Automatic control device based on neural network algorithm model

Abstract

The invention relates to the technical field of industrial automation control, in particular to an automation control device based on a neural network algorithm model, which comprises a signal acquisition module, a neural network control module, a driving output module and an online learning module, one end of the signal acquisition module is detachably connected with an industrial field sensor, the other end of the signal acquisition module is electrically connected with the neural network control module and the online learning module respectively, and the signal acquisition module is used for acquiring a process variable PV of a controlled object and a set value SP set by a user in real time; according to the method, a nonlinear time-varying system can be effectively adapted without establishing an accurate mathematical model of a controlled object through the nonlinear mapping characteristic of the multi-layer feedforward neural network. Experiments prove that in a thermostat temperature control scene, compared with a traditional PID controller, the control precision of the device is improved by 35%, the steady-state error is smaller than or equal to + / -0.2 DEG C, the overshoot is smaller than or equal to 3%, and the steady-state error of + / -0.5 DEG C and the overshoot of 8% of the PID controller are far better.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial automation control, in particular to an automatic control device based on a neural network algorithm model. BACKGROUND

[0002] With the advancement of the industrial 4.0 process, the dynamic characteristics of the controlled object are increasingly complex, and generally exhibit nonlinearity, time-varying, large delay and other characteristics. The traditional automatic control device is mainly based on a PID controller, and its core principle is to realize control based on the linear combination of the proportion (P), integral (I) and differential (D) of the error. In a simple linear system, it can meet the basic requirements, but it has significant limitations in complex scenarios:

[0003] Parameter tuning depends on experience and has poor adaptability: the P, I and D parameters of the PID controller need to be manually tuned according to the characteristics of the controlled object. For example, for temperature control of a thermostat, the parameters need to be repeatedly tested and adjusted to balance the response speed and overshoot. Once the characteristics of the controlled object change (such as a decrease in heating efficiency due to equipment aging or fluctuations in ambient temperature), the original parameters cannot be adapted and need to be re-tuned, which not only consumes time and effort, but also may cause the control performance to deteriorate sharply, resulting in excessive temperature overshoot or excessive steady-state error.

[0004] Insufficient non-linear processing capability: the PID controller is essentially a linear control algorithm and needs to be designed based on a linear approximation mathematical model of the controlled object. However, it is difficult to establish an accurate linear model for actual industrial systems (such as the non-linear relationship between reaction rate and temperature or pressure in a chemical reaction kettle), resulting in low PID control precision and weak anti-interference capability. For example, in multi-joint robot motion control, the coupling nonlinearity between joints can cause trajectory deviation in PID control, affecting the work accuracy.

[0005] Lack of self-learning and dynamic adjustment capability: the traditional control device cannot sense system changes in real time, and when faced with external disturbances (such as unstable heating power caused by fluctuations in power grid voltage) or internal parameter drift, it cannot adjust the control strategy online and can only rely on manual intervention, which makes it difficult to meet the requirements of industrial production for control stability and continuity. Therefore, an automatic control device based on a neural network algorithm model is proposed. SUMMARY

[0006] In view of the above, the present application provides an automatic control device based on a neural network algorithm model to solve or alleviate the technical problems in the prior art and at least provide a beneficial option.

[0007] The technical solution of the present application is as follows: an automatic control device based on a neural network algorithm model, comprising a signal acquisition module, a neural network control module, a drive output module and an online learning module, which form a closed-loop control link in sequence: The signal acquisition module is detachably connected with an industrial field sensor at one end and is electrically connected with the neural network control module and the online learning module at the other end, and is used for collecting process variables PV of a controlled object and a set value SP set by a user in real time, calculating a deviation e(t) = SP - PV, extracting deviation historical data e(t-1), e(t-2), and synchronously transmitting the above data to the neural network control module and the online learning module; The neural network control module is a core control unit, and a multilayer feedforward neural network model is internally constructed, wherein the model comprises an input layer, a hidden layer and an output layer, the input layer receives deviation data transmitted by the signal acquisition module, the hidden layer adopts a ReLU nonlinear activation function to process nonlinear characteristics of a system, and the output layer outputs a control signal u(t), wherein the u(t) is an increment Δu(t) or a total value, and the module can load a pre-trained model at an initial stage or start online training from random weights; The drive output module is electrically connected with the neural network control module at one end and is detachably connected with an industrial actuator at the other end, and is used for converting the control signal u(t) output by the neural network control module into a standard industrial signal such as a 4-20 mA current signal or a 0-10 V voltage signal, and driving the actuator to act. The online learning module is bidirectionally electrically connected with the signal acquisition module and the neural network control module, and comprises a performance evaluation unit and a weight updating unit, the performance evaluation unit calculates control performance indexes such as mean square error (MSE) according to deviation data transmitted by the signal acquisition module, and the weight updating unit adopts a back propagation algorithm or a recursive least square method to dynamically adjust connection weights and bias parameters of a neural network in the neural network control module, with the aim of reducing deviation.

[0008] Further preferably, the signal acquisition module comprises a signal filtering unit and a data buffering unit, the signal filtering unit adopts a Kalman filtering algorithm to perform noise reduction processing on process variables PV collected by a sensor, and the data buffering unit is used for storing deviation historical data e(t-n) of at least 3 sampling periods, wherein n ≥ 2.

[0009] Further preferably, the input layer of the multilayer feedforward neural network model has 3-5 nodes, and node inputs include a current deviation e(t) and deviation historical data of the previous 1-3 sampling periods; the hidden layer has 1-2 layers, each layer has 5-10 nodes, and the activation function is a ReLU function; the output layer has 1 node, and the activation function is a Sigmoid function, which limits the output range of the control signal u(t) to 0-1.

[0010] Further preferably, the weight updating unit of the online learning module is provided with a learning rate adjustment mechanism, when the control performance index MSE is greater than a preset threshold, the learning rate is adjusted to 0.01-0.05; when MSE is less than or equal to the preset threshold, the learning rate is reduced to 0.001-0.01, to avoid parameter oscillation.

[0011] Further preferably, the driving output module comprises a signal conversion unit and a fault detection unit, the fault detection unit monitors the feedback signal of the actuator in real time, and when detecting actuator failure, sends an interrupt signal to the neural network control module, suspends the control output and triggers an alarm.

[0012] Further preferably, a human-computer interaction module is further included, which is electrically connected with the signal acquisition module and the neural network control module, for displaying process variables PV, set values SP, control signals u(t) and control performance indexes, and supporting user manual adjustment of neural network parameters and control thresholds.

[0013] The automatic control method based on the neural network algorithm model comprises the following steps: S2: the signal acquisition module starts sensor data acquisition, obtains process variables PV after filtering processing, calculates the deviation e(t)=SP-PV, and calls historical deviation data e(t-1), e(t-2); S3: the neural network control module receives the deviation data, and calculates the output control signal u(t) through a multi-layer feedforward neural network; S4: the driving output module converts the control signal u(t) into a standard industrial signal, drives the actuator to act, and adjusts the state of the controlled object; S5: the online learning module calculates the control performance index according to the deviation data, if the index does not meet the preset requirement, the weight updating algorithm is started to adjust the neural network connection weight and bias; S6: steps S2-S5 are repeated to realize closed-loop adaptive control.

[0014] Further preferably, the control performance index in step S5 comprises mean square error MSE, overshoot σ% and response time t_s, when MSE≤0.5℃² (temperature control scenario), σ%≤5% and t_s≤10min, it is determined that the control performance meets the preset requirement.

[0015] The embodiment of the application has the following advantages due to the adoption of the above technical solutions: One, the nonlinear mapping characteristics of the multilayer feedforward neural network can effectively adapt to nonlinear and time-varying systems without establishing an accurate mathematical model of the controlled object. Experiments have verified that in the temperature control scene of the thermostat, the control precision of the device is improved by 35% compared with the traditional PID controller, the steady-state error is ≤±0.2℃, and the overshoot is ≤3%, which is much better than the ±0.5℃ steady-state error and 8% overshoot of the PID controller.

[0016] Two, the online learning module of the application can monitor the control performance in real time, automatically adjust the neural network parameters, adapt to the changes in the characteristics of the controlled object and external disturbances, and through the modular design and parameter adjustable function, it can be widely used in temperature, pressure, motion and other industrial control scenes. When the controlled object is replaced, only the neural network parameters need to be adjusted or the corresponding pre-trained model needs to be loaded, without the need to redesign the hardware. At the same time, the visual interface and mode switching function of the human-computer interaction module reduce the operation threshold and adapt to operators with different technical levels.

[0017] The above summary is only for the purpose of the description and is not intended to limit in any way. In addition to the illustrative aspects, embodiments and features described above, further aspects, embodiments and features of the present application will be apparent from the drawings and the following detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0019] Fig. 1 The principle structure block diagram of the automatic control device of the present application; Fig. 2 The model structure schematic diagram of the neural network control module of the present application. DETAILED DESCRIPTION

[0020] In the following, only some exemplary embodiments are simply described. As those skilled in the art can recognize, the described embodiments can be modified in various different ways without departing from the spirit or scope of the present application. Therefore, the drawings and the description are considered to be exemplary in nature rather than limiting.

[0021] The embodiments of the present application will be described in detail below with reference to the drawings.

[0022] As Figs. 1-2As shown, the embodiment of the present application provides an automatic control device based on a neural network algorithm model, which comprises a signal acquisition module, a neural network control module, a driving output module and an online learning module, and each module is sequentially cooperated to form a closed-loop control link: The signal acquisition module is detachably connected with an industrial field sensor at one end and is electrically connected with the neural network control module and the online learning module at the other end, is used for collecting process variables PV of a controlled object and set values SP set by a user in real time, calculating a deviation e(t) = SP - PV, extracting deviation historical data e(t-1), e(t-2), and synchronously transmitting the above data to the neural network control module and the online learning module; The neural network control module is a core control unit, internally constructs a multi-layer feedforward neural network model, the model comprises an input layer, a hidden layer and an output layer, the input layer receives the deviation data transmitted by the signal acquisition module, the hidden layer adopts a ReLU nonlinear activation function to process the nonlinear characteristics of the system, and the output layer outputs a control signal u(t), u(t) is an increment Δu(t) or a total value, and the module can load a pre-trained model at an initial stage or start online training from random weights; The driving output module is electrically connected with the neural network control module at one end and is detachably connected with an industrial execution mechanism at the other end, is used for converting the control signal u(t) output by the neural network control module into a standard industrial signal such as a 4-20 mA current signal or a 0-10 V voltage signal, and driving the execution mechanism to act; The online learning module is bidirectionally electrically connected with the signal acquisition module and the neural network control module, comprises a performance evaluation unit and a weight updating unit, the performance evaluation unit calculates control performance indexes such as mean square error MSE according to the deviation data transmitted by the signal acquisition module, the weight updating unit adopts a back propagation algorithm or a recursive least square method to dynamically adjust the connection weights and bias parameters of the neural network in the neural network control module with the optimization target of reducing the deviation.

[0023] In one embodiment, the signal acquisition module comprises a signal filtering unit and a data buffering unit, the signal filtering unit adopts a Kalman filtering algorithm to perform noise reduction processing on the process variables PV collected by the sensor, and the data buffering unit is used for storing deviation historical data e(t-n) of at least 3 sampling periods, n≥2.

[0024] In one embodiment, the input layer of the multi-layer feedforward neural network model has 3-5 nodes, the node input includes the current deviation e(t) and the deviation historical data of the previous 1-3 sampling periods, the hidden layer is provided with 1-2 layers, each layer has 5-10 nodes, and the activation function is a ReLU function; the output layer has 1 node, the activation function is a Sigmoid function, and the output range of the control signal u(t) is limited to 0-1.

[0025] In one embodiment, the weight updating unit of the online learning module sets a learning rate adjustment mechanism, when the control performance index MSE > preset threshold, the learning rate is adjusted to 0.01-0.05; when MSE ≤ preset threshold, the learning rate is reduced to 0.001-0.01, avoiding parameter oscillation.

[0026] In one embodiment, the drive output module includes a signal conversion unit and a fault detection unit, the fault detection unit monitors the feedback signal of the actuator in real time, and sends an interrupt signal to the neural network control module when the actuator fails, suspends the control output and triggers an alarm.

[0027] In one embodiment, a human-computer interaction module is further included, which is electrically connected with the signal acquisition module and the neural network control module, for displaying process variables PV, set values SP, control signals u(t) and control performance indexes, and supporting user manual adjustment of neural network parameters and control thresholds.

[0028] The automatic control method based on the neural network algorithm model, S1: device initialization, selecting a neural network pre-training model according to the type of the controlled object or setting random initial weights, inputting a set value SP through a human-computer interaction module; S2: the signal acquisition module starts sensor data acquisition, obtains process variables PV after filtering processing, calculates the deviation e(t)=SP-PV, and calls historical deviation data e(t-1), e(t-2); S3: the neural network control module receives the deviation data, calculates the output control signal u(t) through a multi-layer feedforward neural network; S4: the drive output module converts the control signal u(t) into a standard industrial signal to drive the actuator to act and adjust the state of the controlled object; S5: the online learning module calculates the control performance index according to the deviation data, if the index does not meet the preset requirement, the weight updating algorithm is started to adjust the neural network connection weight and bias; S6: repeating steps S2-S5, realizing closed-loop adaptive control.

[0029] In one embodiment, the control performance index in step S5 includes mean square error MSE, overshoot σ% and response time t_s, when MSE ≤ 0.5℃² (temperature control scenario), σ% ≤ 5% and t_s ≤ 10min, it is determined that the control performance meets the preset requirement.

[0030] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An automated control device based on a neural network algorithm model, characterized in that: It includes a signal acquisition module, a neural network control module, a drive output module, and an online learning module. These modules work together in sequence to form a closed-loop control chain. Signal acquisition module: One end is detachably connected to the industrial field sensor, and the other end is electrically connected to the neural network control module and the online learning module respectively. It is used to acquire the process variable PV of the controlled object and the user-set value SP in real time, calculate the deviation e(t)=SP-PV, and extract the historical deviation data e(t-1) and e(t-2). The above data are synchronously transmitted to the neural network control module and the online learning module. Neural network control module: As the core control unit, it internally constructs a multi-layer feedforward neural network model, which includes an input layer, a hidden layer and an output layer. The input layer receives the deviation data transmitted by the signal acquisition module, the hidden layer uses the ReLU nonlinear activation function to process the nonlinear characteristics of the system, and the output layer outputs the control signal u(t), where u(t) is the increment Δu(t) or the full value. In the initial stage, this module can load a pre-trained model or start online training from random weights. Drive output module: One end is electrically connected to the neural network control module, and the other end is detachably connected to the industrial actuator. It is used to convert the control signal u(t) output by the neural network control module into standard industrial signals such as 4-20mA current signal or 0-10V voltage signal to drive the actuator to move. Online learning module: It is bidirectionally electrically connected to the signal acquisition module and the neural network control module, and includes a performance evaluation unit and a weight update unit. The performance evaluation unit calculates control performance indicators such as mean square error (MSE) based on the deviation data transmitted by the signal acquisition module. The weight update unit uses the backpropagation algorithm or recursive least squares method to dynamically adjust the connection weights and bias parameters of the neural network in the neural network control module with the goal of reducing deviation.

2. The automated control device based on a neural network algorithm model according to claim 1, characterized in that: The signal acquisition module includes a signal filtering unit and a data buffer unit. The signal filtering unit uses the Kalman filtering algorithm to reduce noise in the process variable PV acquired by the sensor. The data buffer unit is used to store the historical deviation data e(tn) for at least 3 sampling periods, where n≥2.

3. The automated control device based on a neural network algorithm model according to claim 1, characterized in that: The multilayer feedforward neural network model has 3 to 5 nodes in the input layer, and the node inputs include the current deviation e(t) and the historical deviation data of the previous 1 to 3 sampling periods; the hidden layer has 1 to 2 layers, each with 5 to 10 nodes, and the activation function is the ReLU function; the output layer has 1 node, and the activation function is the Sigmoid function, which limits the output range of the control signal u(t) to between 0 and 1.

4. The automated control device based on a neural network algorithm model according to claim 1, characterized in that: The weight update unit of the online learning module is equipped with a learning rate adjustment mechanism. When the control performance index MSE is greater than the preset threshold, the learning rate is adjusted to 0.01-0.05; when MSE is less than or equal to the preset threshold, the learning rate is reduced to 0.001-0.01 to avoid parameter oscillation.

5. The automated control device based on a neural network algorithm model according to claim 1, characterized in that: The drive output module includes a signal conversion unit and a fault detection unit. The fault detection unit monitors the feedback signal of the actuator in real time. When a fault is detected in the actuator, it sends an interrupt signal to the neural network control module to suspend the control output and trigger an alarm.

6. The automated control device based on a neural network algorithm model according to claim 1, characterized in that: It also includes a human-computer interaction module, which is electrically connected to the signal acquisition module and the neural network control module, respectively. It is used to display process variables PV, setpoints SP, control signals u(t) and control performance indicators, and supports users to manually adjust neural network parameters and control thresholds.

7. An automated control method based on a neural network algorithm model, coupled with an automated control device based on a neural network algorithm model as described in any one of claims 1-6, characterized in that: S1: Device initialization, select a neural network pre-trained model or set random initial weights according to the type of controlled object, and input the set value SP through the human-computer interaction module; S2: The signal acquisition module starts sensor data acquisition, obtains process variable PV after filtering, calculates deviation e(t)=SP-PV, and calls historical deviation data e(t-1) and e(t-2). S3: The neural network control module receives the deviation data and calculates the output control signal u(t) through a multi-layer feedforward neural network. S4: The drive output module converts the control signal u(t) into a standard industrial signal to drive the actuator and adjust the state of the controlled object; S5: The online learning module calculates control performance indicators based on deviation data. If the indicators do not meet the preset requirements, the weight update algorithm is activated to adjust the neural network connection weights and biases. S6: Repeat steps S2-S5 to achieve closed-loop adaptive control.

8. The automatic control method based on a neural network algorithm model according to claim 7, characterized in that: The control performance indicators mentioned in step S5 include mean square error (MSE), overshoot σ%, and response time t_s. When MSE ≤ 0.5℃² (temperature control scenario), σ% ≤ 5%, and t_s ≤ 10min, the control performance is determined to meet the preset requirements.

Images