A material distribution task generation method, device, terminal equipment and computer program product

Abstract

The present application provides a material distribution task generation method, device, terminal equipment and computer program product, which are applicable to the field of production line production and processing technology. The method includes: collecting the status data of the production line; inputting the status data of the production line into a first model, and using the multi-channel deep residual network model structure of the first model to predict the downtime waiting time of the cache storage location of the production line; inputting the status data of the production line into a trained second model, and combining the downtime waiting time to generate a material distribution task. The present application can improve the generation method of material distribution tasks within the production line, reduce the chaos of material preparation logistics and the accumulation of redundant materials, improve the production order of the production line, and improve production efficiency.

Description

Technical Field

[0001] The present application belongs to the field of production line production and processing technology, and in particular, relates to a material distribution task generation method, device, terminal equipment and computer program product. Background Art

[0002] In the manufacturing process of new energy batteries, the assembly of battery packs involves multiple processing stages. Each process stage has line-side cache locations of different capacities, the inventory consumption rate is uneven, and the distribution frequency varies greatly. In order to avoid waiting for incoming materials, the conventional method is to adopt a safety inventory strategy, but frequent production line cutting and pulling changes lead to chaotic material preparation logistics and redundant material accumulation, affecting production order. Therefore, the generation method of material distribution tasks within the production line of the prior art needs to be optimized. Summary of the invention

[0003] The embodiments of the present application provide a material distribution task generation method, device, terminal device and computer program product, which can solve the technical problem of low production efficiency of production lines in the prior art;

[0004] In a first aspect, an embodiment of the present application provides a method for generating a material distribution task, comprising:

[0005] Collect status data of production lines;

[0006] Inputting the status data of the production line into the first model, and using the multi-channel deep residual network model structure of the first model to predict the downtime waiting time of the buffer storage location of the production line;

[0007] Inputting the state data of the production line into the trained second model, and combining the downtime waiting time to generate a material distribution task;

[0008] The step of inputting the state data of the production line into the first model and using the multi-channel deep residual network model structure of the first model to predict the downtime waiting time of the buffer storage location of the production line includes:

[0009] Processing the state data of the production line according to a multi-channel coding representation method to obtain multi-channel input data;

[0010] Fitting the multi-channel input data using the multi-channel deep residual network model structure of the first model to predict the downtime waiting time;

[0011] The multi-channel deep residual network model structure using the first model is used to fit the multi-channel input data to predict the downtime waiting time, including:

[0012] The multi-channel input data is input into a multi-channel deep residual network model structure to learn the relationship between the status data of the production line and the downtime waiting time, and output a downtime waiting time prediction vector, wherein the length of the downtime waiting time prediction vector corresponds to the number of the cache storage locations, and each element of the downtime waiting time prediction vector represents the downtime waiting time prediction value of the corresponding target cache storage location.

[0013] In the embodiments of the present application, the method of generating material distribution tasks within the production line can be improved, the chaos in material preparation logistics and the accumulation of redundant materials can be reduced, the production order of the production line can be improved, and the production efficiency can be improved.

[0014] In some embodiments, the status data of the production line includes first production line data and second production line data, the first production line data at least includes level indicators of different types of storage locations, and the second production line data at least includes storage location consumption time and work-in-progress consumption time; wherein, the second production line data is obtained based on the first production line data.

[0015] In the embodiment of the present application, the first data of the production line and the second data of the production line can more comprehensively reflect the status of the production line, so that the material distribution task finally generated is more in line with the actual situation of the production line.

[0016] In some embodiments, the step of inputting the multi-channel input data into a multi-channel deep residual network model structure to learn the relationship between the state data of the production line and the downtime waiting time, and outputting a downtime waiting time prediction vector, comprises:

[0017] Performing channel convolution processing on the multi-channel input data to extract feature information of input data in each channel of the multi-channel input data, and obtaining first data containing each feature information;

[0018] Performing a short connection operation on the first data and the multi-channel input data through a short connection layer to obtain data subjected to the short connection operation;

[0019] The data after the short connection operation is input as the target input data into the stacked layer group, and each network layer in the stacked layer group operates on the target input data in turn to learn the relationship between the status data of the production line and the downtime waiting time, and output the downtime waiting time prediction vector.

[0020] In an embodiment of the present application, an independent convolution operation is first performed on the input data in each channel of the multi-channel input data, and then the model fitting ability is enhanced through a short connection layer, and each network layer in the stacked layer group operates on the data after the short connection operation in turn, so that the multi-channel deep residual network model structure of this embodiment can perform deep fitting and learning on the status data of the production line, and then realize accurate prediction of the downtime waiting time in a complex and changeable production line environment, and optimize material distribution on this basis to improve production line efficiency.

[0021] In some embodiments, performing a short connection operation on the first data and the multi-channel input data through a short connection layer to obtain data subjected to the short connection operation includes:

[0022] Performing channel convolution processing on the multi-channel input data to extract feature information of input data in each channel of the multi-channel input data, and obtaining first data containing each feature information;

[0023] Reshape the first data so that the dimension of each feature information in the first data matches the input data in each channel of the multi-channel input data, to obtain reshaped second data, and concatenate the second data with the multi-channel input data to obtain third data;

[0024] Performing a full connection operation on the third data through a fully connected layer to calculate a weight value of each channel according to each feature information and the original data of the multi-channel input data, and obtaining a target weight vector, wherein the target weight vector is composed of each weight value, and each weight value represents the dimension size of each channel;

[0025] superimposing the target weight vector onto the multi-channel input data to obtain fourth data, wherein the fourth data is used to distinguish weights of different channels;

[0026] Performing a short connection operation on the fourth data and the multi-channel input data through a short connection layer, so that the fourth data and the multi-channel input data are added to obtain fifth data;

[0027] The data after the short connection operation is input as the target input data into the stacked layer group, and each network layer in the stacked layer group sequentially operates on the target input data to learn the relationship between the state data of the production line and the downtime waiting time, and output the downtime waiting time prediction vector, including:

[0028] The target weight vector and the fifth data are input as target input data into the stacked layer group, and each network layer in the stacked layer group operates on the target input data in turn, so that the output of the previous network layer in the stacked layer group is used as the output of the subsequent network layer, and the terminal network layer of the stacked layer group outputs the downtime waiting time prediction vector; wherein the stacked layer includes a plurality of network layers arranged layer by layer in a preset order, and each of the network layers has the fully connected layer and the short connected layer.

[0029] In the embodiment of the present application, each network layer has a full connection operation function and a short connection operation function. The stacked layer group obtained by superposition and training of multiple layers continuously learns and adjusts the network parameters, and finally outputs the downtime waiting time prediction vector. This vector integrates the downtime waiting time prediction results of all storage locations, and the obtained downtime waiting time prediction results are more accurate and reliable.

[0030] In some embodiments, the step of inputting the state data of the production line into the trained second model and combining the downtime waiting time to generate a material distribution task includes:

[0031] Input the downtime waiting time and multi-channel input data at the tth moment into the third model to obtain the state value prediction value under the production line state data at the tth moment; wherein the third model is a network model that adjusts the input and output structure of the first model;

[0032] Combined with the predicted value of the state value at the tth moment, outputting the replenishment decision action of each of the cache locations at the tth moment through the fully connected network structure of the second model;

[0033] The material delivery task at the t-th moment is generated based on the material replenishment decision action.

[0034] In the embodiment of the present application, the downtime waiting time provides an estimate of the downtime waiting time that may occur under the current storage location state. The third model will output a state value prediction value based on the current storage location state (including information such as the predicted value of the downtime waiting time), which reflects the environmental feedback that can be obtained by executing a certain replenishment decision action under the current state, and the state value prediction value further considers the impact of these downtime waiting times on the overall production efficiency, as well as the possible improvements brought about by the replenishment decision. Then, the second model combined with the third model will make a replenishment decision action based on this state value prediction value and other information, further providing an accurate and important basis for the subsequent material distribution task generation.

[0035] In some embodiments, the decision content of the replenishment decision action is a binomial distribution probability value, wherein the binomial distribution probability value is a first value or a second value, the first value indicates that the cache location takes replenishment, and the second value indicates that the cache location does not take replenishment;

[0036] The generating of the material distribution task at the time t based on the material replenishment decision action includes:

[0037] When the decision content of the material replenishment decision action at the t-th moment is the first value, a material distribution task for the cache location at the t-th moment is generated.

[0038] In the embodiment of the present application, the decision content of the feeding decision action is a binomial distribution probability value, which can accurately determine whether the feeding decision action output by the second model needs to take feeding.

[0039] In some embodiments, the method further comprises:

[0040] After executing the refill decision action at the tth time, determining the state data of the production line at the t+1th time and the execution result at the t+1th time, wherein the execution result represents the environmental feedback after the refill decision action at the tth time is executed;

[0041] The refill decision action at the tth moment is evaluated based on the execution result to obtain an execution evaluation result, and the execution evaluation result is used to adjust the training parameters of the third model, and the training parameters include the level index of the storage location in the first data of the production line.

[0042] In the embodiment of the present application, the state value prediction value can evaluate the decision value under the current storage location state, can estimate the execution results that can be obtained for all future replenishment decisions, and can guide the strategy update in the reinforcement learning process.

[0043] In some embodiments, the method further comprises:

[0044] Calculate the timing target according to the predicted value of the state value at the time t+1 and the execution result at the time t;

[0045] The timing error is calculated based on the predicted value of the state value at the tth moment and the timing target, and the timing error is used as the execution evaluation result to update the training parameters of the second model and the third model.

[0046] In the embodiment of the present application, during the training process of the second model and the third model, the second model will continuously adjust its decision-making strategy based on environmental feedback and the evaluation of the third model to maximize the long-term execution results. The timing error obtained in this embodiment will serve as the basis for the second model to adjust the decision-making strategy, and will also be used to update the parameters of the third model itself to improve the accuracy of the evaluation.

[0047] In a second aspect, the present application further proposes a material distribution task generating device, comprising:

[0048] Status collection module, used to collect status data of the production line;

[0049] A downtime waiting time prediction module, used for inputting the state data of the production line into the first model, and using the multi-channel deep residual network model structure of the first model to predict the downtime waiting time of the cache storage location of the production line;

[0050] The material distribution task production module is used to input the status data of the production line into the trained second model, and generate a material distribution task in combination with the downtime waiting time for materials.

[0051] In a third aspect, the present application further proposes a terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in the first aspect above when executing the computer program.

[0052] In a fourth aspect, the present application further proposes a computer program product, wherein the computer program product comprises a computer program, and when the computer program is executed by a processor, the method described in the first aspect is implemented.

[0053] It can be understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0055] Figure 1 A schematic diagram of a material distribution task generation method embodiment provided in this application;

[0056] Figure 2 The embodiment of the present application provides a schematic diagram of a multi-channel deep residual network structure of a first model;

[0057] Figure 3 A schematic diagram of another embodiment of a material distribution task generation method provided in this application;

[0058] Figure 4 A schematic diagram of multi-channel input data formed by multi-channel encoding representation of the state data of the production line provided in the present application;

[0059] Figure 5 A schematic diagram of a process of encoding input data provided in an embodiment of the present application into a first model for prediction;

[0060] Figure 6 A schematic diagram of another embodiment of the process of encoding input data provided by the present application into a first model for prediction;

[0061] Figure 7 The embodiment of the present application provides a schematic diagram of the network structure of the second model;

[0062] Figure 8 A schematic diagram of another embodiment of the present application for characterizing the state data of the production line by multi-channel encoding;

[0063] Fig. 9 The embodiment of the present application provides a schematic diagram of the network structure of the third model;

[0064] Fig.10 A schematic diagram of modeling the second model and the third model using a reinforcement learning framework in an embodiment of the present application;

[0065] Fig.11 A flowchart of an embodiment of a specific implementation method of generating a material delivery task in an embodiment of the present application;

[0066] Fig.12 A schematic diagram of the structure of an embodiment of a terminal device for generating a material distribution task provided by the present application;

[0067] Fig.13 It is a structural block diagram of a material distribution task generation device provided in this application. DETAILED DESCRIPTION

[0068] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by technicians in the technical field to which this application belongs; the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" in the specification and claims of this application and the above-mentioned figure descriptions and any variations thereof are intended to cover non-exclusive inclusions.

[0069] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0070] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0071] In the description of the embodiments of the present application, the term "plurality" refers to more than two (including two), unless otherwise clearly and specifically defined.

[0072] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0073] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0074] The inventor of this application noticed that in the manufacturing process of new energy batteries, the assembly of battery packs involves multiple processing stages, and each process stage has line-side cache storage locations of different capacities, the inventory consumption rate is uneven, and the distribution frequency varies greatly. In order to avoid waiting for incoming materials, the conventional method is to adopt a safety inventory strategy, but frequent production line cutting and pulling changes lead to chaotic material preparation logistics and accumulation of redundant materials, affecting production order.

[0075] The application scenario of the above problem is a multi-stage flow shop (MSFS), which includes multiple important processes of module / PACK sections such as plasma cleaning, gluing, laser welding, and mechanical parts installation. Each process section has a number of line-side cache locations with different capacities and locations. Considering factors such as the rhythm of each section of the production line, WIP (Work in Process) inventory level and capacity, cutting and pulling changeover, material BOM (Bill of Material, BOM), assembly excellence, process loss, etc., the actual inventory consumption speed in the cache location shows a non-uniform distribution phenomenon, and the corresponding material delivery frequency varies greatly.

[0076] In the actual production process, in order to avoid waiting for incoming materials, a safety inventory strategy is often adopted. The safety inventory strategy needs to control the inventory balance to always be above the safety level. Considering the frequent cutting and pulling of production lines, this leads to problems such as chaotic material preparation logistics and accumulation of redundant materials in some cache locations. In serious cases, it even affects the production order on site. Therefore, the generation strategy of on-site material distribution tasks needs to be optimized and improved. At present, the improvement points of on-site inventory mainly consider the following points:

[0077] (1) Optimize the material delivery window and improve the timeliness of material delivery.

[0078] (2) Materials are delivered on demand to reduce the accumulation of redundant materials in storage.

[0079] (3) Establish an event-driven material distribution task generation mechanism to avoid waiting for incoming materials and reduce the situation where material distribution tasks are too dispersed.

[0080] In order to solve the problem of material distribution within the field, the embodiments of the present application provide a material distribution task generation method, device, terminal equipment and computer program product. In order to illustrate the technical solution proposed in the embodiments of the present application, it is described below through specific embodiments.

[0081] First, see Figure 1 , Figure 1This is a flow chart of the first embodiment of a material distribution task generation method provided in the present application. The material distribution task generation method in the embodiment of the present application can be applied to terminal devices on the production line. It should be noted that the terminal device in the embodiment of the present application is an edge device deployed on the processing production line, which belongs to the execution subject of the material distribution task generation method of the present invention. The edge device can have a built-in processor, and can also have onboard analysis capabilities or artificial intelligence analysis capabilities. In some real-time scenarios, there may be multiple different processing production lines in a factory, each of which can be deployed with an edge device, and there will be multiple processing links in a production line; in a specific embodiment of the present invention, different processing links can be controlled by the same edge device to perform processing control. Of course, different processing links can correspond to different edge devices for control. The method includes steps S10 to S30:

[0082] Step S10: Collecting status data of the production line;

[0083] The status data of the production line includes first data of the production line and second data of the production line, the first data of the production line at least includes level indicators of different types of storage locations, and the second data of the production line at least includes storage location consumption time and work-in-progress consumption time;

[0084] In the specific implementation, the first data of the production line mainly considers the basic parameters in the production-logistics system of the production line. The first data of the production line in the embodiment of the present application includes: the current line-side storage level, the production rhythm of the work section, the production rhythm of the previous work section, the bill of materials (BOM) ratio, the previous work-in-process inventory level, the previous work-in-process inventory consumption ratio, the minimum storage level, the maximum storage level, and the average equipment utilization rate, which can more comprehensively reflect the status of the production line. Among them, Work In Process (WIP) refers to products that are being processed and produced by industrial enterprises but have not yet been manufactured. Among them, "line side" refers to the side of the production line, and "storage location" refers to the regional location used to store materials required during the production line processing; accordingly, the specific meanings of the above first data of the production line include:

[0085] Lineside storage: refers to the storage area for materials set up next to the production line. These storage areas are usually used to temporarily store materials needed for production, so that workers can use them at any time to ensure the smooth operation of the production line.

[0086] Storage location level: refers to the number of materials currently stored in the line-side storage location. It can represent the current abundance of materials in each storage location to remind the relevant operators of the production line whether materials need to be replenished.

[0087] Section production rhythm: refers to the time required for a section on the production line to complete a product or a batch of products. It is an important indicator to measure production efficiency, which helps to arrange production plans and ensure the balanced operation of the production line.

[0088] The production rhythm of the previous section: refers to the time required for the section before the current section to complete a product or a batch of products. Understanding the production rhythm of the previous section helps to better coordinate the production rhythm between the sections and avoid material backlogs or shortages on the production line.

[0089] Bill of Materials (BOM) Ratio: refers to a list of the proportions of different materials in the entire product or assembly process. It lists all the materials required to produce a product and the quantity of each material, and is an important basis for production planning, material procurement and cost control.

[0090] Previous WIP inventory level: refers to the inventory of WIP (unfinished products) before the current section on the production line. This data can help production line operators understand the operating status of the production line and predict the demand for materials in the future.

[0091] Previous WIP inventory consumption ratio: refers to the ratio of the amount of materials consumed by the WIP before the current section on the production line to the total amount of materials. It reflects the use of materials on the production line, helps to discover the law of material consumption, and reasonably arrange material procurement and inventory management.

[0092] Minimum storage level: refers to the minimum requirement for the number of materials stored in the line storage. When the storage level is lower than this minimum requirement, materials need to be replenished in time to ensure the continuous operation of the production line.

[0093] Maximum storage level: refers to the maximum limit of the number of materials stored in the line storage. In order to avoid material backlog and waste, the storage level should not exceed this maximum limit. Reasonable setting of the maximum storage level helps optimize inventory management and reduce inventory costs.

[0094] Average equipment utilization rate: refers to the ratio of the actual operating time of the equipment to the total operating time within a period of time. It reflects the utilization efficiency of the equipment and is an important indicator for measuring the production capacity of the production line. By increasing the equipment utilization rate, production efficiency can be improved and production costs can be reduced.

[0095] Correspondingly, the second data of the production line is obtained based on the first data of the production line, that is, the second data of the production line is processed based on the first data of the production line to obtain the production line task data that is more helpful for the model to judge the shutdown waiting event. Among them, the second data of the production line includes: (line side) storage location consumption time and work-in-process inventory consumption time. The calculation methods of the two parameters are shown in formula (1) and formula (2):

[0096]

[0097] in, ,

[0098] Step S20: inputting the state data of the production line into the first model, and using the multi-channel deep residual network model structure of the first model to predict the downtime waiting time of the buffer storage location of the production line;

[0099] It is understandable that in a flow shop with a certain degree of flexibility, the triggers of the production line downtime waiting for materials event mainly consider the balance of the production line beat and the level of work-in-process inventory, the replenishment strategy and the level of line-side warehouses. To this end, the modeling of the production line downtime waiting for materials problem in the embodiment of the present application mainly considers the relationship between the consumption of work-in-process inventory and the consumption of line-side warehouse inventory, and based on this, constructs a first model for predicting the downtime waiting time of the production line, such as Figure 2 As shown, the first model is a multi-channel deep residual network model to predict the downtime waiting time. Figure 2 It is the specific structure of a multi-channel deep residual network model. Its structural characteristics include: it has the characteristics of a multi-layer configuration network, which can enhance the fitting ability; it has the characteristics of an attention mechanism, which can distinguish multi-channel weights; it has residual connection stacking BN and ReLU activation function layers, which has the function of reducing the risk of overfitting, gradient disappearance or explosion; it has the characteristics of a small number of hyperparameters, which is easy to adjust and converges quickly.

[0100] Specifically, Figure 3 As shown, step S20 of the embodiment of the present application, "inputting the state data of the production line into the first model to predict the downtime waiting time of the production line", further includes the following sub-steps: S201 to S202:

[0101] Step S201 (data encoding): performing multi-channel encoding representation on the state data of the production line to form multi-channel input data;

[0102] It is understandable that the process time between calling for material and implementing material replenishment of the production line is regarded as a preset fixed level, and the state data of the production line is used as the original data before inputting the first model. Now, the full-scale characteristic factors of the original data are characterized by multi-channel encoding. The encoding table is as follows: Figure 4 shown.

[0103] In the specific implementation, Figure 4The encoding method shown combines multi-channel input data: including the first data of the production line (the first data of the production line in the embodiment of the present application may include at least nine factors, namely: current storage level, section production rhythm, previous section production rhythm, BOM ratio, previous WIP inventory level, previous WIP consumption ratio, minimum storage level, maximum storage level, average equipment utilization rate) + the second data of the production line (the second data of the production line in the embodiment of the present application includes at least two factors, namely: storage consumption time and WIP consumption time) to form a multi-channel input data ,in is the sample batch, 11 is the number of channels, is the number of storage locations. Figure 4 As shown, each factor includes multiple locations (location 1, location 2, location 3, location 4, location 5, location 6…). Multi-channel input data is a three-dimensional matrix array, which is represented by Figure 4 A multi-channel encoding characterization scheme is provided, which describes in detail the combination of factors of the first data of the production line and the second data of the production line.

[0104] Step S202 (encoding input data to the first model for prediction): using the multi-channel deep residual network model structure of the first model to fit the multi-channel input data to predict the downtime waiting time.

[0105] It is understandable that the use of a multi-channel deep residual network model to fit the encoded features can be to input the multi-channel input data into the multi-channel deep residual network model structure to learn the relationship between the state of the production line and the downtime waiting time, and output the downtime waiting time prediction vector, the length of the downtime waiting time prediction vector corresponds to the number of cache locations, and each element of the downtime waiting time prediction vector represents the downtime waiting time prediction value of the corresponding target cache location. The structure of the multi-channel deep residual network model includes a multi-layer configuration network, an attention mechanism, a residual connection stacking BN and a ReLU activation function layer; the first model enhances the fitting ability through a multi-layer configuration network, uses an attention mechanism to distinguish multi-channel weights, and reduces the risk of overfitting, gradient disappearance or explosion through residual connection stacking BN and ReLU activation function layers.

[0106] In the specific implementation, refer to Figure 5 , the step S202 includes the following steps S2021 to S2023:

[0107] Step S2021: performing channel convolution processing on the multi-channel input data to extract feature information of the input data in each channel of the multi-channel input data, and obtaining first data containing each of the feature information;

[0108] Specifically, the embodiments of the present application can extract feature information of the input data in each channel by performing intra-channel convolution processing on multi-channel input data.

[0109] Step S2022: performing a short connection operation on the first data and the multi-channel input data through a short connection layer to obtain data subjected to the short connection operation;

[0110] It is understandable that the embodiments of the present application can perform a short connection operation on the data processed by feature extraction and the multi-channel input data through a short connection layer to enhance the model fitting ability and prevent the gradient disappearance or explosion problem.

[0111] Step S2023: Input the data that has undergone the short connection operation as target input data into the stacked layer group, and each network layer in the stacked layer group operates on the target input data in turn to learn the relationship between the status data of the production line and the downtime waiting time, and output the downtime waiting time prediction vector.

[0112] The above steps S2021 to S2023 of the embodiment of the present application first perform independent convolution operations on the input data in each channel of the multi-channel input data, and then enhance the model fitting ability through the short connection layer, and each network layer in the stacked layer group operates on the data after the short connection operation in turn, so that the multi-channel deep residual network model structure of this embodiment can deeply fit and learn the status data of the production line, and then realize accurate prediction of the downtime waiting time in a complex and changeable production line environment, and optimize material distribution based on this, thereby improving production line efficiency.

[0113] In a specific implementation, the above steps S2021 to S2023 may specifically include the following sub-steps A1 to A6:

[0114] Sub-step A1 (intra-channel convolution, corresponding to step S2021): perform convolution processing on the multi-channel input data to extract the feature information of the input data in each channel of the multi-channel input data, and obtain the first data containing each of the feature information. It is understandable that each input data sample of the multi-channel input data is a two-dimensional vector, and the input data in each channel of the multi-channel input data is independently subjected to channel-wise convolution operation, and each inner convolution operation is independently performed on the input data of each channel content, with the purpose of extracting the feature information in each channel. Sub-step A1 helps to enhance the model's sensitivity and recognition ability to different channel features. Among them, the relevant model operation process and the scale of the intermediate vector in sub-step A1 are as follows:

[0115] Input data for the L-1th sample: ;

[0116] In-channel convolution: ,

[0117] Where L represents the number of channels, Indicates the number of storage locations.

[0118] Sub-step A2 (Reshape): Reshape the first data obtained by the inner convolution operation so that the dimension of each feature information in the first data matches the input data in each channel of the multi-channel input data, thereby obtaining reshaped second data, and splicing the second data with the multi-channel input data (i.e., the original input data) to obtain (spliced) third data.

[0119] It is understandable that the Reshape operation is a data dimension transformation operation that can rearrange a multidimensional array (such as a matrix or tensor) into a multidimensional array of another shape without changing its data content. It is understandable that since the input data is encoded in the form of multiple channels, a multidimensional array (tensor) is formed, and then its dimensional shape is changed through the Reshape operation to adapt to the processing needs of the network model. In this process, the order of elements in the original array remains unchanged, but is reorganized according to the new shape. This helps to combine the convolved features with the original features and improve the prediction accuracy of the model.

[0120] Among them, the model operation process related to Reshape splicing in sub-step A2 is:

[0121]

[0122] The results of the convolution operation of the present invention need to be spliced ​​through the Reshape operation. Specifically, the output of the intra-channel convolution (whose shape is Before performing a Hadamard product operation with a weight vector (obtained through two fully connected operations), the convolution result needs to be reshaped in order to perform the operation with the weight vector.

[0123] Sub-step A3 (full connection operation): perform a full connection operation on the third data after Reshape splicing through a fully connected layer to calculate the weight value of each channel according to each of the feature information and the original data of the multi-channel input data, and obtain a target weight vector, wherein the target weight vector is composed of each of the weight values, and each of the weight values ​​represents the dimensional size of each channel.

[0124] It is understandable that the new data set is fully connected at least twice using the fully connected layer to further extract and integrate feature information. The output of the fully connected operation is a weight vector of channel lengths, which reflects the importance of different channels in predicting downtime waiting time;

[0125] The model operation process of the two full connections in sub-step A3 is as follows:

[0126]

[0127] The target weight vector is used to adjust or emphasize the importance of different channels in subsequent steps, thereby affecting the final downtime waiting time prediction.

[0128] Sub-step A4 (Hadamard matrix product superposition weight vector): superimpose the target weight vector on the multi-channel input data using the Hadamard matrix product method to obtain fourth data, wherein the fourth data is used to distinguish weights of different channels;

[0129] It is understandable that using the Hadamard product method to superimpose the weight vector on the original input data to obtain a new data set, namely the fourth data, helps to adjust the feature representation of the original input data according to the weight vector and improve the prediction accuracy of the model.

[0130] Sub-step A5 (short connection, corresponding to step S2022): perform a short connection operation on the fourth data and the multi-channel input data through a short connection layer, so that the fourth data is added to the multi-channel input data to obtain fifth data, so as to retain the original information and enhance the fitting ability of the first model, and prevent problems such as gradient disappearance or explosion during the training process.

[0131] In the specific implementation of this process, the fourth data will be processed by batch normalization (BN) to improve the training speed and stability of the neural network. It normalizes the input of each small batch of data so that its mean is close to 0 and its variance is close to 1. Next, the activation function ReLU (Rectified Linear Unit) can be applied. ReLU is a nonlinear activation function that sets all negative values ​​to 0 and keeps positive values ​​unchanged. This step is to introduce nonlinear factors so that the model can learn and simulate more complex functional relationships.

[0132] Then short connection (ShortcutConnection) processing is performed: the processed data will be combined with the fourth data obtained in the previous step through a short connection (also called residual connection). This connection method allows the model to directly access the output of the previous layer (or the previous few layers) and add it to the output of the current layer, which can help alleviate the gradient vanishing problem in deep networks and improve the training efficiency and performance of the network. Finally, the data processed by BN and ReLU is combined with the superposition results of the previous step to form new input data for further processing by the next network layer. This layer-by-layer superposition and layer-by-layer processing method enables the model to gradually extract the deep features of the input data, thereby realizing accurate prediction of the production line downtime waiting time;

[0133] Among them, the model operation process related to "Hadamard superposition weight vector + short connection" is:

[0134]

[0135] Sub-step A6 (multiple module layer blocks are superimposed, corresponding to the step S2023): the target weight vector obtained in sub-step A3 and the fifth data obtained in sub-step A5 are input as target input data into the stacked layer group, and each network layer in the stacked layer group operates on the target input data in turn, so that the output of the previous network layer in the stacked layer group is used as the output of the subsequent network layer, so as to learn the relationship between the state of the production line and the downtime waiting time, and finally the terminal network layer of the stacked layer group outputs the downtime waiting time prediction vector through continuous learning and adjustment of network parameters; the length of the downtime waiting time prediction vector corresponds to the number of cache locations, and each element of the downtime waiting time prediction vector represents the downtime waiting time prediction value of the corresponding target cache location; wherein the stacked layer includes a plurality of network layer blocks arranged layer by layer in a preset order, and each of the network layer blocks has the fully connected layer and the short connected layer.

[0136] Among them, the model operation process related to the "full connection operation + short connection operation" of the stacked layer is:

[0137]

[0138] The model operation process related to "fully connected output" is:

[0139]

[0140] That is to say, each network layer block has the full connection operation function of sub-step A3 and the short connection operation function of sub-step A5. Through the stacked layer group obtained by superposition and training of multiple layers of blocks, the model can learn the complex relationship between the production line status and the downtime waiting time. The final output is the downtime waiting time prediction vector, which integrates the downtime waiting time prediction results of all warehouse locations, providing an important basis for the subsequent material distribution task generation.

[0141] Specifically, the stacked layer group of sub-step A6 is a deep network structure formed by stacking multiple layers of blocks. It performs a series of complex operations and transformations on the input data, and finally outputs a downtime waiting time prediction vector corresponding to the number of storage locations. This vector accurately reflects the downtime waiting time of each storage location, providing strong support for optimizing material distribution tasks.

[0142] Step S30: inputting the state data of the production line into the trained second model, and combining the downtime waiting time to generate a material distribution task.

[0143] In order to cope with various complex and dynamically changing production environments on the production line and to realize real-time material distribution task decisions, the embodiment of the present application enters the current inventory levels of each storage location of the production line (i.e., the level indicators of the storage location in the first data of the production line, which can be regarded as hyperparameters to be optimized) into the input data to make real-time decisions on replenishing materials.

[0144] The input data of the replenishment decision follows the data encoding scheme of step S201 involved in the first model; in a specific implementation, the second model obtained by reinforcement learning training can be used to make material distribution task decisions. In the embodiment of the present application, a second model for users to implement replenishment decisions can be established using a fully connected network. The model structure of the second model is as follows: Figure 7 As shown, the final activation function of the model network of the second model constructed in the embodiment of the present application uses the Softmax method, and the decision content is: whether the current (line edge) storage location level is replenished (the output of the corresponding network is a binomial distribution probability value, wherein the binomial distribution probability value is a first value or a second value, the first value represents that the cache storage location takes replenishment, and the second value represents that the cache storage location does not take replenishment).

[0145] It is understandable that if Figure 7As shown, the second model constructed in the embodiment of the present application may include an input vector unit (Input vector), a flattening processing unit (Flatten), multiple fully connected layers (Multi-layer), a Softmax function calculation unit, and an output action vector unit (Output action vector). The second model receives various feature information that affects the feeding decision through the input vector, and after being processed by multiple layers of fully connected layers and converted by the Softmax function, outputs a probability distribution representing the feeding decision (i.e., an output action vector). Specifically:

[0146] Input vector unit: The input vector is the input part of the second model, representing various factors or features that affect the restocking decision of the second model. These features may include the first data of the production line and the second data of the production line mentioned above; the first data of the production line may include the current (line side) storage level, the production rhythm of the section, the production rhythm of the previous section, the bill of materials (BOM) ratio, the previous WIP inventory level, the previous WIP consumption ratio, the minimum storage level, the maximum storage level, the average equipment utilization rate, etc. The second data of the production line may include abstract features such as storage consumption time and work-in-process inventory WIP consumption time. After encoding and processing, these features form input vectors for use by the second model.

[0147] Flatten processing unit (Flatten): Flattening can be understood as an operation that converts multi-dimensional input data into one-dimensional data so that the subsequent fully connected layer can process it. For the second model's feeding decision network, if the input data is multi-dimensional (for example, composed of multiple feature channels), it needs to be converted into a one-dimensional vector through a flattening operation so that it can be input into the fully connected layer.

[0148] Multiple fully connected layers (Multi-layer): Multi-layer refers to multiple fully connected layers in the model, which perform linear transformations on the input data through weight matrices and bias vectors, and then apply activation functions (such as ReLU) to increase the nonlinear capabilities of the model. For the second model's feeding decision network, multiple layers of fully connected layers are used to extract deep features of the input data and make feeding decisions based on these features.

[0149] Softmax function calculation unit: The Softmax used in the embodiment of the present application converts the input vector (usually the output of the last fully connected layer of the model) into a probability distribution, in which the value of each element is between 0 and 1, and the sum of the values ​​of all elements is 1. For the feeding decision network of the second model, the Softmax function is used to convert the output of the model into a probability value of feeding (yes, corresponding to the first value) or not feeding (no, corresponding to the second value).

[0150] Output action vector unit: The output action vector is the final output of the second model, representing the feeding decision made by the second model based on the input data. For the feeding decision network of the second model, the output action vector is usually a two-dimensional vector, in which one element represents the probability of feeding (a value output by the Softmax function) and the other element represents the probability of not feeding (usually 1 minus the probability of feeding). However, in practical applications, in order to simplify the decision-making process, decisions are usually made directly based on the probability value of feeding (for example, if the probability of feeding is greater than a certain threshold, then decide to feed; otherwise, decide not to feed).

[0151] In some embodiments, for the final generation of the material distribution task, the second model and the third model are also used to make a material distribution task decision;

[0152] The third model is a network model that adjusts the input and output structure of the first model. The specific transformation method is: the output structure of the first model is adjusted from the downtime waiting time prediction value to the decision value evaluation value under the current storage state, which is used to evaluate the comprehensive effect after the execution of the replenishment decision, mainly considering the downtime waiting time and the replenishment frequency; that is, the transformation part of the first model into the third model only includes the input and output structure, and the output structure is changed from the output of the first model Became ; Input from the first model input become , an additional action representation channel is added to indicate whether each storage location is replenished, such as Figure 8 As shown, Fig. 9 It is a schematic diagram of the network structure of the third model;

[0153] It is understandable that the first model ( Figure 2 ) is transformed into the third model ( Fig. 9 ) mainly transformed the input and output structure. The first model ( Figure 2 ) is , is a vector form, representing the predicted value of downtime waiting time corresponding to each storage location; the third model ( Fig. 9 ) , is a scalar, representing the predicted value under the current production line status, and is an estimate of the sum of rewards that can be obtained from all future replenishment decisions ("reward" can be understood as the environmental feedback after the execution of the replenishment decision).

[0154] Specifically, the first model ( Figure 2 ) to the third model ( Fig. 9 The change in the output form of ) can be reflected by the following expression:

[0155] The output structure of the first model is:

[0156] ;

[0157] The output structure of the third model is:

[0158] ;

[0159] The changes in the input form of the two models can be reflected by the following expressions:

[0160] The input structure of the first model is: ;

[0161] The output structure of the third model is: , an additional action representation channel is added to indicate whether each storage location is replenished. The other structures are the same.

[0162] Furthermore, after the third model is obtained by transforming the first model, the embodiment of the present application can combine the third model and the second model into a dynamic prediction decision model based on the Actor-Critic model, so as to realize the prediction of downtime waiting time in the production-logistics system and the real-time generation of optimization decisions for material distribution tasks. Specifically, the first model is used as a Critic to evaluate the decision value under the current storage location status; the second model is used as an Actor model to quickly make replenishment decisions and produce material distribution tasks. In this solution, the default replenishment strategy is to replenish the inventory from the current storage location level at one time. Fig.10 Schematic diagram of modeling the second and third models using the Actor-Critic reinforcement learning framework.

[0163] Accordingly, in step S30, a material delivery task is generated at the current time t, which specifically includes sub-steps S301 to S303:

[0164] Sub-step S301: input the downtime waiting for materials and multi-channel input data at the tth moment into the third model to obtain the state value prediction value under the production line state data at the tth moment; it can be understood that the downtime waiting for materials provides an estimate of the downtime waiting for materials that may occur under the current storage location state. The third model will output a state value prediction value based on the current storage location state (including information such as the predicted value of the downtime waiting for materials), which reflects the environmental feedback that can be obtained by executing a certain replenishment decision action under the current state. The state value prediction value further considers the impact of these downtime waiting for materials on the overall production efficiency, as well as the improvement that may be brought about by the replenishment decision.

[0165] Sub-step S302: In combination with the predicted value of the state value at the t-th moment, the replenishment decision action of each of the cache locations at the t-th moment is output through the fully connected network structure of the second model; wherein the decision content of the replenishment decision action is a binomial distribution probability value, wherein the binomial distribution probability value is a first value or a second value, the first value represents that the cache location takes replenishment, and the second value represents that the cache location does not take replenishment.

[0166] Sub-step S303: Generate the material distribution task at the tth moment based on the material replenishment decision action. That is, when the decision content of the material replenishment decision action at the tth moment is the first value, generate the material distribution task of the cache location at the tth moment.

[0167] It is understandable that in the specific implementation, in order to use the reinforcement learning framework for modeling, the decision-making problem of stopping for material and replenishing material is abstracted as follows: at the current time t, the relationship between the consumption of work-in-process inventory in the production line and the consumption of the inventory at the line side is considered as the production line state data s, including the trigger of the production line stopping for material events, which mainly considers the balance of production line beat and work-in-process inventory level, replenishment strategy and line side warehouse level, BOM ratio, etc. Figure 2 The encoding representation method (i.e., multi-channel input data) shown in the figure inputs the multi-channel input data representing the production line status data s as input data to the first model Predict (downtime waiting time prediction model), the third model Critic (value network model), and the second model Actor (strategy network model), respectively obtaining the predicted value yt of the downtime waiting time of each cache location under the current production line status data s and the predicted value of the state value under the current production line status s. , and the replenishment decision actions at each storage location a That is to say, the embodiment of the present application first collects the production line status data s (including various information mentioned in the production line first data and the production line second data) and inputs it into the first model to obtain the predicted value yt of the downtime waiting time. Then, this predicted value yt and some relevant information in the production line status data s will be input into the third model to obtain the state value prediction value Finally, the second model predicts the value based on the state value And some relevant information in the production line status data s to make a refill decision action a , thus generating the final material distribution task. The three models work closely together to achieve the optimal generation of material distribution tasks.

[0168] Further, after executing the feed decision action obtained in the above sub-step S302, a Afterwards, determine that the production line is at t+The state data s+1 at time 1, and the t+ Execution results at time 1 r The execution result represents the execution of the feeding decision action at the tth moment. a The subsequent environmental feedback;

[0169] Based on the execution results r Evaluate the feeding decision action at the tth moment a , obtaining an execution evaluation result, wherein the execution evaluation result is used to adjust the training parameters of the third model, wherein the training parameters include the level index of the storage location in the first data of the production line.

[0170] Among them, the training parameters of the third model can be understood as hyperparameters to be optimized. "Hyperparameters to be optimized" refer to parameters that need to be optimized and adjusted in the network model related to the material distribution task decision. These parameters directly affect the decision results of material distribution. Hyperparameters to be optimized (such as the inventory level corresponding to the material type in the warehouse area) are actually an important part of multi-channel input data. In the decision model, these hyperparameters are used as part of the input data to generate replenishment decisions.

[0171] Through the multi-channel encoding representation scheme, the hyperparameters to be optimized are effectively integrated into the input data, enabling the model to understand the production line status more comprehensively and make more accurate feeding decisions accordingly.

[0172] It is understandable that the embodiment of the present application can be a certain assembly link in the battery pack assembly process of the battery module. Taking a certain assembly link scenario of the battery pack as an example, the setting of the on-site cache warehouse area is determined according to the working station on the pulling line. According to the material characteristics, the types of on-site cache warehouse areas mainly consider large packaging loading areas, large material loading areas and small material loading areas. Among them, the material types of the large packaging loading areas and the large material loading areas are single. There may be multiple types of materials stored in the small material loading area at the same time, such as bars, high-voltage connectors, pressure strips, assembly mechanical parts, wiring harnesses, standard parts, etc. The embodiment of the present application can perform hyperparameter characterization on the relevant status data of the material distribution task in this scenario. First, consider the large packaging loading area and the large material loading area, and take the material level in the warehouse as the decision point of the material distribution task. Since the material type in each warehouse area is single, the level of each warehouse area in the large packaging loading area and the large material loading area is separately regarded as a hyperparameter to be optimized. Secondly, some key materials in the small-piece loading area (such as panels, wire harnesses, etc.) are considered, and the inventory levels corresponding to each material type in the warehouse are identified as different hyperparameters to be optimized.

[0173] In the specific implementation, the feeding decision action is executed a , and then observe the next moment ( t+1 ) status data s+1 and execution resultr , execution results r Indicates the execution of the refill decision action a The environmental feedback obtained after the decision is a judgment on the quality of a decision made in the current state. The embodiment of the present application defines the execution result by the following formula r The calculation formula is:

[0174]

[0175] in, and It represents the weight factor of the task execution result mechanism on the downtime waiting for material and the refueling frequency.

[0176] Finally by the moment t and time t+ The state value prediction value and execution result of the third model output under 1 are used to calculate the timing error , used to update the trainable parameters of the third model.

[0177] It is understandable that in the embodiment of the present application, model training will be performed for the "second model Actor (strategy network model)" and the "third model Critic (value network model)". Among them, the "second model Actor (strategy network model)" is responsible for making the feeding decision, that is, whether to feed; the "third model Critic (value network model)" is responsible for evaluating the decision value in the current state, that is, predicting the execution result that may be obtained after executing a certain feeding decision action (the execution result represents the environmental feedback after the execution of the feeding decision action).

[0178] It should be noted that the "first model Predict (downtime waiting time prediction model)" is used as an auxiliary tool in the model training process to calculate the possible downtime waiting time when the refueling decision is not executed, and to feed this information back to the user or for other decision-making processes. The "first model Predict (downtime waiting time prediction model)" itself does not directly participate in the parameter update of the model training process.

[0179] In a specific implementation, the embodiment of the present application introduces a trust region strategy to optimize the training process of the model. For each time step t, the following training process is performed:

[0180] Step C1: Observe the current state of the production line at time t, use the second model (i.e., the policy network model Actor) to make decisions, output and execute actions. Specifically, observe the current state data of the production line , use the second model to make decisions: Execute the refill decision action ;

[0181] Step C2: Observe the execution action from the current state environment The execution result and new state after the environment observation are obtained, and the combination of state-action-execution result-new state is obtained.

[0182] Step C3: No actual action is performed, only the third model (Critic) is used to score the current state. Use the second model to make a decision: , no action is performed ;

[0183] Step C4: Calculate the time series error (TD error) between the target value (i.e. the estimated value of the future execution result) and the actual value. Use the following formula to achieve the use value network scoring:

[0184] Current state value prediction value (time t) , the predicted value of the state value at the next moment (t+1 moment) is .

[0185] Step C5: Based on the predicted value of the state value at the t+1th moment , the execution result at time t Calculating Timing Goals ; According to the state value prediction value at the tth moment and the timing goals A timing error is calculated, and the timing error is used as an evaluation result to update the training parameters of the second model and the third model.

[0186] Specifically, we can first calculate the TD target (timing target) and TD error (timing error): Timing target , timing error And update the value network: .

[0187] Step C6: Update the parameters of the policy network model using approximation and maximization methods to improve the decision-making process.

[0188] Specifically, the formula using the approximate method is:

[0189]

[0190] The formula using the maximization method is:

[0191]

[0192] in, Indicates the current step status. Indicates the current step action. Indicates the execution result. represents the discount factor, Represents timing error (TD error), represents the learning rate, represents the policy network of the second model, The value network representing the third model

[0193] According to the timing error The trainable parameters of the third model and the second model are updated, that is, the level indicators of the storage locations in the first data of the production line are optimized to guide the strategy update in the reinforcement learning process, which is helpful for updating future replenishment decisions.

[0194] Furthermore, in other embodiments, the safety stock strategy may be introduced and integrated with the modeling scheme of the Actor-Critic reinforcement learning framework in the above embodiment, further including steps Step 1 to Step 5:

[0195] Step 1: Collect the current production line status data s. The production line status data takes into account the current (line side) storage level, production rhythm of the work section, production rhythm of the previous work section, bill of materials (BOM) ratio, previous WIP inventory level, previous WIP consumption ratio, minimum storage level, maximum storage level, average equipment utilization rate, storage consumption time and WIP consumption time, plus replenishment decision, a total of 12 dimensions.

[0196] Step 2: Trigger the safety stock check condition and check the inventory level in the current state. If the current inventory level is in the cut-off zone, replenish the material directly; otherwise, proceed to the next step.

[0197] Step 3: Input the current production line status data s into the second model for decision making and output the refill decision action a , the action space is fixed, all of which are either replenishing or not replenishing.

[0198] Step 4: Execute the feeding decision action a ,like a To make a refill decision, the execution action is obtained from the production line status environment. a The execution result after r And the updated production line environment status s - .

[0199] Step 5: If the action a In order to execute the decision not to refill materials, the first model is used to calculate the downtime waiting time for materials and the time is fed back to the terminal equipment used by the relevant staff of the production line.

[0200] On the other hand, see Fig.12 , Fig.12 This is a schematic diagram of an embodiment of a terminal device for generating a material distribution task provided in the present application. The terminal device may be an edge device deployed in a processing production line, such as Fig.12 As shown, the terminal device of this embodiment includes: at least one processor 10 ( Fig.12 Only one is shown), a memory 11 and a computer program 12 stored in the memory 11 and executable on the at least one processor 10, and when the processor 10 executes the computer program 12, the steps in the embodiment of the material distribution task generation method of the present application are implemented.

[0201] Fig.12 The terminal device shown may include, but is not limited to, a processor 10 and a memory 11. Those skilled in the art will appreciate that Fig.12 It is only an example of a terminal device and does not constitute a limitation on the terminal device. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, it may also include input and output devices, network access devices, etc.

[0202] The processor 10 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0203] In some embodiments, the memory 11 may be an internal storage unit of the terminal device, such as a hard disk or memory of the terminal device. In other embodiments, the memory 11 may also be an external storage device of the terminal device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal device. Further, the memory 11 may also include both an internal storage unit of the terminal device and an external storage device. The memory 11 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program. The memory 11 may also be used to temporarily store data that has been output or is to be output.

[0204] Further, in one embodiment, if Fig.13 As shown, the present invention also provides a material distribution task generating device, comprising:

[0205] A status collection module 10 is used to collect status data of the production line;

[0206] A downtime waiting time prediction module 20 is used to input the state data of the production line into the first model, and use the multi-channel deep residual network model structure of the first model to predict the downtime waiting time of the cache storage location of the production line;

[0207] The material distribution task production module 30 is used to input the status data of the production line into the trained second model, and generate a material distribution task in combination with the downtime waiting time.

[0208] It should be noted that the material delivery task generation device can be understood as a virtual device that can be loaded into the terminal device of the aforementioned embodiment. The terminal device calls the material delivery task generation device through the processor, and then runs the specific implementation plan in the above material delivery task generation method embodiment. The information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiment of this application. The specific functions and technical effects brought about by them can be specifically referred to the method embodiment part.

[0209] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0210] An embodiment of the present application further provides a storage medium, which is a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.

[0211] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.

[0212] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium can at least include: any entity or device that can carry the computer program code to the camera / terminal device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, RandomAccess Memory), electric carrier signal, telecommunication signal and software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0213] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0214] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A material distribution task generation method, characterized in that: include: Collect status data of production lines; Inputting the status data of the production line into the first model, and using the multi-channel deep residual network model structure of the first model to predict the downtime waiting time of the buffer storage location of the production line; Inputting the status data of the production line into the trained second model, and combining the downtime waiting time to generate a material distribution task; The step of inputting the state data of the production line into the first model and using the multi-channel deep residual network model structure of the first model to predict the downtime waiting time of the buffer storage location of the production line includes: Processing the state data of the production line according to a multi-channel coding representation method to obtain multi-channel input data; Fitting the multi-channel input data using the multi-channel deep residual network model structure of the first model to predict the downtime waiting time; The multi-channel deep residual network model structure using the first model is used to fit the multi-channel input data to predict the downtime waiting time, including: The multi-channel input data is input into a multi-channel deep residual network model structure to learn the relationship between the status data of the production line and the downtime waiting time, and output a downtime waiting time prediction vector, wherein the length of the downtime waiting time prediction vector corresponds to the number of the cache storage locations, and each element of the downtime waiting time prediction vector represents the downtime waiting time prediction value of the corresponding target cache storage location.

2. The method according to claim 1, characterized in that The status data of the production line includes first production line data and second production line data, the first production line data at least includes level indicators of different types of storage locations, and the second production line data at least includes storage location consumption time and work-in-progress consumption time; wherein the second production line data is obtained based on the first production line data.

3. The method according to claim 2, characterized in that The multi-channel input data is input into a multi-channel deep residual network model structure to learn the relationship between the state data of the production line and the downtime waiting time, and output a downtime waiting time prediction vector, including: Performing channel convolution processing on the multi-channel input data to extract feature information of input data in each channel of the multi-channel input data, and obtaining first data containing each feature information; Performing a short connection operation on the first data and the multi-channel input data through a short connection layer to obtain data subjected to the short connection operation; The data after the short connection operation is input as the target input data into the stacked layer group, and each network layer in the stacked layer group operates on the target input data in turn to learn the relationship between the status data of the production line and the downtime waiting time, and output the downtime waiting time prediction vector.

4. The method according to claim 3, characterized in that The step of performing a short connection operation on the first data and the multi-channel input data through a short connection layer to obtain data subjected to the short connection operation includes: Reshape the first data so that the dimension of each feature information in the first data matches the input data in each channel of the multi-channel input data, to obtain reshaped second data, and concatenate the second data with the multi-channel input data to obtain third data; Performing a full connection operation on the third data through a fully connected layer to calculate a weight value of each channel according to each feature information and the original data of the multi-channel input data, and obtaining a target weight vector, wherein the target weight vector is composed of each weight value, and each weight value represents the dimension size of each channel; superimposing the target weight vector onto the multi-channel input data to obtain fourth data, wherein the fourth data is used to distinguish weights of different channels; Performing a short connection operation on the fourth data and the multi-channel input data through a short connection layer, so that the fourth data and the multi-channel input data are added to obtain fifth data; The data after the short connection operation is input as the target input data into the stacked layer group, and each network layer in the stacked layer group sequentially operates on the target input data to learn the relationship between the state data of the production line and the downtime waiting time, and output the downtime waiting time prediction vector, including: The target weight vector and the fifth data are input as target input data into the stacked layer group, and each network layer in the stacked layer group operates on the target input data in turn, so that the output of the previous network layer in the stacked layer group is used as the output of the subsequent network layer, and the terminal network layer of the stacked layer group outputs the downtime waiting time prediction vector; wherein the stacked layer includes a plurality of network layers arranged layer by layer in a preset order, and each of the network layers has the fully connected layer and the short connected layer.

5. The method according to any one of claims 2 to 4, characterized in that: The step of inputting the state data of the production line into the trained second model and combining the downtime waiting time to generate a material distribution task includes: Input the downtime waiting time and multi-channel input data at the tth moment into the third model to obtain the state value prediction value under the production line state data at the tth moment; wherein the third model is a network model that adjusts the input and output structure of the first model; Combined with the predicted value of the state value at the tth moment, outputting the replenishment decision action of each of the cache locations at the tth moment through the fully connected network structure of the second model; The material delivery task at the t-th moment is generated based on the material replenishment decision action.

6. The method according to claim 5, characterized in that The decision content of the replenishment decision action is a binomial distribution probability value, wherein the binomial distribution probability value is a first value or a second value, the first value indicates that the cache location takes replenishment, and the second value indicates that the cache location does not take replenishment; The generating of the material distribution task at the time t based on the material replenishment decision action includes: When the decision content of the material replenishment decision action at the t-th moment is the first value, a material distribution task for the cache location at the t-th moment is generated.

7. The method according to claim 6, characterized in that The method further comprises: After executing the refill decision action at the tth time, determining the state data of the production line at the t+1th time and the execution result at the t+1th time, wherein the execution result represents the environmental feedback after the refill decision action at the tth time is executed; The refill decision action at the tth moment is evaluated based on the execution result to obtain an execution evaluation result, and the execution evaluation result is used to adjust the training parameters of the third model, and the training parameters include the level index of the storage location in the first data of the production line.

8. The method according to claim 7, characterized in that The method further comprises: Calculate the timing target according to the predicted value of the state value at the time t+1 and the execution result at the time t; The timing error is calculated based on the predicted value of the state value at the tth moment and the timing target, and the timing error is used as the execution evaluation result to update the training parameters of the second model and the third model.

9. A material distribution task generation device, characterized in that: include: Status collection module, used to collect status data of the production line; A downtime waiting time prediction module is used to process the state data of the production line according to a multi-channel coding representation method to obtain multi-channel input data; The multi-channel input data is fitted using the multi-channel deep residual network model structure of the first model to predict the downtime waiting time of the cache storage location of the production line, including: inputting the multi-channel input data into the multi-channel deep residual network model structure to learn the relationship between the state data of the production line and the downtime waiting time, and outputting a downtime waiting time prediction vector, wherein the length of the downtime waiting time prediction vector corresponds to the number of the cache storage locations, and each element of the downtime waiting time prediction vector represents the downtime waiting time prediction value of the corresponding target cache storage location; The material distribution task production module is used to input the status data of the production line into the trained second model, and generate a material distribution task in combination with the downtime waiting time for materials.

10. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 8 is implemented.

11. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images