Warehousing asset intelligent management and control optimization method and system based on machine learning

Abstract

The invention discloses a warehouse asset intelligent management and control optimization method and system based on machine learning. According to the method, asset item image data is acquired through a visual device, an object overall area and an image code area are identified by using a CNN model, overall and local features are extracted, graphic code identification is performed, and first identification information is obtained; meanwhile, a classification model based on a decision tree is constructed, historical asset article features are used for training, multi-level classification nodes are set, overall regional features are classified, and second identification information is generated. Asset authentication is carried out through the first identification information and the second identification information, and an identification rate and an identification efficiency value are calculated. And if the recognition rate and the efficiency value are lower than a preset range, dynamically adjusting classification nodes, removing part of nodes, resetting model parameters, and training to obtain a second classification model applied to a next storage cycle. The method improves the accuracy and efficiency of asset identification, improves the dynamic adjustment capability of a storage platform, and further improves the adaptability of storage management.

Description

Technical Field

[0001] The present invention relates to the field of warehouse management, and more specifically, to an intelligent control and optimization method and system for warehouse assets based on machine learning. Background Art

[0002] With the rapid development of logistics and supply chain, as one of the core links, the efficiency and accuracy of warehouse management directly affect the operating costs and service quality of enterprises. However, in warehouse management, traditional technologies often lack dynamic adjustment of material identification models, etc., and it is difficult to achieve dynamic control of management and identification modes, making it difficult to meet the requirements of warehouse inbound and outbound management in multiple scenarios. The adaptability of the management system is poor, and manual intervention is often required for inspection in the later stage of management, greatly increasing the input of human and material resources, increasing management costs, and resulting in low management efficiency. Therefore, there is an urgent need for an efficient intelligent control and optimization method for warehouse assets at present. Summary of the Invention

[0003] The present invention overcomes the defects of the prior art and provides an intelligent control and optimization method and system for warehouse assets based on machine learning.

[0004] The first aspect of the present invention provides an intelligent control and optimization method for warehouse assets based on machine learning, including: During a warehouse operation cycle, image data of asset items is obtained through a vision device, and the image data is preprocessed; Through a CNN recognition model, object overall region recognition and image code region recognition are performed on the image data, overall region features and local region features are extracted, and graphic code recognition and asset authentication are performed on the local region features to obtain first recognition information; A classification model based on a decision tree is constructed, historical asset item features are imported into the classification model as training data, multiple classification nodes are set, by calculating the classification effects of different nodes, multi-level classification nodes are set, and the classification nodes include first-level classification nodes, second-level classification nodes and third-level classification nodes, and a trained classification model is formed; The overall region features are imported into the classification model for appearance feature classification and recognition of asset items, second recognition information is generated according to the classification results, the first recognition information and the second recognition information are used for asset item authentication, the recognition rate is calculated based on the number of successfully authenticated assets, and during the recognition process, the data processing times of the recognition model and the classification model are counted, and a recognition efficiency value is calculated; During a warehousing operation cycle, if both the recognition rate and the recognition efficiency value are lower than the preset range, the classification nodes in the classification model are dynamically adjusted. A preset number of classification nodes are randomly selected and removed, and the classification model parameters are reset based on the existing node state pattern after removal. The accuracy of the classification model after removing the corresponding nodes is tested, and a second classification model is obtained. Using the second recognition information as the training target, the second classification model is trained and the nodes are adjusted; Apply the second classification model to the next warehousing operation cycle.

[0005] In this solution, during a warehousing operation cycle, image data of asset items is acquired through a vision device, and the image data is preprocessed, specifically as follows: Set the warehousing operation cycle; During a warehousing operation cycle, multiple sampling points are set in the warehousing area, and each sampling point includes at least one vision device; Based on the vision device, image data of asset items entering and leaving the warehousing area is acquired; Perform image denoising and enhancement preprocessing on the image data.

[0006] In this solution, through the CNN recognition model, the overall object area and the image code area of the image data are recognized, and the overall area feature and the local area feature are extracted. The graphic code of the local area feature is recognized and asset authentication is performed to obtain the first recognition information, specifically as follows: Construct a recognition model based on CNN; Through the recognition model, the overall object area and the image code area of the image data are recognized, and the overall area feature and the local area feature are extracted respectively; In the overall area recognition, specifically through the Roberts operator, edge contour detection and contour feature extraction are performed on the image data to obtain the overall area feature; Perform graphic code recognition on the local area feature, and combine with the asset database for asset authentication to generate the first recognition information.

[0007] In this solution, a classification model based on a decision tree is constructed. The historical asset item features are imported into the classification model as training data, multiple classification nodes are set, and through calculating the classification effects of different nodes, multi-level classification nodes are set. The classification nodes include first-level classification nodes, second-level classification nodes and third-level classification nodes, and a trained classification model is formed, specifically as follows: Construct a classification model based on a decision tree; Import the historical asset item features into the classification model as training data, classify the historical asset item features using a clustering model, and obtain multiple classification feature sets; For each set of classification features, multiple classification nodes are constructed, the information gain of each classification node is calculated, and the classification effect of each classification node is evaluated through the information gain; Based on the heuristic algorithm and multiple classification nodes, the root node of the classification model is determined and internal nodes and leaf nodes are generated; For the classification nodes in the classification model, hierarchical classification is performed based on the distance from the classification node to the root node, and multiple levels of classification nodes are set. The classification nodes include first-level classification nodes, second-level classification nodes, and third-level classification nodes; In the classification model, a preset amount of features are randomly selected from the historical asset item features for classification testing, the classification accuracy is evaluated based on the classification results, and the branch points of the classification model are adjusted to optimize the classification model.

[0008] In this solution, the overall area features are imported into the classification model for appearance feature classification and recognition of asset items, the second recognition information is generated according to the classification results, the first recognition information and the second recognition information are used for asset item authentication, and the recognition rate is calculated based on the number of successfully authenticated assets. During the recognition process, the data processing times of the recognition model and the classification model are statistically counted, and the recognition efficiency value is calculated. Specifically: The overall area features are imported into the classification model for appearance feature classification and recognition of asset items to obtain the overall area classification results; According to the overall area classification results, combined with the asset database, item classification confirmation is performed to generate the second recognition information; The first recognition information and the second recognition information are used for dual authentication of asset items, and based on a warehousing operation cycle, the number of successfully authenticated assets and the corresponding recognition rate are calculated; In the recognition model and the classification model, the corresponding data processing times T1 and T2 during the process of obtaining the first recognition information and the second recognition information are statistically counted; T1 and T2 are weighted and averaged to obtain the time mean value, and the recognition efficiency value is evaluated based on the time mean value.

[0009] In this solution, within a warehousing operation cycle, if both the recognition rate and the recognition efficiency value are lower than the preset range, the classification nodes in the classification model are dynamically adjusted. A preset number of classification nodes are randomly selected and removed, and the classification model parameters are reset based on the existing node state mode after removal. The accuracy of the classification model after removing the corresponding nodes is tested, and a second classification model is obtained. Using the second recognition information as the training target, the second classification model is trained and the nodes are adjusted. Specifically: During a warehousing operation cycle, if both the recognition rate and the recognition efficiency value are lower than the preset range, a preset number of classification nodes are randomly selected and marked as selected nodes. Based on the levels of the selected nodes, they are deleted in sequence. For each deleted selected node, the classification model status and parameters are reset based on the existing node status mode to obtain an optimized classification model. Use the image data that has been successfully recognized during a warehousing operation cycle as test data. Import the test data into the optimized classification model to calculate the classification accuracy. Mark the optimized classification model with the highest accuracy as the second classification model. Use the second recognition information as the training target to train and adjust the nodes of the second classification model until the accuracy of the second classification model reaches the expectation.

[0010] In this solution, applying the second classification model to the next warehousing operation cycle further includes: storing the model parameters and node status information in the second classification model in a model database.

[0011] In this solution, the vision device is connected to the platform terminal through the Internet of Things.

[0012] The second aspect of the present invention also provides an intelligent control and optimization system for warehousing assets based on machine learning. The system includes: a memory and a processor. The memory includes an intelligent control and optimization program for warehousing assets based on machine learning. When the intelligent control and optimization program for warehousing assets based on machine learning is executed by the processor, the following steps are implemented: During a warehousing operation cycle, use a vision device to obtain image data of asset items and preprocess the image data. Use a CNN recognition model to identify the overall area of the object and the image code area of the image data, extract the overall area features and local area features, perform graphic code recognition and asset authentication on the local area features to obtain the first recognition information. Construct a classification model based on a decision tree. Import the historical asset item features as training data into the classification model. Set multiple classification nodes. By calculating the classification effects of different nodes, set multi-level classification nodes, including primary classification nodes, secondary classification nodes, and tertiary classification nodes, and form a trained classification model. Import the overall area features into the classification model to classify and identify the appearance features of asset items. Generate the second recognition information according to the classification results. Authenticate the asset items with the first recognition information and the second recognition information. Calculate the recognition rate based on the number of successfully authenticated assets. During the recognition process, count the data processing times of the recognition model and the classification model and calculate the recognition efficiency value. During a warehousing operation cycle, if both the recognition rate and the recognition efficiency value are lower than the preset range, the classification nodes in the classification model are dynamically adjusted. A preset number of classification nodes are randomly selected for removal, and the classification model parameters are reset based on the existing node state pattern after removal. The accuracy of the classification model after removing the corresponding nodes is tested, and a second classification model is obtained. Using the second recognition information as the training target, the second classification model is trained and the nodes are adjusted; Apply the second classification model to the next warehousing operation cycle.

[0013] The third aspect of the present invention also provides a computer-readable storage medium, which includes an optimization program for intelligent control of warehousing assets based on machine learning. When the optimization program for intelligent control of warehousing assets based on machine learning is executed by a processor, the steps of the optimization method for intelligent control of warehousing assets based on machine learning as described in any one of the above are implemented.

[0014] The present invention discloses an optimization method and system for intelligent control of warehousing assets based on machine learning. Asset item image data is obtained through a vision device, and a CNN model is used to identify the overall area and the image code area of the object, extract the overall and local features, and perform graphic code recognition to obtain the first recognition information. At the same time, a classification model based on a decision tree is constructed, trained using historical asset item features, multi-level classification nodes are set, the overall area features are classified, and the second recognition information is generated. Asset authentication is performed through the first and second recognition information, and the recognition rate and the recognition efficiency value are calculated. If the recognition rate and the efficiency value are lower than the preset range, the classification nodes are dynamically adjusted, some nodes are removed and the model parameters are reset, and a second classification model is trained and applied to the next warehousing cycle. This method improves the accuracy and efficiency of asset recognition, enhances the dynamic adjustment ability of the warehousing platform, and further improves the adaptability of warehousing management. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Shows a flowchart of an optimization method for intelligent control of warehousing assets based on machine learning according to the present invention; Figure 2 Shows a block diagram of an optimization system for intelligent control of warehousing assets based on machine learning according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] In order to more clearly understand the above objects, features and advantages of the present invention, the present invention will be further described in detail below with reference to the drawings and specific embodiments. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

[0017] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited by the specific embodiments disclosed below.

[0018] Figure 1 The flowchart of an intelligent control and optimization method for warehouse assets based on machine learning according to the present invention is shown.

[0019] As Figure 1 shown, in a first aspect of the present invention, there is provided an intelligent control and optimization method for warehouse assets based on machine learning, including: S102, within a warehouse operation cycle, through a visual device, obtain image data of asset items and preprocess the image data; S104, through a CNN recognition model, perform object overall area recognition and image code area recognition on the image data, extract overall area features and local area features, perform graphic code recognition and asset authentication on the local area features, and obtain first recognition information; S106, construct a classification model based on a decision tree, import historical asset item features as training data into the classification model, set multiple classification nodes, set multi-level classification nodes by calculating the classification effects of different nodes, where the classification nodes include first-level classification nodes, second-level classification nodes, and third-level classification nodes, and form a trained classification model; S108, import the overall area features into the classification model for appearance feature classification and recognition of asset items, generate second recognition information according to the classification results, perform asset item authentication on the first recognition information and the second recognition information, calculate the recognition rate based on the number of successfully authenticated assets, during the recognition process, count the data processing times of the recognition model and the classification model, and calculate the recognition efficiency value; S110, within a warehouse operation cycle, if both the recognition rate and the recognition efficiency value are lower than a preset range, dynamically adjust the classification nodes in the classification model, randomly select a preset number of classification nodes for removal, and re-set the classification model parameters based on the existing node state mode after removal, test the accuracy of the classification model after removing the corresponding nodes, and obtain a second classification model. Using the second recognition information as the training target, train and adjust the nodes of the second classification model; S112, apply the second classification model to the next warehouse operation cycle.

[0020] It should be noted that both the recognition model and the classification model are used as machine learning models for multi-dimensional data recognition.

[0021] According to an embodiment of the present invention, within one warehousing operation cycle, image data of asset items is acquired through a vision device, and preprocessing is performed on the image data, specifically as follows: Set the warehousing operation cycle; Within one warehousing operation cycle, set multiple sampling points in the warehousing area, and each sampling point includes at least one vision device; Based on the vision device, acquire image data of asset items entering and leaving the warehousing area; Perform image denoising and enhancement preprocessing on the image data.

[0022] It should be noted that the warehousing operation cycle is a preset time period.

[0023] According to an embodiment of the present invention, through the CNN recognition model, object overall area recognition and image code area recognition are performed on the image data, and overall area features and local area features are extracted. Graphic code recognition and asset authentication are performed on the local area features to obtain the first recognition information, specifically as follows: Construct a recognition model based on CNN; Through the recognition model, object overall area recognition and image code area recognition are performed on the image data, and overall area features and local area features are respectively extracted; In the overall area recognition, specifically through the Roberts operator, edge contour detection and contour feature extraction are performed on the image data to obtain the overall area features; Perform graphic code recognition on the local area features, and combine with the asset database for asset authentication to generate the first recognition information.

[0024] It should be noted that the image data includes multi-angle image data of asset items, which is used for item recognition, authentication, inbound and outbound record keeping, and management. Image code area recognition generally performs feature extraction and graphic code recognition based on the fixed position of asset items. Graphic codes include two-dimensional codes, barcodes, etc.

[0025] According to an embodiment of the present invention, construct a classification model based on a decision tree, import historical asset item features as training data into the classification model, set multiple classification nodes, set multi-level classification nodes by calculating the classification effects of different nodes, and the classification nodes include first-level classification nodes, second-level classification nodes, and third-level classification nodes, and form a trained classification model, specifically as follows: Construct a classification model based on a decision tree; Import historical asset item features as training data into the classification model, classify the historical asset item features using a clustering model, and obtain multiple classification feature sets; For each classification feature set, multiple classification nodes are constructed, the information gain of each classification node is calculated, and the classification effect of each classification node is evaluated through the information gain; Based on the heuristic algorithm and multiple classification nodes, the root node of the classification model is determined, and internal nodes and leaf nodes are generated; For the classification nodes in the classification model, hierarchical classification is performed based on the distance from the classification node to the root node, and multiple levels of classification nodes are set. The classification nodes include first-level classification nodes, second-level classification nodes, and third-level classification nodes; In the classification model, a preset amount of features is randomly selected from the historical asset item features for classification testing, the classification accuracy is evaluated based on the classification results, and the branch points of the classification model are adjusted to optimize the classification model.

[0026] It should be noted that the clustering model can adopt clustering models such as DBSCAN, which is used to divide different types of feature sets and determine the classification conditions for different types (i.e., the classification nodes on the decision tree). The first-level classification node is the closest to the root node, generally being the child node of the root node. The second-level classification node is the child node of the first-level classification node and is farther from the root node compared to the first-level node, and so on. Based on the classification requirements of the classification model, multiple levels of nodes can be set. Here, three levels of classification nodes are set for illustration.

[0027] According to the embodiments of the present invention, the overall area features are imported into the classification model for the appearance feature classification and recognition of asset items, the second recognition information is generated according to the classification results, the first recognition information and the second recognition information are used for asset item authentication, and the recognition rate is calculated based on the number of successfully authenticated assets. During the recognition process, the data processing times of the recognition model and the classification model are statistically counted, and the recognition efficiency value is calculated, specifically as follows: The overall area features are imported into the classification model for the appearance feature classification and recognition of asset items to obtain the overall area classification results; According to the overall area classification results, combined with the asset database, item classification confirmation is performed to generate the second recognition information; The first recognition information and the second recognition information are used for the dual authentication of asset items, and based on a warehousing operation cycle, the number of successfully authenticated assets and the corresponding recognition rate are calculated; In the recognition model and the classification model, the corresponding data processing times T1 and T2 during the process of obtaining the first recognition information and the second recognition information are statistically counted; T1 and T2 are weighted and averaged to obtain the time mean value, and the recognition efficiency value is evaluated based on the time mean value.

[0028] It should be noted that the larger the number of successfully authenticated assets, the higher the recognition rate. The process of obtaining the first recognition information and the second recognition information is the recognition process. In the weighted average, the weights of T1 and T2 are Q1 and Q2 respectively, and Q1 < Q2. The recognition efficiency value is inversely proportional to the time average.

[0029] According to an embodiment of the present invention, within a warehousing operation cycle, if both the recognition rate and the recognition efficiency value are lower than a preset range, the classification nodes in the classification model are dynamically adjusted. A preset number of classification nodes are randomly selected for removal, and the classification model parameters are reset based on the existing node state mode after removal. The accuracy rate of the classification model after removing the corresponding nodes is tested, and a second classification model is obtained. Using the second recognition information as the training target, the second classification model is trained and the nodes are adjusted, specifically as follows: Within a warehousing operation cycle, if both the recognition rate and the recognition efficiency value are lower than a preset range, a preset number of classification nodes are randomly selected and marked as selected nodes. Based on the levels of the selected nodes, they are deleted in sequence. Each time a selected node is deleted, the classification model state and parameters are reset based on the existing node state mode to obtain an optimized classification model; The image data that has been successfully recognized within a warehousing operation cycle is used as test data. The test data is imported into the optimized classification model to calculate the classification accuracy rate, and the optimized classification model with the highest accuracy rate is marked as the second classification model; Using the second recognition information as the training target, the second classification model is trained and the nodes are adjusted until the accuracy rate of the second classification model reaches the expectation.

[0030] It should be noted that the preset range includes a preset recognition rate and a preset recognition efficiency value. In the sequential deletion based on the levels of the selected nodes, the nodes with larger levels are preferentially deleted. The present invention adopts the form of gradually deleting nodes. Each time a selected node is deleted, the classification model state and parameters are reset based on the existing node state mode to obtain a corresponding optimized classification model. Based on multiple deletion times, a corresponding number of optimized classification models are obtained, and the optimal second classification model is selected from them. In this process, the dynamic adjustment of the classification model parameters and node states is realized to meet the classification task requirements in complex scenarios and the user's management and query requirements for asset data.

[0031] It is worth mentioning here that in warehouse management, traditional technologies often lack dynamic adjustments to material identification models, etc., and it is difficult to achieve dynamic regulation of management efficiency. It is difficult to adapt to the needs of warehouse entry and exit management in multiple scenarios, and the management system has poor adaptability. For example, the process of asset entry and exit in warehouse is too fast, there are many assets, or many types of assets and materials need to be identified and entered and exited in a short period of time. This situation can easily lead to abnormalities or errors in identification, and lead to the need for manual intervention and inspection in the later stage, which greatly increases the investment of manpower and material resources, increases management costs, and at the same time, low management efficiency. The present invention can effectively solve the above problems.

[0032] According to an embodiment of the present invention, applying the second classification model to the next warehouse operation cycle further includes: storing model parameters and node status information in the second classification model in a model database.

[0033] According to an embodiment of the present invention, the visual device is connected to the platform terminal via the Internet of Things.

[0034] It should be noted that the visual device is connected to the platform terminal through the Internet of Things to achieve data transmission, data analysis, and data storage functions.

[0035] According to an embodiment of the present invention, in a warehousing operation cycle, if the recognition rate and the recognition efficiency value are both lower than a preset range, the classification nodes in the classification model are dynamically adjusted, and further comprising: Real-time collection of image data that failed to be identified during a warehouse operation cycle and marking them as image data to be tested; The image data to be tested is used as the test classification data. In the classification model, the test classification data is classified and tested for each classification node, and the information gain of each node is recalculated based on the data state before and after classification, and the record is marked as P; Set the classification node level identifier to D, calculate the classification weight of each node based on the information gain P and D, and set the node priority table according to the classification weight; The removal number N is set, and based on the node priority table, N classification nodes with lower priorities are removed, and the classification model parameters and states are reset based on the existing node state mode after the removal to obtain a second classification model.

[0036] It should be noted that during the process of regulating the classification model, based on a warehousing platform with low computing power, classification nodes can be randomly selected for removal and the model can be optimized. For a warehousing platform with high computing power, the classification weights can be set based on the information gain of the secondary analysis of the nodes, combined with the original level division. The node priority is set based on the classification weights. The greater the classification weight, the greater the node priority, and the corresponding node has a higher classification efficiency for the overall data. This process can effectively combine the actual recognition situation for dynamic setting of the model, and dynamically improve the data processing efficiency while ensuring the classification efficiency.

[0037] In the node priority table, the greater the classification weight W, the higher the priority.

[0038] The calculation formula for the classification weight W is as follows: W = K × P + D; Where K is a preset correction coefficient, P is the information gain, and D is the classification node level.

[0039] According to the embodiments of the present invention, it further includes: Dividing an M warehousing operation cycle into M time periods; For each time period, based on the first recognition information and the second recognition information, an asset item authentication process is carried out, and the corresponding recognition rate and recognition efficiency value are calculated in real time; Setting a certain current time period as the analysis object, calculating the weighted average of the recognition rate and the recognition efficiency value to obtain the recognition evaluation mean value, and performing a linear change analysis on the recognition evaluation mean value to obtain the linear change rate; Generating a two-dimensional vector based on the two-dimensional values of the recognition rate and the recognition efficiency value, and performing a difference evaluation on the two-dimensional vectors of the current time period and the previous time period. The difference evaluation is calculated based on the standard Euclidean distance to obtain the difference value; If the linear change rate is negative and the difference value is greater than the preset value, then mark the current time period as an abnormal time period; Performing real-time statistics on the anomalies for each time period, integrating and marking the image data corresponding to the abnormal time periods to obtain an abnormal image set; If the number of abnormal time periods is greater than the preset time length, then extract the features of the abnormal image set, and use a clustering model to cluster the extracted features and construct new classification nodes, and insert the new classification nodes as leaf nodes into the classification model to form a new classification model.

[0040] It should be noted that a negative linear change rate corresponds to a linear decrease in the data for the corresponding time period. It is worth mentioning here that in complex scenarios of asset item inbound and outbound, the analysis for identifying anomalies often relies on analyzing each piece of data one by one, and it is difficult to determine the time period when anomalies occur, resulting in low efficiency. Especially in large-scale identification and authentication and warehouse management analysis, traditional technologies are difficult to meet the corresponding management requirements. Based on this, the present invention comprehensively analyzes the recognition rate and recognition efficiency value of the model, analyzes the linear change situation and divides the abnormal time interval (i.e., the abnormal time period), further collects the abnormal data set in real time. When the abnormal interval is greater than the preset length, it represents that the abnormal situation meets the early warning situation. Based on the data set collected in real time, the classification model is dynamically adjusted, thereby reducing the impact of subsequent recognition tasks, improving the dynamic adjustment ability of the platform, and improving the efficiency of warehouse automation management.

[0041] Figure 2 Fig. shows a block diagram of an intelligent control and optimization system for warehouse assets based on machine learning according to the present invention.

[0042] In a second aspect of the present invention, there is also provided an intelligent control and optimization system 2 for warehouse assets based on machine learning. The system includes: a memory 21 and a processor 22. The memory 21 includes an intelligent control and optimization program for warehouse assets based on machine learning. When the intelligent control and optimization program for warehouse assets based on machine learning is executed by the processor 22, the following steps are implemented: During a warehouse operation cycle, image data of asset items is obtained through a vision device, and the image data is preprocessed. Through a CNN recognition model, the overall area of the object and the image code area of the image data are recognized, and the overall area feature and the local area feature are extracted. The graphic code of the local area feature is recognized and asset authentication is performed to obtain the first recognition information. A classification model based on a decision tree is constructed. The historical asset item features are used as training data and imported into the classification model. Multiple classification nodes are set. By calculating the classification effects of different nodes, multi-level classification nodes are set. The classification nodes include primary classification nodes, secondary classification nodes, and tertiary classification nodes, and a trained classification model is formed. The overall area feature is imported into the classification model for appearance feature classification and recognition of asset items. According to the classification result, the second recognition information is generated. The first recognition information and the second recognition information are used for asset item authentication. Based on the number of successfully authenticated assets, the recognition rate is calculated. During the recognition process, the data processing times of the recognition model and the classification model are statistically calculated, and the recognition efficiency value is calculated. During a warehousing operation cycle, if both the recognition rate and the recognition efficiency value are lower than the preset range, the classification nodes in the classification model are dynamically adjusted. A preset number of classification nodes are randomly selected and removed, and the classification model parameters are reset based on the existing node state pattern after removal. The accuracy of the classification model after removing the corresponding nodes is tested, and a second classification model is obtained. Using the second recognition information as the training target, the second classification model is trained and the nodes are adjusted; Apply the second classification model to the next warehousing operation cycle.

[0043] It should be noted that both the recognition model and the classification model are used as machine learning models for multi-dimensional data recognition.

[0044] According to an embodiment of the present invention, during a warehousing operation cycle, image data of asset items is acquired through a vision device, and the image data is preprocessed, specifically as follows: Set the warehousing operation cycle; During a warehousing operation cycle, multiple sampling points are set in the warehousing area, and each sampling point includes at least one vision device; Based on the vision device, acquire the image data of asset items entering and leaving the warehousing area; Perform image noise reduction and enhancement preprocessing on the image data.

[0045] It should be noted that the warehousing operation cycle is a preset time period.

[0046] According to an embodiment of the present invention, through the CNN recognition model, the overall object area and the image code area of the image data are recognized, the overall area feature and the local area feature are extracted, and the local area feature is used for graphic code recognition and asset authentication to obtain the first recognition information, specifically as follows: Construct a recognition model based on CNN; Through the recognition model, the overall object area and the image code area of the image data are recognized, and the overall area feature and the local area feature are extracted respectively; In the overall area recognition, specifically through the Roberts operator, edge contour detection and contour feature extraction are performed on the image data to obtain the overall area feature; Perform graphic code recognition on the local area feature, and combine with the asset database for asset authentication to generate the first recognition information.

[0047] It should be noted that the image data includes multi-angle image data of asset items for item recognition, authentication, inbound and outbound record keeping, and management. The image code area recognition generally performs feature extraction and graphic code recognition based on the fixed position of the asset item. The graphic code includes two-dimensional code, bar code, etc.

[0048] According to an embodiment of the present invention, to construct a classification model based on a decision tree, historical asset item features are imported into the classification model as training data, multiple classification nodes are set, and by calculating the classification effects of different nodes, multi-level classification nodes are set. The classification nodes include first-level classification nodes, second-level classification nodes, and third-level classification nodes, and a trained classification model is formed. Specifically: Construct a classification model based on a decision tree; Import historical asset item features into the classification model as training data, classify the historical asset item features using a clustering model, and obtain multiple classification feature sets; For each classification feature set, construct multiple classification nodes, calculate the information gain of each classification node, and evaluate the classification effect of each classification node through the information gain; Based on a heuristic algorithm and multiple classification nodes, determine the root node of the classification model and generate internal nodes and leaf nodes; For the classification nodes in the classification model, perform level division based on the distance from the classification node to the root node, set multi-level classification nodes, and the classification nodes include first-level classification nodes, second-level classification nodes, and third-level classification nodes; In the classification model, randomly select a preset amount of features from the historical asset item features for classification testing, evaluate the classification accuracy based on the classification results, and adjust the branch points of the classification model to optimize the classification model.

[0049] It should be noted that the clustering model can adopt clustering models such as DBSCAN, which is used to divide different types of feature sets and determine the classification conditions of different types (i.e., the classification nodes on the decision tree). The first-level classification node is the closest to the root node, generally being the child node of the root node. The second-level classification node is the child node of the first-level classification node and is farther from the root node compared to the first-level node, and so on. Based on the classification requirements of the classification model, multiple levels of nodes can be set, and here three levels of classification nodes are set as an example.

[0050] According to an embodiment of the present invention, to import the overall area features into the classification model for appearance feature classification and recognition of asset items, generate second recognition information based on the classification results, perform asset item authentication on the first recognition information and the second recognition information, calculate the recognition rate based on the number of successfully authenticated assets, and during the recognition process, count the data processing times of the recognition model and the classification model and calculate the recognition efficiency value. Specifically: Import the overall area features into the classification model for appearance feature classification and recognition of asset items to obtain the overall area classification result; According to the overall area classification result, combine with the asset database to confirm item classification and generate second recognition information; Perform dual authentication of asset items for the first identification information and the second identification information, and calculate the number of successfully authenticated assets and the corresponding recognition rate based on a warehousing operation cycle. In the identification model and the classification model, statistically obtain the corresponding data processing times T1 and T2 during the process of obtaining the first identification information and the second identification information. Perform a weighted average on T1 and T2 to obtain a time average value, and evaluate the recognition efficiency value based on the time average value.

[0051] It should be noted that the larger the number of successfully authenticated assets, the higher the recognition rate. The process of obtaining the first identification information and the second identification information is the recognition process. In the weighted average, the weights of T1 and T2 are Q1 and Q2 respectively, and Q1 < Q2. The recognition efficiency value is inversely proportional to the time average value.

[0052] According to an embodiment of the present invention, within a warehousing operation cycle, if both the recognition rate and the recognition efficiency value are lower than a preset range, then dynamically adjust the classification nodes in the classification model, randomly select a preset number of classification nodes for removal, and reset the classification model parameters based on the existing node state mode after removal, test the accuracy rate of the classification model after removing the corresponding nodes, and obtain a second classification model. Using the second identification information as the training target, train and adjust the nodes of the second classification model, specifically: Within a warehousing operation cycle, if both the recognition rate and the recognition efficiency value are lower than a preset range, then randomly select a preset number of classification nodes, mark them as selected nodes, and sequentially delete them based on the levels of the selected nodes. For each deleted selected node, reset the classification model state and parameters based on the existing node state mode to obtain an optimized classification model. Use the image data that has been successfully recognized within a warehousing operation cycle as test data, import the test data into the optimized classification model for calculating the classification accuracy rate, and mark the optimized classification model with the highest accuracy rate as the second classification model. Using the second identification information as the training target, train and adjust the nodes of the second classification model until the accuracy rate of the second classification model reaches the expectation.

[0053] It should be noted that the preset range includes a preset recognition rate and a preset recognition efficiency value. In the sequential deletion based on the levels of the selected nodes, the nodes with larger levels are preferentially deleted. The present invention adopts the form of gradually deleting nodes. For each deleted selected node, reset the classification model state and parameters based on the existing node state mode to obtain a corresponding optimized classification model. Based on multiple deletion times, obtain corresponding multiple optimized classification models, and screen out the optimal second classification model from them. In this process, the dynamic adjustment of the parameters and node states of the classification model is realized to meet the classification task requirements in complex scenarios and the user's management and query requirements for asset data.

[0054] It is worth mentioning here that in warehouse management, traditional technologies often lack dynamic adjustment of material identification models, etc., making it difficult to achieve dynamic regulation of management efficiency and difficult to adapt to the warehouse inbound and outbound management requirements in multiple scenarios, resulting in poor adaptability of the management system. For example, when the inbound and outbound process of assets and items in the warehouse is too fast, there are many assets and items, or a relatively large number of different types of asset materials need to be identified and inbound and outbound in a short period of time, this situation is likely to lead to abnormal or incorrect identification, and requires manual intervention for inspection later, greatly increasing the investment in manpower and material resources, increasing management costs, and at the same time, the management efficiency is low. However, the present invention can effectively solve the above problems.

[0055] According to an embodiment of the present invention, applying the second classification model to the next warehouse operation cycle further includes: storing the model parameters and node state information in the second classification model in a model database.

[0056] According to an embodiment of the present invention, the vision device is connected to the platform terminal through the Internet of Things.

[0057] It should be noted that the vision device is connected to the platform terminal through the Internet of Things to achieve data transmission, data analysis, and data storage functions.

[0058] The third aspect of the present invention also provides a computer-readable storage medium, which includes an intelligent control and optimization program for warehouse assets based on machine learning. When the intelligent control and optimization program for warehouse assets based on machine learning is executed by a processor, the steps of the intelligent control and optimization method for warehouse assets based on machine learning as described in any one of the above are realized.

[0059] The present invention discloses an intelligent control and optimization method and system for warehouse assets based on machine learning. Image data of assets and items is obtained through a vision device, and a CNN model is used to identify the overall area and image code area of the object, extract overall and local features, and perform graphic code recognition to obtain first recognition information. At the same time, a classification model based on a decision tree is constructed, trained using historical asset item features, multiple classification nodes are set, the overall area features are classified, and second recognition information is generated. Asset authentication is performed through the first and second recognition information, and the recognition rate and recognition efficiency value are calculated. If the recognition rate and efficiency value are lower than the preset range, the classification nodes are dynamically adjusted, some nodes are removed and the model parameters are reset, and a second classification model is trained and applied to the next warehouse cycle. This method improves the accuracy and efficiency of asset identification, enhances the dynamic adjustment ability of the warehouse platform, and further improves the adaptability of warehouse management.

[0060] In several embodiments provided by this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the couplings, direct couplings, or communication connections between the various components shown or discussed can be through some interfaces. The indirect couplings or communication connections of devices or units can be electrical, mechanical, or other forms.

[0061] The units described above as separate components may or may not be physically separated. The components shown as units may or may not be physical units. They can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0062] In addition, each functional unit in the embodiments of the present invention can be all integrated in a processing unit, or each unit can be separately used as a unit, or two or more units can be integrated in one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of hardware plus software functional units.

[0063] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including the above method embodiments. The foregoing storage medium includes: removable storage devices, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disks, or optical disks and other various media that can store program codes.

[0064] Alternatively, if the above-mentioned integrated units of the present invention are implemented in the form of software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: removable storage devices, ROM, RAM, magnetic disks, or optical disks and other various media that can store program codes.

[0065] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

Claims

1. A method for optimizing intelligent management and control of warehouse assets based on machine learning, characterized in that: include: In a warehouse operation cycle, the image data of assets and items are obtained through visual devices and pre-processed; Through the CNN recognition model, the image data is subjected to overall object area recognition and image code area recognition, and overall area features and local area features are extracted. The local area features are subjected to graphic code recognition and asset authentication to obtain the first recognition information; Construct a classification model based on a decision tree, import the characteristics of historical assets into the classification model as training data, set multiple classification nodes, calculate the classification effects of different nodes, set multi-level classification nodes, and form a trained classification model; Import the overall regional features into the classification model to classify and identify the appearance features of the asset items, generate the second identification information according to the classification results, authenticate the asset items with the first identification information and the second identification information, calculate the recognition rate based on the number of successfully authenticated assets, and in the recognition process, count the data processing time of the recognition model and the classification model, and calculate the recognition efficiency value; In a warehouse operation cycle, if the recognition rate and recognition efficiency values ​​are both lower than the preset range, the classification nodes in the classification model are dynamically adjusted, a preset number of classification nodes are randomly selected for removal, and the classification model parameters are reset based on the existing node status mode after removal, and the accuracy of the classification model after the corresponding node is removed is tested, and a second classification model is obtained. The second recognition information is used as the training target, and the second classification model is trained and the nodes are adjusted; Apply the second classification model to the next warehouse operation cycle.

2. According to the method for optimizing intelligent management and control of warehouse assets based on machine learning according to claim 1, it is characterized in that: In a warehousing operation cycle, the image data of the asset items is obtained through the visual device, and the image data is pre-processed, specifically: Set the warehouse operation cycle; In a storage operation cycle, multiple sampling points are set in the storage area, and each sampling point includes at least one visual device; Based on visual devices, image data of assets entering and leaving the storage area is obtained; The image data is pre-processed by image noise reduction and enhancement.

3. According to the method for optimizing intelligent management and control of warehouse assets based on machine learning according to claim 1, it is characterized in that: The CNN recognition model is used to perform overall object area recognition and image code area recognition on the image data, and overall area features and local area features are extracted. Graphic code recognition and asset authentication are performed on the local area features to obtain the first recognition information, which is specifically: Build a CNN-based recognition model; Through the recognition model, the image data is recognized for the overall area of ​​the object and the image code area, and the overall area features and local area features are extracted respectively; In the overall region recognition, the edge contour detection and contour feature extraction of the image data are performed by Roberts operator to obtain the overall region features; Graphic code recognition is performed on local area features, and asset authentication is performed in combination with an asset database to generate first identification information.

4. According to the method of intelligent management and optimization of warehouse assets based on machine learning in claim 3, it is characterized in that: The classification model based on the decision tree is constructed, and the historical asset item features are imported into the classification model as training data, and multiple classification nodes are set. By calculating the classification effects of different nodes, multi-level classification nodes are set, and the classification nodes include primary classification nodes, secondary classification nodes and tertiary classification nodes, and the trained classification model is formed, specifically: Build a decision tree-based classification model; Import the historical asset item features as training data into the classification model, use the clustering model to classify the historical asset item features, and obtain multiple classification feature sets; For each classification feature set, multiple classification nodes are constructed, the information gain of each classification node is calculated, and the classification effect of each classification node is evaluated by the information gain; Based on the heuristic algorithm and multiple classification nodes, the root node of the classification model is determined and the internal nodes and leaf nodes are generated; For the classification nodes in the classification model, the classification nodes are divided into levels according to the distance from the classification nodes to the root nodes, and multi-level classification nodes are set. The classification nodes include first-level classification nodes, second-level classification nodes and third-level classification nodes; In the classification model, features of a preset data volume are randomly selected from the historical asset item features for classification testing, the classification accuracy is evaluated based on the classification results, and the branch points of the classification model are adjusted to optimize the classification model.

5. According to claim 4, a method for optimizing intelligent management and control of warehouse assets based on machine learning is characterized in that: The overall regional features are introduced into the classification model to classify and identify the appearance features of the asset items, and the second identification information is generated according to the classification result. The first identification information and the second identification information are used to authenticate the asset items, and the recognition rate is calculated based on the number of successfully authenticated assets. In the recognition process, the data processing time of the recognition model and the classification model is counted, and the recognition efficiency value is calculated, which is specifically: Import the overall regional features into the classification model to classify and identify the appearance features of assets and items, and obtain the overall regional classification results; According to the overall regional classification results, the item classification is confirmed in combination with the asset database to generate the second identification information; The first identification information and the second identification information are used to perform dual authentication of assets and items, and based on a warehouse operation cycle, the number of successfully authenticated assets and the corresponding recognition rate are calculated; In the recognition model and the classification model, the corresponding data processing times T1 and T2 in the process of obtaining the first recognition information and the second recognition information are counted; The weighted average of T1 and T2 is obtained to obtain the time mean, and the recognition efficiency value is evaluated based on the time mean.

6. The method for optimizing intelligent management and control of warehouse assets based on machine learning according to claim 5 is characterized in that: In the above-mentioned storage operation cycle, if the recognition rate and the recognition efficiency value are both lower than the preset range, the classification nodes in the classification model are dynamically adjusted, a preset number of classification nodes are randomly selected for removal, and the classification model parameters are reset based on the existing node state mode after removal, the accuracy of the classification model after the corresponding node is removed is tested, and a second classification model is obtained. The second recognition information is used as the training target, and the second classification model is trained and the nodes are adjusted, specifically: In a warehouse operation cycle, if the recognition rate and recognition efficiency values ​​are both lower than the preset range, a preset number of classification nodes are randomly selected and marked as selected nodes, and are deleted in sequence based on the levels of the selected nodes. Each time a selected node is deleted, the classification model state and parameters are reset based on the existing node state mode to obtain an optimized classification model. The image data that has been successfully identified in a warehouse operation cycle is used as test data, and the test data is imported into the optimized classification model to calculate the classification accuracy, and the optimized classification model with the highest accuracy is marked as the second classification model; The second identification information is used as a training target to train the second classification model and adjust its nodes until the accuracy of the second classification model reaches the expected level.

7. The method for optimizing intelligent management and control of warehouse assets based on machine learning according to claim 1 is characterized in that: The applying the second classification model to the next storage operation cycle also includes: storing the model parameters and node status information in the second classification model in the model database.

8. The method for optimizing intelligent management and control of warehouse assets based on machine learning according to claim 1 is characterized in that: The visual device is connected to the platform terminal via the Internet of Things.

9. A warehouse asset intelligent management and control optimization system based on machine learning, characterized in that: The system includes: a memory and a processor, wherein the memory includes a warehousing asset intelligent management and control optimization program based on machine learning, and when the warehousing asset intelligent management and control optimization program based on machine learning is executed by the processor, the following steps are implemented: In a warehouse operation cycle, the image data of assets and items are obtained through visual devices and pre-processed; Through the CNN recognition model, the image data is subjected to overall object area recognition and image code area recognition, and overall area features and local area features are extracted. The local area features are subjected to graphic code recognition and asset authentication to obtain the first recognition information; Construct a classification model based on a decision tree, import the characteristics of historical assets into the classification model as training data, set multiple classification nodes, calculate the classification effects of different nodes, set multi-level classification nodes, and form a trained classification model; Import the overall regional features into the classification model to classify and identify the appearance features of the asset items, generate the second identification information according to the classification results, authenticate the asset items with the first identification information and the second identification information, calculate the recognition rate based on the number of successfully authenticated assets, and in the recognition process, count the data processing time of the recognition model and the classification model, and calculate the recognition efficiency value; In a warehouse operation cycle, if the recognition rate and recognition efficiency values ​​are both lower than the preset range, the classification nodes in the classification model are dynamically adjusted, a preset number of classification nodes are randomly selected for removal, and the classification model parameters are reset based on the existing node status mode after removal, and the accuracy of the classification model after the corresponding node is removed is tested, and a second classification model is obtained. The second recognition information is used as the training target, and the second classification model is trained and the nodes are adjusted; Apply the second classification model to the next warehouse operation cycle.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a warehouse asset intelligent management and control optimization program based on machine learning. When the warehouse asset intelligent management and control optimization program based on machine learning is executed by a processor, the steps of the warehouse asset intelligent management and control optimization method based on machine learning as described in any one of claims 1 to 8 are implemented.

Images