TWI815592B - Yield estimation apparatus and method - Google Patents

Abstract

A yield estimation apparatus and method. The apparatus stores target data and non-target data corresponding to a predicted target, and the non-target data includes a plurality of feature fields and a plurality of temporal feature fields. Based on the feature fields and the temporal feature fields contained in the non-target data, the apparatus trains a gate recursive unit encoder to generate a non-target feature extraction result. Based on the non-target feature extraction result and the target data, the apparatus trains a gate recursive unit decoder to generate a yield estimation result, and the target data corresponds to a historical sales volume of the predicted target.

Description

Production volume estimation device and method

The present invention relates to a production volume prediction device and method. Specifically, the present invention relates to a production volume prediction device and method that can improve the production volume prediction accuracy.

In recent years, technologies and applications related to big data have developed rapidly. Enterprise-side supply chains often estimate production volume data by building forecast models, so that the delivery time of finished products can be shortened to the delivery expected by customers. period process.

However, due to the existence of time-related sequences in actual data (such as product historical sales data), which are highly unstable and non-linear, companies must rely only on simple linear models (such as SVM, ARIMA) or Basic machine learning models (such as LSTM, CNN_LSTM) are usually difficult to capture the correlation and changes in data over time, and therefore cannot accurately predict the production output forecast results, often leading to prediction errors.

In view of this, how to provide a production volume prediction technology that can improve the accuracy of production volume prediction is an urgent goal for the industry.

One object of the present invention is to provide a production volume prediction device. The production volume prediction device includes a storage and a processor, and the processor is electrically connected to the storage. The storage is used to store target data corresponding to a prediction target and non-target data, wherein the non-target data includes a plurality of feature fields and a plurality of temporal feature fields. The processor trains a gate recursive unit encoder to generate a non-target feature extraction result based on the feature fields and the temporal feature fields included in the non-target data. The processor trains a gated recursive unit decoder to generate a production volume estimation result based on the non-target feature extraction result and the target data, wherein the target data corresponds to one of the historical sales volumes of the forecast target.

Another object of the present invention is to provide a production volume prediction method for an electronic device. The production volume prediction method is executed by the electronic device and includes the following steps: based on non-target data corresponding to a prediction target. A plurality of feature fields and a feature convolution are used to train a gated recursive unit encoder to generate an extraction feature value in a non-target feature extraction result; based on the plurality of temporal feature fields contained in the non-target data and a temporal convolution, training the gated recursive unit encoder to generate an extracted temporal feature value in the non-target feature extraction result; and based on the non-target feature extraction result and a target data, training a gated recursive unit decoder The machine is used to generate a production forecast result, in which the target data corresponds to one of the historical sales volumes of the forecast target.

The production volume prediction technology (including at least a device and a method) provided by the present invention generates non-target feature extraction results for the feature fields and the temporal feature fields included in the non-target data corresponding to the prediction target, respectively. And generate target feature extraction results based on the target data corresponding to the predicted target. By simultaneously considering the non-target feature extraction results and the target feature extraction results, the throughput estimation results are generated. Since the production volume prediction technology provided by the present invention simultaneously considers the correlation and changes in time of the characteristics of target data and non-target data, it can improve the accuracy of production volume prediction and solve the problem of inaccuracy in the conventional technology.

The following describes the detailed technology and implementation of the present invention in conjunction with the drawings, so that those with ordinary knowledge in the technical field to which the present invention belongs can understand the technical features of the claimed invention.

The following will explain a production volume prediction device and method provided by the present invention through implementation examples. However, these embodiments are not intended to limit the invention to be implemented in any environment, application or manner as described in these embodiments. Therefore, the description of the embodiments is only for the purpose of explaining the present invention and is not used to limit the scope of the present invention. It should be understood that in the following embodiments and drawings, elements not directly related to the present invention have been omitted and not shown, and the size of each element and the size ratio between elements are only for illustration and are not intended to limit the present invention. range.

The first embodiment of the present invention is a production volume prediction device 1, the schematic structural diagram of which is depicted in Figure 1. The production volume prediction device 1 includes a storage 11 and a processor 13 . The processor 13 is electrically connected to the storage 11 . The storage 11 can be a memory, a Universal Serial Bus (USB) disk, a hard disk, an optical disk, a pen drive, or any other storage that is known to those with ordinary skill in the technical field of the present invention and has the same function. media or circuit. The processor 13 may be various processing units, a central processing unit (Central Processing Unit; CPU), a microprocessor, or other computing devices known to those with ordinary skill in the technical field to which this invention belongs.

In some embodiments, the production volume estimating device 1 further includes a transceiver interface to receive data. Specifically, the transceiver interface is an interface that can receive and transmit data or other interfaces that can receive and transmit data known to those with ordinary knowledge in the technical field to which the present invention belongs. The transceiver interface can be through, for example: external devices, external web pages, external Applications and other sources receive data.

In this implementation, as shown in Figure 1, the storage 11 is used to store target data TD and non-target data NTD corresponding to the prediction target, where the non-target data NTD includes a plurality of feature fields and a plurality of Temporal feature fields.

For example, when the prediction target of the production volume estimating device 1 is a beverage A, the target data TD stored in the storage 11 is the product sales volume of beverage A in the past period of time (for example: the sales data of each day in the past week), Non-target data NTD are other parameters that may affect the sales of beverage A, such as: weather conditions, temperature, body temperature, rainfall, rainfall probability, relative humidity, comfort, etc.

For ease of understanding, take a practical example as an example, please refer to the non-target data NTD diagram in Figure 2. As shown in Figure 2, the characteristic fields in the non-target data NTD include fields for date, time, weather conditions, temperature, perceived temperature, rainfall probability, relative humidity, comfort, etc. The temporal characteristic fields in non-target data NTD include multiple fields corresponding to time series, for example: various time intervals for Saturday, July 17, Sunday, July 18, etc.

The operation of the first embodiment of the present invention will be briefly described first. The present invention mainly includes two stages of operation, namely the encoder operation stage and the decoder operation stage. In this implementation, the encoder operation stage mainly extracts features from the information in the non-target data NTD, and inputs the non-target feature extraction results to the decoder. The decoder operation stage further adds the feature extraction information of the target data TD. , the production volume estimation result is generated based on the non-target feature extraction results and the target feature extraction results.

It should be noted that in this implementation, the present invention will use an encoder (for example: GRU Encoder) and a decoder (for example: GRU Decoder) to find out the temporal correlation and change between the non-target data NTD and the target data TD. Detailed characteristics and use these characteristics to predict the production volume estimation results at future time points.

In some embodiments, the processor 13 can further combine the features extracted by the encoder and the decoder as the input layer of a convolutional neural network (Convolutional Neural Network; CNN), through convolution layers (Convolution Layers) and pooling. After the features are automatically extracted from the Pooling Layers, they are then sent to the Fully Connected layers (FC), and finally the final production volume estimate is obtained in the output layer.

In this embodiment, the processor 13 first performs feature extraction on the non-target data NTD using feature convolution (Feature Convolution) and temporal convolution (Temporal Convolution). It should be understood that the characteristic fields refer to different types of information in the non-target data NTD, and the temporal characteristic fields refer to various time points in the non-target data NTD. The processor 13 performs feature extraction on the feature fields and temporal feature fields of the non-target data NTD, so that the model can see each type of the non-target data NTD (for example: weather conditions, temperature, body temperature, rainfall probability, relative humidity, comfort, etc.) over a period of time, and between each species at the same point in time.

It should be noted that due to the traditional method of directly convolving non-target data, the model may ignore certain details by considering all the data at once. In comparison, the operation of the present invention can see more detailed characteristics, and can pay attention to characteristics that differ for types and time.

For ease of understanding, please refer to the operational flow diagram in Figure 4 in the following paragraphs. It should be understood that Figure 4 is only used to illustrate some embodiments of the present invention and to explain the technical features of the present invention, but is not used to limit the scope and scope of the present invention.

Specifically, in this implementation, the processor 13 trains the gated recurrent unit (GRU) encoder EN to generate non-target data based on the feature fields and temporal feature fields included in the non-target data NTD. Feature extraction results. Then, the processor 13 trains the gate recursive unit decoder DE to generate a production volume estimation result based on the non-target feature extraction result and the target data TD, where the target data TD corresponds to the historical sales volume of the predicted target.

In some embodiments, the gated recurrent unit encoder EN includes a feature convolution FC and a temporal convolution TC.

Specifically, the processor 13 performs feature convolution FC on the feature fields in the non-target data NTD to generate the first extracted feature values corresponding to each of the feature fields. The processor 13 performs temporal convolution TC on the temporal feature fields in the non-target data NTD to generate the first extracted temporal feature value corresponding to each of the temporal feature fields.

For ease of understanding, please refer to Figure 3A. The processor 13 can divide the non-target data NTD into a one-dimensional vector based on each feature field F (for example: feature F1, feature F2, feature F3), and then multiply them by By applying the weight value learned by the model (for example: convolution kernel K1), you can get the features of each feature field (for example: channel CH1, channel CH2 and channel CH3). Finally, by merging the extracted features, you can get Non-target data features extracted for feature fields.

In addition, please refer to Figure 3B, the processor 13 can divide the non-target data NTD into a one-dimensional vector (for example: feature F1, feature F2, feature F3) based on each temporal feature field TF, and then multiply them By applying the weight value learned by the model (for example: convolution kernel K1), the features of each temporal feature field (for example: channel CH1, channel CH2 and channel CH3) can be obtained. Finally, the extracted features are merged to obtain Non-target data features extracted for temporal feature fields can be obtained.

It should be noted that in this example, the convolution operation is to highlight the enhanced feature relationship. The 1x1 filter (i.e., convolution kernel K1) is selected because it is necessary to emphasize and focus on each block without being biased. Therefore, Filters are also called weights of convolutional layers.

In some embodiments, the processor 13 can use the query, key, and value generated by the convolution operation to calculate the correlation between each feature, and remove noise on the extracted features. ization to improve the model’s focus on highly relevant features.

Specifically, the processor 13 performs a first self-attention operation SA (Self-Attention) on the first extraction feature values to generate a plurality of second extraction feature values. In addition, the processor 13 performs a second self-attention operation SA on the first extraction temporal characteristic values to generate a plurality of second extraction temporal characteristic values.

In some embodiments, as shown in Figure 4, the feature convolution operation FCO performed by the processor 13 includes the feature convolution FC and the self-attention operation SA, and the temporal convolution operation TCO includes the temporal convolution TC and the self-attention operation SA. .

For example, the processor 13 can first multiply the query (query) and the key (key), scale it first to avoid the multiplied value being too large, and then use the softmax function to calculate the correlation between the two. The value is usually Between 0 and 1. Finally, by multiplying the calculated value by the original value, you can reduce and weaken the values with small correlations between features, and increase and strengthen the values with large correlations.

In some embodiments, in order to achieve a higher extraction effect, the processor 13 deepens the extraction method multiple times and performs a plurality of iterations (for example: setting a default iteration of N times, where N is a positive integer). Specifically, the processor 13 performs a plurality of feature convolutions FC and a plurality of first self-attention operations SA to generate a plurality of second extracted feature values. In addition, the processor 13 performs a plurality of temporal convolutions TC and a plurality of second self-attention operations SA to generate a plurality of second extracted temporal feature values.

In some implementations, as shown in Figure 4, in order to avoid the gradient explosion problem caused by too many model layers, a residual method can be added to compare the features extracted by multiple layers with the original non-target data. Add to enhance original features.

In some implementations, in order to prevent the added value from being too large, batch normalization is added. After the value is normalized, it is then thrown into the gate recursive unit GRU to allow the model to learn changes in continuous time. Characteristics. Specifically, the processor 13 performs a first batch normalization operation BN and a first gate return unit GRU operation on the second extraction feature values to generate a plurality of third extraction feature values. In addition, the processor 13 performs a second batch normalization operation BN and a second gate return unit GRU operation for the second extraction temporal characteristic values to generate a plurality of third extraction temporal characteristic values.

In some embodiments, the processor 13 generates the non-target feature extraction result based on the third extraction feature values and the third extraction temporal feature values.

Subsequently, the processor 13 adds the two branches and passes them to the decoder DE for subsequent training. In some implementations, in order to extract features more accurately, the decoder DE can further process the part of the non-target data NTD, and throw the non-target data NTD into the same model structure as the encoder EN. As shown in Figure 4, the processor 13 can combine the two data and extract features again in the extraction operation EX.

Specifically, the non-target data NTD is thrown into the extraction operation EX with the same model architecture as the encoder EN, and through pooling (Pooling) on the way, the maximum pooling method is used to obtain the maximum value (that is, to find the most relevant part ), without losing important information, and then flatten the data through fully connected layers (Fully Connected layers), and convert all feature matrices into vectors, which can be one or more days of data. Subsequently, the generated data is merged with the target data TD, and finally the gate return unit GRU is entered to observe the short-term and long-term feature correlation again. It should be understood that the output of the decoder DE can be one or more pieces of data.

In some embodiments, for the target data TD, the processor 13 uses a gate recursion to return the target data TD n times before the unit GRU training (ie, time points T1, T2, ..., Tn). Then, at the Tn+1 time point, the features extracted from the non-target data NTD are added and merged based on the relationship of the time series data, so that the model can learn the features between the non-target data NTD and the target data TD. Improve prediction accuracy.

Specifically, in some embodiments, the gate return unit decoder includes a plurality of third gate return units GRU, and the processor 13 inputs the historical sales volume corresponding to a plurality of time points into these respectively. The third gate recursively unit GRU to produce target feature extraction results.

Specifically, in some embodiments, the gate recursive unit decoder further includes a fourth gate recursive unit GRU, and the processor 13 inputs the non-target feature extraction results and the target feature extraction results to the fourth gate recursive unit. Back to the unit GRU to generate production volume estimates.

It should be noted that in this disclosure, the non-target data NTD will enter the encoder EN and the decoder DE to extract its features, including the correlation between various features and the correlation between time series, for example: between seasons and off-peak seasons. There is some correlation, there may be a correlation between the demand in the next three months and the order cycle three months later, and there may be a correlation between the current inventory and the purchase cost of each item, etc. In addition, the target data TD will be returned to the unit GRU through multiple layers of gates to capture feature changes over a long and short period of time. For example, the product sales volume in the past period may be related to the product sales volume of the day, last summer may be related to this summer, etc. wait.

In some implementations, the output data of the decoder DE can be further input into the model architecture of an additional convolutional neural network (CNN) to enhance feature relationships through convolution operations, and use global average pooling through the pooling layer to optimize the results. More evenly, the data is flattened through a fully connected layer, converting all feature matrices into vectors (for example: data for one or more days). Finally, the neural network is fed into the fully connected layer for classification, and the product sales volume is predicted for one or more days in the future. It should be understood that convolutional neural networks learn the weights (ie, filters) of the convolutional layer, which are mainly used to highlight features. The purpose of pooling is to compress to extract the strongest features. The fully connected layer is the hidden layer and output layer of the multi-layer perceptron, which can be used for classification.

It can be seen from the above description that the production volume prediction device 1 provided by the present invention generates non-target feature extraction results for the feature fields and the temporal feature fields included in the non-target data corresponding to the prediction target, and Target feature extraction results are generated based on target data corresponding to the predicted target. By simultaneously considering the non-target feature extraction results and the target feature extraction results, the throughput estimation results are generated. Since the production volume prediction technology provided by the present invention simultaneously considers the correlation and changes in time of the characteristics of target data and non-target data, it can improve the accuracy of production volume prediction and solve the problem of inaccuracy in the conventional technology.

The second embodiment of the present invention is a production output estimation method, the flow chart of which is depicted in Figure 5 . The production volume prediction method 500 is suitable for an electronic device, such as the production volume prediction device 1 described in the first embodiment. The production volume prediction method 500 generates production volume prediction results through steps S501 to S505.

In step S501, the electronic device trains a gated recursive unit encoder based on a plurality of feature fields and feature convolutions included in the non-target data corresponding to the prediction target to generate extraction feature values in the non-target feature extraction results. Next, in step S503, the electronic device trains the gated recursive unit encoder based on the plurality of temporal feature fields and temporal convolutions included in the non-target data to generate the extracted temporal features in the non-target feature extraction results. value.

Finally, in step S505, the electronic device trains the gate return unit decoder to generate a production volume estimation result based on the non-target feature extraction results and target data, where the target data corresponds to the historical sales volume of the predicted target.

In some embodiments, the gated recurrent unit encoder includes feature convolution and temporal convolution.

In some embodiments, the production volume estimation method 500 further includes the following steps: performing feature convolution on feature fields in the non-target data to generate first extracted feature values corresponding to each feature field; and For the temporal feature fields in the non-target data, perform temporal convolution to generate the first extracted temporal feature value corresponding to each temporal feature field.

In some embodiments, the throughput estimation method 500 further includes the following steps: for the first extraction characteristic value, performing a first self-attention operation to generate a plurality of second extraction characteristic values; and for the first extraction temporal characteristic value, and perform a second self-attention operation to generate a plurality of second extracted temporal feature values.

In some embodiments, the throughput estimation method 500 further includes the following steps: performing the feature convolution and the first self-attention operation a plurality of times to generate a plurality of second extracted feature values; and performing the time convolution a plurality of times. The product and a plurality of second self-attention operations are performed to generate a plurality of second extraction temporal characteristic values.

In some embodiments, the production volume estimation method 500 further includes the following steps: for the second extraction characteristic value, perform a first batch standardization operation and a first gate return unit operation to generate a plurality of third extraction characteristic values; And for the second extraction temporal characteristic values, a second batch normalization operation and a second gate return unit operation are performed to generate a plurality of third extraction temporal characteristic values.

In some embodiments, the throughput estimation method 500 further includes the following steps: generating a non-target feature extraction result based on the third extraction feature value and the third extraction temporal feature value.

In some embodiments, the gate return unit decoder includes a plurality of third gate return units, and the production volume estimation method 500 further includes the following steps: inputting historical sales volumes corresponding to a plurality of time points into The third gate recursively unites to produce target feature extraction results.

In some embodiments, the gate return unit decoder further includes a fourth gate return unit, and the throughput estimation method 500 further includes the following steps: inputting the non-target feature extraction results and the target feature extraction results into Four gates are returned to the unit to produce throughput estimates.

In addition to the above steps, the second embodiment can also perform all operations and steps of the production volume prediction device 1 described in the first embodiment, has the same functions, and achieves the same technical effects. Those with ordinary skill in the technical field of the present invention can directly understand how the second embodiment performs these operations and steps based on the above-mentioned first embodiment, has the same functions, and achieves the same technical effects, so no further description is given.

It should be noted that in the patent specification and patent application scope of this invention, certain terms (including: extraction characteristic value, extraction time characteristic value, self-attention operation, batch standardization operation, gate return unit operation, etc.) are preceded by "First", "Second", "Third" or "Fourth" are used only to distinguish different terms. For example: "First" and "Second" in the first self-attention operation and the second self-attention operation are only used to indicate the self-attention operations used in different operations.

To sum up, the production volume prediction technology (including at least a device and a method) provided by the present invention generates non-target feature extraction results for the feature fields and temporal feature fields contained in the non-target data corresponding to the prediction target. , and generate target feature extraction results based on the target data corresponding to the predicted target. By simultaneously considering the non-target feature extraction results and the target feature extraction results, the throughput estimation results are generated. Since the production volume prediction technology provided by the present invention simultaneously considers the correlation and changes in time of the characteristics of target data and non-target data, it can improve the accuracy of production volume prediction and solve the problem of inaccuracy in the conventional technology.

The above embodiments are only used to illustrate some implementation aspects of the present invention and to illustrate the technical features of the present invention, but are not intended to limit the scope and scope of the present invention. Any changes or equivalence arrangements that can be easily accomplished by those with ordinary skill in the technical field to which the present invention belongs fall within the scope claimed by the present invention, and the scope of rights protection of the present invention shall be subject to the scope of the patent application.

1: Production volume estimation device 11:Storage 13: Processor TD: target data NTD: non-target data F: Feature field TF: Temporal feature field F1, F2, F3: Features K1: convolution kernel CH1, CH2, CH3: Channel EN: Encoder DN: decoder FCO: Feature convolution operation FC: Feature convolution TCO: temporal convolution operation TC: temporal convolution SA: self-attention operation BN: batch standardized operation GRU: Gate return unit EX: Feature extraction operation T1, T2,..., Tn: time point 500:Production volume estimation method S501, S503, S505: steps

Figure 1 is a schematic diagram depicting the structure of the production volume estimating device according to the first embodiment; Figure 2 is a schematic diagram depicting non-target data in the first embodiment; Figure 3A is a schematic diagram depicting feature convolution of the first embodiment; Figure 3B is a schematic diagram depicting the temporal convolution of the first embodiment; Figure 4 is a schematic operational flow diagram depicting certain embodiments; and FIG. 5 is a partial flowchart depicting the production volume estimation method of the second embodiment.

Domestic storage information (please note in order of storage institution, date and number) without Overseas storage information (please note in order of storage country, institution, date, and number) without

500:Production volume estimation method

S501, S503, S505: steps

Claims (10)

A production volume forecasting device, including: A storage device for storing target data corresponding to a prediction target and non-target data, wherein the non-target data includes a plurality of feature fields and a plurality of temporal feature fields; and A processor electrically connected to the storage and used to perform the following steps: Based on the feature fields and the temporal feature fields included in the non-target data, train a gated recursive unit encoder to generate a non-target feature extraction result; and Based on the non-target feature extraction results and the target data, a gated recursive unit decoder is trained to generate a production volume estimation result, wherein the target data corresponds to one of the historical sales volumes of the forecast target. The production volume prediction device of claim 1, wherein the gate recursive unit encoder includes a feature convolution and a temporal convolution. The production volume estimation device as described in claim 2, wherein the processor further performs the following operations: Perform the feature convolution on the feature fields in the non-target data to generate a first extracted feature value corresponding to each of the feature fields; and The temporal convolution is performed on the temporal feature fields in the non-target data to generate a first extracted temporal feature value corresponding to each of the temporal feature fields. The production volume estimation device as described in claim 3, wherein the processor further performs the following operations: For the first extraction feature values, perform a first self-attention operation to generate a plurality of second extraction feature values; and For the first extraction temporal characteristic values, a second self-attention operation is performed to generate a plurality of second extraction temporal characteristic values. The production volume estimation device as described in claim 3, wherein the processor further performs the following operations: Perform a plurality of times of feature convolution and a plurality of first self-attention operations to generate a plurality of second extracted feature values; and The temporal convolution and the second self-attention operation are performed a plurality of times to generate a plurality of second extracted temporal feature values. The production volume estimation device as described in claim 4, wherein the processor further performs the following operations: For the second extraction feature values, perform a first batch normalization operation and a first gate return unit operation to generate a plurality of third extraction feature values; and For the second extraction temporal characteristic values, a second batch normalization operation and a second gate recursive unit operation are performed to generate a plurality of third extraction temporal characteristic values. The production volume estimation device as described in claim 6, wherein the processor further performs the following operations: The non-target feature extraction result is generated based on the third extraction feature values and the third extraction temporal feature values. The production volume estimation device of claim 1, wherein the gate return unit decoder includes a plurality of third gate return units, and the processor further performs the following operations: The historical sales volumes corresponding to multiple time points are respectively input into the third gate return units to generate a target feature extraction result. The production volume estimation device of claim 8, wherein the gate return unit decoder further includes a fourth gate return unit, and the processor further performs the following operations: The non-target feature extraction result and the target feature extraction result are input into a fourth gate return unit to generate the production volume estimation result. A production volume prediction method is used for an electronic device. The production volume prediction method is executed by the electronic device and includes the following steps: Based on a plurality of feature fields and a feature convolution included in the non-target data corresponding to a prediction target, train a gated recursive unit encoder to generate an extracted feature value in a non-target feature extraction result; Based on a plurality of temporal feature fields and a temporal convolution included in the non-target data, train the gated recursive unit encoder to generate an extracted temporal feature value in the non-target feature extraction result; and Based on the non-target feature extraction results and a target data, a gated recursive unit decoder is trained to generate a production volume prediction result, wherein the target data corresponds to a historical sales volume of the prediction target.

Images