjusts the data format according to the input requirements of the network. For example, if the system control parameter path data is multi-dimensional time series data, the server will organize it into a tensor form that can be accepted by the input layer of the network. This may involve operations such as dimension conversion of the data and unification of data types. Moreover, the server will ensure the integrity of the data during the loading process, perform necessary verification on the data to prevent errors or missing data during transmission or storage.

[0018] Step S120, use the first graph generation module to perform graph auto-encoding representation to generate a graph auto-encoding representation vector sequence corresponding to the feed production control system.

[0019] In this embodiment, first, the obtained system control parameter path data corresponding to the feed production control system is preprocessed. Taking a specific feed production process as an example, before the raw materials enter the processing link, it is necessary to control parameters such as their humidity and particle size. There are complex correlation relationships between these parameters. After the server receives these path data containing multiple control parameters, it first performs data cleaning to remove some obviously incorrect data, such as extremely high or low values caused by sensor failures. Then, data normalization is performed so that parameter values in different ranges can be processed on the same scale. For example, the raw material humidity value (which may be between 0 - 100%) and the equipment pressure value (which may be between 0 - 1000 pascals) are both normalized to the range of 0 - 1. Then, data smoothing is performed to reduce noise interference in the data. Just like the raw material feeding amount may have slight fluctuations due to the slight vibration of the equipment, and through data smoothing, the data can be made more stable.

[0020] Based on the preprocessed system control parameter path data, the server constructs a graph structure representing the operating state of the feed production control system. In this graph structure, nodes represent control parameters or combinations of parameters. For example, a node can represent the combination of raw material humidity and particle size in the raw material feeding process, as they jointly affect the subsequent processing effect. Edges represent the association or dependency relationships between parameters. For instance, there is an association between raw material humidity and mixing time because different humidities may require different mixing times to achieve the best mixing effect, so there will be an edge connecting the nodes representing these two parameters in the graph structure.

[0021] The server extracts the graph structure feature information of the graph structure using graph traversal algorithms or graph embedding algorithms. For example, through graph traversal algorithms, the degree of each node can be calculated, which is the number of edges connected to the node. For the node representing the raw material feeding amount, it may be connected to nodes such as raw material type and feeding time, so its degree reflects its connection complexity in the entire graph structure. The neighbor information of nodes can also be extracted. For example, for the node representing equipment temperature, its neighbor nodes may be the equipment pressure node and the equipment rotation speed node, and this neighbor information helps to understand the status of this node in the entire system. The path length between nodes is also important feature information. For example, the path length from the node in the raw material feeding process to the node in the final product quality inspection process reflects the complexity of the entire production process.

[0022] The server converts the extracted graph structure feature information into an initial vector representation. This means assigning a feature vector with a fixed dimension to each node of the graph structure. For example, the feature information of the node representing raw material humidity, such as the numerical range of humidity and the influence weight on subsequent processing links, is encoded into a vector with a fixed dimension (such as 10 dimensions). Then, this initial vector representation is input into the encoder part of the first graph generation module. The encoder is a multi-layer neural network that, through layer-by-layer compression and feature extraction, converts the high-dimensional initial vector representation into a low-dimensional latent space representation. For example, the initial 10-dimensional vector may be compressed into a 5-dimensional latent space representation. This latent space representation serves as the intermediate representation of the first graph generation module.

[0023] Finally, the server inputs this latent space representation into the decoder part of the first graph generation module. The decoder is also a multi-layer neural network that decodes and reconstructs this latent space representation layer by layer to generate a reconstructed graph structure representation, and this reconstructed graph structure representation is the graph auto-encoding representation vector sequence corresponding to the feed production control system. This vector sequence contains important feature information about the operating state of the feed production control system and exists in a form suitable for subsequent analysis and processing.

[0024] Step S130: Using the second graph generation module, based on the feature focusing unit, perform multi-round feature focusing weight allocation on the graph auto-encoding representation vector sequence to generate a first feature focusing vector sequence. The second graph generation module is obtained by cascading an encoder and the feature focusing unit.

[0025] In this embodiment, first, the graph auto-encoding representation vector sequence generated in step S120 is loaded into the first encoder of the second graph generation module for compression processing. Taking the mixing and stirring link in feed production as an example, the graph auto-encoding representation vector sequence contains information on relevant parameters such as the rotation speed of the stirring paddle, the stirring time, and the raw material mixing ratio. The first encoder will compress this information. Assuming that the original graph auto-encoding representation vector sequence is a 50-dimensional vector, after the compression processing by the first encoder, a 30-dimensional compressed graph auto-encoding representation vector sequence may be generated. This process is like refining the information related to the complex mixing and stirring link, removing some redundant information, and retaining the most crucial part.

[0026] Next, the server loads the graph auto-encoding representation vector sequence into the feature focusing unit. In the feed production control system, the influence degrees of different parameters on the final product quality are different. For example, the proportion of the main nutrient components of the raw materials has a crucial impact on the feed quality, while the feeding time of some auxiliary additives, although also having an impact, is relatively small. The feature focusing unit will generate feature focusing coefficients that reflect the influence weights of the respective second blending vectors in the graph auto-encoding representation vector sequence. For the vectors corresponding to those key parameters, their feature focusing coefficients will be relatively high, while the feature focusing coefficients of the vectors corresponding to relatively minor parameters will be relatively low.

[0027] Then, the server performs optimization processing on each second blending vector in the graph auto-encoding representation vector sequence according to the feature focusing coefficients. Suppose a certain vector in the graph auto-encoding representation vector sequence represents the proportion of the main nutrient components of the raw materials. Due to its importance, according to the feature focusing coefficient, the server will increase the weight of this vector in the overall feature representation. The specific operation may include performing weight normalization on the graph auto-encoding representation vector sequence with initial weights, and fine-tuning the normalized weights according to a preset sensitivity parameter. For example, if the preset sensitivity parameter is 0.1, for important vectors, their weights may increase by 0.1 after normalization. Then, use the weighted summation method to perform feature fusion on the graph auto-encoding representation vector sequence with fine-tuned weights. For each dimension, calculate the weighted sum of all second blending vectors in this dimension to obtain the fused feature. For example, in the two dimensions of raw material nutrient components and additive dosage, calculate the weighted sum according to their respective weights to obtain the fused feature. Then, perform preliminary optimization on the fused feature through smoothing filtering to remove some small fluctuations generated during the fusion process, and obtain a preliminarily optimized fused feature vector sequence.

[0028] The server then uses information gain to evaluate the importance of each fused feature vector in the fused feature vector sequence to obtain a new feature importance score. For example, for the fused feature vectors representing the nutritional components of the raw materials and the amount of additives added, their importance is re-evaluated by calculating information gain. The initially optimized fused feature vector sequence is subjected to a secondary weight adjustment based on the new feature importance score. If the importance score of a feature vector is improved, its weight will increase again. The neighborhood information is then used to enhance the features of the fused feature vector sequence after the secondary weight adjustment. For the feature vector representing the nutritional components of the raw materials, its neighborhood features in the graph autoencoder representation are searched, such as features such as the origin of the raw materials related to the nutritional components of the raw materials, and the weighted sum of these neighborhood features is calculated as the enhanced value of the feature vector. The enhanced value is fused with the original feature value of the feature vector to obtain an enhanced feature vector sequence.

[0029] Finally, the server updates the vector channel of the initial feature focused vector sequence based on the second encoder of the second graph generation module to generate a first feature focused vector sequence. The second encoder adjusts the vector channel according to the optimized feature information obtained by the previous processing, so that the first feature focused vector sequence can better reflect the importance and relationship of each parameter in the feed production control system.

[0030] Step S140: blending the graph self-encoding representation vector sequence with the first feature focus vector sequence to generate a first blended vector sequence.

[0031] In the feed production control system, the server needs to blend the graph self-encoding representation vector sequence with the first feature focus vector sequence. Taking the finished product quality inspection link in the production process as an example, the graph self-encoding representation vector sequence contains various parameter information from raw material input to the processing process, such as the initial quality of the raw materials, the temperature and pressure during the processing, etc. The first feature focus vector sequence is a vector sequence after the weight adjustment and optimization of each parameter after the previous step, and it focuses more on reflecting the influence of key parameters.

[0032] The server blends the two vector sequences through a specific blending algorithm. For example, a weighted blending method can be used to weight each vector in the graph self-encoding representation vector sequence and the corresponding vector in the first feature focus vector sequence according to a certain weight ratio. Assume that a vector in the graph self-encoding representation vector sequence represents the temperature of the equipment during the processing, and its value is 80 degrees (the vector value after encoding). After the corresponding vector in the first feature focus vector sequence is adjusted by weight, it further emphasizes the importance of the equipment temperature to the quality of the finished product. The server performs a weighted sum of the two vectors according to the blending algorithm to obtain the blended vector.

[0033] After performing such a blending operation on all vectors in the graph auto-encoding representation vector sequence and the first feature focusing vector sequence, a first blended vector sequence is generated. This first blended vector sequence combines the comprehensive information in the original graph auto-encoding representation vector sequence and the feature information with adjusted weights in the first feature focusing vector sequence, more comprehensively and accurately reflecting the state of the feed production control system and providing richer feature information for subsequent anomaly classification.

[0034] Step S150: Load the first blended vector sequence into the fully connected mapping module for anomaly classification to generate anomaly classification data for the system control parameter path data.

[0035] In the feed production control system, the server loads the first blended vector sequence into the fully connected mapping module for anomaly classification. The fully connected mapping module is a pre-trained neural network structure specifically used for classification based on the input vector features.

[0036] Taking the raw material supply link in the feed production process as an example, if there are anomalies in the raw material supply, such as raw material supply interruption or sudden decline in raw material quality, this will be reflected in the system control parameter path data. The first blended vector sequence contains parameter information related to raw material supply, such as the feature information of the raw material inventory level, raw material transportation time, etc. after being processed in the previous steps.

[0037] The fully connected mapping module processes the first blended vector sequence according to its internal neuron connections and weight parameters. For example, the neurons in the module perform operations such as weighted summation and activation function calculation on the input vectors. If the input features received by a certain neuron are significantly different from the features in the normal situation, a relatively large value may be output after calculation, indicating that there may be an anomaly.

[0038] After being calculated by the fully connected mapping module, anomaly classification data for the system control parameter path data will be generated. For example, the anomaly classification data may be a binary classification result (normal or abnormal), or a multi-classification result (such as raw material supply anomaly, processing equipment anomaly, finished product quality anomaly, etc.). Based on this anomaly classification data, the server can take corresponding measures in a timely manner. If it is a raw material supply anomaly, the server can notify the relevant department to check the raw material supply channel; if it is a processing equipment anomaly, the server can trigger the equipment maintenance process, etc., so as to ensure the normal operation of the feed production control system.

[0039] Based on the above steps, in the embodiment of the present application, by introducing a target deep learning network, efficient analysis and anomaly recognition of system control parameter path data are achieved. First, the first graph generation module is used to perform graph auto-encoding representation on the system control parameter path data, generating a graph auto-encoding representation vector sequence with characterization ability. Subsequently, through the feature focusing unit in the second graph generation module, multi-round feature focusing weight assignment is performed on the graph auto-encoding representation vector sequence, highlighting key features and generating a more discriminative first feature focusing vector sequence. After fusing the two, anomaly classification is performed through the fully connected mapping module, which can accurately identify the anomaly classification data in the system control parameter path data, not only improving the accuracy and efficiency of anomaly detection, but also enhancing the stability and reliability of the feed production control system.

[0040] In a possible implementation manner, the target deep learning network further includes a recurrent neural module located between the second graph generation module and the fully connected mapping module. After step S140, the method further includes: Step A110, using the recurrent neural module, according to the transfer sequence information of multiple first fusion vectors in the first fusion vector sequence, perform feature focusing weight assignment on the multiple first fusion vectors to generate a second feature focusing vector sequence.

[0041] Step A120, integrate based on the first fusion vector sequence and the second feature focusing vector sequence to generate a first integrated vector sequence.

[0042] Step S150 includes: loading the first integrated vector sequence into the fully connected mapping module for anomaly classification to generate the anomaly classification data of the system control parameter path data.

[0043] In this embodiment, first, the transfer sequence information in the first fusion vector sequence is clarified. In feed production, taking the process from raw material processing to finished product packaging as an example, the first fusion vectors corresponding to the parameters in different stages have a specific transfer sequence. For example, the first fusion vector corresponding to the parameters in the preliminary raw material processing stage precedes the first fusion vector corresponding to the parameters in the mixing and stirring stage. The server will operate according to this inherent transfer sequence information.

[0044] In the recurrent neural module, according to the transfer sequence information of each first fusion vector in the first fusion vector sequence, determine the retrieval vector corresponding to the last transfer node and at least one memory vector corresponding to other transfer nodes from the first fusion vector sequence. For example, in the feed production process, the first fusion vector corresponding to the finished product packaging stage may be determined as the retrieval vector corresponding to the last transfer node, while the first fusion vectors corresponding to the previous stages such as raw material processing and mixing and stirring are used as memory vectors.

[0045] Next, the retrieval vector is successively used to calculate the feature distance from the memory vector of each transfer node, so as to generate the attention weight value corresponding to the memory vector of each transfer node. In the feed production scenario, this means that if there is a large feature difference between a certain parameter (related to the retrieval vector) in the finished product packaging stage and a certain parameter (related to the memory vector) in the raw material processing stage, it will be reflected when calculating the feature distance, so as to obtain the corresponding attention weight value, which accurately reflects the correlation between the memory vector and the retrieval vector of each transfer node.

[0046] Then, each first fusion vector in the corresponding memory vector is updated according to the attention weight value of each memory vector to generate a second feature focus vector sequence. For example, if the memory vector corresponding to a certain parameter in the raw material processing stage has a low correlation with the retrieval vector in the finished product packaging stage, according to the calculated attention weight value, the server will adjust the representation of this parameter in the memory vector in the raw material processing stage. After such processing of all memory vectors, a second feature focus vector sequence is generated.

[0047] After that, integration is performed based on the first fusion vector sequence and the second feature focus vector sequence to generate a first integration vector sequence. In the feed production control system, this is equivalent to integrating the original first fusion vector sequence containing comprehensive information and the second feature focus vector sequence that is more focused on the transfer sequence relationship and correlation adjustment after being processed by the recurrent neural module. For example, the parameter information of the raw material processing stage in the first fusion vector sequence is integrated with the adjusted parameter information of the raw material processing stage in the second feature focus vector sequence. After performing such integration operations on the parameter information of all stages, the first integration vector sequence is obtained.

[0048] Finally, the first integration vector sequence is loaded into the fully connected mapping module for anomaly classification to generate anomaly classification data for the system control parameter path data. Just like the various links in the feed production process mentioned above, the fully connected mapping module will analyze according to the comprehensive information of the relevant parameters of each link such as raw material supply, processing process, and finished product packaging in the first integration vector sequence. For example, if the information in the first integration vector sequence of the raw material supply link shows that the raw material inventory level suddenly drops and remains below the safety threshold, and at the same time, the equipment operation parameters in the processing process also show abnormal fluctuations, after internal neuron calculations by the fully connected mapping module, it will accurately determine that this is an abnormal situation, so as to generate anomaly classification data for the system control parameter path data. This anomaly classification data can clarify specific situations such as anomalies in both the raw material supply and the processing process, so that the server can take corresponding measures according to this result, such as notifying the raw material procurement department to replenish raw materials and arranging equipment maintenance personnel to check the processing equipment.

[0049] In a possible implementation manner, step A110 includes: Step A111: Using the recurrent neural module, according to the flow order information of each first fusion vector in the first fusion vector sequence, determine the retrieval vector corresponding to the last flow node and at least one memory vector corresponding to other flow nodes from the first fusion vector sequence.

[0050] Step A112: Calculate the feature distance between the retrieval vector and the memory vector of each flow node in sequence, and generate an attention weight value corresponding to the memory vector of each flow node. The attention weight value reflects the correlation between the memory vector of each flow node and the retrieval vector.

[0051] Step A113: Update each first fusion vector in the corresponding memory vector according to the attention weight value of each memory vector to generate a second feature focusing vector sequence.

[0052] In a possible implementation manner, step A113 includes: Step A1131: Update each first fusion vector in the corresponding memory vector according to the attention weight value of each memory vector to generate an updated first fusion vector.

[0053] Step A1132: For each vector channel of the first fusion vector sequence, fuse the first fusion vectors of different flow nodes corresponding to the vector channel to generate a target fusion vector corresponding to the vector channel.

[0054] Step A1133: Generate a second feature focusing vector sequence based on the multiple target fusion vectors corresponding to multiple vector channels.

[0055] In this embodiment, first, the server uses a recurrent neural module to determine, based on the transfer sequence information of each first fusion vector in the first fusion vector sequence, the retrieval vector corresponding to the last transfer node and at least one memory vector corresponding to other transfer nodes from the first fusion vector sequence. During the feed production process, the parameter information corresponding to each link in the entire process from raw material procurement to final product packaging is encoded in the first fusion vector sequence. Taking a complete feed production process as an example, the raw material procurement link involves parameters such as the type, quality, and supplier of raw materials. The processing link includes operating parameters of equipment such as temperature, pressure, and rotation speed. The mixing and stirring link has parameters such as stirring time, speed, and sequence. And the final finished product packaging link includes parameters such as packaging materials and packaging specifications. These links are carried out in sequence according to the production order, and the corresponding parameter vectors also have a specific transfer sequence. For example, if the finished product packaging link is at the end of the entire process, then the first fusion vector corresponding to this link is determined as the retrieval vector corresponding to the last transfer node. And the first fusion vectors corresponding to the previous links such as raw material procurement, processing, and mixing and stirring become the memory vectors corresponding to other transfer nodes. The server accurately identifies these different types of vectors by precisely analyzing the transfer sequence information in the first fusion vector sequence, which lays the foundation for subsequent operations.

[0056] Next, the server sequentially calculates the feature distance between the retrieval vector and the memory vector of each transfer node, generating an attention weight value corresponding to the memory vector of each transfer node. The attention weight value reflects the correlation between the memory vector of each transfer node and the retrieval vector. In the feed production scenario, assume that the retrieval vector of the finished product packaging link contains parameter information such as packaging specifications and packaging weights. For the memory vector of the raw material procurement link, it contains parameter information such as the type and quality of raw materials. The server calculates the feature distance between these two vectors through a specific algorithm. For example, if there is a certain correlation between the packaging specifications of the finished product and the type of raw materials, such as a specific raw material being suitable for a specific packaging specification, then this correlation will be reflected in the distance value when calculating the feature distance. If the two are closely related, the distance value is small; otherwise, it is large. Through such calculations, an attention weight value corresponding to the memory vector of the raw material procurement link is generated. In the same way, the server calculates the feature distance between the memory vectors of each transfer node such as the processing link and the mixing and stirring link and the retrieval vector of the finished product packaging link, thereby obtaining the attention weight value corresponding to the memory vector of each transfer node. These attention weight values accurately reflect the correlation between each memory vector and the retrieval vector, that is, the degree of feature association between each link and the final finished product packaging link.

[0057] Then, the server updates each first fusion vector in the corresponding memory vector according to the attention weight value of each memory vector to generate a second feature focus vector sequence. Specifically, each first fusion vector in the corresponding memory vector is updated according to the attention weight value of each memory vector to generate an updated first fusion vector. For example, in the memory vector of the raw material procurement link, assume that one of the first fusion vectors represents the quality parameters of the raw material. If the attention weight value of this memory vector and the retrieval vector of the finished product packaging link indicates a strong correlation between the two, then the server will update the first fusion vector corresponding to the raw material quality parameter according to this weight value. This update may involve operations such as adjusting the numerical values in the vector or re-weighting the feature information contained in the vector, so as to generate an updated first fusion vector.

[0058] For each vector channel of the first fusion vector sequence, the first fusion vectors of different transfer nodes corresponding to the vector channel are fused to generate a target fusion vector corresponding to the vector channel. In the feed production control system, the first fusion vector sequence may contain multiple vector channels, and each channel corresponds to different types of parameter information. For example, one vector channel may specifically correspond to parameters related to raw materials, such as the quality, humidity, and nutritional components of the raw materials. The server will fuse the first fusion vectors of different transfer nodes such as the raw material procurement link, processing link, and mixing and stirring link in this vector channel. Assume that the raw material quality in the first fusion vector of the raw material procurement link is 80 points (the value after being updated in the previous steps), the vector related to the raw material quality in the processing link is adjusted to 82 points (due to possible changes in the raw material quality during the processing), and the mixing and stirring link is 81 points. The server fuses the first fusion vectors of these different transfer nodes through a specific fusion algorithm (such as weighted average, etc.) to obtain the target fusion vector corresponding to this vector channel. For example, the calculated raw material quality in the target fusion vector is 81 points (assuming a simple weighted average algorithm is used).

[0059] Finally, a second feature focusing vector sequence is generated based on the multiple target fusion vectors corresponding to multiple vector channels. In the feed production system, in addition to the vector channels related to raw materials, there may also be vector channels related to equipment, packaging-related vector channels, etc. The server performs the above-mentioned fusion operation on each vector channel to obtain the target fusion vectors corresponding to each vector channel. For example, the vector channel related to equipment obtains a target fusion vector reflecting the overall condition of the equipment and its association with the final product packaging after fusion, and the packaging-related vector channel also obtains the corresponding target fusion vector. Combining the target fusion vectors corresponding to these different vector channels generates the second feature focusing vector sequence. This second feature focusing vector sequence synthesizes the association information of each link with the final product packaging link and is effectively fused on each vector channel, providing more targeted and accurate feature information for subsequent integration based on the first fusion vector sequence and the second feature focusing vector sequence and the final anomaly classification.

[0060] In a possible implementation manner, step A111 includes: Step A1111, receiving the first fusion vector sequence generated after the fusion of the first graph generation module and the second graph generation module, and the flow order information of each first fusion vector in the first fusion vector sequence, where the flow order information is represented by a data structure of a timestamp or a sequence index.

[0061] Step A1112, performing a unique node identification on each first fusion vector in the first fusion vector sequence, and outputting the first fusion vector sequence carrying the unique node identification and its corresponding flow order information.

[0062] Step A1113, constructing a flow path of the first fusion vector sequence based on the extracted flow order information, where the flow path is represented as a directed graph or an undirected graph, and the nodes represent the first fusion vectors and the edges represent the flow relationships between the vectors.

[0063] Step A1114, identifying the last flow node in the constructed flow path, where the last flow node is the end point of the flow path or the last processed node, and outputting the flow path carrying the last node identification and the retrieval vector corresponding to the last flow node.

[0064] Step A1115, after identifying the last flow node, backtracking the flow path to find all the precursor nodes directly pointing to the last flow node. The precursor nodes are the nodes processed before the last flow node in the flow path. For each precursor node, initialize a memory vector set for storing the feature information of the precursor node and its subsequent nodes. The memory vector is the first fusion vector corresponding to the precursor node or the result obtained after transforming the first fusion vector.

[0065] Step A1116: For each predecessor node in the transfer path, traverse its successor nodes in reverse order of the transfer sequence, and update the memory vector set of the predecessor node. The update process includes adding the feature information of the successor node to the memory vector of the predecessor node, or transforming the memory vector of the predecessor node to fuse the information of the successor node, thereby generating an updated memory vector set, and the transfer relationship based on the transfer sequence between the updated memory vector sets.

[0066] Step A1117: Optimize the updated memory vector set based on the correlation between the memory vector and the retrieval vector corresponding to the last transfer node, and the redundancy between the memory vectors, to generate an optimized memory vector set.

[0067] Step A1118: Establish an association relationship between the optimized memory vector set and the retrieval vector corresponding to the last transfer node. The association relationship is established by calculating the similarity or distance metric between the memory vector and the retrieval vector.

[0068] In this embodiment, in the feed production scenario, the first fusion vector sequence contains parameter information related to each link from the start of raw material processing to the output of the finished product. For example, parameters such as the type, quality, and humidity of the raw materials are encoded in the first fusion vector sequence after passing through the previous graph generation module and fusion operations. The transfer sequence information clarifies the order of these parameters in the production process. For example, the first fusion vector corresponding to the parameters of the raw material processing link precedes the first fusion vector corresponding to the parameters of the processing link, and the timestamp or sequence index can accurately record this order relationship.

[0069] Next, the server assigns a unique node identifier to each first fusion vector in the first fusion vector sequence. In feed production, each first fusion vector corresponding to each link has its unique meaning. For example, the first fusion vector corresponding to the equipment operation parameters in the processing link can be distinguished from the vectors of other links through the unique node identifier. The server outputs the first fusion vector sequence carrying the unique node identifier and its corresponding transfer sequence information, which enables each first fusion vector to have a clear identity identifier and be closely associated with the transfer sequence information.

[0070] Then, the server constructs the transfer path of the first fusion vector sequence based on the extracted transfer sequence information. The transfer path is represented as a directed graph or an undirected graph, where nodes represent the first fusion vectors and edges represent the transfer relationships between the vectors. In the feed production system, taking the links such as raw material input, processing, mixing and stirring, and finished product packaging as examples, the first fusion vector corresponding to the raw material input link is used as one node, and the first fusion vector corresponding to the processing link is used as another node. If the parameters of the raw material input link affect the parameters of the processing link, then there will be an edge in the transfer path graph pointing from the raw material input node to the processing node, indicating this transfer relationship. In this way, a complete transfer path graph reflecting the transfer relationship of the parameter vectors in each link of the feed production process is constructed.

[0071] In the constructed transfer path, the server identifies the last transfer node. The last transfer node is the end point of the transfer path or the last processed node. Then, it outputs the transfer path carrying the last node identifier and the retrieval vector corresponding to the last transfer node. For example, in the feed production process, the finished product packaging link is the last link, so the first fusion vector corresponding to this link is the retrieval vector corresponding to the last transfer node. This retrieval vector contains parameter information related to the finished product packaging link, such as packaging materials, packaging specifications, etc.

[0072] After identifying the last transfer node, the server traces back the transfer path to find all the precursor nodes that directly point to the last transfer node. A precursor node is a node that was processed before the last transfer node in the transfer path. For each precursor node, the server initializes a memory vector set to store the feature information of the precursor node and its subsequent nodes. The memory vector is the first fusion vector corresponding to the precursor node, or the result obtained after transforming this first fusion vector. In the feed production scenario, taking the finished product packaging link as the last transfer node, then the mixing and stirring link is one of the precursor nodes. For the precursor node of the mixing and stirring link, the server initializes a memory vector set, which will store the feature information of the mixing and stirring link and the subsequent links (if any) that may affect the finished product packaging link. The first fusion vector corresponding to the mixing and stirring link or the result after a certain transformation becomes the initial content in this memory vector set.

[0073] For each predecessor node in the transfer path, the server traverses its successor nodes in reverse order of the transfer sequence and updates the memory vector set of the predecessor node. The update process includes adding the feature information of the successor node to the memory vector of the predecessor node, or transforming the memory vector of the predecessor node to fuse the information of the successor node, thereby generating an updated memory vector set, as well as the transfer relationship based on the transfer sequence between the updated memory vector sets. For example, for the predecessor node of the mixing and stirring process, traverse the successor nodes in reverse order (assuming only the finished product packaging process is the successor node). If a certain parameter of the finished product packaging process (such as the packaging specification) is related to the stirring speed of the mixing and stirring process, then the server will add the feature information of the packaging specification to the memory vector of the mixing and stirring process, or perform a certain transformation on the memory vector of the mixing and stirring process (such as adjusting the numerical weights in the vector according to their correlation) to fuse the information of the packaging specification, so as to update the memory vector set of the mixing and stirring process. At the same time, if there are multiple predecessor nodes, such as the processing process and the mixing and stirring process are both predecessor nodes of the finished product packaging process, then after updating their respective memory vector sets, a transfer relationship will be established between these updated memory vector sets based on the transfer sequence. For example, some information in the memory vector set of the processing process may be transferred to the memory vector set of the mixing and stirring process according to the transfer sequence to reflect the coherence of the entire production process.

[0074] After that, the server optimizes the updated memory vector set based on the correlation between the memory vector and the retrieval vector corresponding to the last transfer node, and the redundancy between the memory vectors, generating an optimized memory vector set. In feed production, taking the retrieval vector of the finished product packaging process and the memory vector of the mixing and stirring process as an example, if some parameters in the memory vector of the mixing and stirring process (such as the stirring speed) have a high correlation with the packaging specification in the retrieval vector of the finished product packaging process, then the weights of these parameters in the memory vector may be increased. At the same time, if there is some redundant information in the memory vector, such as a certain parameter has been included by other parameters or has a minimal impact on the final result, the server will reduce the weight of this redundant information or directly delete it according to the redundancy between the memory vectors, thereby optimizing the memory vector set.

[0075] Finally, the server establishes an association relationship between the optimized memory vector set and the retrieval vector corresponding to the last flow node. The association relationship is established by calculating the similarity or distance measurement between the memory vector and the retrieval vector. In the feed production system, for the retrieval vector of the finished product packaging link and the memory vector of the optimized mixing link, the server calculates the similarity or distance measurement between the two through a specific algorithm. If the similarity between the two is high, it means that the mixing link and the finished product packaging link have a strong correlation in terms of features. The establishment of this correlation relationship helps to perform more accurate feature focus weight allocation and other operations on the first fusion vector in the recurrent neural module, thereby providing a more accurate feature information basis for tasks such as abnormal classification of the entire feed production control system.

[0076] In a possible implementation, step S130 includes: Step S131: Load the graph self-encoding representation vector sequence into the first encoder of the second graph generation module for compression processing to generate a compressed graph self-encoding representation vector sequence.

[0077] Step S132: Load the graph self-encoding representation vector sequence into a feature focusing unit, and generate a feature focusing coefficient reflecting the influence weight of each second blending vector in the graph self-encoding representation vector sequence.

[0078] Step S133, optimizing each of the second blending vectors in the graph self-encoding representation vector sequence according to the feature focusing coefficient to generate an initial feature focusing vector sequence.

[0079] Step S134: Based on the second encoder of the second image generation module, the vector channel of the initial feature focus vector sequence is updated to generate a first feature focus vector sequence.

[0080] In a possible implementation, step S133 includes: Step S1331, receiving a feature focusing coefficient sequence generated by a feature focusing unit, wherein each feature focusing coefficient in the feature focusing coefficient sequence corresponds one-to-one to a second fusion vector in the graph self-encoding representation vector sequence, indicating the importance of the second fusion vector in the overall feature representation.

[0081] Step S132, initialize a weight vector of the same dimension as the graph self-encoding representation vector sequence, wherein the initial value of each weight vector is set to the corresponding feature focusing coefficient, which is used as the basis for initial weight adjustment, so that the initial weight of each second fusion vector is proportional to its feature focusing coefficient, thereby reflecting the different importance of each second fusion vector in the feature representation.

[0082] Step S1333: Perform weight normalization on the graph auto-encoding representation vector sequence with initial weights. After fine-tuning the normalized weights according to a preset sensitivity parameter, use weighted summation to perform feature fusion on the graph auto-encoding representation vector sequence with fine-tuned weights. Specifically, for each dimension, calculate the weighted sum of all second fusion vectors in this dimension to obtain the fused feature, and perform preliminary optimization on the fused feature through smoothing filtering to obtain a preliminarily optimized fused feature vector sequence.

[0083] Step S1334: Use information gain to evaluate the importance of each fused feature vector in the fused feature vector sequence to obtain a new feature importance score. After performing secondary weight adjustment on the preliminarily optimized fused feature vector sequence according to the new feature importance score, use neighborhood information to perform feature enhancement on the fused feature vector sequence after secondary weight adjustment. Specifically, for each fused feature vector, find its neighborhood features in the graph auto-encoding representation of this fused feature vector, and calculate the weighted sum of the neighborhood features as the enhancement value of this fused feature vector. Fuse the enhancement value with the original feature value of this fused feature vector to obtain an enhanced feature vector sequence.

[0084] Step S1335: Perform feature dimensionality reduction on the enhanced feature vector sequence through linear discriminant analysis to generate an initial feature focus vector sequence.

[0085] In this embodiment, in the feed production scenario, the graph auto-encoding representation vector sequence contains information on numerous parameters in each link of the feed production process. For example, parameters such as raw material types, humidity, and nutrient components in the raw material link, parameters such as equipment temperature, pressure, and running duration in the processing link, and parameters such as stirring speed and stirring time in the mixing and stirring link are all encoded in this vector sequence. The first encoder compresses this complex vector sequence, just like refining a large amount of raw data. Suppose the graph auto-encoding representation vector sequence originally contained information in 100 dimensions. After being compressed by the first encoder, it may be reduced to 60 dimensions, generating a compressed graph auto-encoding representation vector sequence, which can remove some redundant information and retain the most critical feature information for subsequent processing.

[0086] Next, in the feed production system, different parameters have varying degrees of influence on the quality of the final feed product and production efficiency. For example, the proportion of nutrient components in raw materials has a crucial impact on feed quality, while minor changes in the operating parameters of certain auxiliary equipment may have relatively less impact. The feature focusing unit generates corresponding feature focusing coefficients for each second blending vector in the graph auto-encoding representation vector sequence based on the internal relationships and importance of these parameters. These coefficients accurately reflect the importance of each second blending vector in the overall feature representation. For example, the feature focusing coefficient of the second blending vector corresponding to the nutrient components of raw materials may be relatively high, while the feature focusing coefficient of the second blending vector corresponding to the operating parameters of some secondary equipment may be relatively low.

[0087] Then, the server first receives the sequence of feature focusing coefficients generated by the feature focusing unit. Each feature focusing coefficient in this sequence corresponds one-to-one with a second blending vector in the graph auto-encoding representation vector sequence, indicating the importance of this second blending vector in the overall feature representation. Then, the server initializes a weight vector with the same dimension as the graph auto-encoding representation vector sequence, where the initial value of each weight vector is set to the corresponding feature focusing coefficient, serving as the basis for initial weight adjustment, so that the initial weight of each second blending vector is proportional to its feature focusing coefficient, thereby reflecting the different importance of each second blending vector in the feature representation. For example, for an important second blending vector (such as the vector corresponding to the nutrient components of raw materials), its feature focusing coefficient is relatively high, so in the initialized weight vector, the weight corresponding to this vector is also relatively high, highlighting its importance in the overall feature representation.

[0088] After that, the server performs weight normalization on the graph auto-encoding representation vector sequence with initial weights, and then fine-tunes the normalized weights according to the preset sensitivity parameter. After that, the server performs feature fusion on the graph auto-encoding representation vector sequence with fine-tuned weights using the weighted summation method. In feed production, weight normalization is to adjust each weight value to a suitable range to ensure that all weights are compared and calculated on the same scale. Suppose the preset sensitivity parameter is 0.1. If the weight of a certain second blending vector needs to be fine-tuned according to its importance after normalization, then it is adjusted according to this sensitivity parameter. Then, for each dimension, the weighted sum of all second blending vectors in this dimension is calculated to obtain the fused feature. For example, in the two dimensions of raw material nutrient components and equipment temperature, the weighted sum is calculated according to their respective weights to obtain the fused feature. And, the fused feature is preliminarily optimized through smoothing filtering to remove some small fluctuations generated during the fusion process, obtaining a preliminarily optimized fused feature vector sequence.

[0089] Subsequently, the server uses information gain to evaluate the importance of each fused feature vector in the fused feature vector sequence, obtaining new feature importance scores. After performing a secondary weight adjustment on the preliminarily optimized fused feature vector sequence according to the new feature importance scores, the server uses neighborhood information to enhance the fused feature vector sequence after the secondary weight adjustment. In the feed production scenario, information gain can measure the contribution degree of each fused feature vector to the final result (such as feed quality or production efficiency), thereby obtaining new feature importance scores. If the new feature importance score of a certain fused feature vector is relatively high, it indicates that it has a greater impact on the final result. Then, during the secondary weight adjustment, its weight will be increased. Then, for each fused feature vector, the neighborhood features in the self - encoding representation of the fused feature vector graph are found, and the weighted sum of the neighborhood features is calculated as the enhancement value of the fused feature vector. The enhancement value is fused with the original feature value of the fused feature vector to obtain an enhanced feature vector sequence. For example, for the fused feature vector representing the nutritional components of raw materials, its neighborhood features may include information such as the origin of raw materials. By calculating the weighted sum of these neighborhood features and fusing it with the original feature value, the representation ability of the fused feature vector can be enhanced.

[0090] Finally, the enhanced feature vector sequence may still contain a relatively large amount of dimensional information, and there may be some redundant or highly correlated dimensions among them. Linear discriminant analysis can find the feature dimensions that can best distinguish different categories (such as feeds of different quality grades) based on the differences between different categories, thereby reducing the dimensions of the enhanced feature vector sequence. For example, the enhanced feature vector sequence that may originally contain 30 dimensions is reduced to 20 dimensions, generating an initial feature - focused vector sequence. This initial feature - focused vector sequence reduces the data complexity while retaining the key feature information, providing a more suitable data basis for the second encoder based on the second graph generation module to update the vector channels of the initial feature - focused vector sequence to generate the first feature - focused vector sequence.

[0091] In a possible implementation manner, step S120 includes: Step S121, receiving the system control parameter path data corresponding to the feed production control system, where the system control parameter path data includes the time - series values of each control parameter during the operation of the feed production control system and the correlation relationships between them.

[0092] Step S122, performing data cleaning, data normalization, and data smoothing on the system control parameter path data to generate pre - processed system control parameter path data. Based on the pre - processed system control parameter path data, a graph structure representing the operating state of the feed production control system is constructed. In the graph structure, nodes represent control parameters or parameter combinations, and edges represent the correlation relationships or dependency relationships between parameters.

[0093] Step S123: Use a graph traversal algorithm or a graph embedding algorithm to extract the graph structure feature information of the graph structure. The feature information includes the degree of nodes, the neighbor information of nodes, and the path length between nodes.

[0094] Step S124: Convert the extracted graph structure feature information into an initial vector representation, that is, assign a feature vector with a fixed dimension to each node of the graph structure.

[0095] Step S125: Input the initial vector representation into the encoder part of the first graph generation module, and perform layer-by-layer compression and feature extraction through a multi-layer neural network to generate a low-dimensional latent space representation, which serves as the intermediate representation of the first graph generation module.

[0096] Step S126: Input the latent space representation into the decoder part of the first graph generation module, and perform layer-by-layer decoding and reconstruction on the latent space representation through a multi-layer neural network to generate a reconstructed graph structure representation, which serves as the graph auto-encoding representation vector sequence corresponding to the feed production control system.

[0097] In this embodiment, in the feed production scenario, the first fusion vector sequence contains parameter information related to each link in the process from raw material processing to finished product output. For example, parameters such as the type, quality, and humidity of raw materials are encoded in the first fusion vector sequence after passing through the previous graph generation module and the fusion operation. The transfer sequence information clarifies the order of these parameters in the production process. For example, the first fusion vector corresponding to the parameters in the raw material processing link precedes the first fusion vector corresponding to the parameters in the processing link, and the timestamp or sequence index can accurately record this order relationship.

[0098] Next, in feed production, each first fusion vector corresponding to each link has its unique meaning. For example, the first fusion vector corresponding to the equipment operation parameters in the processing link can be distinguished from the vectors of other links through a unique node identifier. The server outputs the first fusion vector sequence carrying the unique node identifier and its corresponding transfer sequence information, which makes each first fusion vector have a clear identity identifier and is closely associated with the transfer sequence information.

[0099] Then, taking the links such as raw material input, processing, mixing and stirring, and finished product packaging as examples, the first fusion vector corresponding to the raw material input link is used as one node, and the first fusion vector corresponding to the processing link is used as another node. If the parameters in the raw material input link affect the parameters in the processing link, then there will be an edge in the flow path diagram pointing from the raw material input node to the processing node, indicating this flow relationship. In this way, a complete flow path diagram reflecting the vector flow relationship of parameters in each link of the feed production process is constructed.

[0100] In the constructed flow path, the server identifies the last flow node. The last flow node is the end point of the flow path or the last processed node, and then outputs the flow path carrying the last node identifier and the retrieval vector corresponding to the last flow node. For example, in the feed production process, the finished product packaging link is the last link, so the first fusion vector corresponding to this link is the retrieval vector corresponding to the last flow node. This retrieval vector contains parameter information related to the finished product packaging link, such as packaging materials, packaging specifications, etc.

[0101] After identifying the last flow node, the server traces back the flow path to find all the precursor nodes directly pointing to the last flow node. A precursor node is a node processed before the last flow node in the flow path. For each precursor node, the server initializes a set of memory vectors to store the feature information of the precursor node and its subsequent nodes. The memory vector is the first fusion vector corresponding to the precursor node or the result obtained after transforming this first fusion vector. In the feed production scenario, taking the finished product packaging link as the last flow node, then the mixing and stirring link is one of the precursor nodes. For the precursor node of the mixing and stirring link, the server initializes a set of memory vectors, and this set will store the feature information of the mixing and stirring link and the subsequent links (if any) that may affect the finished product packaging link. The first fusion vector corresponding to the mixing and stirring link or the result after a certain transformation becomes the initial content in this set of memory vectors.

[0102] For each predecessor node in the flow path, the server traverses its successor nodes in reverse order of the flow sequence and updates the memory vector set of the predecessor node. The update process includes adding the feature information of the successor node to the memory vector of the predecessor node, or transforming the memory vector of the predecessor node to fuse the information of the successor node, thereby generating an updated memory vector set and the transfer relationship based on the flow sequence among the updated memory vector sets. For example, for the predecessor node of the mixing and stirring process, traverse the successor nodes in reverse order (assuming only the finished product packaging process is the successor node). If a certain parameter of the finished product packaging process (such as the packaging specification) is related to the stirring speed of the mixing and stirring process, then the server will add the feature information of the packaging specification to the memory vector of the mixing and stirring process, or perform a certain transformation on the memory vector of the mixing and stirring process (such as adjusting the numerical weight in the vector according to their correlation) to fuse the information of the packaging specification, thereby updating the memory vector set of the mixing and stirring process. At the same time, if there are multiple predecessor nodes, such as the processing process and the mixing and stirring process are both predecessor nodes of the finished product packaging process, then after updating their respective memory vector sets, a transfer relationship based on the flow sequence will be established among these updated memory vector sets. For example, some information in the memory vector set of the processing process may be transferred to the memory vector set of the mixing and stirring process according to the flow sequence to reflect the coherence of the entire production process.

[0103] After that, the server optimizes the updated memory vector set based on the correlation between the memory vector and the retrieval vector corresponding to the last flow node, and the redundancy among the memory vectors, generating an optimized memory vector set. In feed production, taking the retrieval vector of the finished product packaging process and the memory vector of the mixing and stirring process as an example, if some parameters in the memory vector of the mixing and stirring process (such as the stirring speed) have a high correlation with the packaging specification in the retrieval vector of the finished product packaging process, then the weights of these parameters in the memory vector may be increased. At the same time, if there is some redundant information in the memory vector, such as a certain parameter has been included by other parameters or has little impact on the final result, the server will reduce the weight of this redundant information or directly delete it according to the redundancy among the memory vectors, thereby optimizing the memory vector set.

[0104] Finally, the server establishes an association relationship between the optimized memory vector set and the retrieval vector corresponding to the last transfer node. The association relationship is established by calculating the similarity or distance metric between the memory vector and the retrieval vector. In the feed production system, for the retrieval vector of the finished product packaging link and the memory vector of the optimized mixing and stirring link, the server calculates the similarity or distance metric between the two through a specific algorithm. If the similarity between the two is high, it indicates that there is a strong association relationship between the mixing and stirring link and the finished product packaging link. The establishment of this association relationship helps to perform more accurate feature focusing weight allocation and other operations on the first fusion vector in the recurrent neural module subsequently, thereby providing a more accurate feature information basis for tasks such as anomaly classification in the entire feed production control system.

[0105] In a possible implementation manner, the training steps of the target deep learning network include: Obtain a first deep learning network trained under a first processing scenario.

[0106] Determine, from the first deep learning network, target topological dimensions corresponding to each set topological dimension in the reference set deep learning network of a second processing scenario.

[0107] Obtain the neuron parameter information of each target topological dimension, and convert the neuron parameter information into the set topological dimension corresponding to the reference set deep learning network to generate an initialized deep learning network with updated neuron parameter information.

[0108] Obtain the sample system control parameter path data sequence corresponding to the second processing scenario, and perform network parameter learning on the initialized deep learning network according to the sample system control parameter path data sequence to generate a target deep learning network.

[0109] In this embodiment, the first processing scenario may be the production process of a specific type of feed, such as the production of chicken feed. In this scenario, the first deep learning network has been trained with a large amount of data. This network is constructed and trained based on various system control parameter path data in the chicken feed production process. These data include information on various links from raw material procurement (such as parameters of the quality, quantity, supply time, etc. of raw materials such as corn and soybean meal), processing (such as the values of parameters such as equipment temperature, pressure, stirring speed, etc. at different times and their association relationships), to finished product packaging (such as parameters of packaging specifications, weight, speed, etc.). This trained first deep learning network can effectively analyze and process various situations in the chicken feed production process, such as identifying abnormal situations in the production process and predicting production efficiency.

[0110] Next, assume that the second processing scenario is the production of pig feed. Since there are similarities and differences between the production processes of pig feed and chicken feed, a deep learning network for pig feed production needs to be constructed. Refer to the initial network structure specifically designed for pig feed production. It has specific set topological dimensions. For example, the number of neurons in the input layer corresponds to the number of raw material types in pig feed production, and the number of hidden layers and neurons is set according to the estimated complexity of the pig feed production process, etc. The server needs to find the target topological dimensions in the first deep learning network (related to chicken feed production) that correspond to each set topological dimension in the reference set deep learning network for pig feed production. For example, if the number of neurons in the input layer of the reference set deep learning network is set according to 10 main raw materials of pig feed, then the dimension corresponding to the number of neurons in the input layer related to raw materials in the first deep learning network (chicken feed production) is one of the target topological dimensions to be determined. The purpose of this step is to utilize the existing structural information in the first deep learning network to build the network structure basis applicable to the second processing scenario (pig feed production).

[0111] Then, the neuron parameter information includes parameters such as weights and biases. For example, for the weight and bias parameters of the hidden layer neurons related to the temperature control of chicken feed processing equipment in the first deep learning network (obtained during the training of the chicken feed production process), the server will convert them into the set topological dimensions related to the temperature control of processing equipment in the reference set deep learning network for pig feed production. By converting the neuron parameter information for each target topological dimension in this way, an initialized deep learning network is generated. This network is based on the reference set deep learning network for pig feed production in terms of structure, but incorporates some information from the first deep learning network (related to chicken feed production) in terms of neuron parameters, providing an initial model basis for the subsequent further training for pig feed production.

[0112] Finally, the sample system control parameter path data sequence contains a large amount of actual data on the pig feed production process. In terms of raw materials, there are detailed parameters for various pig feed raw materials (such as corn, bran, fish meal, etc.), including the quality fluctuations of raw materials provided by different suppliers, the impact of raw material prices in different seasons on the purchase volume, etc.; in the processing link, the operating parameters of equipment (such as the change of crushing particle size of crushers, the mixing uniformity of mixers, the temperature and pressure of pellet mills over time and their correlation relationships); in the finished product link, the relationships between parameters such as the nutritional components, particle size, and packaging form of pig feed and the parameters of the previous production links. The server uses these rich sample system control parameter path data sequences to train the initialized deep learning network. During the training process, the network continuously adjusts parameters such as the connection weights and biases between neurons according to the patterns and rules in the data to improve its analysis and processing capabilities for the pig feed production process. For example, if a specific relationship is found between the change in the quality of a certain raw material and the nutritional components of the finished product, the network will adjust the parameters of the relevant neurons to better capture this relationship. After this process, a target deep learning network is finally generated, which can accurately analyze, predict, and detect anomalies in various situations in the pig feed production control system.

[0113] Furthermore, Figure 2 shows a schematic hardware structure diagram of an anomaly analysis system 100 of a feed production control system for implementing the method provided in the embodiments of the present application. As Figure 2 shown, the anomaly analysis system 100 of the feed production control system may include at least one processor 102 (the processor 102 may include, but is not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, a transmission device 106 for communication functions, and a controller 108. Those of ordinary skill in the art can understand that Figure 2 the structure shown is only schematic and does not limit the structure of the anomaly analysis system 100 of the feed production control system. For example, the anomaly analysis system 100 of the feed production control system may also include more or fewer components than Figure 2 shown, or have a different configuration from Figure 2 shown.

[0114] The memory 104 can be used to store software programs and modules of application software, such as the program instructions corresponding to the above method embodiments in the embodiments of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, to implement the above abnormal analysis method of a feed production control system. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some instances, the memory 104 may further include a memory remotely disposed relative to the processor 102, and these remote memories can be connected to the abnormal analysis system 100 of the feed production control system through a network. Examples of the above network include but are not limited to the Internet, enterprise intranet, local area network, mobile communication network, and combinations thereof.

[0115] The transmission device 106 is used to obtain or send data via a network. Specific examples of the above network may include a wireless network provided by a communication provider of the abnormal analysis system 100 of the feed production control system. In one instance, the transmission device 106 includes a network adapter, which can be connected to other network devices through a base station and thus can communicate with the Internet. In one instance, the transmission device 106 can be a radio frequency module, which is used to communicate with the Internet wirelessly.

[0116] It should be noted that: the above sequence of the embodiments of the present application is only for description and does not represent the superiority or inferiority of the embodiments. And the above specific embodiments of the present application have been described. Other embodiments are within the scope of the appended claims. In some cases, the abnormalities or steps recorded in the claims can be executed in a different order from that in the embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0117] The various embodiments in the embodiments of the present application are all described in a progressive manner. For the parts that are consistent and similar among the various embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the above different embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and reference can be made to the partial description of the method embodiments for the relevant parts.

[0118] Those of ordinary skill in the art can understand that all or part of the steps to implement the above embodiments can be completed by hardware, or can be completed by a program instructing relevant hardware. The above program can be stored in a computer-readable storage medium, and the above-mentioned storage medium can be a read-only memory, a magnetic disk, an optical disk, etc.

Claims

1. A method for abnormal analysis of a feed production control system, characterized in that: The method comprises: Obtaining system control parameter path data corresponding to the feed production control system, and loading the system control parameter path data into a target deep learning network, wherein the target deep learning network includes a first graph generation module, a second graph generation module, and a fully connected mapping module; Using the first graph generation module to perform graph self-encoding representation, generating a graph self-encoding representation vector sequence corresponding to the feed production control system; Using the second graph generation module to perform multiple rounds of feature focusing weight allocation on the graph self-encoding representation vector sequence according to the feature focusing unit to generate a first feature focusing vector sequence, wherein the second graph generation module is obtained by cascading an encoder and the feature focusing unit; Blending the graph autoencoder representation vector sequence with the first feature focus vector sequence to generate a first blended vector sequence; The first blending vector sequence is loaded into the fully connected mapping module for abnormal classification, so as to generate abnormal classification data of the system control parameter path data.

2. The abnormality analysis method of the feed production control system according to claim 1, characterized in that: The target deep learning network further includes a recurrent neural module located between the second graph generation module and the fully connected mapping module. After the step of blending the graph autoencoder representation vector sequence with the first feature focus vector sequence to generate a first blended vector sequence, the method further includes: Using the recurrent neural module, according to the flow order information of the plurality of first blending vectors in the first blending vector sequence, feature focusing weights are assigned to the plurality of first blending vectors to generate a second feature focusing vector sequence; Integrate the first blending vector sequence and the second feature focusing vector sequence to generate a first integrated vector sequence; Then, the step of loading the first blending vector sequence into the fully connected mapping module for abnormal classification to generate abnormal classification data of the system control parameter path data includes: The first integrated vector sequence is loaded into the fully connected mapping module for abnormal classification, so as to generate abnormal classification data of the system control parameter path data.

3. The abnormality analysis method of the feed production control system according to claim 2, characterized in that: The method of using the recurrent neural module to allocate feature focusing weights to the plurality of first blending vectors in the first blending vector sequence according to the flow order information of the plurality of first blending vectors to generate a second feature focusing vector sequence includes: Using the recurrent neural module, according to the flow order information of each of the first blending vectors in the first blending vector sequence, determining the search vector corresponding to the last flow node and at least one memory vector corresponding to other flow nodes from the first blending vector sequence; Performing feature distance calculations on the search vector and the memory vector of each flow node in turn, generating an attention weight value corresponding to the memory vector of each flow node, wherein the attention weight value reflects the correlation between the memory vector of each flow node and the search vector; The first blending vectors in the corresponding memory vectors are updated according to the attention weight value of each memory vector to generate a second feature focus vector sequence.

4. The abnormality analysis method of the feed production control system according to claim 3, characterized in that: The updating of each of the first blending vectors in the corresponding memory vector according to the attention weight value of each of the memory vectors to generate a second feature focus vector sequence includes: updating each of the first blending vectors in the corresponding memory vector according to the attention weight value of each of the memory vectors to generate an updated first blending vector; For each vector channel of the first blending vector sequence, first blending vectors of different flow nodes corresponding to the vector channel are merged to generate a target blending vector corresponding to the vector channel; A second feature focus vector sequence is generated based on the multiple target blending vectors corresponding to the multiple vector channels.

5. The abnormality analysis method of the feed production control system according to claim 3, characterized in that: The step of using the recurrent neural module to determine, from the first blending vector sequence, a search vector corresponding to the last flow node and at least one memory vector corresponding to other flow nodes according to the flow order information of each of the first blending vectors in the first blending vector sequence, comprises: Receiving a first blended vector sequence generated by blending the first graph generation module and the second graph generation module, and flow sequence information of each first blended vector in the first blended vector sequence, wherein the flow sequence information is represented by a data structure of a timestamp or a sequence index; Performing a unique node identification on each first blending vector in the first blending vector sequence, and outputting a first blending vector sequence carrying the unique node identification and corresponding flow order information; Based on the extracted flow sequence information, construct a flow path of the first blending vector sequence, wherein the flow path is represented as a directed graph or an undirected graph, wherein nodes represent first blending vectors and edges represent flow relationships between vectors; Identify the last flow node in the constructed flow path, the last flow node being the end point of the flow path or the last node processed, and output the flow path carrying the last node identifier and the search vector corresponding to the last flow node; After the last flow node is identified, the flow path is traced back to find all predecessor nodes directly pointing to the last flow node, the predecessor node is a node in the flow path that is processed before the last flow node, and for each predecessor node, a memory vector set is initialized to store feature information of the predecessor node and its subsequent nodes, the memory vector is the first blending vector corresponding to the predecessor node, or a result obtained by transforming the first blending vector; For each predecessor node in the flow path, traverse its successor nodes in reverse order of the flow order, and update the memory vector set of the predecessor node, wherein the updating process includes adding feature information of the successor node to the memory vector of the predecessor node, or transforming the memory vector of the predecessor node to fuse the information of the successor node, thereby generating an updated memory vector set, and a transfer relationship between the updated memory vector sets based on the flow order; Based on the correlation between the memory vector and the retrieval vector corresponding to the last flow node and the redundancy between the memory vectors, the updated memory vector set is optimized to generate an optimized memory vector set; An association relationship is established between the optimized memory vector set and the search vector corresponding to the last flow node, wherein the association relationship is established by calculating the similarity or distance measurement between the memory vector and the search vector.

6. The abnormality analysis method of the feed production control system according to claim 1, characterized in that: The step of using the second graph generation module to perform multiple rounds of feature focusing weight allocation on the graph self-encoding representation vector sequence according to the feature focusing unit to generate a first feature focusing vector sequence includes: Loading the graph self-encoding representation vector sequence into the first encoder of the second graph generation module for compression processing to generate a compressed graph self-encoding representation vector sequence; Loading the graph self-encoding representation vector sequence into a feature focusing unit, generating a feature focusing coefficient reflecting the influence weight of each second blending vector in the graph self-encoding representation vector sequence; Optimizing each of the second blending vectors in the graph autoencoder representation vector sequence according to the feature focusing coefficient to generate an initial feature focusing vector sequence; The second encoder based on the second image generation module updates the vector channel of the initial feature focus vector sequence to generate a first feature focus vector sequence.

7. The abnormality analysis method of the feed production control system according to claim 6, characterized in that: The step of optimizing each of the second blending vectors in the graph autoencoding representation vector sequence according to the feature focusing coefficient to generate an initial feature focusing vector sequence comprises: receiving a feature focusing coefficient sequence generated by a feature focusing unit, wherein each feature focusing coefficient in the feature focusing coefficient sequence corresponds one-to-one to a second blending vector in the graph self-encoding representation vector sequence, indicating the importance of the second blending vector in the overall feature representation; Initialize a weight vector with the same dimension as the graph autoencoder representation vector sequence, wherein the initial value of each weight vector is set to the corresponding feature focusing coefficient, which is used as the basis for initial weight adjustment, so that the initial weight of each second fusion vector is proportional to its feature focusing coefficient, thereby reflecting the different importance of each second fusion vector in the feature representation; The graph autoencoder representation vector sequence carrying the initial weight is weight-normalized, and the normalized weight is fine-tuned according to the preset sensitivity parameter, and then the weighted summation method is used to perform feature fusion on the graph autoencoder representation vector sequence after weight fine-tuning, wherein, for each dimension, the weighted sum of all second fusion vectors on the dimension is calculated to obtain the fused feature, and the fused feature is preliminarily optimized by smoothing filtering to obtain a preliminarily optimized fused feature vector sequence; The importance of each fused feature vector in the fused feature vector sequence is evaluated by using information gain to obtain a new feature importance score, and the fused feature vector sequence after the preliminary optimization is adjusted for a second weight according to the new feature importance score, and the neighborhood information is used to enhance the features of the fused feature vector sequence after the second weight adjustment. Specifically, for each fused feature vector, the neighborhood features in the self-encoding representation of the fused feature vector image are searched, and the weighted sum of the neighborhood features is calculated as the enhanced value of the fused feature vector, and the enhanced value is fused with the original feature value of the fused feature vector to obtain an enhanced feature vector sequence; The enhanced feature vector sequence is subjected to feature dimension reduction processing through linear discriminant analysis to generate an initial feature focus vector sequence.

8. The abnormality analysis method of the feed production control system according to claim 1, characterized in that: The step of using the first graph generation module to perform graph self-encoding representation to generate a graph self-encoding representation vector sequence corresponding to the feed production control system includes: Receive system control parameter path data corresponding to the feed production control system, where the system control parameter path data includes time series values ​​of various control parameters when the feed production control system is running and correlation relationships between them; Performing data cleaning, data normalization and data smoothing on the system control parameter path data to generate preprocessed system control parameter path data, and constructing a graph structure representing the operating state of the feed production control system based on the preprocessed system control parameter path data, wherein nodes in the graph structure represent control parameters or parameter combinations, and edges represent associations or dependencies between parameters; Extracting graph structure feature information of the graph structure using a graph traversal algorithm or a graph embedding algorithm, wherein the feature information includes the degree of a node, neighbor information of a node, and path length between nodes; Converting the extracted graph structure feature information into an initial vector representation, that is, assigning a feature vector of a fixed dimension to each node of the graph structure; Inputting the initial vector representation into the encoder part of the first graph generation module, performing layer-by-layer compression and feature extraction through a multi-layer neural network to generate a low-dimensional latent space representation, and the latent space representation is used as an intermediate representation of the first graph generation module; The latent space representation is input into the decoder part of the first graph generation module, and the latent space representation is decoded and reconstructed layer by layer through a multi-layer neural network to generate a reconstructed graph structure representation as a graph autoencoder representation vector sequence corresponding to the feed production control system.

9. The abnormality analysis method of the feed production control system according to claim 1, characterized in that: The training steps of the target deep learning network include: Obtaining a first deep learning network trained in a first processing scenario; Determining, from the first deep learning network, target topological dimensions corresponding to each set topological dimension in a reference set deep learning network of a second processing scenario; Obtaining neuron parameter information of each target topological dimension, and converting the neuron parameter information into the set topological dimension corresponding to the reference set deep learning network, to generate an initialized deep learning network after updating the neuron parameter information; Acquire a sample system control parameter path data sequence corresponding to the second processing scenario, and perform network parameter learning on the initialized deep learning network based on the sample system control parameter path data sequence to generate a target deep learning network.

10. An abnormality analysis system for a feed production control system, characterized in that: The abnormality analysis system of the feed production control system includes a processor and a readable storage medium, wherein the readable storage medium stores a program, and when the program is executed by the processor, the abnormality analysis method of the feed production control system according to any one of claims 1 to 9 is implemented.

Images