A method and system for automatic batching control of PVC pipe production

Abstract

The present invention provides an automatic batching control method and system for PVC pipe production, which relates to the field of production control technology. The method includes: obtaining the amount of raw materials added, opening a feed channel, a vibrator, and an air pump; weighing the weight of the materials in the hopper; training a feed amount prediction model for the final state of the previous control cycle, obtaining a feed amount prediction model for the intermediate state of the current control cycle, and predicting a first power of the vibrator and a second power of the air pump; closing the feed channel, and training a feed amount prediction model for the intermediate state of the current control cycle, obtaining a feed amount prediction model for the final state of the current control cycle. According to the present invention, when controlling the feed amount, the influence of the weight of the material in the hopper on the feed speed can be considered, thereby setting appropriate power for the vibrator and air pump, and the feed amount prediction model can be continuously updated to improve the accuracy of the power of the vibrator and air pump, thereby improving the control accuracy of the feed amount of the raw materials.

Description

Technical Field

[0001] The present invention relates to the technical field of production control, and in particular to an automatic batching control method and system for PVC pipe production. Background Art

[0002] In related technologies, raw materials for PVC pipe production are mixed using devices such as hoppers, and the raw materials are drawn from the hoppers by equipment such as air pumps. However, this technology has difficulties in precisely controlling the power of these devices, as well as the feed rate, making it difficult to improve the quality of PVC pipes.

[0003] The information disclosed in the background technology section of this application is only intended to deepen the understanding of the general background technology of this application, and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to those skilled in the art. Summary of the Invention

[0004] The present invention provides an automatic batching control method and system for PVC pipe production, which can solve the technical problem in related technologies that it is difficult to accurately control the power of an air pump and thus difficult to accurately control the feed amount.

[0005] According to a first aspect of the present invention, a method for automatic batching control of PVC pipe production is provided, comprising: obtaining, at the start of a current control cycle, the addition amounts of a plurality of raw materials in the current control cycle; opening the feed channels corresponding to the plurality of raw materials, and opening the vibrators and air pumps of each feed channel; weighing the weight of the materials in the hopper corresponding to each raw material at a first preset number of moments before the current control cycle; training a feed amount prediction model of the final state of the previous control cycle based on the rated power of the vibrators and air pumps and the weight of the materials, and obtaining a feed amount prediction model of the intermediate state of the current control cycle; and training a feed amount prediction model of the intermediate state of the current control cycle based on the feed amount prediction model of the intermediate state of the current control cycle after the first preset number of moments in the current control cycle. The weight of the material in the hopper at each moment, and the feed quantity prediction model of the intermediate state of the current control cycle are used to determine the first power of the vibrator and the second power of the air pump at each moment after the first preset number of moments in the current control cycle; when the weight reduction of the raw material in the hopper reaches the added amount during the current control cycle, or at the end of the current control cycle, the feed channel is closed; at the end of the current control cycle, the feed quantity prediction model of the intermediate state of the current control cycle is trained according to the weight of the material in the hopper at the first preset number of moments and each moment thereafter, as well as the first power and the second power, to obtain the feed quantity prediction model of the final state of the current control cycle.

[0006] According to the present invention, the feed amount prediction model of the final state of the previous control cycle is trained according to the rated power and material weight of the vibrator and the air pump to obtain the feed amount prediction model of the intermediate state of the current control cycle, including: the training input vector of the i-th moment composed of the rated power of the vibrator, the rated power of the air pump, and the material weight at the i-th moment, and the training feature vector of the i-1-th moment are input into the i-th feature extraction level of the feed amount prediction model of the final state of the previous control cycle to obtain the training feature vector of the i-th moment, wherein i is a positive integer less than a first preset number, when i=1, the training feature vector of the i-1-th moment is obtained. The training feature vector at the i-th moment is a zero vector; the training feature vector at the i-th moment is input into the first fully connected layer and the first activation layer of the feed quantity prediction model of the final state of the previous control cycle for processing to obtain the predicted feed quantity between the i-th moment and the i+1-th moment; according to the predicted feed quantity between the i-th moment and the i+1-th moment, the material weight at the i-th moment and the material weight at the i+1-th moment, the first loss function of the feed quantity prediction model is determined; according to the first loss function, the feed quantity prediction model of the final state of the previous control cycle is trained to obtain the feed quantity prediction model of the intermediate state of the current control cycle.

[0007] According to the present invention, the first loss function of the feed amount prediction model is determined based on the predicted feed amount between the i-th moment and the i+1-th moment, the material weight at the i-th moment, and the material weight at the i+1-th moment, including: according to the formula , determine the first loss function of the feed rate prediction model ,in, is the predicted feed quantity between the i-th moment and the i+1-th moment, is the material weight at the i-th moment, is the material weight at the i+1th moment, is the first preset number, i and All are positive integers.

[0008] According to the present invention, according to the weight of the material in the hopper at each moment after the first preset number of moments of the current control cycle and the feed amount prediction model of the intermediate state of the current control cycle, the first power of the vibrator and the second power of the air pump at each moment after the first preset number of moments of the current control cycle are determined, including: The weight of the material in the hopper at each moment, the rated power of the vibrator, and the rated power of the air pump constitute the weight of the material in the hopper at each moment. The first state vector at the moment is input into each feature extraction level of the feed quantity prediction model of the intermediate state of the current control cycle to obtain the first The first eigenvector at the moment; get the The number of remaining time periods between the moment and the end of the current control cycle; Get the The weight of the added material before the moment, and obtain the first difference between the weight of the added material and the amount of raw material added; the remaining time period quantity and the first difference form the first The second state vector at the moment is processed by the second fully connected layer to obtain the second state vector The second eigenvector at the moment; The second eigenvector at the moment and the The first eigenvectors of the moments are spliced ​​together to obtain the The third eigenvector of the moment; The third eigenvector at the moment is input into the third fully connected layer and the second activation layer to obtain the The first power of the vibrator and the second power of the air pump at the moment; the first power of the vibrator and the second power of the air pump at the moment The weight of the material in the hopper at the moment, the weight of the vibrator at the moment 1+j to the moment The rated power of the vibrator at the moment From moment to The first power at the moment, the air pump at the 1+j moment to the The rated power of the air pump at the moment From moment to The second power at the moment constitutes the power from the 1+jth moment to the The first state vector at the moment is input into each feature extraction level of the feed quantity prediction model of the intermediate state of the current control cycle to obtain the first The first eigenvector at the moment, j is an integer greater than or equal to 0, and , N is the sequence number of the end time of the current control cycle, N is a positive integer; get the The number of remaining time periods between the moment and the end of the current control cycle; Get the The weight of the added material before the moment, and obtain the first difference between the weight of the added material and the amount of raw material added; the remaining time period quantity and the first difference form the first The second state vector at the moment is processed by the second fully connected layer to obtain the second state vector The second eigenvector at the moment; The second eigenvector at the moment and the The first eigenvectors of the moments are spliced ​​together to obtain the The third eigenvector of the moment; The third eigenvector at the moment is input into the third fully connected layer and the second activation layer to obtain the The first power of the vibrator and the second power of the air pump at a certain moment.

[0009] According to the present invention, the method further comprises: The first feature vector at the moment is input into the first fully connected layer and the first activation layer for processing, and the first feature vector is obtained. From moment to The predicted feed volume between the moments; The weight of the material in the hopper at the first moment is From moment to The predicted feed amount between the moments is obtained The predicted material weight within a moment; The predicted material weight, the first power of the vibrator and the second power of the air pump at the moment are obtained. The first predicted feature vector at the moment; From moment to The predicted feed amount between moments is obtained The predicted weight of the added material before the moment, and obtain the first predicted difference between the predicted weight of the added material and the amount of raw material added; The number of remaining time periods between the moment and the end of the control cycle and the first prediction difference constitute the first The second predicted state vector at the moment is processed by the second fully connected layer to obtain the second state vector The second predicted feature vector at the moment The first predicted feature vector at the moment The second prediction feature vectors of the moment are spliced ​​to obtain the The third predicted feature vector at the moment; The third predicted feature vector at the moment is input into the third fully connected layer and the second activation layer to obtain the The first predicted power of the vibrator and the second predicted power of the air pump at a moment; iteratively execute the steps of obtaining the first predicted power and the second predicted power to obtain the first predicted power and the second predicted power. The first predicted power of the vibrator and the second predicted power of the air pump from the first moment to the Nth moment, as well as the predicted feed amount between adjacent moments; The predicted feed amount between each adjacent moment in the time period from moment 1 to moment N is used to determine the The second loss function corresponding to the multiple moments is used to train the second fully connected layer and the third fully connected layer according to the second loss functions corresponding to the multiple moments.

[0010] According to the present invention, according to The predicted feed amount between each adjacent moment in the time period from moment 1 to moment N is used to determine the The second loss function corresponding to the moment includes: According to the formula , determine the The second loss function corresponding to the moment ,in, is the predicted feed quantity between the kth moment and the k+1th moment, , is the feed amount between the sth moment and the s+1th moment, , is the amount of raw material added, and s and k are both positive integers.

[0011] According to the present invention, at the end of the current control cycle, the feed amount prediction model of the intermediate state of the current control cycle is trained based on the weight of the material in the hopper at each moment after the first preset number of moments, as well as the first power and the second power, to obtain the feed amount prediction model of the final state of the current control cycle, including: determining the predicted feed amount between each moment after the first preset number of moments based on the weight of the material in the hopper at each moment after the first preset number of moments, the first power and the second power, and the feed amount prediction model of the intermediate state of the current control cycle; according to the formula , obtain the third loss function of the feed quantity prediction model in the intermediate state of the current control cycle ,in, is the predicted feed quantity between the hth moment and the h+1th moment, is the material weight at the hth moment, is the material weight at the h+1th moment, , N is the sequence number of the end time of the current control cycle, is the first preset number, h is a positive integer; according to the third loss function, the feed quantity prediction model of the intermediate state of the current control cycle is trained to obtain the feed quantity prediction model of the final state of the current control cycle.

[0012] According to a second aspect of the present invention, there is provided an automatic batching control system for PVC pipe production, comprising: an addition amount module for obtaining the addition amount of a plurality of raw materials in the current control cycle at the start moment of the current control cycle; an opening module for opening the feed channels corresponding to the plurality of raw materials, and opening the vibrators and air pumps of the respective feed channels; a weighing module for weighing the weight of the materials in the hopper corresponding to the respective raw materials at the first preset number of moments before the current control cycle; a first training module for training the feed amount prediction model of the final state of the previous control cycle according to the rated power of the vibrator and the air pump and the weight of the materials, and obtaining the feed amount prediction model of the intermediate state of the current control cycle; a determination module for training the feed amount prediction model of the final state of the previous control cycle according to the rated power of the vibrator and the air pump and the weight of the materials, and obtaining the feed amount prediction model of the intermediate state of the current control cycle; and a determination module for determining the feed amount prediction model of the intermediate state of the current control cycle according to the feed amount prediction model of the intermediate state of the current control cycle. The weight of the material in the hopper at each moment after the first preset number of moments, and the feed quantity prediction model of the intermediate state of the current control cycle are used to determine the first power of the vibrator and the second power of the air pump at each moment after the first preset number of moments in the current control cycle; a closing module is used to close the feed channel when the weight reduction of the raw material in the hopper reaches the said addition amount during the current control cycle, or at the end of the current control cycle; a second training module is used to train the feed quantity prediction model of the intermediate state of the current control cycle at the end of the current control cycle based on the weight of the material in the hopper at the first preset number of moments and each moment thereafter, as well as the first power and the second power, to obtain the feed quantity prediction model of the final state of the current control cycle.

[0013] Technical effect: According to the present invention, when controlling the feed rate, the influence of the weight of the material in the hopper on the feed rate can be considered, thereby setting appropriate power for the vibrator and the air pump. The feed rate prediction model can also be continuously updated based on data such as power and material weight obtained during the control cycle, thereby improving the accuracy of the power of the vibrator and the air pump, and improving the accuracy of the feed rate prediction model, so as to improve the control accuracy of the feed rate of raw materials and improve the quality of PVC pipes. When training the feed rate prediction model of the final state of the previous control cycle, the data of multiple past moments can be fused to accurately predict the predicted feed rate between the i-th moment and the i+1-th moment, and when setting the first loss function, the weight is set according to the number of moments, thereby accurately reflecting the influence of the accuracy of the predicted data at that moment on the accuracy of subsequent predictions, improving training accuracy and training efficiency, and improving the accuracy of the feed rate prediction model. When training the second and third fully connected layers, after each moment, the first predicted power, second predicted power, and predicted feed amount at subsequent moments can be solved iteratively, thereby determining the predicted total feed amount within the control cycle using the predicted feed amount. Furthermore, by setting a second loss function, the error between the predicted total feed amount and the amount of raw material added can be reduced to train the second and third fully connected layers, thereby continuously improving the accuracy of the second and third fully connected layers, as well as the accuracy of the predictions of the first power of the vibrator and the second power of the air pump, and reducing the error between the predicted total feed amount and the amount of raw material added. When training the feed amount prediction model for the intermediate state of the current control cycle, the feed amount prediction model is trained not only using data from the first preset number of moments, but also, after the control cycle ends, further training can be continued using data from multiple moments after the first preset number of moments. This maximizes the utilization of measured and predicted data within the control cycle, improves the training efficiency of the feed amount prediction model, improves prediction accuracy, and provides an accurate data basis for predicting the first power of the vibrator and the second power of the air pump.

[0014] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and not limiting of the present invention. Other features and aspects of the present invention will become more apparent from the following detailed description of exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can derive other embodiments based on these drawings without inventive efforts.

[0016] Figure 1A flow chart of an automatic batching control method for PVC pipe production according to an embodiment of the present invention is exemplarily shown;

[0017] Figure 2 A schematic diagram exemplarily shows a feed rate prediction model according to an embodiment of the present invention;

[0018] Figure 3 The automatic batching control system for PVC pipe production according to an embodiment of the present invention is exemplarily shown. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0020] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0021] Figure 1 A flowchart of an automatic batching control method for PVC pipe production according to an embodiment of the present invention is exemplarily shown, the method comprising: step S1, at the start moment of a current control cycle, obtaining the addition amount of multiple raw materials in the current control cycle; step S2, opening the feed channels corresponding to the multiple raw materials, and opening the vibrators and air pumps of each feed channel; step S3, at the first preset number of moments before the current control cycle, weighing the weight of the material in the hopper corresponding to each raw material; step S4, training the feed amount prediction model of the final state of the previous control cycle according to the rated power of the vibrator and the air pump and the material weight, and obtaining the feed amount prediction model of the intermediate state of the current control cycle; step S5, training the feed amount prediction model of the intermediate state of the current control cycle according to the current control cycle The weight of the material in the hopper at each moment after the first preset number of moments, and the feed quantity prediction model of the intermediate state of the current control cycle are used to determine the first power of the vibrator and the second power of the air pump at each moment after the first preset number of moments in the current control cycle; step S6, when the weight reduction of the raw material in the hopper reaches the said addition amount during the current control cycle, or at the end of the current control cycle, the feed channel is closed; step S7, at the end of the current control cycle, according to the weight of the material in the hopper at the first preset number of moments and at each moment thereafter, as well as the first power and the second power, the feed quantity prediction model of the intermediate state of the current control cycle is trained to obtain the feed quantity prediction model of the final state of the current control cycle.

[0022] According to the automatic batching control method for PVC pipe production in an embodiment of the present invention, when controlling the feed amount, the influence of the material weight in the hopper on the feed speed can be considered, thereby setting appropriate power for the vibrator and the air pump. The feed amount prediction model can also be continuously updated based on data such as power and material weight obtained during the control cycle, thereby improving the accuracy of the power of the vibrator and the air pump, and improving the accuracy of the feed amount prediction model, thereby improving the control accuracy of the feed amount of raw materials and improving the quality of PVC pipes.

[0023] According to one embodiment of the present invention, when producing PVC pipes, a variety of raw materials are required, such as base resin, heat stabilizer, lubricant and other raw materials. Different hoppers can be used to hold the multiple raw materials, and each hopper can be connected to a feed channel. Each hopper can weigh the weight of the remaining raw materials in the hopper.

[0024] According to one embodiment of the present invention, in step S1, the order and amount of raw materials added vary. Therefore, multiple control cycles can be set, and the amounts added within each control cycle are determined based on the order in which the raw materials are added. For example, if raw materials A and B need to be added first, in amounts of 10 kg and 2 kg, respectively, and in the next process, raw materials C and D need to be added in amounts of 5 kg and 1 kg, respectively, then raw materials A and B can be added in the first control cycle (i.e., the feed channel connecting the hoppers for raw materials A and B is opened), and the amounts of raw materials A and B can be set to 10 kg and 2 kg, respectively. Raw materials C and D can be added in the second control cycle (i.e., the feed channel connecting the hoppers for raw materials C and D is opened), and the amounts of raw materials C and D can be set to 5 kg and 1 kg, respectively. In this example, each control cycle lasts 5 minutes and includes multiple moments, with the time interval between adjacent moments being 10 seconds. The present invention does not limit the specific values ​​of the control cycle duration or the time interval between adjacent moments.

[0025] According to one embodiment of the present invention, in step S2, as described above, after determining the type and amount of raw materials that need to be added in the current control cycle, the feed channels corresponding to the multiple raw materials can be opened, and since the raw materials are usually powdered, the powdered raw materials may block the feed channel due to the interaction force between the particles. Therefore, equipment such as vibrators and air pumps can be used to help feed. The vibrator can be set on the hopper to vibrate the hopper to destroy the interaction force between the particles and help the powdered raw materials fall into the feed channel from the discharge port of the hopper. The air pump is set in the feed channel to suck the raw materials in the feed channel into the container, thereby mixing them in the container and then proceeding to the next processing step.

[0026] According to one embodiment of the present invention, in step S3, after the current control cycle begins, the vibrator and air pump can be set to rated power for a first preset number of moments (e.g., 10), and the weight of the material in the hopper corresponding to each raw material can be measured. The material weight at each moment can be used to determine the feed amount between the moments. That is, the material weight at two moments can be subtracted to obtain the feed amount between the two moments. Furthermore, the effect of the weight of the raw material remaining in the hopper on the feed amount can be determined, assuming the power of the vibrator and air pump remains unchanged. For example, a larger amount of raw material remaining in the hopper results in a higher pressure at the hopper outlet, which has a positive impact on the feed amount between the two moments. That is, the more raw material remaining in the hopper, the greater the feed amount between the two moments. This pattern can then be summarized and the power of the vibrator and air pump at subsequent moments adjusted based on this pattern. For example, if the feed amount between two moments is small and the feed speed is slow, the power of the vibrator and air pump can be increased to increase the feed speed.

[0027] According to one embodiment of the present invention, in step S4, data such as material weight and rated power at a first preset number of moments before each control cycle can be used to train a feed rate prediction model for the final state of the previous control cycle, thereby obtaining a feed rate prediction model for the intermediate state of the current control cycle. The feed rate prediction model for the intermediate state of the current control cycle is then used to determine the first power of the vibrator and the second power of the air pump for the remaining moments of the current control cycle. The material weight within the hopper can also be monitored at all times during the remaining moments. At the end of the current control cycle, the feed rate prediction model for the intermediate state of the current control cycle is trained using the first power, second power, and material weight at the remaining moments of the current control cycle to obtain a feed rate prediction model for the final state of the current control cycle. Similarly, the above steps can be iteratively performed for the next control cycle. Furthermore, if the current control cycle is the first control cycle, a pre-trained feed rate prediction model (a feed rate prediction model trained using data from a control cycle specifically used for training) or an initial feed rate prediction model (e.g., a model whose parameters are randomly initialized) can be used.

[0028] Figure 2 A schematic diagram of a feed rate prediction model according to an embodiment of the present invention is exemplarily shown.

[0029] According to one embodiment of the present invention, in step S4, the feed amount prediction model of the final state of the previous control cycle is trained according to the rated power of the vibrator and the air pump and the material weight to obtain the feed amount prediction model of the intermediate state of the current control cycle, including: inputting the training input vector of the i-th moment composed of the rated power of the vibrator, the rated power of the air pump, and the material weight at the i-th moment, and the training feature vector of the i-1-th moment into the i-th feature extraction level of the feed amount prediction model of the final state of the previous control cycle to obtain the training feature vector of the i-th moment, wherein i is a positive integer less than the first preset number, and when i=1 , the training feature vector at the i-1th moment is a zero vector; the training feature vector at the i-th moment is input into the first fully connected layer and the first activation layer of the feed quantity prediction model of the final state of the previous control cycle for processing to obtain the predicted feed quantity between the i-th moment and the i+1-th moment; according to the predicted feed quantity between the i-th moment and the i+1-th moment, the material weight at the i-th moment and the material weight at the i+1-th moment, the first loss function of the feed quantity prediction model is determined; according to the first loss function, the feed quantity prediction model of the final state of the previous control cycle is trained to obtain the feed quantity prediction model of the intermediate state of the current control cycle.

[0030] According to one embodiment of the present invention, the feed rate prediction model can be a recurrent neural network model or an LSTM model, etc., and can include multiple processing blocks. Each processing block can include a feature extraction layer that can take the current input vector and the feature vector output by the previous processing block as input and output the current feature vector. For example, the training input vector at the i-th moment is composed of the rated power of the vibrator, the rated power of the air pump, and the weight of the material at the i-th moment. The output of the i-1th processing block is the training feature vector at the i-1th moment. The training input vector at the i-th moment and the training feature vector at the i-1th moment serve as the input of the i-th processing block. After processing by the i-th block, the training feature vector at the i-th moment is obtained. If i = 1, the training feature vector at the i-1th moment is a zero vector, or it can be a randomly initialized vector (i.e., each data in the vector is a random value). The processing blocks can be reusable. For example, a feed rate prediction model can include only one processing block. When receiving a training input vector at the first moment and a training feature vector at the 0th moment (e.g., a zero vector), the processing block can serve as the first processing block (its feature extraction level is the first feature extraction level). When receiving a training input vector at the second moment and a training feature vector at the first moment, the processing block can serve as the second processing block. ... When receiving a training input vector at the i-th moment and a training feature vector at the i-1-th moment, the processing block can serve as the i-th processing block. When inputting the training input vector at each moment, it can be directly input into the feature extraction level of the processing block, or it can first be input into a fully connected layer to obtain a higher-dimensional vector, thereby describing more comprehensive feature information, and then input into the feature extraction level.

[0031] According to one embodiment of the present invention, based on the above feed rate prediction model, a training feature vector output at the current moment can be obtained by performing comprehensive processing based on the training input vector at the current moment and the training feature vector output at the previous moment. The feed rate is correlated with the weight of the material in the hopper and is also related to the power of the vibrator and air pump. Therefore, when the power of the vibrator and air pump remains unchanged, the feed rate over a period of time changes with the change in the weight of the material in the hopper. Therefore, the predicted feed rate for future time periods can be predicted based on data from past moments. For example, the training feature vector at the i-th moment is obtained based on the training input vector at the i-th moment and the training feature vector at the i-1-th moment, while the training feature vector at the i-1-th moment is obtained based on the training input vector at the i-1-th moment and the training feature vector at the i-2-th moment... Therefore, the training feature vector at the i-th moment incorporates data from multiple past moments. Therefore, the training feature vector at the i-th moment can be used to predict the feed rate between the i-th moment and the i+1-th moment based on data from multiple past moments. The training feature vector at the i-th moment is input into the first fully connected layer, and the output data of the first fully connected layer is input into the first activation layer. Processing is performed using an activation function (e.g., a RELU activation function) to obtain the predicted feed rate between the i-th moment and the i+1-th moment. In other words, the data from multiple past moments can represent the feed rate during the time period between each moment when the vibrator and air pump are at a certain power (e.g., rated power). Based on this, the predicted feed rate between the i-th moment and the i+1-th moment can be determined when the vibrator and air pump are at their rated power.

[0032] According to one embodiment of the present invention, there may be errors in the predicted feed amount between the i-th moment and the i+1-th moment, and the accurate feed amount can be obtained by the material weight at the i-th moment and the material weight at the i+1-th moment, that is, the material weight at the i-th moment and the material weight at the i+1-th moment are subtracted to determine the accurate feed amount, and the error between the accurate feed amount and the predicted feed amount can be determined, thereby determining the first loss function and training the feed amount prediction model of the final state of the previous control cycle.

[0033] According to one embodiment of the present invention, the first loss function of the feed amount prediction model is determined based on the predicted feed amount between the i-th moment and the i+1-th moment, the material weight at the i-th moment, and the material weight at the i+1-th moment, including: determining the first loss function of the feed amount prediction model according to formula (1): ,

[0034] (1)

[0035] in, is the predicted feed quantity between the i-th moment and the i+1-th moment, is the material weight at the i-th moment, is the material weight at the i+1th moment, is the first preset number, i and All are positive integers.

[0036] According to one embodiment of the present invention, in formula (1), is the exact feed amount between the i-th moment and the i+1-th moment, so, The error between the predicted feed amount and the accurate feed amount between the i-th moment and the i+1-th moment is, and the larger the moment number i is, the closer the moment is to the end of the first preset number of moments, and the greater the influence of the accuracy of the predicted feed amount between the moment and the previous moment on the accuracy of subsequent predictions. Therefore, a higher weight can be set, for example, As the weight of the error in the predicted feed quantity between the i-th moment and the previous moment, the errors in the predicted feed quantity between each moment and the previous moment can be weighted and summed to obtain the first loss function of the feed quantity prediction model.

[0037] According to one embodiment of the present invention, after obtaining the first loss function, backpropagation can be performed on the first loss function to adjust the parameters of the feed rate prediction model for the final state of the previous control cycle using a gradient descent method to perform training and obtain a feed rate prediction model for the intermediate state of the current control cycle. Compared to the feed rate prediction model for the final state of the previous control cycle, the feed rate prediction model for the intermediate state of the current control cycle is more accurate in predicting feed rate data between adjacent moments.

[0038] In this way, data from multiple moments in the past can be fused to accurately predict the feed volume between the i-th moment and the i+1-th moment. When setting the first loss function, the weight is set according to the number of moments to accurately reflect the impact of the accuracy of the predicted data at that moment on the accuracy of subsequent predictions, thereby improving training accuracy and training efficiency, and improving the accuracy of the feed volume prediction model.

[0039] According to one embodiment of the present invention, in step S5, the feed quantity prediction model of the intermediate state of the current control cycle that has been trained as described above can be used to determine the first power of the vibrator and the second power of the air pump at subsequent moments, so that the feed quantity of the raw material can reach the predetermined addition quantity within the current control cycle.

[0040] According to one embodiment of the present invention, in step S5, according to the weight of the material in the hopper at each moment after the first preset number of moments of the current control cycle and the feed amount prediction model of the intermediate state of the current control cycle, the first power of the vibrator and the second power of the air pump at each moment after the first preset number of moments of the current control cycle are determined, including: The weight of the material in the hopper at each moment, the rated power of the vibrator, and the rated power of the air pump constitute the weight of the material in the hopper at each moment. The first state vector at the moment is input into each feature extraction level of the feed quantity prediction model of the intermediate state of the current control cycle to obtain the first The first eigenvector at the moment; get the The number of remaining time periods between the moment and the end of the current control cycle; Get the The weight of the added material before the moment, and obtain the first difference between the weight of the added material and the amount of raw material added; the remaining time period quantity and the first difference form the first The second state vector at the moment is processed by the second fully connected layer to obtain the second state vector The second eigenvector at the moment; The second eigenvector at the moment and the The first eigenvectors of the moments are spliced ​​together to obtain the The third eigenvector of the moment; The third eigenvector at the moment is input into the third fully connected layer and the second activation layer to obtain the The first power of the vibrator and the second power of the air pump at the moment; the first power of the vibrator and the second power of the air pump at the moment The weight of the material in the hopper at the moment, the weight of the vibrator at the moment 1+j to the moment The rated power of the vibrator at the moment From moment to The first power at the moment, the air pump at the 1+j moment to the The rated power of the air pump at the moment From moment to The second power at the moment constitutes the power from the 1+jth moment to the The first state vector at the moment is input into each feature extraction level of the feed quantity prediction model of the intermediate state of the current control cycle to obtain the first The first eigenvector at the moment, j is an integer greater than or equal to 0, and , N is the sequence number of the end time of the current control cycle, N is a positive integer; get the The number of remaining time periods between the moment and the end of the current control cycle; Get the The weight of the added material before the moment, and obtain the first difference between the weight of the added material and the amount of raw material added; the remaining time period quantity and the first difference form the first The second state vector at the moment is processed by the second fully connected layer to obtain the second state vector The second eigenvector at the moment; The second eigenvector at the moment and the The first eigenvectors of the moments are spliced ​​together to obtain the The third eigenvector of the moment; The third eigenvector at the moment is input into the third fully connected layer and the second activation layer to obtain the The first power of the vibrator and the second power of the air pump at a certain moment.

[0041] According to one embodiment of the present invention, from the first moment to the The first state vector at each moment is obtained in a similar way to the above-mentioned training input vector, which is a combination of the weight of the material in the hopper, the rated power of the vibrator, and the rated power of the air pump. However, due to the difference in parameters between the feed quantity prediction model of the intermediate state of the current control cycle and the feed quantity prediction model of the final state of the previous control cycle, the characteristic vector of the output at each moment obtained by processing the feed quantity prediction model of the intermediate state of the current control cycle is different from the characteristic vector of the output at each moment obtained by processing the feed quantity prediction model of the final state of the previous control cycle. The feature vector output by each processing block can be integrated from the first moment to the The data at the moment is used as the The first eigenvector at the moment can be used for subsequent predictions.

[0042] According to one embodiment of the present invention, the The number of remaining time periods between the first moment and the end moment of the current control cycle. For example, if the sequence number of the last moment of the current control cycle is N, then the sequence number of the The number of remaining time periods between the moment and the end of the current control cycle is . You can get The weight of the added material before the moment, that is, The amount of material fed between each two adjacent moments before the moment is summed to obtain the weight of the added material. A first difference between the weight of the added material and the amount of raw material added can also be obtained, that is, the first difference is obtained by subtracting the amount of addition from the weight of the added material.

[0043] According to one embodiment of the present invention, the number of remaining time periods and the first difference form the first The second state vector at each moment can be obtained, and the second state vector can be upgraded through the second fully connected layer to obtain a second eigenvector with a higher dimension, so as to provide a more comprehensive description of the state at that moment, that is, to describe the remaining feeding demand and the remaining feeding time, thereby providing a data basis for predicting the first power of the vibrator and the second power of the air pump at subsequent moments.

[0044] According to one embodiment of the present invention, the The second eigenvector at the moment and the The first eigenvectors of the moments are spliced ​​together to obtain the The third eigenvector of the moment, where The first eigenvector at the moment can be used to describe the The weight of the material in the hopper at the first moment, and the weight of the material in the hopper at the next time period (i.e., the first time period) when the power of the vibrator and the air pump is rated From moment to The second eigenvector is used to describe the remaining feeding demand and the remaining feeding time. Therefore, the first eigenvector can be used as a reference and the second eigenvector can be used as a description of the feeding demand to determine how to set the power of the vibrator and the air pump in the subsequent time period so that the total feeding amount in the remaining time can meet the feeding demand. Therefore, the third eigenvector obtained by concatenating the first and second eigenvectors can be used for solution. The third eigenvector is processed by the third fully connected layer and the second activation layer to obtain the third eigenvector. That is, the third fully connected layer and the second activation layer can determine the first power of the vibrator and the second power of the air pump based on the first eigenvector. From moment to The predicted feed amount between the moments, thus predicting the The weight of the material in the hopper at a certain moment is used to predict how to set the first power of the vibrator and the second power of the air pump in the subsequent time period.

[0045] According to one embodiment of the present invention, At this moment, the first power of the vibrator and the second power of the air pump predicted above are used for setting, and the first power can be actually measured. The weight of the material in the hopper at the moment can be used to form the The first state vector at the moment. The feed rate prediction model may include processing blocks, so the second moment to the The first state vector at the moment is input into each processing block in sequence, and finally the first state vector is obtained. The first eigenvector at the moment. Generally, we can get The first eigenvector at the moment. Similar to the above, the first eigenvector can be determined The number of remaining time periods between the first moment and the end moment of the current control cycle, and the The first difference between the weight of the added material before the moment and the weight of the added material and the amount of raw material added, and the remaining time period and the first difference form the first The second state vector at the moment is obtained, and then the second eigenvector is obtained, and it is spliced ​​with the first eigenvector to obtain the The third eigenvector at the moment is then obtained through the third fully connected layer and the second activation layer. The first power of the vibrator and the second power of the air pump at each moment can be obtained. Similarly, after each moment, the first power of the vibrator and the second power of the air pump at the next moment can be obtained, thereby obtaining the first power of the vibrator and the second power of the air pump at each moment in the current control cycle, and making the total feed amount in the current control cycle reach the addition amount of the raw material.

[0046] According to one embodiment of the present invention, the second fully connected layer, the third fully connected layer and the second activation layer may be trained. The method further comprises: The first feature vector at the moment is input into the first fully connected layer and the first activation layer for processing, and the first feature vector is obtained. From moment to The predicted feed volume between the moments; The weight of the material in the hopper at the first moment is From moment to The predicted feed amount between the moments is obtained The predicted material weight within a moment; The predicted material weight, the first power of the vibrator and the second power of the air pump at the moment are obtained. The first predicted feature vector at the moment; From moment to The predicted feed amount between moments is obtained The predicted weight of the added material before the moment, and obtain the first predicted difference between the predicted weight of the added material and the amount of raw material added; The number of remaining time periods between the moment and the end of the control cycle and the first prediction difference constitute the first The second predicted state vector at the moment is processed by the second fully connected layer to obtain the second state vector The second predicted feature vector at the moment The first predicted feature vector at the moment The second prediction feature vectors of the moment are spliced ​​to obtain the The third predicted feature vector at the moment; The third predicted feature vector at the moment is input into the third fully connected layer and the second activation layer to obtain the The first predicted power of the vibrator and the second predicted power of the air pump at a moment; iteratively execute the steps of obtaining the first predicted power and the second predicted power to obtain the first predicted power and the second predicted power. The first predicted power of the vibrator and the second predicted power of the air pump from the first moment to the Nth moment, as well as the predicted feed amount between adjacent moments; The predicted feed amount between each adjacent moment in the time period from moment 1 to moment N is used to determine the The second loss function corresponding to the multiple moments is used to train the second fully connected layer and the third fully connected layer according to the second loss functions corresponding to the multiple moments.

[0047] According to one embodiment of the present invention, the above training steps may be performed after each moment, thereby gradually improving the accuracy of the second fully connected layer and the third fully connected layer, and also improving the accuracy of determining the first power of the vibrator and the second power of the air pump.

[0048] According to one embodiment of the present invention, the first fully connected layer and the first activation layer are used to connect the first The first eigenvector of the moment can be processed to obtain the From moment to The predicted feed amount between the moments. The weight of the material in the hopper at the first moment is From moment to Subtract the predicted feed amount between the moments to obtain the The predicted material weight within a moment, so that the The predicted material weight at the moment, the first power of the vibrator and the second power of the air pump are combined to obtain the The first predicted feature vector at the moment.

[0049] According to one embodiment of the present invention, The total amount of feed before the moment (the weight of the material in the hopper at the beginning of the control cycle is the same as the weight of the material in the hopper at the beginning of the control cycle). The difference between the weight of the material in the hopper at the first moment) and the From moment to The predicted feed amount between the moments is summed up to obtain the The predicted weight of the added material before the moment is subtracted from the added amount of the raw material to obtain the first predicted difference. The number of remaining time periods between the moment and the end of the control cycle and the first prediction difference constitute the first The second predicted state vector at the moment is processed by the second fully connected layer to obtain the second state vector The second predicted feature vector at the moment The first predicted feature vectors at each moment are spliced ​​to obtain the The third predicted feature vector at the moment is obtained through the third fully connected layer and the second activation layer. The third predicted feature vector of the moment is processed to obtain the The first predicted power of the vibrator and the second predicted power of the air pump at a certain moment.

[0050] According to one embodiment of the present invention, the above process can be performed iteratively, thereby enabling backward prediction. The first predicted feature vector at the moment is input into the first fully connected layer and the first activation layer for processing, and the first prediction feature vector is obtained. From moment to The predicted feed quantity between the moments; The predicted material weight at the moment is subtracted from the From moment to The predicted feed amount between the moments is obtained The predicted material weight within a moment; The predicted material weight at a certain moment, the first predicted power of the vibrator and the second predicted power of the air pump constitute the first The first predicted feature vector at the moment The total amount of feed before the moment, and From moment to The predicted feed volume between the moments and the From moment to The predicted feed amount between the moments is summed up to get the The predicted weight of the added material before the moment, and obtain the first predicted difference between the predicted weight of the added material and the amount of raw material added; The number of remaining time periods between the moment and the end of the control cycle and the first prediction difference constitute the first The second predicted state vector at the moment is processed by the second fully connected layer to obtain the second state vector The second predicted feature vector at the moment The first predicted feature vectors at each moment are spliced ​​to obtain the first The third predicted feature vector at the moment is obtained through the third fully connected layer and the second activation layer. The first predicted power of the vibrator and the second predicted power of the air pump at the moment, and so on, can be obtained. The first predicted power of the vibrator and the second predicted power of the air pump from the moment to the Nth moment, as well as the predicted feed amount between adjacent moments. At each moment, the first predicted power and the second predicted power at the subsequent moments, as well as the predicted material weight at the subsequent moments, can be obtained through the above derivation method (instead of obtaining the actual data after the end of the control cycle). Therefore, the second loss function can be calculated after each moment, and training can be performed after each moment, so that the second fully connected layer and the third fully connected layer can be continuously trained within the control cycle, thereby improving the accuracy of determining the first power and the second power at the subsequent moments.

[0051] According to one embodiment of the present invention, The predicted feed amount between each adjacent moment in the time period from moment 1 to moment N is used to determine the The second loss function corresponding to the moment includes: determining the first loss function according to formula (2) The second loss function corresponding to the moment ,

[0052] (2)

[0053] in, is the predicted feed quantity between the kth moment and the k+1th moment, , is the feed amount between the sth moment and the s+1th moment, , is the amount of raw material added, and s and k are both positive integers.

[0054] According to one embodiment of the present invention, That is, from the start of the control cycle to the The total amount of feed at a moment, For the The sum of the predicted feed amounts between each adjacent moment in the time period from moment 1 to moment N is the total predicted feed amount within the control period. The difference between the amount of raw material added and the total amount of feed predicted within the control period can be used as the first A second loss function corresponding to each moment is obtained, and the second fully connected layer and the third fully connected layer are trained using the second loss function. For example, the second fully connected layer and the third fully connected layer are trained by backpropagating the second loss function. The above training can be performed after each moment, thereby improving the accuracy of the second fully connected layer and the third fully connected layer, and continuously improving the accuracy of the predicted first power of the vibrator and the second power of the air pump, so that the total amount of raw material fed within the control period is as close as possible to the amount of raw material added.

[0055] In this way, after each moment, the first predicted power, the second predicted power and the predicted feed amount at the subsequent moments can be solved iteratively, so that the predicted total feed amount within the control period can be determined by the predicted feed amount, and then the error between the predicted total feed amount and the amount of raw material added can be reduced by setting the second loss function to train the second fully connected layer and the third fully connected layer, thereby continuously improving the accuracy of the second fully connected layer and the third fully connected layer, as well as the prediction accuracy of the first power of the vibrator and the second power of the air pump, and reducing the error between the predicted total feed amount and the amount of raw material added.

[0056] According to one embodiment of the present invention, in step S6, before the end of the current control cycle, if the weight reduction in the raw material hopper reaches the added amount, that is, the total amount of raw material fed in the current control cycle reaches the added amount, the feed channel can be closed, or the feed channel can be closed at the end of the current control cycle. Furthermore, if the total amount of raw material fed does not reach the added amount at the end of the current control cycle, a prompt message can be generated to prompt the staff to refill the material. Furthermore, the aforementioned feed amount prediction model, the second fully connected layer, and the third fully connected layer can be continuously trained in subsequent control cycles to continuously improve the model accuracy, so that the total amount of raw material fed in a control cycle is as close as possible to the added amount, and the duration of the process in which the total amount of raw material fed reaches the added amount is as close as possible to the duration of the control cycle.

[0057] According to one embodiment of the present invention, in step S7, at the end of the current control cycle, a feed quantity prediction model of the intermediate state of the current control cycle is trained based on the weight of the material in the hopper at each moment after the first preset number of moments, as well as the first power and the second power, to obtain a feed quantity prediction model of the final state of the current control cycle, including: determining the predicted feed quantity between each moment after the first preset number of moments based on the weight of the material in the hopper at each moment after the first preset number of moments, the first power and the second power, and the feed quantity prediction model of the intermediate state of the current control cycle; obtaining a third loss function of the feed quantity prediction model of the intermediate state of the current control cycle according to formula (3) ,

[0058] (3)

[0059] in, is the predicted feed quantity between the hth moment and the h+1th moment, is the material weight at the hth moment, is the material weight at the h+1th moment, , N is the sequence number of the end time of the current control cycle, is the first preset number, h is a positive integer; according to the third loss function, the feed quantity prediction model of the intermediate state of the current control cycle is trained to obtain the feed quantity prediction model of the final state of the current control cycle.

[0060] According to one embodiment of the present invention, as described above, in the above steps, the first power and the second power at each moment after the first preset number are obtained, and the weight of the material in the hopper can also be actually weighed at each moment, so that the input vector of each moment can be composed based on the first power, the second power and the material weight at each moment, combined with the output vector of the previous moment, and jointly input into the feed quantity prediction model of the intermediate state of the current control cycle to obtain the output vector of the moment, and processed through the first fully connected layer and the first activation layer to obtain the predicted feed quantity in the time period between the moment and the next moment. Furthermore, since the above-mentioned current control cycle has ended, the actual material weight in the hopper at all moments in the current control cycle can be obtained, so that the actual material weights of adjacent moments can be subtracted to obtain the actual feed quantity. In formula (3), That is, the measured feed amount between the hth moment and the h+1th moment, is the predicted feed amount between the hth moment and the h+1th moment, and the error between the two is By summing the errors at each moment, we can obtain a third loss function. We then perform backpropagation on the third loss function and adjust the parameters of the feed rate prediction model for the intermediate state of the current control cycle in a direction that reduces the third loss function. This yields a feed rate prediction model for the final state of the current control cycle. This model is then used in the next control cycle.

[0061] In this way, not only the data within the first preset number of moments are used to train the feed quantity prediction model, but also after the control cycle ends, the data of multiple moments after the first preset number of moments can be used to continue training the feed quantity prediction model, thereby maximizing the use of the measured and predicted data within the control cycle, improving the training efficiency of the feed quantity prediction model, improving the prediction accuracy, and providing an accurate data basis for predicting the first power of the vibrator and the second power of the air pump.

[0062] According to the automatic batching control method for PVC pipe production of an embodiment of the present invention, when controlling the feed rate, the influence of the weight of the material in the hopper on the feed rate can be considered, thereby setting appropriate power for the vibrator and air pump. The feed rate prediction model can also be continuously updated based on data such as power and material weight obtained during the control cycle, thereby improving the accuracy of the power of the vibrator and air pump, improving the accuracy of the feed rate prediction model, and thus improving the control accuracy of the feed rate of the raw materials and improving the quality of the PVC pipe. When training the feed rate prediction model for the final state of the previous control cycle, data from multiple past moments can be fused to accurately predict the predicted feed rate between the i-th moment and the i+1-th moment. When setting the first loss function, the weight is set according to the number of moments, thereby accurately reflecting the influence of the accuracy of the predicted data at that moment on the accuracy of subsequent predictions, improving training accuracy and training efficiency, and improving the accuracy of the feed rate prediction model. When training the second and third fully connected layers, after each moment, the first predicted power, second predicted power, and predicted feed amount at subsequent moments can be solved iteratively, thereby determining the predicted total feed amount within the control cycle using the predicted feed amount. Furthermore, by setting a second loss function, the error between the predicted total feed amount and the amount of raw material added can be reduced to train the second and third fully connected layers, thereby continuously improving the accuracy of the second and third fully connected layers, as well as the accuracy of the predictions of the first power of the vibrator and the second power of the air pump, and reducing the error between the predicted total feed amount and the amount of raw material added. When training the feed amount prediction model for the intermediate state of the current control cycle, the feed amount prediction model is trained not only using data from the first preset number of moments, but also, after the control cycle ends, further training can be continued using data from multiple moments after the first preset number of moments. This maximizes the utilization of measured and predicted data within the control cycle, improves the training efficiency of the feed amount prediction model, improves prediction accuracy, and provides an accurate data basis for predicting the first power of the vibrator and the second power of the air pump.

[0063] Figure 3The automatic batching control system for PVC pipe production according to an embodiment of the present invention is exemplarily shown, and the system includes: an addition amount module, which is used to obtain the addition amount of multiple raw materials in the current control cycle at the beginning of the current control cycle; an opening module, which is used to open the feeding channels corresponding to the multiple raw materials and open the vibrators and air pumps of each feeding channel; a weighing module, which is used to weigh the material weight in the hopper corresponding to each raw material at the first preset number of times before the current control cycle; a first training module, which is used to train the feeding amount prediction model of the final state of the previous control cycle according to the rated power of the vibrator and the air pump and the material weight, and obtain the feeding amount prediction model of the intermediate state of the current control cycle; a determination module, which is used to train the feeding amount prediction model of the final state of the previous control cycle according to the rated power of the vibrator and the air pump and the material weight, and obtain the feeding amount prediction model of the intermediate state of the current control cycle; and a determination module, which is used to determine the feeding amount prediction model of the intermediate state of the current control cycle according to the current control cycle. The first power of the vibrator and the second power of the air pump at each moment after the first preset number of moments in the current control cycle are determined based on the weight of the material in the hopper at each moment after the first preset number of moments in the current control cycle, and the feed quantity prediction model at the intermediate state of the current control cycle; a closing module is used to close the feed channel when the weight reduction of the raw material in the hopper reaches the added amount during the current control cycle, or at the end of the current control cycle; a second training module is used to train the feed quantity prediction model at the intermediate state of the current control cycle at the end of the current control cycle based on the weight of the material in the hopper at each moment after the first preset number of moments, as well as the first power and the second power, to obtain the feed quantity prediction model at the final state of the current control cycle.

[0064] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0065] Those skilled in the art will appreciate that the embodiments of the present invention described above and shown in the accompanying drawings are intended to be illustrative only and are not intended to limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functional and structural principles of the present invention have been demonstrated and illustrated in the embodiments. Any variations or modifications may be made to the embodiments of the present invention without departing from the principles described.

Claims

1. A method for controlling automatic batching in PVC pipe production, characterized in that: include: At the beginning of the current control cycle, the addition amount of multiple raw materials in the current control cycle is obtained; Open the feeding channels corresponding to the multiple raw materials, and open the vibrators and air pumps of each feeding channel; weigh the weight of the materials in the hopper corresponding to each raw material at the first preset number of moments before the current control cycle; train the feeding amount prediction model of the final state of the previous control cycle according to the rated power of the vibrator and the air pump and the material weight, and obtain the feeding amount prediction model of the intermediate state of the current control cycle; determine the first power of the vibrator and the second power of the air pump at each moment after the first preset number of moments of the current control cycle according to the material weight in the hopper at each moment after the first preset number of moments of the current control cycle, and the feeding amount prediction model of the intermediate state of the current control cycle; close the feeding channel when the weight reduction of the raw material in the hopper reaches the added amount during the current control cycle, or at the end of the current control cycle; at the end of the current control cycle, train the feeding amount prediction model of the intermediate state of the current control cycle according to the material weight in the hopper at each moment after the first preset number of moments of the current control cycle, as well as the first power and the second power, and obtain the feeding amount prediction model of the final state of the current control cycle; According to the rated power and material weight of the vibrator and the air pump, the feed amount prediction model of the final state of the previous control cycle is trained to obtain the feed amount prediction model of the intermediate state of the current control cycle, including: inputting the training input vector of the i-th moment composed of the rated power of the vibrator, the rated power of the air pump, and the material weight at the i-th moment, and the training feature vector at the i-1-th moment into the i-th feature extraction level of the feed amount prediction model of the final state of the previous control cycle to obtain the training feature vector at the i-th moment, wherein i is a positive integer less than a first preset number, and when i=1, the training feature vector at the i-1-th moment is The training feature vector is a zero vector; the training feature vector at the i-th moment is input into the first fully connected layer and the first activation layer of the feed quantity prediction model of the final state of the previous control cycle for processing to obtain the predicted feed quantity between the i-th moment and the i+1-th moment; according to the predicted feed quantity between the i-th moment and the i+1-th moment, the material weight at the i-th moment and the material weight at the i+1-th moment, the first loss function of the feed quantity prediction model is determined; according to the first loss function, the feed quantity prediction model of the final state of the previous control cycle is trained to obtain the feed quantity prediction model of the intermediate state of the current control cycle; According to the weight of the material in the hopper at each moment after the first preset number of moments of the current control cycle and the feed amount prediction model of the intermediate state of the current control cycle, the first power of the vibrator and the second power of the air pump at each moment after the first preset number of moments of the current control cycle are determined, including: The weight of the material in the hopper at each moment, the rated power of the vibrator, and the rated power of the air pump constitute the weight of the material in the hopper at each moment. The first state vector at the moment is input into each feature extraction level of the feed quantity prediction model of the intermediate state of the current control cycle to obtain the first The first eigenvector at the moment; get the The number of remaining time periods between the moment and the end of the current control cycle; Get the The weight of the added material before the moment, and obtain the first difference between the weight of the added material and the amount of raw material added; the remaining time period quantity and the first difference form the first The second state vector at the moment is processed by the second fully connected layer to obtain the second state vector The second eigenvector at the moment; The second eigenvector at the moment and the The first eigenvectors of the moments are spliced ​​together to obtain the The third eigenvector of the moment; The third eigenvector at the moment is input into the third fully connected layer and the second activation layer to obtain the The first power of the vibrator and the second power of the air pump at the moment; the first power of the vibrator and the second power of the air pump at the moment The weight of the material in the hopper at the moment, the weight of the vibrator at the moment 1+j to the moment The rated power of the vibrator at the moment From moment to The first power at the moment, the air pump at the 1+j moment to the The rated power of the air pump at the moment From moment to The second power at the moment constitutes the power from the 1+jth moment to the The first state vector at the moment is input into each feature extraction level of the feed quantity prediction model of the intermediate state of the current control cycle to obtain the first The first eigenvector at the moment, j is an integer greater than or equal to 0, and , N is the sequence number of the end time of the current control cycle, N is a positive integer; get the The number of remaining time periods between the moment and the end of the current control cycle; Get the The weight of the added material before the moment, and obtain the first difference between the weight of the added material and the amount of raw material added; the remaining time period quantity and the first difference form the first The second state vector at the moment is processed by the second fully connected layer to obtain the second state vector The second eigenvector at the moment; The second eigenvector at the moment and the The first eigenvectors of the moments are spliced ​​together to obtain the The third eigenvector of the moment; The third eigenvector at the moment is input into the third fully connected layer and the second activation layer to obtain the The first power of the vibrator and the second power of the air pump at a certain moment.

2. The automatic batching control method for PVC pipe production according to claim 1, characterized in that: According to the predicted feed quantity between the i-th moment and the i+1-th moment, the material weight at the i-th moment and the material weight at the i+1-th moment, the first loss function of the feed quantity prediction model is determined, including: according to the formula , determine the first loss function of the feed rate prediction model ,in, is the predicted feed quantity between the i-th moment and the i+1-th moment, is the material weight at the i-th moment, is the material weight at the i+1th moment, is the first preset number, i and All are positive integers.

3. The automatic batching control method for PVC pipe production according to claim 1, characterized in that: The method further comprises: The first feature vector at the moment is input into the first fully connected layer and the first activation layer for processing, and the first feature vector is obtained. From moment to The predicted feed volume between the moments; The weight of the material in the hopper at the first moment is From moment to The predicted feed amount between the moments is obtained The predicted material weight within a moment; The predicted material weight, the first power of the vibrator and the second power of the air pump at the moment are obtained. The first predicted feature vector at the moment; From moment to The predicted feed amount between moments is obtained The predicted weight of the added material before the moment, and obtain the first predicted difference between the predicted weight of the added material and the amount of raw material added; The number of remaining time periods between the moment and the end of the control cycle and the first prediction difference constitute the first The second predicted state vector at the moment is processed by the second fully connected layer to obtain the second predicted state vector The second predicted feature vector at the moment The first predicted feature vector at the moment The second prediction feature vectors of the moment are spliced ​​to obtain the The third predicted feature vector at the moment; The third predicted feature vector at the moment is input into the third fully connected layer and the second activation layer to obtain the The first predicted power of the vibrator and the second predicted power of the air pump at a moment; iteratively execute the steps of obtaining the first predicted power and the second predicted power to obtain the first predicted power and the second predicted power. The first predicted power of the vibrator and the second predicted power of the air pump from the first moment to the Nth moment, as well as the predicted feed amount between adjacent moments; The predicted feed amount between each adjacent moment in the time period from moment 1 to moment N is used to determine the The second loss function corresponding to the multiple moments is used to train the second fully connected layer and the third fully connected layer according to the second loss functions corresponding to the multiple moments.

4. The automatic batching control method for PVC pipe production according to claim 3, characterized in that: According to The predicted feed amount between each adjacent moment in the time period from moment 1 to moment N is used to determine the The second loss function corresponding to the moment includes: According to the formula , determine the The second loss function corresponding to the moment ,in, is the predicted feed quantity between the kth moment and the k+1th moment, , is the feed amount between the sth moment and the s+1th moment, , is the amount of raw material added, and s and k are both positive integers.

5. The automatic batching control method for PVC pipe production according to claim 1, characterized in that: At the end of the current control cycle, the feed amount prediction model of the intermediate state of the current control cycle is trained according to the weight of the material in the hopper at each moment after the first preset number of moments, as well as the first power and the second power, to obtain the feed amount prediction model of the final state of the current control cycle, including: determining the predicted feed amount between each moment after the first preset number of moments according to the weight of the material in the hopper at each moment after the first preset number of moments, the first power and the second power, and the feed amount prediction model of the intermediate state of the current control cycle; according to the formula , obtain the third loss function of the feed quantity prediction model in the intermediate state of the current control cycle ,in, is the predicted feed quantity between the hth moment and the h+1th moment, is the material weight at the hth moment, is the material weight at the h+1th moment, , N is the sequence number of the end time of the current control cycle, is the first preset number, h is a positive integer; according to the third loss function, the feed quantity prediction model of the intermediate state of the current control cycle is trained to obtain the feed quantity prediction model of the final state of the current control cycle.

6. An automatic batching control system for PVC pipe production, used to execute the method according to any one of claims 1 to 5, characterized in that: include: The addition amount module is used to obtain the addition amount of various raw materials in the current control cycle at the beginning of the current control cycle; The opening module is used to open the feeding channels corresponding to various raw materials and to start the vibrators and air pumps of each feeding channel; A weighing module, configured to weigh the weight of the materials in the hopper corresponding to each raw material at a first preset number of moments before the current control cycle; a first training module for training a feed quantity prediction model of a final state of a previous control cycle according to the rated power of the vibrator and the air pump and the weight of the material, and obtaining a feed quantity prediction model of an intermediate state of a current control cycle; a determination module for determining a first power of the vibrator and a second power of the air pump at each moment after a first preset number of moments of the current control cycle according to the weight of the material in the hopper at each moment after a first preset number of moments of the current control cycle and the feed quantity prediction model of the intermediate state of the current control cycle; a closing module for closing the feed channel when the weight reduction of the raw material in the hopper reaches the added amount during the current control cycle, or when the current control cycle ends; The second training module is used to train the feed quantity prediction model of the intermediate state of the current control cycle at the end of the current control cycle based on the weight of the material in the hopper at and after the first preset number of moments, as well as the first power and the second power, to obtain the feed quantity prediction model of the final state of the current control cycle.

Images