Workshop capacity prediction system

Abstract

This invention discloses a workshop capacity prediction system, relating to the field of monitoring technology. It includes an edge processing module and a capacity prediction module. The edge processing module integrates material inventory quantity and operational data into data packets, captures the packet transmission environment, and calculates anomaly scores using an isolated forest model. It employs a priority transmission strategy to split the data packets into priority data packets and supplementary data packets, determining the transmission order and duration. After receiving the priority data packets, the capacity prediction module, based on the priority prediction strategy, chooses to wait until the end of the current cycle or combine the priority data packets and some supplementary data packets from the previous cycle. It then uses the backbone network and auxiliary network for joint prediction and updates parameters. At the end of the current cycle, based on the reception status of the supplementary data packets, it chooses to perform joint prediction through the backbone network and auxiliary network or use the backbone network alone for prediction, thus achieving capacity prediction that considers latency and congestion.

Description

Technical Field

[0001] This invention relates to the field of monitoring technology, and in particular to a workshop capacity prediction system. Background Technology

[0002] In modern manufacturing, capacity forecasting plays a crucial role in a company's production planning, resource allocation, and market competitiveness. Accurate capacity forecasting helps companies rationally arrange production plans, effectively avoid overcapacity or undercapacity, optimize resource allocation, thereby reducing production costs and enhancing the company's market responsiveness and economic benefits.

[0003] With the rapid development of the Industrial Internet, the significant increase in the number of data collection devices in workshops has generated massive amounts of production data. Existing capacity forecasting technologies often face latency and network congestion during massive data transmission, making it impossible to deliver complete data to the processing center in a timely manner. This results in a significant reduction in the real-time performance and accuracy of capacity forecasts, making it difficult for enterprises to make rapid and scientific decisions based on actual production conditions. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by proposing a workshop capacity prediction system to achieve timely capacity prediction that takes into account data transmission latency and network congestion.

[0005] The technical solution to achieve the purpose of this invention is as follows: The workshop capacity forecasting system includes an edge processing module and a capacity forecasting module; The edge processing module calls the material inventory quantity at the start of the cycle. and running data The data is then integrated into a data packet, the packet transmission environment is captured, and anomaly scores are calculated using the isolated forest model. A priority transmission strategy is adopted to split data packets into priority data packets and supplementary data packets, based on anomaly scoring. The transmission order and duration of the data packets to be transmitted are determined based on the remaining transmission status of the previous cycle. To transmit performance data, the data packets to be transmitted include priority data packets, supplementary data packets, and remaining data packets from the previous period; After receiving all the priority data packets for the current cycle, the capacity forecasting module identifies the next data packet to be received based on the priority forecasting strategy. If it is a remaining data packet from the previous cycle, it combines the priority data packets and some supplementary data packets transmitted in the previous cycle to perform joint forecasting using the backbone network and auxiliary network. A progressive strategy is then adopted based on the joint forecasting results and the total capacity for the current cycle. Update the parameters of the backbone network and auxiliary network, and wait until the end of the current cycle. If the remaining data packets are not from the previous cycle, they will also wait until the end of the current cycle. Based on the reception of supplementary data packets in this cycle, decide to jointly predict and generate the collaborative total capacity vector for this cycle through the backbone network and auxiliary network. Alternatively, the total backbone capacity vector can be used to predict the production cost cycle. The backbone network is a temporal reasoning network, and the auxiliary network is a convolutional network.

[0006] Furthermore, the edge processing module includes a processing unit, a docking unit, and a transmission unit; At the start of each cycle, the processing unit triggers the docking unit to obtain the material inventory quantity. and running data And package it into a data package; Once the docking unit is triggered, it calls the material inventory quantity via the API interface. and running data And feedback processing unit, wherein, For the first Operating data of the production equipment, including the first Taiwan production equipment capacity Operating rate ,temperature and power ; The transmission unit captures packets in the transmission environment to obtain transmission performance data. And input the isolated forest model to calculate the anomaly score. Based on anomaly scores The priority transmission strategy is enabled to split data packets and transmit them to the capacity forecasting module, transmitting performance data. Including throughput Packet loss rate and retransmission rate .

[0007] Furthermore, transmission performance data Placed in the isolated forest model The root node of an isolated tree, based on transmission performance data The feature values ​​in each feature dimension and the first The node splitting values ​​of the isolated tree are compared to determine whether to assign the node to the left or right child node, until the transmission performance data is obtained. Classified as the first When dealing with leaf nodes of an isolated tree, statistical transmission performance data is obtained. In the Path length of an isolated tree And calculate the average path length Average path length With correction factor The ratio is used as an exponent to calculate the power of 0.5, resulting in transmission performance data. Abnormal scores , This represents the total number of trees.

[0008] Furthermore, constructing an isolated forest model requires collecting transmission performance data samples as the original dataset, and then randomly drawing transmission performance data samples with replacement from the original dataset based on bootstrapping. Next, get The training subset will be the first The training subset is used as the first The root node of an isolated tree is randomly selected from a transmission performance data sample, along with a feature of one dimension and its corresponding splitting value. The root node is then split based on the selected feature's splitting value. The resulting nodes are then randomly split again using randomly selected features and splitting values, continuing this process until the th... In an isolated tree, if any of the unsplit nodes contain only a single transmission performance data sample or all of the existing transmission performance data samples are identical, then the first... The isolated tree training is complete, and unsplit nodes are considered leaf nodes. The isolated forest model is completed when the training of each isolated tree is finished.

[0009] Furthermore, the priority transmission strategy includes the following steps: Pack the high-priority data in the current cycle's data packets and add corresponding tags to both ends to generate the current cycle's priority data packets. Pack the remaining low-priority data and add corresponding tags to both ends to generate the current cycle's supplementary data packets. The high-priority data includes material inventory quantities. , The capacity and operating rate of the production equipment in Taiwan, low-priority data includes The temperature and power of the production equipment are indicated by markings indicating the data packet's ownership period and priority (priority or supplementary). Based on the length of the priority data packet and the supplementary data packet, and transmission performance data and abnormal scores Calculate the priority transmission duration for this period. and supplementary transmission time ; Confirm the delay transmission duration of the previous cycle. The priority transmission duration for this cycle and supplementary transmission time The total transmission duration for this cycle is obtained by superposition. And determine whether it is less than or equal to the period duration. ; If less than or equal to the period duration Based on the data packet priority, the existing data packets to be transmitted are transmitted to the capacity prediction module in sequence. The data packet priority is that the priority data packet of the current period is greater than the remaining data packet of the previous period, which is greater than the supplementary data packet of the current period. If it is greater than the period duration The total transmission time for this cycle With cycle duration The difference is used as the delay transmission duration for this period. Based on the priority of data packets, the existing data packets to be transmitted are transmitted to the capacity prediction module in sequence, and the transmission stops at the beginning of the next cycle.

[0010] Furthermore, the priority prediction strategy includes the following steps: When it is detected that the priority data packets of the current cycle have been received, it is determined whether the next data packet is a remaining data packet of the previous cycle. The identification and judgment of the data packets both depend on the tag. If the data packet is not a remaining data packet from the previous cycle, wait until the end of this cycle. If it is a remaining data packet from the previous cycle, after all data is received, it will be merged with the supplementary data packets already transmitted in the previous cycle to form a supplementary data packet for the previous cycle, and the previous cycle and earlier cycles will be invoked. The priority data packets and supplementary data packets of each cycle are concatenated to form the backbone sequence and auxiliary sequence of the previous cycle, respectively. These are then input into the backbone network and auxiliary network to jointly generate the previous cycle. Total collaborative capacity for each cycle; Extract the total collaborative capacity for this cycle and the total capacity of the main trunk line in this cycle Extract from the priority data packets of this period The total production capacity for this period is obtained by summing the production capacity of each production unit. Among them, the total capacity of the main trunk line in this cycle Derived from the previous period predicted and generated by the backbone network in the previous period. Total main capacity of each cycle; Calculate the total capacity of the main trunk line for this period based on a progressive strategy. With total collaborative capacity Main trunk coordination error and total coordination capacity With total production capacity The actual collaborative error is used to update the backbone network parameters and auxiliary network parameters respectively using the network optimization algorithm, and then waits until the end of the current cycle. At the end of this cycle, determine whether all supplementary data packets for this cycle have been received; If the data has not been fully received, call the current cycle and previous cycles. Prioritized data packets from each period are concatenated to form the backbone sequence for this period. The backbone network is then used to predict the generation of data for the next period. The total backbone capacity of each cycle is combined to form the vector of the total backbone capacity of the current cycle. ; If reception is complete, call the current cycle and previous cycles. The priority data packets and supplementary data packets of each period are concatenated to form the backbone sequence and auxiliary sequence of the current period. The backbone network and auxiliary network are used to jointly predict and generate the next period. The total collaborative capacity of each cycle is concatenated to form the total collaborative capacity vector for the current cycle. .

[0011] Specifically, the backbone network uses a Long Short-Term Memory (LSTM) network to sequentially extract the hidden states corresponding to the backbone vector for each cycle from the backbone sequence, generating a hidden state sequence. A multi-head attention mechanism is then used to monitor the temporal dependencies of the hidden state sequence. Finally, linear modulation, batch normalization, and ReLU function mapping are applied to generate the total backbone capacity vector, which includes the material inventory quantity. as well as The production capacity and operating rate of the equipment.

[0012] Specifically, the auxiliary network uses a 3x3 convolutional kernel to extract features twice from each auxiliary vector in the auxiliary sequence. Then, it performs pointwise convolution with a 1x1 convolutional kernel and generates high-level features corresponding to each auxiliary vector through ReLU mapping. Lightweight temporal convolution is used to model the temporal relationships between high-level features. Causal convolutional kernels with different dilation rates are used to perform temporal convolution on high-level features to capture the periodic dependence information of high-level features over different period spans. Pointwise convolution is used to fuse and generate a temporal enhanced feature sequence, which is then linearly modulated with the backbone total capacity vector output by the backbone network. A cooperative total capacity vector is generated through ReLU activation. The auxiliary vectors include... Temperature and power of the production equipment.

[0013] Compared with existing technologies, this invention sets up an edge processing module and a capacity prediction module, which effectively solves the problem of capacity prediction under transmission latency and network congestion. At the beginning of the cycle, the edge processing module integrates the material inventory quantity and operation data, and calculates the anomaly score by combining the transmission environment and the isolated forest model. Through the priority transmission strategy, it ensures the timely transmission of key data to the greatest extent. The capacity prediction module flexibly adjusts the prediction application of the backbone network and auxiliary network according to the priority prediction strategy and the progressive strategy, and updates the network parameters in real time, which significantly improves the real-time performance and accuracy of capacity prediction. Attached Figure Description

[0014] Figure 1 Here is a flowchart of the workshop capacity forecasting system; Figure 2 A flowchart for the priority transmission strategy; Figure 3 A flowchart for the priority prediction strategy; Figure 4 This is a model diagram of the backbone network and auxiliary networks. Detailed Implementation

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0016] like Figure 1 As shown, a specific embodiment of the present invention discloses a workshop capacity prediction system, including an edge processing module and a capacity prediction module; The edge processing module connects to and calls the material inventory quantity at the beginning of each cycle. and Operating data of production equipment The data is then integrated into a data packet, and anomaly scores are calculated using the UAExpert packet capture environment combined with the Isolation Forest model. A priority transmission strategy is adopted to split data packets into priority data packets and supplementary data packets, based on anomaly scoring. The transmission order and duration of all data packets to be transmitted are dynamically determined based on the remaining transmission status of the previous cycle. The transmission performance data obtained from packet capture includes priority packets, supplementary packets, and remaining packets from the previous cycle. This represents the total number of production equipment, set according to the actual situation of the factory. After receiving all the priority data packets for the current cycle, the capacity forecasting module identifies the next data packet to be received based on the priority forecasting strategy. If it is a remaining data packet from the previous cycle, after receiving it, it combines the priority data packets from the previous cycle with the partially transmitted supplementary data packets, uses the backbone network and auxiliary network to perform joint forecasting, and obtains the total collaborative capacity for the current cycle from the forecasting results. Calculate the total capacity for this period. The parameters of the backbone network and auxiliary network are updated based on a progressive strategy of approximating collaboration from the backbone and approximating reality from collaboration. If it is a supplementary data packet for the current cycle, no processing is performed until the end of the current cycle. Based on the reception status of the supplementary data packets for the current cycle, a decision is made to jointly predict and generate the total collaborative capacity vector for the current cycle through the backbone network and auxiliary network. Alternatively, the total backbone capacity vector can be used to predict the production cost cycle. The backbone network is a temporal reasoning network, and the auxiliary network is a convolutional network.

[0017] Furthermore, the edge processing module includes a processing unit, a docking unit, and a transmission unit; At the start of each cycle, the processing unit triggers a data call to the docking unit to obtain the material inventory quantity. and Operating data of production equipment And package it into a data package; The interface unit is triggered and controlled by the processing unit. It connects to the Enterprise Resource Planning (ERP) system and the Manufacturing Execution System (MES) via API interfaces to retrieve material inventory quantities. and Operating data of production equipment And feedback processing unit, wherein, For the first Operating data of the production equipment, including the first Taiwan production equipment capacity Operating rate ,temperature and power Production capacity and operating rate For the first A single production line The ratio of actual production capacity to operating time within the facility; The transmission unit uses UAExpert to capture packets in the transmission environment to obtain transmission performance data. And input the isolated forest model to calculate the anomaly score. Based on anomaly scores A priority transmission strategy is enabled to split data packets and transmit them to the capacity forecasting module in batches. , and These are throughput, packet loss rate, and retransmission rate, respectively.

[0018] Furthermore, the transmission unit will transmit performance data. After inputting the isolated forest model, performance data will be transferred. Placed in the isolated forest model The root node of an isolated tree, based on transmission performance data The feature values ​​in each feature dimension and the first In an isolated tree, the node splitting value is compared to decide whether to assign the node to the left or right child node. Each assignment affects the transmission performance data. In the Path length of an isolated tree Increment by 1 until the transmission performance data is available. Classified as the first When the leaf node of an isolated tree is accessed, transmission performance data is obtained. In the Path length of an isolated tree Calculate isolated forests Average path length of isolated trees Calculate transmission performance data Abnormal scores The details are as follows: , , in, This represents an exponential function with base 2. It is the natural logarithm function. Here, is Euler's constant, with a value of 0.577. To correct the coefficients and avoid the total number of isolated trees The difference in average path length Statistical differences, ensuring outlier scores Stability.

[0019] For example, the total number of trees Set to 100, collect 5000 transmission performance data samples as the original dataset. Based on bootstrapping, randomly draw 500 samples with replacement from the original dataset each time, for a total of 100 times, resulting in 100 training subsets, each with 500 samples. Adjust the coefficients... Specifically, it is 7.787.

[0020] Furthermore, building an isolated forest model requires pre-collecting a number of trees that are far greater than the total number of trees. Using transmission performance data samples as the original dataset, a number of samples with replacement are randomly drawn from the original dataset with replacement based on bootstrapping, exceeding the total number of trees. Furthermore, transmission performance data samples smaller than the total number of samples in the original dataset are repeatedly sampled autonomously. This time received The training subset, for the , The corresponding training subset is used to construct the training subset. The isolated tree, will be the first The training subset is used as the first The root node of an isolated tree is randomly selected. A feature of one dimension from a transmission performance data sample is randomly chosen, and a splitting value is randomly selected within the feature's range. The root node is then split based on the selected splitting value. For each split node, a feature and its splitting value are randomly selected again for further splitting, until the th node is reached. In an isolated tree, if all unsplit nodes have only a single transmission performance data sample, or all transmission performance data samples in all unsplit nodes are identical, then the first... After training the isolated trees, the nodes that have not been split are considered as the first... The leaf node of an isolated tree, when The isolated forest model is complete when all isolated trees have been trained.

[0021] For example, the correction factor The value is 7.787, based on the transmission performance data of the 20th cycle. For example, assuming throughput Packet loss rate and retransmission rate Given speeds of 12.5 MB / s, 0.03 MB / s, and 0.01 MB / s respectively, and 100 isolated trees as input, calculate the average path length. The score is 8.5, resulting in an anomaly score. It is 0.47.

[0022] like Figure 2 As shown, the priority transmission strategy further includes the following steps: Extract the high-priority data from the data packets for this period and package it. Add corresponding tags to both ends to generate the priority data packets for this period. The high-priority data consists of predefined data directly related to capacity forecasting, including material inventory quantities. and The capacity and operating rate of the production equipment in Taiwan are marked with ownership cycle and level, with the level including priority and supplement; The remaining low-priority data in the current period's data packets is packaged and marked with corresponding tags at both ends to generate a supplementary data packet for the current period. This low-priority data is used to assist high-priority data in capacity forecasting to improve forecast accuracy. The low-priority data includes... Temperature and power of the production equipment; Statistics on the length of priority data packets in this period and supplementary data packet length Based on throughput Packet loss rate retransmission rate with abnormal scores Fine-tune the priority transmission duration for this cycle. and supplementary transmission time The details are as follows: , , in, This is the adjustment coefficient. Based on empirical settings or determined by fitting historical data, it is determined whether there are any remaining data packets from the previous period. In this embodiment, the adjustment coefficient... It is 1.3; If there are remaining data packets from the previous period, obtain the delayed transmission duration from the previous period. and the priority transmission duration of this cycle. and supplementary transmission time The total transmission duration for this cycle is obtained by summing the results. Determine the total transmission duration of this cycle. Is it less than or equal to the period duration? , If less than or equal to the period duration This indicates that the priority data packets and supplementary data packets of this cycle, as well as the remaining data packets of the previous cycle, can all be transmitted within this cycle. They are transmitted to the capacity prediction module in sequence according to the data packet priority, where the data packet priority is: priority data packets of this cycle > remaining data packets of the previous cycle > supplementary data packets of this cycle. If it is greater than the period duration This indicates that it is impossible to transmit all the priority data packets and supplementary data packets of this cycle, as well as the remaining data packets from the previous cycle, within this cycle, thus reducing the total transmission time of this cycle. With cycle duration The difference is used as the delay transmission duration for this period. Data packets are transmitted sequentially to the capacity prediction module according to their priority and transmission stops at the start of the next cycle. If there are no remaining data packets from the previous period, the delay transmission duration of the previous period. Total transmission duration for this cycle Determine the total transmission duration of this cycle. Is it less than or equal to the period duration? ; If less than or equal to the period duration This indicates that both the priority data packets and supplementary data packets for this cycle can be transmitted within this cycle, and are transmitted to the capacity prediction module in sequence according to the data packet priority. If it is greater than the period duration This indicates that it is impossible to transmit all the priority data packets and supplementary data packets for this period within this cycle, thus reducing the total transmission time for this period. With cycle duration The difference is used as the delay transmission duration for this period. The data packets are transmitted sequentially to the capacity prediction module according to their priority and transmission stops at the start of the next cycle.

[0023] For example, assume the total number of production equipment The inventory quantity of packaged materials is 10 and the current attribution period is 20. The capacity and operating rate of 10 production machines are marked as "20+ Priority", and a priority data packet length is generated. For a 5MB priority data packet, package the temperature and power of 10 production devices, mark it "20+ Supplement", and generate a supplementary data packet of length. For a supplementary data packet of 15MB, the period duration is... The interval is set to 3 seconds. Since there are no remaining data packets in the 19th cycle, the transmission performance data for the 20th cycle is obtained. and abnormal scores Given a value of 0.47, calculate the priority transmission duration for the 20th cycle. The time is 0.794 seconds, as detailed below: , Due to the supplementary data packet length Prioritize packet length Three times that, therefore the supplementary transmission time for the 20th cycle is... Also three times, or 2.382 seconds; since there are no remaining data packets in the 19th cycle, the delay transmission time for the 19th cycle is... The summation is 0, and we get the total transmission duration of the 20th cycle. That is, 3.176 seconds, the total transmission duration of the 20th cycle. It has exceeded the cycle length. The 3-second delay in transmission occurs during the 20th cycle. It takes 0.176 seconds.

[0024] like Figure 3 As shown, the further preferred prediction strategy includes the following steps: If the "Current Period + Priority" marker is detected for the second time within this period, it is determined that the priority data packet of this period has been received, and it is then determined whether the "Current Period + Supplement" marker is present at the beginning of the next data packet. If it does not exist, it indicates that the next data packet is the remaining data packet of the previous cycle. It waits to identify the "previous cycle + supplement" marker, determines that the remaining data packet of the previous cycle has been transmitted, merges it with the transmitted portion of the supplement data packet of the previous cycle to form the supplement data packet of the previous cycle, and calls the previous cycle and previous cycles. The supplementary data packets for each period are concatenated into the auxiliary sequence for the previous period, and the previous and earlier cycles are called. The priority data packets of each cycle are concatenated to form the backbone sequence of the previous cycle. The backbone sequence and auxiliary sequence of the previous cycle are then input into the backbone network and auxiliary network respectively to jointly generate the collaborative total capacity vector of the previous cycle. The total collaborative capacity vector of the previous cycle Including the last cycle In this embodiment, the total collaborative capacity for each cycle is... The value is 8.

[0025] Extract the total collaborative capacity vector from the previous period Total collaborative capacity in the middle cycle And from the main trunk total capacity vector of the previous cycle Extract the main total capacity of this cycle Extract from the priority data packets of this period The total capacity for this period is obtained by summing the capacity of each production unit in this cycle. Among them, the total capacity vector of the main trunk in the previous cycle Generated solely through the backbone network predictions within the previous period, including periods after the previous period. Total main capacity of each cycle; Calculate the total capacity of the main trunk line for this period based on a progressive strategy. With total collaborative capacity Main trunk coordination error and total coordination capacity With total production capacity The collaborative actual error is used to update the parameters of the backbone network and the auxiliary network based on the backbone collaborative error and the collaborative actual error, respectively. This is to make the result predicted by the backbone network alone approximate the result predicted by the backbone network and the auxiliary network together, and at the same time make the result predicted by the backbone network and the auxiliary network together approximate the actual result, so as to achieve a progressive approximation effect. Then wait until the end of the current cycle. The network optimization algorithm includes stochastic gradient descent and Adam. If the "Current Period + Supplement" flag exists, it indicates that there are no remaining data packets in the previous period, and the next data packet is a supplementary data packet for this period. No processing is performed, and the packet waits until the end of the current period. At the end of the current cycle, determine whether the "current cycle + supplement" marker has been identified a second time; If the "Current Cycle + Supplement" marker is not detected a second time, it indicates that the supplementary data packet for the current cycle was not fully transmitted, and the current cycle and previous cycles are called. Priority data packets from each period are concatenated to form the backbone sequence for the current period. This backbone sequence is then input into the backbone network to predict and generate the total backbone capacity vector for the current period. The main trunk total capacity vector for this cycle Including this cycle and beyond Total main capacity of each cycle; If the "Current Cycle + Supplement" marker is detected a second time, it indicates that the supplementary data packet for the current cycle has been transmitted, and the current cycle and previous cycles are then invoked. Priority data packets from each period are concatenated into the backbone sequence for this period, and previous data packets from this period are called. The supplementary data packets for each period are merged with the supplementary data packets for the current period to form the auxiliary sequence for the current period. The backbone sequence and auxiliary sequence for the current period are then input into the backbone network and auxiliary network, respectively, to jointly predict and generate the collaborative total capacity vector for the current period. The total collaborative capacity vector for this cycle Including this cycle and beyond Total collaborative capacity for each cycle.

[0026] like Figure 4 As shown, specifically, a single-period backbone sequence includes a single period and the preceding period. The main vector for each period is composed of the material inventory quantity for that period. as well as The production capacity and operating rate of each production equipment are generated by concatenating columns. When the backbone sequence of a single period is input into the backbone network, the backbone network first uses a Long Short-Term Memory (LSTM) network to extract the backbone vector according to the period order. Based on the combined operation of forget gate, input gate and output gate, the hidden state corresponding to each backbone vector is calculated to obtain the hidden state sequence corresponding to the backbone sequence of a single period. Through the multi-head attention mechanism, the temporal dependency relationship between the hidden states of different periods in the hidden state sequence is focused on from different angles to filter out irrelevant information and generate an enhanced hidden state sequence. Through linear modulation, batch normalization and ReLU function, nonlinear mapping is performed to generate the total backbone capacity vector of a single period.

[0027] For example, when The value is 8 and the total number of production equipment When the value is 10, the core sequence dimension of each period is 8×21, where 8 represents the period number, and the core vector of each period is 21-dimensional, including the material inventory quantity of each period. The system takes the capacity and operating rate of 10 production machines as input. The backbone sequence for each cycle is input into the backbone network. First, it passes through a Long Short-Term Memory (LSTM) network with a hidden layer dimension of 64. When an 8×21 dimensional backbone sequence is input, the sequence length is kept at 8, and the feature dimension is increased to 64 to preserve the long-term temporal correlation between capacity and operating rate. This generates an 8×64 dimensional hidden state sequence, which is then enhanced by a multi-head attention mechanism. The number of heads is set to 4, and each head has a dimension of 16. The multi-head attention mechanism linearly projects the 8×64 dimensional hidden state sequence into 8×16 dimensional subspace features. By scaling dot product attention, the complementarity of different cycles within different subspaces is mined, irrelevant noise is filtered out, and the temporal correlation expression of the hidden state is enhanced. The output of each head is 8×16 dimensional and concatenated along the feature dimension to obtain an 8×64 dimensional enhanced hidden state sequence. Finally, two linear layers reduce the dimension from 8×64 dimensional to 8×1 dimensional. After batch normalization and ReLU function mapping, an 8×1 dimensional backbone total capacity vector for each cycle is obtained.

[0028] like Figure 4 As shown, specifically, single-period auxiliary sequences include single-period and preceding sequences. Each period has an auxiliary vector, and the auxiliary vector for each period is composed of the auxiliary vector for each period. The temperature and power of each production equipment are generated by concatenating columns. When a single-cycle auxiliary sequence is input into the auxiliary network, the network performs feature extraction three times consecutively on each auxiliary vector. The first feature extraction uses a convolution kernel of size 3 with a stride of 1 to convolve with each auxiliary vector, initially extracting the local features of each auxiliary vector. The second feature extraction again uses a convolution kernel of size 3 with a stride of 1 to convolve with each local feature again to uncover deeper and more complex local features. The third feature convolution uses a convolution kernel of size 1 with a stride of 1 to convolve point-by-point with each complex local feature to uncover the correlation between details and pass the ReLU function. To enhance nonlinear expressive power, high-level features corresponding to each auxiliary vector are generated and concatenated into a high-level feature sequence. Lightweight temporal convolution is used to perform temporal modeling on the high-level feature sequence. Lightweight temporal convolution sets causal convolution kernels with different dilation rates to change the receptive field. High-level features are convolved temporally by causal convolution kernels with different dilation rates to capture the periodic dependence information of high-level features over different period spans. The convolution results of all causal convolution kernels are fused by pointwise convolution to generate a temporally enhanced feature sequence and linearly modulate it with the backbone total capacity vector output by the backbone network. A single-period cooperative total capacity vector is generated by activation using the ReLU function.

[0029] For example, when The value is 8 and the total number of production equipment When the number is 10, the auxiliary sequence dimension for each cycle is 8×20, where 8 represents the cycle number. The auxiliary vector for each cycle is 20-dimensional, including the temperature and power of the 10 production devices in each cycle. The auxiliary sequence for each cycle is input into the auxiliary network. First, it is convolved with each auxiliary vector using a convolutional kernel of size 3 with a stride of 1, resulting in a 1×32-dimensional local feature corresponding to each auxiliary vector. Then, a convolutional kernel of size 3 is used again with a stride of 1 to convolve with each local feature to explore the nonlinear coupling relationship between temperature and power and enhance the expression of complex patterns in local features. The reason why the output dimensions of convolutional kernels of the same size are inconsistent is because the output channels of the first two convolutions are 32 and 64 respectively, while the third convolution uses a 1D convolutional kernel of size 1 with a stride of 1 to convolve with each complex feature. Pointwise convolution is performed on the features to uncover the correlations between details, and the nonlinear expressive power is enhanced by the ReLU function to generate a 1×64-dimensional high-level feature corresponding to each auxiliary vector. The high-level features of 8 periods are concatenated to obtain an 8×64-dimensional high-level feature sequence. Lightweight temporal convolution is then performed with convolution kernels of size 2 and dilation rates of 1, 2 and 4 to capture adjacent period dependencies, interval period dependencies and 4-period dependencies. Multi-scale fusion is performed again by pointwise convolution to obtain an 8×64-dimensional temporal enhanced feature sequence. The dimensionality is reduced to 8×1 dimension by a linear layer and weighted by the main total capacity vector of each period of the 8×1 dimension through a linear layer. Finally, the ReLU function is used to map and obtain the cooperative total capacity vector of each period of the 8×1 dimension.

[0030] Furthermore, a two-stage training strategy is adopted for the backbone network and auxiliary network. The training dataset is constructed based on historical data from 1000 consecutive cycles of production equipment. A 21-dimensional backbone vector and a 20-dimensional auxiliary vector are extracted from each cycle. These are concatenated into an 8×21-dimensional backbone sequence and an 8×20-dimensional auxiliary sequence after 8 consecutive cycles. The corresponding label is the actual total production capacity of the 8th cycle. A total of 993 samples are generated and divided into 794 training samples, 99 validation samples, and 100 test samples in an 8:1:1 ratio. The training loss adopts a weighted combined loss function, where the backbone prediction loss and the collaborative prediction loss are both constructed based on L1 loss. The total loss is equal to the backbone prediction loss multiplied by 0.4 plus the collaborative prediction loss multiplied by 0.6. The Adam algorithm is selected for optimization. The initial learning rate, two momentum parameters, weight decay coefficients used to suppress overfitting, batch size, and maximum number of training rounds are set to 0.01, 0.9, 0.999, 0.001, 32, and 200, respectively. An early stopping strategy is enabled. Training stops when the total loss on the validation set fails to decrease for 10 consecutive rounds. In the first phase (rounds 1-50), the auxiliary network is frozen, and only the backbone sequence is input into the backbone network for one-stage training. This continues until the backbone prediction loss on the validation set is below 0.05 for 5 consecutive rounds, at which point the first-stage training stops and the second-stage training begins. In the second phase (rounds 51-200), the backbone and auxiliary sequences are input simultaneously, and the parameters of the backbone and auxiliary networks are jointly updated to minimize the total loss. The training ends when early stopping is triggered, the maximum number of rounds is reached, or the total loss on the training set is below 0.01 and the total loss on the validation set is below 0.015. The test set is used for validation. The mean absolute error of the backbone prediction loss of the backbone network is required to be less than or equal to 0.03, and the mean absolute error of the co-prediction loss of the backbone network and the auxiliary network is required to be less than or equal to 0.02. If these criteria are not met, the learning rate is adjusted to half of the initial learning rate, and the second-stage training is repeated. Once the criteria are met, the optimal model parameters on the validation set are fixed and saved, and the two-stage training of the backbone and auxiliary networks is completed.

[0031] This invention discloses a workshop capacity prediction system, which sets up an edge processing module and a capacity prediction module. It effectively solves the capacity prediction problem under the conditions of transmission latency and network congestion. At the beginning of the cycle, the edge processing module integrates the material inventory quantity and operation data, and calculates the anomaly score by combining the transmission environment and the isolated forest model. The priority transmission strategy ensures the timely transmission of key data to the greatest extent. The capacity prediction module flexibly adjusts the prediction application of the backbone network and the auxiliary network according to the priority prediction strategy and the progressive strategy, and updates the network parameters in real time, which significantly improves the real-time performance and accuracy of capacity prediction.

[0032] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A workshop capacity prediction system, characterized in that, Includes an edge processing module and a capacity forecasting module; At the start of the cycle, the edge processing module calls up the material inventory quantity and operation data and integrates them into a data packet. It captures the packet transmission environment and calculates the anomaly score using the isolated forest model. It then uses a priority transmission strategy to split the data packet into priority data packets and supplementary data packets. Based on the anomaly score and the remaining transmission status of the previous cycle, it decides the transmission order and duration of the data packets to be transmitted. The data packets to be transmitted include priority data packets, supplementary data packets, and the remaining data packets from the previous cycle. After the capacity prediction module finishes receiving the priority data packets for the current cycle, it selects to wait directly until the end of the current cycle based on the priority prediction strategy, depending on whether the next received data packet is a supplementary data packet for the current cycle or a remaining data packet from the previous cycle. Alternatively, it selects to perform joint prediction using the priority data packets and some supplementary data packets transmitted in the previous cycle, and through the backbone network and auxiliary network. It then updates the parameters of the backbone network and auxiliary network based on the joint prediction results and the total capacity for the current cycle using a progressive strategy. At the end of the current cycle, based on the reception status of the supplementary data packets for the current cycle, it selects to generate the collaborative total capacity vector for the current cycle through joint prediction using the backbone network and auxiliary network, or to generate the backbone total capacity vector for the current cycle using the backbone network. The backbone network and auxiliary network are a temporal inference network and a convolutional network, respectively.

2. The workshop capacity prediction system as described in claim 1, characterized in that, The priority transmission strategy includes the following steps: Pack the material inventory quantity, capacity and operating rate of all production equipment in the data package for this cycle, and add corresponding tags at both ends to generate the priority data package for this cycle. Pack the temperature and power of all production equipment, and add corresponding tags at both ends to generate the supplementary data package for this cycle. The tags indicate the cycle and level of the data package as priority or supplementary. Based on the length of the priority data packet and the supplementary data packet, the transmission performance data, and the anomaly score, calculate the priority transmission duration and the supplementary transmission duration for this period. Add these to the delayed transmission duration of the previous period to obtain the total transmission duration for this period and determine whether it is less than or equal to the period duration. If the data packet length is less than or equal to the cycle length, the data packets to be transmitted are transmitted to the capacity prediction module in sequence according to the data packet priority. The data packet priority is: the priority data packet of this cycle is greater than the remaining data packet of the previous cycle, which is greater than the supplementary data packet of this cycle. If the transmission duration exceeds the cycle length, the difference between the total transmission duration of this cycle and the cycle length will be used as the delayed transmission duration for this cycle. Based on the data packet priority, the existing data packets to be transmitted will be transmitted to the capacity prediction module in sequence, and transmission will stop at the beginning of the next cycle.

3. The workshop capacity prediction system as described in claim 1, characterized in that, The priority prediction strategy includes the following steps: Upon receiving the priority data packets for the current cycle, it checks whether the next data packet is a remaining data packet from the previous cycle. If not, it waits until the end of the current cycle. If it is, after receiving the next data packet, it merges it with the supplementary data packets already transmitted in the previous cycle to form the supplementary data packets for the previous cycle, and then calls upon the previous cycle and earlier data packets. The priority data packets and supplementary data packets of each cycle are concatenated and input into the backbone network and auxiliary network respectively, and then jointly generated to produce the next cycle. Total collaborative capacity for each cycle; Extract the total collaborative capacity of this cycle and the total backbone capacity of this cycle predicted by the backbone network in the previous cycle. Extract the capacity of all production equipment from the priority data packets of this cycle and sum them to obtain the total capacity of this cycle. Calculate the backbone collaborative error and the actual collaborative error based on the progressive strategy and update the backbone network parameters and auxiliary network parameters respectively in combination with the network optimization algorithm. Wait until the end of this cycle. At the end of this cycle, determine whether all supplementary data packets for this cycle have been received. If not, call the functions of this cycle and previous cycles. Prioritized data packets for each cycle are concatenated and input into the backbone network to generate the data for the next cycle. Total main capacity of each cycle; If reception is complete, call the current cycle and previous cycles. Priority data packets and supplementary data packets for each cycle are concatenated and input into the backbone network and auxiliary network respectively, and joint prediction is generated after this cycle. Total collaborative capacity for each cycle.

4. The workshop capacity prediction system as described in claim 1, characterized in that, The edge processing module includes a transmission unit; The transmission unit captures packets in the transmission environment to obtain transmission performance data and inputs it into the isolated forest model to calculate anomaly scores. Based on the anomaly scores, it enables a priority transmission strategy to split data packets and transmit them to the capacity prediction module. The transmission performance data includes throughput, packet loss rate, and retransmission rate.

5. The workshop capacity prediction system as described in claim 4, characterized in that, The transmission unit places the transmission performance data at the root node of each isolated tree in the isolated forest model. Based on the feature values ​​of the transmission performance data in each feature dimension, it compares them with the node splitting values ​​of each isolated tree to decide whether to assign the data to the left or right child node of each isolated tree. This process continues until the transmission performance data is assigned to the leaf node of each isolated tree. Then, it calculates the path length of the transmission performance data in each isolated tree and the average path length. The ratio of the average path length to the correction coefficient is used as an exponent to calculate the power of 0.5 to obtain the anomaly score of the transmission performance data. The correction coefficient is directly related to the total number of trees.

6. The workshop capacity prediction system as described in claim 4 or 5, characterized in that, When constructing the Isolation Forest model, transmission performance data samples are collected as the original dataset. Transmission performance data samples are then randomly drawn with replacement from the original dataset using bootstrapping. Next, get Each training subset is used as the root node of each isolated tree. A feature and corresponding split value of one dimension are randomly selected from the transmission performance data samples. The root node is split according to the split value of the selected feature. The split nodes are then randomly selected again for splitting, until all unsplit nodes in each isolated tree contain only a single transmission performance data sample or all existing transmission performance data samples are identical. Once the isolated tree is trained, the unsplit nodes are considered leaf nodes, and the isolated forest model is completed.

7. The workshop capacity prediction system as described in claim 1 or 3, characterized in that, The backbone network uses a long short-term memory network to sequentially extract the hidden states corresponding to the backbone vector of each cycle from the backbone sequence, generate a hidden state sequence, and use a multi-head attention mechanism to focus on the temporal dependencies of the hidden state sequence. Then, it generates the total backbone capacity vector through linear modulation, batch normalization, and ReLU function mapping. The backbone vector includes the material inventory quantity and the capacity and operating rate of all production equipment.

8. The workshop capacity prediction system as described in claim 1 or 3, characterized in that, The auxiliary network uses convolution to extract features twice for each auxiliary vector in the auxiliary sequence. Then, it generates high-level features corresponding to each auxiliary vector through pointwise convolution and ReLU function mapping. Lightweight temporal convolution is used to model the temporal relationship between high-level features. Causal convolution kernels with different dilation rates are used to perform temporal convolution on high-level features to capture the periodic dependence information of high-level features over different period spans. Pointwise convolution is used to fuse and generate a temporal enhanced feature sequence, which is linearly modulated with the backbone total capacity vector output by the backbone network. The ReLU function is used to activate and generate a collaborative total capacity vector. The auxiliary vectors include the temperature and power of all production equipment.

9. The workshop capacity prediction system as described in claim 1, characterized in that, The edge processing module also includes a processing unit, which triggers the docking unit at the beginning of each cycle to obtain the material inventory quantity and operating data and package them into a data package.

10. The workshop capacity prediction system as described in claim 1, characterized in that, The edge processing module also includes a docking unit. When the docking unit is triggered, it retrieves material inventory quantities and operating data from the enterprise resource planning system and production management system through the API interface and feeds them back to the processing unit. The operating data includes the production capacity, operating rate, temperature and power of the production equipment.

Images