Rotary table PID adjustment method based on deep learning and self-training

Abstract

The invention discloses a rotary table PID adjustment method based on deep learning and self-training, and relates to the technical field of automatic control, the rotary table after PID adjustment is marked, the type of the current rotary table is recorded, the current rotary table is operated, data of the current rotary table is collected through a data collection device, and the collected data is classified according to the type of the rotary table; and establishing a deep learning model, and carrying out self-training by using the operation data of the current turntable. According to the method, the deep learning model is set, the adjusted turntable data is used for training, a model capable of outputting good PID parameters is constructed, the defect that PID parameter adjustment depends on experience can be overcome, the turntable performance is improved, and the working efficiency is higher.

Description

Technical Field

[0001] The present invention relates to the field of automatic control, and in particular to a turntable PID adjustment method based on deep learning and self-training Background Art

[0002] In the turntable system, the accuracy and efficiency of PID control play an important role in the precise control of the turntable. The application of traditional PID control methods in turntable control systems faces many challenges, especially in terms of sensitivity and efficiency when dealing with more and more variable load demands and environmental conditions. The traditional PID control method is to adjust a set of PID parameters according to the requirements of the current usage scenario by experienced professional commissioning personnel, which in many cases will lead to the situation that the existing PID parameters are not enough to meet the needs of the new usage scenario.

[0003] With the rapid development of artificial intelligence technology and its great application in industrial control, deep learning technology has achieved great results in the participation of large-scale industrial PID automatic control of the same type of products. However, for large industrial control products such as turntables with medium and small scales and a wide variety of models, turntables cannot provide a large number of parameters for deep learning to build high-performance big data models, which makes it difficult to achieve efficient application. Summary of the invention

[0004] In response to the above technical problems, the present invention provides a turntable PID adjustment method based on deep learning and self-training, which marks the turntable after PID adjustment, records the current turntable type, runs the current turntable, collects the current turntable data through a data acquisition device, and classifies the collected data according to the turntable type; establishes a deep learning model, and uses the current turntable operation data for self-training.

[0005] The training process includes the following steps:

[0006] Step S1, obtain the support set for deep learning model training, where the support set is the data set of the current PID regulation target turntable, and the support set includes the operating data of turntables of different models. Use the PID-regulated data set of the same type as the current PID regulation target turntable as the query set to perform preliminary parameter adjustment on the trained deep learning model.

[0007] In step S2, after the P, I, and D values ​​generated by the deep model after preliminary parameter adjustment are input into the PID controller, the turntable is controlled and self-training is performed based on the feedback of the turntable control.

[0008] Step S3, predict the final P, I, and D values ​​using the self-trained deep learning model, and input the values ​​into the PID controller as the final PID values ​​for the current usage environment.

[0009] Furthermore, the deep learning model includes an embedding module and a relationship module, the embedding module provides feature mappings for samples in the query set and samples in the sample set; is used to convert the input samples into specific feature representations to generate feature mappings and; the feature mappings and are combined with operators, and the operator is the relationship score of the query.

[0010] Furthermore, the relationship module consists of two convolutional blocks and two fully connected layers.

[0011] The beneficial effect of the present invention compared with the prior art is: the present invention constructs a model that can output good PID parameters by setting a deep learning model and using the adjusted turntable data for training, which can solve the shortcoming that PID parameter adjustment relies on experience, improve turntable performance, and make work efficiency more efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is an exemplary step of the adjustment method of the present invention.

[0013] Figure 2 Schematic diagram of the relationship network structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the invention described is a part of the invention of the present invention, not the whole invention. Based on the invention in this embodiment, all other inventions obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0015] Example: Figure 1 A turntable PID adjustment method based on deep learning and self-training is shown, which marks the turntable after PID adjustment, records the current turntable type, runs the current turntable, collects the current turntable data through a data acquisition device, and classifies the collected data according to the turntable type; establishes a deep learning model, uses the current turntable operation data, and performs self-training;

[0016] The training process includes the following steps:

[0017] Step S1, obtain the support set for deep learning model training, the support set is the data set of the current PID adjustment target turntable, the support set includes the operating data of turntables of different models, and use the PID-adjusted data set of the same type as the current PID adjustment target turntable as the query set to perform preliminary parameter adjustment on the trained deep learning model.

[0018] In step S2, after the P, I, and D values ​​generated by the deep model after preliminary parameter adjustment are input into the PID controller, the turntable is controlled and self-training is performed based on the feedback of the turntable control.

[0019] Step S3, predict the final P, I, and D values ​​using the self-trained deep learning model, and input the values ​​into the PID controller as the final PID values ​​for the current usage environment.

[0020] like Figure 2 As shown in FIG. 1 , in this embodiment, the deep learning model consists of two modules: an embedding module f and a relationship module g. j and sample x in sample set S i It is provided by the embedding module f, which generates the feature map f(x i ) and f(x j ). Feature map f(x i ) and f(x j ) and the operator C(f(x i ),f(x j ))combined.

[0021] In this example, we assume that C(·) is the concatenation of deep feature maps, although other choices are possible. The combined feature map of the sample and the query is input into the relation module g, which ultimately generates a scalar of 0 to 1, representing x i and x j The similarity between them is called the relation score. Therefore, in the one-shot setting of C-way, for a query input x j and training sample set example x i The relationship between them generates a relationship score C i,j .

[0022]

[0023] The model is trained using mean squared error loss and the relationship score r i,j Regression: matched pairs have similarity 1, mismatched pairs have similarity 0:

[0024]

[0025] The relation module consists of two convolutional blocks and two fully connected layers. Each convolutional block is a 3*3 convolution with 64 filters, followed by batch normalization, ReLU linearization, and 2*2 maximum pooling.

[0026] All fully connected layers are ReLU except the output layer which is Sigmoid in order to generate relation scores within a reasonable range for all versions of the network architecture.

[0027] Specifically, this paper proposes a two-branch relational network (Two-branch RN), which achieves few-shot recognition by learning to compare the query dataset and the few-shot labeled sample set.

[0028] First, the embedding module generates representations of query and training images.

[0029] These embeddings are then compared through a relation module that determines whether they are from matching categories.

[0030] A scenario-based strategy is defined, and the embedding and relation modules are meta-learned end-to-end to support few-shot learning.

[0031] The error e between the target value Xtarget and the output value Xout and the rate of change of e As input to the method of the present invention;

[0032] In this embodiment, PID parameter training is optimized by self-training. Given a set of training sets {x i} n And the deep network f θ (·) is parameterized by θ, the method of the present invention records the training target {t i} n So, for all data points, we first get the predictions of the deep network:

[0033] p i =ρ(f θ (x i )) (1)

[0034] where ρ(·) is a normalization function. Then, the training objective tracks all historical model predictions during training and is updated by an exponential moving average scheme such as:

[0035] t i ←α×t i +(1-α)×p i (2)

[0036] The exponential moving average scheme in equation (2) alleviates the instability problem of model prediction and smoothes t during training. i process, and enables the algorithm of the present invention to completely change the training labels. The momentum term α controls the weight of the model prediction. Finally, the present invention can update the deep network f based on stochastic gradient θ The weight θ of each training iteration is the loss function L(p i ,t i ,f θ )’s decline.

[0037] The method of this embodiment predicts p, i, and d based on the input and gives a set of p, i, and d values;

[0038] Then, motion control is performed, and self-training is performed according to the self-training data composed of the current p, i, and d values; the above steps are repeated in this way; finally, the PID values ​​kp, ki, and kd predicted and output by this method are input into the PID controller as the final PID values.

[0039] The above contents are merely examples and explanations of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the invention or exceed the scope defined by the claims, they shall all fall within the protection scope of the present invention.

Claims

1. A turntable PID adjustment method based on deep learning and self-training, characterized in that: Mark the turntable after PID adjustment, record the current turntable type, run the current turntable, collect the current turntable data through the data acquisition device, and classify the collected data according to the turntable type; Establish a deep learning model and use the current turntable operation data for self-training; The training process includes the following steps: Step S1, obtaining a support set for deep learning model training, wherein the support set is a data set of the current PID regulation target turntable, and the support set includes operating data of turntables of different models, and a data set of the same type as the current PID regulation target turntable that has been PID regulated is used as a query set to perform preliminary parameter adjustment on the trained deep learning model; Step S2, after the P, I, and D values ​​generated by the deep model after preliminary parameter adjustment are input into the PID controller, the turntable is controlled, and self-training is performed based on the feedback of the turntable control; Step S3, predict the final P, I, and D values ​​using the self-trained deep learning model, and input the values ​​into the PID controller as the final PID values ​​for the current usage environment.

2. The turntable PID adjustment method based on deep learning and self-training according to claim 1, characterized in that: The deep learning model includes an embedding module and a relationship module. The embedding module provides feature mappings for samples in a query set and samples in a sample set; is used to convert input samples into specific feature representations to generate feature mappings and; the feature mappings and are combined with an operator, and the operator is a relationship score for the query.

3. The turntable PID adjustment method based on deep learning and self-training according to claim 2 is characterized in that: The relation module consists of two convolutional blocks and two fully connected layers.

Images