Four-way thermoelectric temperature control system based on residual neural network adaptive PID

Abstract

The application relates to the field of automatic control and embedded systems, and discloses a four-way thermoelectric temperature control system based on a residual neural network adaptive PID, which comprises a main control unit, a temperature sensing unit, a power driving unit, a heat dissipation control unit, a Res-NNPID control algorithm module, a man-machine interaction unit and a remote monitoring unit. The main control unit uses the Res-NNPID module to independently maintain neural network weights for four channels, uses a network topology structure containing residual skip connections to real-time set PID parameters, and combines a real-time operating system to perform multi-task parallel scheduling. The application optimizes the online training stability of the neural network through the residual structure, effectively alleviates the gradient vanishing problem in the deep network, realizes high-precision adaptive adjustment for different temperature control objects, and ensures accurate temperature control and real-time synchronization of operation data among the channels without interference through the cooperation of the multi-way independent control logic and the MQTT communication protocol.

Description

Technical Field

[0001] This invention relates to the field of automatic control and embedded systems, specifically to a four-channel thermoelectric temperature control system based on residual neural network adaptive PID. Background Technology

[0002] Thermoelectric temperature control technology primarily utilizes the Peltier effect to achieve reversible cooling or heating, and is commonly used in experimental and testing platforms requiring temperature regulation. Existing thermoelectric temperature control systems mostly employ fixed-parameter proportional-integral-derivative (PID) controllers, whose control parameters remain unchanged after initial debugging. However, because the controlled object is subject to nonlinear interference from ambient temperature fluctuations, thermoelectric module aging, and changes in load conditions during actual operation, fixed-parameter controllers struggle to maintain stable control performance in dynamically changing environments, often leading to increased system overshoot, prolonged settling time, and increased steady-state error.

[0003] To address the shortcomings of fixed-parameter controllers, some technical solutions have introduced neural network self-tuning control architectures, attempting to adjust control parameters in real time through online learning. However, in practical applications, conventional feedforward neural networks are prone to gradient vanishing or exploding phenomena as their computational depth increases. This leads to extreme instability in the network training process within the control loop, causing fluctuations in the output parameters and compromising the reliability of temperature control equipment during long-term operation.

[0004] Furthermore, in multi-channel temperature control applications, existing temperature control systems often employ an architecture where each channel shares control parameters or uses a single controller for time-sharing multiplexing. This structure cannot perform differentiated adaptive parameter tuning based on the independent thermodynamic characteristics and target temperature of each channel, making it difficult to achieve true independent control for each channel. At the same time, these traditional multi-channel temperature control devices have weak data presentation capabilities in human-machine interfaces and data interaction capabilities for IoT remote monitoring, making it difficult to meet the needs of multi-channel synchronous control and real-time data traceability in complex experimental scenarios. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a four-channel thermoelectric temperature control system based on residual neural network adaptive PID, which solves the problems of low adjustment accuracy and poor adaptability of traditional fixed parameter PID under nonlinear complex working conditions, as well as the problems of gradient degradation and multi-task resource conflicts that are easy to occur in conventional neural network self-tuning algorithms in embedded multi-channel systems.

[0006] To achieve the above objectives, the present invention provides a four-channel thermoelectric temperature control system based on residual neural network adaptive PID, including a main control unit, a temperature sensing unit, a Res-NNPID control algorithm module, a power drive unit, a heat dissipation control unit, a human-machine interaction unit, and a remote monitoring unit.

[0007] The main control unit, as the core control and scheduling node of the system, achieves parallel management of multiple tasks by running a real-time operating system. The main control unit establishes electrical connections or communication links with the temperature sensing unit, power drive unit, human-machine interface unit, and remote monitoring unit. The Res-NNPID control algorithm module is deployed in the main control unit as logic code. The temperature sensing unit is used to acquire the feedback temperature of the controlled objects through four independent channels in real time. The Res-NNPID control algorithm module achieves online parameter tuning based on the feedback temperature and the characteristic quantities of the set target through a neural network with a residual structure. The power drive unit outputs an adjustable duty cycle drive current to the thermoelectric cooler according to the algorithm decision. The heat dissipation control unit operates synchronously with the power drive unit to perform auxiliary heat exchange. The human-machine interface unit and the remote monitoring unit provide local and remote task configuration and status monitoring interfaces, respectively.

[0008] As a preferred solution, the main control unit employs a dual-core processor, utilizing the processor's multi-core parallel capabilities to allocate hardware resources. Through the task scheduler of the real-time operating system, the system adopts a symmetric multiprocessing (SMP) architecture, automatically and dynamically scheduling cross-core tasks between the two processor cores. The operating system automatically load-balances the temperature control algorithm task, high-frequency acquisition task, network communication task, and graphical interface display task according to a preset priority mechanism. This dual-core dynamic scheduling mechanism maximizes the utilization of hardware computing power, and combined with a priority preemption strategy, effectively avoids sudden loads on the network protocol stack from interfering with the real-time performance of temperature control, ensuring the time determinism of the control loop.

[0009] As a preferred approach, the temperature sensing unit is powered by a two-stage buck regulator topology, utilizing a low-dropout linear regulator to suppress power supply ripple and support the high-resolution sensor's acquisition accuracy. The main control unit, based on the filtered temperature data, extracts the current temperature, target temperature, temperature change rate, and temperature deviation from the previous cycle, and performs normalization processing to construct the input feature vector required by the algorithm.

[0010] As a preferred approach, the Res-NNPID control algorithm module maintains independent neural network weight spaces for each of the four temperature control channels. The core algorithm employs a four-layer neural network topology, introducing residual skip connections in the hidden layers. This structure allows input information to bypass some nonlinear transformation layers and be directly transmitted, effectively mitigating the gradient vanishing problem that easily occurs in deep networks during embedded online training, thereby improving the algorithm's convergence speed and training stability.

[0011] As a preferred solution, the Res-NNPID control algorithm module integrates a confidence assessment function. The system calculates a sampling confidence factor based on a preset temperature measurement accuracy constant and the current instantaneous deviation. This factor is used to dynamically adjust the learning rate and residual weights of the neural network. When the system approaches steady state and the temperature difference is extremely small, the system reduces the learning rate to suppress parameter misadjustments caused by measurement noise, ensuring the stability of control in the steady-state region.

[0012] As a preferred approach, the output layer of the neural network generates normalized adjustment values ​​for the PID parameters using a hyperbolic tangent activation function. The main control unit linearly maps these adjustment values ​​to preset baseline parameter values ​​to generate the actual proportional, integral, and derivative parameters, and performs non-negativity verification to conform to the physical control logic. During backpropagation weight updates, the system performs gradient pruning to limit the magnitude of each weight update. The main control unit continuously monitors the weight value status; if calculation overflow or invalid values ​​occur, a weight reset mechanism is automatically triggered.

[0013] As a preferred solution, the power drive unit employs a bus-controlled pulse width modulation (PWM) extension chip, coupled with four independent H-bridge circuits to achieve current polarity switching and power regulation. The main control unit performs amplitude limiting on the calculated control quantity and determines the operating mode of the thermoelectric cooler based on its positive or negative polarity. The control quantity is mapped to a high-resolution duty cycle signal, which drives the thermoelectric cooler to perform heating or cooling actions through the H-bridge circuits.

[0014] As a preferred solution, the human-machine interface unit utilizes a graphics library to construct a multi-channel monitoring interface on the display medium, used for real-time digital display of the current temperature, target setpoint, and start / stop status of each channel. The remote monitoring unit accesses the cloud platform via a wireless network and uses a message queue transmission protocol to serialize and upload the operating data of each channel. The cloud platform deploys a web monitoring page built on HTML, supporting dynamic rendering and display of historical temperature curves for multiple channels and offline download of operating data. It can also receive control commands from remote terminals in real time, enabling parameter tuning and status switching across geographical locations.

[0015] Through the above technical solution, this invention achieves multi-channel independent thermoelectric temperature control with adaptive capabilities. The residual neural network solves the problem of instability in embedded systems caused by traditional self-tuning algorithms. Combined with multi-task scheduling and high-precision signal chain design, it improves the dynamic response speed, steady-state accuracy, and system integration of the temperature control system under complex disturbance environments.

[0016] This invention provides a four-channel thermoelectric temperature control system based on residual neural network adaptive PID. It has the following beneficial effects: 1. This invention effectively suppresses the vanishing and exploding gradient phenomena during network training by introducing a neural network model with residual skip connections to update PID parameters online in real time. Combined with gradient pruning and weight outlier reset mechanisms, the system can adapt to nonlinear disturbances such as changes in thermoelectric module characteristics and ambient temperature fluctuations. Compared with traditional fixed-parameter PID controllers, this structure shortens the settling time and reduces steady-state error, ensuring the stability of the control loop during long-term operation.

[0017] 2. This invention achieves multi-channel parallel computing and independent control without interference by assigning and maintaining independent neural network weight parameters, integral terms, and control variables for the four temperature control channels. Combined with the real-time operating system and multi-core dynamic scheduling mechanism of the main control unit, it ensures that each channel independently completes parameter tuning and power mapping within the set control cycle, enabling the system to simultaneously perform precise temperature regulation on four controlled objects with different target temperatures.

[0018] 3. This invention integrates a high-precision temperature sensing unit with an IoT communication module to construct a complete closed-loop link from bottom-level data acquisition to cloud-based remote management. The high-precision temperature feedback provided by the front-end sensor provides reliable data for the adaptive learning of the neural network; the back-end, combined with the MQTT protocol and a local touch interface, realizes real-time visualization of the operating status of each channel and command issuance, improving the hardware and software integration of the multi-channel temperature control system and the equipment debugging efficiency. Attached Figure Description

[0019] Figure 1 This is a system overall framework diagram of one embodiment of the present invention.

[0020] Figure 2 This is a top-level layout diagram of a printed circuit board (PCB) according to an embodiment of the present invention.

[0021] Figure 3 This is a system schematic diagram of one embodiment of the present invention.

[0022] Figure 4 This is a photograph of a physical system according to an embodiment of the present invention; Figure 5 This is a flowchart of a four-channel thermoelectric temperature control system based on residual neural network adaptive PID according to an embodiment of the present invention; Figure 6 This is a comparison diagram of the step response of Res-NNPID and conventional PID according to an embodiment of the present invention; Figure 7 This is the operating curve of the first temperature control channel according to an embodiment of the present invention; Figure 8 This is the operating curve of the second temperature control channel according to an embodiment of the present invention; Figure 9 This is the operating curve of the third temperature control channel according to an embodiment of the present invention; Figure 10 This is the operating curve of the fourth temperature control channel according to an embodiment of the present invention; Figure 11 This is a comparison chart of the thermal disturbance resistance performance of one embodiment of the present invention; Among them, 10 is the main control and drive circuit board; 20 is the battery; 30 is the thermoelectric temperature control assembly; 40 is the electrical connection harness; and 50 is the external interface terminal. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] See attached document Figure 1 This invention provides a four-channel thermoelectric temperature control system based on residual neural network adaptive PID, which may include: a main control unit; a temperature sensing unit; a Res-NNPID control algorithm module; a power drive unit; a heat dissipation control unit; a human-machine interaction unit; and a remote monitoring unit.

[0025] The main control unit, as the core scheduling component of the system, is connected to the temperature sensing unit, power drive unit, human-machine interface unit, and remote monitoring unit. The Res-NNPID control algorithm module resides in the main control unit's storage medium as software code. The heat dissipation control unit is located at the execution end of the power drive unit to assist in heat exchange.

[0026] The temperature sensing unit comprises four independent acquisition channels, with each channel's sensor electrically connected to the signal input interface of the main control unit. The power drive unit receives control commands from the main control unit via a communication bus and converts these commands into power signals to drive the thermoelectric cooler. The human-machine interface unit and the remote monitoring unit constitute the system's local and remote configuration interfaces, respectively, supporting the setting of temperature control targets and the reading of their status.

[0027] See attached document Figure 4The battery 20 serves as the system's power supply module, providing basic electrical energy to the main control and drive circuit board 10. The main control and drive circuit board 10 physically integrates the main control unit and the power drive unit. Four thermoelectric temperature control components 30 are independently distributed, corresponding to the system's four temperature control channels, and each contains a thermoelectric cooling chip, a cooling fan, and a temperature sensor. Each thermoelectric temperature control component 30 establishes an electrical connection and signal interaction with the main control and drive circuit board 10 via an electrical connection harness 40. External interface terminals 50 are located at the edge of the circuit board for system debugging and external power supply access.

[0028] See attached document Figure 5 This invention provides a method for operating a four-channel thermoelectric temperature control system based on residual neural network adaptive PID, comprising the following steps: S10, System Initialization. The main control unit powers on and loads the real-time operating system kernel, allocates computing memory for the Res-NNPID control algorithm module, and sets the neural network weights to preset initial values.

[0029] S20, Data Acquisition. The temperature sensing unit acquires the current temperature signal of the controlled object, converts it into real-time temperature data in digital format, and transmits it to the main control unit. The main control unit directly extracts this real-time temperature data for subsequent calculations.

[0030] S30, Feature Vector Construction. The main control unit combines the target temperature data obtained from the human-machine interaction unit or remote monitoring unit, including the current temperature, target temperature, temperature change rate, and temperature deviation from the previous cycle. The main control unit constructs these parameters into the input feature vector required by the algorithm.

[0031] S40, Adaptive Parameter Calculation. The main control unit inputs the input feature vector into the Res-NNPID control algorithm module. The Res-NNPID control algorithm module uses a neural network model with residual jump connections to perform inference calculations and outputs proportional parameters, integral parameters, and derivative parameters.

[0032] S50, Control Variable Mapping and Drive. The main control unit calculates the target control variable based on proportional, integral, and derivative parameters, and generates a PWM signal with the corresponding duty cycle, which is then sent to the power drive unit. The power drive unit adjusts the current flowing through the thermoelectric cooler via an H-bridge circuit to change the temperature of the controlled object.

[0033] S60, adaptive heat dissipation adjustment. The heat dissipation control unit receives a synchronization signal from the main control unit and adjusts the cooling fan speed according to the real-time operating power of the thermoelectric cooler to maintain heat exchange efficiency.

[0034] S70, status feedback and remote synchronization. The main control unit sends the operating data to the human-machine interaction unit 60 for local display, and then uploads the serialized data to the cloud monitoring platform through the remote monitoring unit.

[0035] See attached document Figure 4 The main control unit uses an ESP32-S3 dual-core Xtensa LX7 processor with a clock speed of 240MHz and built-in Wi-Fi and Bluetooth communication modules. The main control unit integrates SRAM memory, Flash memory, and various peripheral interfaces to support the execution of system logic.

[0036] The main control unit runs the FreeRTOS real-time operating system, managing hardware resources through a preemptive scheduling algorithm provided by the kernel. The tick timer of this real-time operating system is configured with a frequency range of 100Hz to 10000Hz, preferably 1000Hz. Under this preferred configuration, the system clock tick is 1ms to ensure that each control task has deterministic timing. The main control unit implements a multi-task parallel scheduling mechanism through the real-time operating system, decomposing the system functions into multiple independent tasks with different priorities. The specific implementation process includes the following steps: S110, Kernel Initialization and Task Creation. After system power-on, the main control unit loads the underlying hardware drivers and calls the task creation function of the real-time operating system, allocating independent stack space and task control blocks at the main program entry point. The created concurrent tasks include: the LVGLLoop task for periodic graphical interface refresh, the displayTask task for interface data update, the btnTask task for key detection, the mySensor task for temperature acquisition, the myControl task for Res-NNPID control algorithm module calculation, and the mqttLoop task for MQTT remote communication.

[0037] The S120 features hardware resource allocation and dynamic scheduling. The main control unit utilizes the parallel computing capabilities of a dual-core processor, relying on the symmetric multiprocessing (SMP) architecture of the real-time operating system to automatically perform cross-core task allocation. Based on the real-time priority of each task and the core's idle state, the system dynamically load balances frequently executed tasks such as the myControl control task and algorithm inference task, as well as tasks with fluctuating loads such as the MQTTLoop communication task and the interactive interface display task. This multi-core dynamic scheduling strategy effectively avoids interference from sudden loads on the network communication protocol stack on the real-time performance of the temperature control loop while fully leveraging hardware computing power, achieving highly efficient parallel processing at the hardware level.

[0038] S130, Task Priority Configuration. The main control unit sets differentiated priorities for different functions. Among them, the mySensor acquisition task and the myControl control task corresponding to the Res-NNPID control algorithm module are given the highest priority; the btnTask button detection task and interface refresh task corresponding to the human-machine interaction unit are given medium priority; and the MQTT communication task corresponding to the remote monitoring unit is given low priority.

[0039] S140, Multi-task periodic scheduling. During system operation, the main control unit schedules each task according to a preset execution cycle. To optimize system resource scheduling and bus bandwidth, the execution cycle Tsense of the mySensor acquisition task and the execution cycle Tcontrol of the myControl control task are synchronized. Their values ​​are set between 100ms and 5000ms, preferably 1000ms (i.e., 1 second). The main control unit ensures that task execution meets the following time constraints: ; in, Indicates the first The computation time of each temperature control channel in the Res-NNPID control algorithm module Indicates the first The signal output time of each temperature control channel in the power drive unit.

[0040] S150, inter-task communication and data synchronization. The main control unit achieves data exchange through globally shared variables. These shared variables include a `temperature[]` array storing four real-time temperatures, a `target_temp[]` array storing the target temperature, and a `btnState[]` array storing button states. During inter-task data transfer, the main control unit invokes the mutex lock mechanism provided by the real-time operating system to prevent memory conflicts when multiple tasks read or write globally shared variables.

[0041] See attached document Figure 2 Regarding the electrical connections between the main control unit and various peripherals, those skilled in the art can use standard PCB routing guidelines and impedance matching techniques to achieve signal transmission. The specific circuit routing schemes are well-known technologies in this field and will not be elaborated here. The main control unit uses its integrated general-purpose input / output interface, serial peripheral interface, and integrated circuit bus interface to achieve low-level hardware control of the temperature sensing unit, power drive unit, and human-machine interaction unit.

[0042] The main control unit's internal Flash memory is used to persistently store basic configuration information such as user-defined temperature control target values ​​and network connection credentials. For the neural network weights and bias parameters of the Res-NNPID control algorithm module, the main control unit clears them to zero and performs random initialization operations within a preset range in the SRAM dynamic storage area each time the system powers on or resets. This mechanism of re-initializing upon each power-on ensures that the control algorithm can perform online adaptive learning from scratch based on the latest thermodynamic conditions, avoiding control overshoot caused by historical weights in the initial state. When the system restarts, the main control unit automatically reads the temperature control target value from a preset address and loads it into the SRAM. After successful Wi-Fi connection, the main control unit starts the NTP time synchronization service to calibrate the system clock.

[0043] The real-time operating system running on the main control unit also includes an idle task to perform processor low-power management when all application tasks enter a blocked state. Through the aforementioned hardware and software collaborative parallel scheduling mechanism, the main control unit can ensure that the control algorithms for the four temperature control channels complete iterations within a 1-second sampling interval.

[0044] The temperature sensing unit includes four independent temperature sensors. (See attached diagram.) Figure 3 The system includes a power management module to provide stable power to the sensors and digital circuits. The external 12V DC power supply is first converted to 5V by a DC-DC step-down converter chip, and then converted to 3.3V by an LDO linear regulator. This two-stage step-down topology, combined with the ripple suppression characteristics of the low-dropout linear regulator, ensures that the power quality meets the accuracy requirements of the sensor's high-resolution data acquisition at 0.0001℃.

[0045] The signal chain implementation and data preprocessing process of the temperature sensing unit includes the following steps: S210, Sensor Arrangement and Thermal Coupling. Four high-precision digital temperature sensors are fixed to the surface of the controlled object corresponding to each channel using thermally conductive silicone grease. The sensors are professionally calibrated to ensure a theoretical temperature measurement accuracy of 0.0001℃ within a preset range. The sensors are electrically connected to the main control unit via a digital bus, converting detected temperature changes into digital signals.

[0046] S220, real-time data acquisition. The mySensor task inside the main control unit triggers the temperature sensing unit to perform data read and write operations according to a 1-second control cycle. The main control unit acquires the raw data stream output by the sensor, performs data bit-width conversion, restores it to a Celsius value, and stores it in a shared temperature[] array variable. Those skilled in the art can write the driver according to the actual selected sensor protocol. The specific timing control is a well-known technology in the field and will not be described in detail here.

[0047] S240, Input Feature Vector Construction and Normalization. The main control unit constructs the 4-dimensional input feature vector required by the algorithm based on the acquired real-time temperature data. The normalization logic is implemented as follows: Current temperature characteristics: The current real-time temperature Divide by 60 to normalize and restrict the result to the [0,1] interval.

[0048] Target temperature characteristics: The set target temperature Divide by 60 to normalize and restrict the result to the [0,1] interval.

[0049] Temperature change rate characteristic: Calculates the temperature change rate within the current acquisition period. We then restrict it to the range of [-5, 5] and then divide it by 5 for normalization.

[0050] Temperature deviation from the previous cycle: Calculate the cumulative deviation term We then restrict it to the range of [-10, 10] and then divide it by 10 to normalize it.

[0051] S250, Data Security Check and Anomaly Repair. Before the feature values ​​are passed to the algorithm module, the main control unit executes anomaly detection logic. If an anomaly such as a non-numeric value (NaN) or positive or negative infinity (Inf) appears in the input features, the main control unit automatically activates the anomaly repair mechanism to replace it with a preset safe default value to prevent numerical overflow during neural network calculations.

[0052] S260, Sampling confidence factor calculation. The Res-NNPID control algorithm module calls its internal confidence evaluation module to calculate the sampling confidence factor based on the current temperature deviation. Its calculation formula is defined as: ; in, A constant coefficient related to the preset temperature measurement accuracy. Sampling confidence factor. The value range is limited to between 0 and 1. When the real-time temperature data is close to the target temperature and the fluctuation is extremely small, The learning rate tends to its maximum value, providing a basis for subsequent dynamic adjustment of the learning rate.

[0053] The communication link between the temperature sensing unit and the main control unit employs impedance matching design to reduce reflection loss. Its physical layer connection complies with industrial-grade electrostatic discharge (ESD) protection standards. Specific interface circuit protection schemes are well-known technologies in the field and will not be elaborated upon here. Through the aforementioned signal chain processing, the system provides high-precision feedback input to the Res-NNPID control algorithm module.

[0054] The Res-NNPID control algorithm module runs inside the main control unit. This module maintains independent neural network weight parameters for each of the four channels, and achieves online adaptive PID parameter tuning by extracting features from real-time temperature feedback data.

[0055] The Res-NNPID control algorithm module employs a four-layer neural network topology, consisting of an input layer, two hidden layers, and an output layer. The input layer contains four neurons, receiving the normalized current temperature, target temperature, temperature change rate, and temperature deviation from the previous cycle, respectively. Each of the two hidden layers has eight neurons, using the ReLU linear rectified function as the activation function. To improve the stability of network training and suppress the vanishing gradient phenomenon in deep networks, the Res-NNPID control algorithm module introduces residual skip connections in each hidden layer.

[0056] The specific logic of residual skip connections is to bypass the nonlinear transformation layer and directly accumulate the input features of the hidden layer to the output of the activation function of that layer. Let the... The input vector of the hidden layer is The weight matrix is The bias term is Then the layer contains the output of the residual structure. The calculation formula is expressed as follows: ; The value 0.3 is a preset residual weight used to adjust the proportion of skip connection signals in the output features. This structure allows the input signal to be directly transmitted to the output, improving the network's convergence speed. To prevent numerical overflow during the calculation process, the calculation results of the intermediate layers are all subjected to amplitude limiting within the range of [-50, 50].

[0057] The output layer of the Res-NN PID control algorithm module contains three neurons, corresponding to the three core parameters of the PID controller. The output layer uses the hyperbolic tangent activation function (tanh function) to output three normalized adjustment values ​​within the interval (-1, 1). , , The Res-NNPID control algorithm module converts the normalized adjustment into actual proportional, integral, and derivative parameters. The conversion logic follows the mapping relationship below: ; ; ; The main control unit calculates the results. , , The parameters undergo non-negativity verification to meet physical control requirements. The specific execution process of the Res-NNPID control algorithm module is as follows: S310, Network Parameter Initialization. Upon system startup, the main control unit randomly initializes the weights and biases of the four neural network channels. The initial values ​​are limited to the range [-0.1, 0.1], preferably ±0.05. A random seed is set in conjunction with the channel number and system time to ensure that the initial weights of each channel are independent.

[0058] S320, Residual Neural Network Forward Inference. The main control unit inputs the constructed 4D normalized feature vector into the network. The Res-NNPID control algorithm module performs network operations according to the topology of the hidden layer and residual jump connections, and outputs the PID normalized adjustment.

[0059] S330, PID control quantity calculation and integral limiting. The main control unit uses the generated... , , PID control calculations are performed. The integral term undergoes saturation limiting, with the limiting range set within the range of [-5, 100], preferably ±20, to avoid system overshoot caused by integral saturation. Total control quantity... It is limited to the range [-100, 100], with positive values ​​corresponding to heating and negative values ​​corresponding to cooling.

[0060] S340, confidence-based learning rate adjustment. The Res-NNPID control algorithm module obtains the sampled confidence factor from the temperature sensing unit output. And use this factor to adjust the learning rate of the neural network. The adjustment formula is: ; In the formula, 0.002 is the baseline learning rate. When the sampling confidence factor... When the learning rate is reduced, the system simultaneously decreases the learning rate to reduce the probability of parameter misadjustment caused by temperature measurement noise.

[0061] S350, online weight backpropagation update. The Res-NNPID control algorithm module uses the current control deviation as the error signal to update the network weights. The output layer error term performs gradient clipping in the range [-10, 10]. The hidden layer error term combines the ReLU derivative (when...) The weight increment is calculated using the residual path gradient (1 if the value is 1, 0 otherwise). The increment of each weight update is limited to the range of 0.001 to 0.2, preferably 0.05.

[0062] S360, online weight backpropagation update. The Res-NNPID control algorithm module uses the current control deviation as the error signal to update the network weights. The output layer error term performs gradient clipping in the range [-10, 10]. The hidden layer error term combines the ReLU derivative (when...) The weight increment is calculated using the residual path gradient (1 if the value is 1, otherwise 0). The increment of each weight update is limited to the range of 0.001 to 0.2, preferably 0.05.

[0063] S370, Weight Anomaly Monitoring and Reset. The Res-NNPID control algorithm module performs numerical verification after each weight update. If the weight parameter is non-numerical (NaN) or infinite (Inf), the main control unit triggers a reset operation to restore the weight to its initial state and maintain the normal operation of the control loop.

[0064] For the differentiation and matrix operations involved in backpropagation of neural networks, those skilled in the art can use basic mathematical operation libraries to implement them. The specific calculation logic is a well-known technology in this field and will not be described in detail here.

[0065] Based on the above logic, the system can generate applicable PID control parameters online according to the temperature feedback characteristics of four independent channels, realize multi-channel independent temperature control, and has a certain anti-interference capability.

[0066] The power drive unit is connected to the main control unit and is used to convert the control quantity calculated by the Res-NNPID control algorithm module into a drive signal for the thermoelectric cooler. The power drive unit includes a PWM signal extension chip and four independent H-bridge circuits. In this embodiment, the PWM signal extension chip is a PCA9685 chip, which uses I... 2 The C-bus receives commands from the main control unit. The four-way H-bridge circuit is constructed using HSOP8 packaged power MOSFETs, which facilitates PCB layout and controls current switching.

[0067] The execution logic and signal conversion process of the power drive unit includes the following steps: S410, Control Input Reception and Limiting. The main control unit periodically transmits the control input output from the Res-NNPID control algorithm module to the power drive unit. Control quantity The range is set to [-100, 100], with positive values ​​representing heating and negative values ​​representing cooling. The main control unit performs a limit check on this control quantity. If the calculated result exceeds this range, the boundary value is taken; if a non-numerical anomaly (NaN) occurs, it is replaced with 0 to prevent abnormal signals from driving the power devices.

[0068] S420, Operating mode determination. The main control unit determines the operating mode based on the control quantity. The positive and negative polarities determine the operating state of the thermoelectric cooling element in each channel. When the control quantity... When the value is positive, the main control unit controls the H-bridge circuit of the corresponding channel to enter heating mode; when the control quantity is positive... When the value is negative, the control H-bridge circuit switches the current direction to enter cooling mode; when the control quantity is negative... When the output is zero, the thermoelectric cooler stops outputting power.

[0069] S430, command issuance and PWM signal generation. The main control unit uses I... 2 The C-bus is configured with a PCA9685 chip, setting the PWM signal frequency to 2kHz. The main control unit then controls the input... The absolute value is mapped to a 12-bit resolution duty cycle value. In this embodiment, the power mapping relationship is: the absolute value of the control quantity is within the range of [1000, 4095], preferably 2000. The specific duty cycle value... The calculation formula is: ; In the formula, This is the integer value written to the PCA9685 chip register. Through this mapping logic, the system limits the maximum output duty cycle to 500Hz to 5000Hz, preferably 2000Hz (the total range of the PCA9685 is 4095), which is about 48.8% of the total range. This is to prevent the thermoelectric cooler from generating excessive heat loss when outputting at full power and to maintain stable system operation.

[0070] The S440 uses an H-bridge circuit for driving. Four H-bridge circuits receive PWM signals from the PCA9685 chip. Each H-bridge circuit converts external DC power into drive current by switching its internal power MOSFETs on and off. The main control unit controls the current flow via output signals: in heating mode, the current flows forward through the thermoelectric cooler; in cooling mode, the H-bridge circuits flip the MOSFET states, causing the current to flow backward through the thermoelectric cooler.

[0071] S450, power regulation. The thermoelectric cooler utilizes the Peltier effect to regulate the temperature of the controlled object based on the direction and magnitude of the drive current. The power drive unit switches between cooling and heating modes via 2kHz pulse width modulation. Upon receiving a new I... 2 Before the C command, the power drive unit maintains the current duty cycle output to ensure the continuity of current output.

[0072] The specific selection and design of the freewheeling diodes, gate resistors, and boost circuits involved in the H-bridge circuit are well-known technologies in this field and will not be elaborated further. The power drive unit's circuit board adopts a high-current copper-plated design and is equipped with filter capacitors to reduce loop impedance.

[0073] I / O between the main control unit and the PWM signal expansion chip 2 The C communication rate is 400kHz, ensuring a control cycle with a command transmission delay of less than 1 second. Through the above hardware configuration and mapping logic, the system converts the algorithm decision results into current output, realizing independent power control of the four temperature control channels.

[0074] The heat dissipation control unit is connected to the main control unit and is used to dissipate the heat generated at the hot end of the thermoelectric cooler during operation. The heat dissipation control unit includes four independently configured cooling fans, each corresponding to a temperature control channel. The cooling fans are driven by independent PWM channels provided by the PCA9685 chip in the power drive unit, reducing heat accumulation at the hot end of the thermoelectric cooler through forced convection and maintaining the heat exchange efficiency of the thermoelectric module.

[0075] The control logic and execution process of the heat dissipation control unit include the following steps: S510, Channel Status Monitoring. The main control unit reads the button status data generated by the btnTask task in the human-machine interface unit. When the main control unit determines that a certain temperature control channel is turned on, it triggers the corresponding channel's cooling fan to run.

[0076] S520, fixed duty cycle control. To ensure the thermoelectric cooler has heat exchange capabilities during operation, the main control unit allocates a fixed power output to the cooling fan during operation. The duty cycle count value of the cooling fan drive signal is configured to a preset high-speed reference value. (Based on 12-bit resolution, total range is 4095). This benchmark value. The duty cycle can be configured within the range of [2000, 4095], with 4000 being the preferred setting. The corresponding duty cycle percentage is calculated using the following formula: ; In the formula, This is the set high-speed reference value. When using the preferred setting value... At that time, the calculated result is approximately 97.6%. With this duty cycle output, the cooling fan maintains high speed operation, controlling the temperature rise of the hot end of the thermoelectric module and ensuring cooling efficiency.

[0077] S530, PWM command issuance and execution. The main control unit uses I / O... 2 The C bus accesses the control register of the PCA9685 chip and writes the value 4000 into the LEDn_OFF register of the corresponding cooling fan channel according to the communication timing. The PWM signal generated by the PCA9685 chip drives the cooling fan motor to rotate after passing through the drive circuit (such as a transistor or MOSFET switching circuit) inside the thermal control unit.

[0078] S540, shutdown control. When the main control unit detects that a certain temperature control channel is closed, or the Res-NNPID control algorithm module triggers system shutdown, the main control unit sends a shutdown command to the PCA9685 chip to clear the PWM duty cycle of the corresponding cooling fan channel to zero, causing the fan to stop rotating, thereby reducing system power consumption and noise.

[0079] S550, drive circuit protection. The drive circuit of the thermal control unit consists of a power switch and a reverse freewheeling diode. The reverse freewheeling diode is connected in parallel across the fan motor to provide a current loop for the inductive load when the PWM signal is off, reducing the impact of back electromotive force on the main control unit interface or drive chip. Since the drive signal of the thermal control unit is also generated by the PCA9685 chip, the PWM frequency configured by the main control unit is consistent with that of the cooling control unit, set in the range of 500Hz to 5000Hz, preferably 2000Hz (achieved by overclocking the internal oscillator of the PCA9685). This frequency configuration can provide a smooth drive current for the inductive fan motor and achieve the best balance between system control accuracy and switching losses.

[0080] The physical mounting structure of the cooling fan and the design of the ventilation duct can be configured by those skilled in the art according to the spatial layout of the temperature control system. The specific mechanical installation method is a well-known technology in the field and will not be described in detail here.

[0081] Through the control logic of the aforementioned heat dissipation control unit, the system achieves synchronous control of heat dissipation operation and channel status. Using fixed duty cycle control, the thermal stability of the four independent temperature control channels is maintained during operation, meeting the system's heat dissipation requirements.

[0082] The human-machine interface (HMI) unit is connected to the main control unit and is used to locally set target temperature data, switch channel states, and display system operating parameters. The HMI unit hardware consists of a TFT color touchscreen display and four physical buttons. On the software side, the interactive interface is built based on the LVGL graphics library and processed through the main control unit's core 0, avoiding the interaction process from consuming the control loop's computing resources.

[0083] The interface rendering and interaction execution process of the human-computer interaction unit includes the following steps: S610, Display Driver and Graphics Library Initialization. After system power-on, the main control unit drives the TFT color touchscreen display to complete hardware reset and parameter configuration. The main control unit allocates two sets of frame buffers for the LVGL graphics library in memory, improving the screen refresh rate through double buffering technology. The main control unit simultaneously enables the direct memory access (DMA) channel of the serial peripheral interface to transmit image data to the display driver chip, providing hardware support for concurrent channel monitoring.

[0084] S620 Monitoring Interface Layout and Object Construction. The human-machine interface unit constructs a global status bar and a channel monitoring area within the display area. The global status bar displays the system time and Wi-Fi signal connection status; the channel monitoring area is divided into four independent areas, each corresponding to one of the four temperature control channels. Each area contains a real-time temperature display label, a target temperature display label, and a working status indicator light. Each label is bound to a corresponding element in a globally shared variable array, including sensor1 to sensor4 storing the measured temperature and set1 to set4 storing the set temperature.

[0085] S630, Physical Button Polling and State Switching. The main control unit runs the btnTask task, which polls the level states of the four physical buttons at a fixed frequency. When a button trigger is detected and debouncing is performed, the btnTask task changes the Boolean value of the corresponding channel in the global variable array btnState[], thus toggling the start or stop state of that channel. The main control unit determines whether to trigger the power output command for the corresponding channel based on this state value.

[0086] The S640 employs a multi-task refresh logic. The software architecture of the human-computer interaction unit includes displayTask and LVGLLoop tasks. The displayTask periodically extracts the latest temperature values ​​and system connection status from globally shared variables and calls the graphics library interface to update text objects on the interface. The LVGLLoop task handles touch input events and the graphical redrawing of controls. By configuring different priorities for display tasks and acquisition / control tasks in FreeRTOS, the underlying control tasks are ensured to remain unaffected during interface rendering.

[0087] S650 features interactive feedback and parameter synchronization. When a user modifies the target temperature setpoint via the touch panel, the human-machine interface unit captures the input event and updates the `target_temp[]` array. The updated value is synchronized to the input of the Res-NNPID control algorithm module, enabling it to perform parameter tuning according to the new target value in the next control cycle. If the system detects a sensor malfunction or Wi-Fi disconnection, the interface switches the relevant icon style to prompt the user.

[0088] For LCD screen backlight brightness adjustment, touch chip I 2 The C communication timing and the hardware debouncing circuit design for the buttons can be implemented by those skilled in the art using standard hardware solutions or general driver libraries. The specific details are well-known in the field and will not be elaborated here.

[0089] Through the interface interaction and button control logic of the aforementioned human-computer interaction unit, the system achieves control of four independent temperature control channels. Utilizing the double buffering mechanism and multi-task scheduling of the LVGL graphics library, the system displays system time, connection status, and temperature parameters for each channel, providing a technical means for the local operation and maintenance of the multi-channel thermoelectric temperature control system.

[0090] The remote monitoring unit is connected to the main control unit and is used to realize data interaction and control between the temperature control system and the external cloud platform. In terms of hardware, the remote monitoring unit utilizes the Wi-Fi radio frequency circuit integrated into the main control unit to access the wireless network, supporting the 802.11b / g / n standard. At the application layer protocol level, the system establishes a communication link through the MQTT protocol.

[0091] The communication and execution process of the remote monitoring unit includes the following steps: S710, Network Initialization and Connection Handshake. After the system powers on, the remote monitoring unit scans the preset SSID and performs authentication to access the local area network. The main control unit assigns a client identifier to the remote monitoring unit and configures the IP address and port number of the MQTT server. The remote monitoring unit sends a connection message to request access to the cloud platform and establishes a connection upon receiving a response message. For the underlying protocol implementation of wireless encryption authentication and TCP / IP handshake, those skilled in the art can use a standard protocol stack; the specific details are well-known in the field and will not be elaborated here.

[0092] S720, data serialization. The remote monitoring unit extracts data from the shared variables of the main control unit at a preset period, ranging from 500ms to 30s, preferably 2s. The system encapsulates the data into a JSON-formatted string. The data packet contains the device ID, timestamp, and data arrays for four temperature control channels. Each channel's data item includes: real-time temperature data (temperature[]), target temperature data (target_temp[]), and the channel's start / stop status (btnState[]). S730, Message Publishing. The remote monitoring unit sends the serialized JSON string as the message payload to the cloud topic via MQTT publish. The system sets the Quality of Service (QoS) level to QoS0 or QoS1. The data frame structure is defined as follows: ; in, This indicates the complete JSON format data packet assembly that has been packaged and is ready to be sent. A unique identifier for the current temperature control device, used by the cloud platform to distinguish different devices; This indicates the timestamp at which the system generated the data packet; to These correspond to the first through fourth independent temperature control channels within the system, respectively. This indicates the real-time temperature data currently being collected by the corresponding channel; This indicates the target temperature data currently set for the corresponding channel; Indicates the current start / stop status of the corresponding channel (Boolean or enumerated value).

[0093] The S740 features cloud-based data processing and web-based visualization interaction. Upon receiving an MQTT message, the cloud platform parses the JSON data and persistently stores it in a database. The cloud platform also hosts a web server frontend written in HTML. Users can access this web interface via a PC or mobile browser. The web frontend extracts stored historical operational data and uses a frontend chart rendering engine to dynamically plot and visualize the temperature change trends of the four temperature control channels. Simultaneously, the web interface includes a data export interface, allowing users to package and download temperature sequences and status data for a specified time period to their local terminal for subsequent model training and offline analysis of control performance.

[0094] S750 Control Command Subscription and Parsing. The remote monitoring unit subscribes to control topics sent from the cloud platform. When a user modifies the target temperature of a channel or triggers a start / stop switch on the web interface, the cloud platform sends a JSON-formatted control command message. The remote monitoring unit receives this message, performs deserialization processing, and parses out the command type, channel index, and target value.

[0095] S760, parameter update. The remote monitoring unit distributes the parsed instructions to the main control unit. The main control unit performs validity checks and numerical limiting on the received parameters. After successful verification, the main control unit updates the corresponding target_temp[] or btnState[] variables. The new parameters participate in the calculation in the next execution cycle of the Res-NNPID control algorithm module, realizing remote control.

[0096] S770, Link Maintenance and Reconnection. The remote monitoring unit runs heartbeat detection logic, sending heartbeat packets to the server at preset intervals to maintain the connection. If a Wi-Fi signal loss or MQTT connection disconnection is detected, the remote monitoring unit triggers a reconnection mechanism based on an exponential backoff algorithm, gradually extending the time interval between reconnection attempts to reduce system resource consumption.

[0097] The hardware circuitry of the remote monitoring unit is equipped with a ceramic antenna or an external gain antenna, supplemented by a filtering network to reduce radio frequency interference. The antenna impedance matching and radio frequency layout design can be adjusted by those skilled in the art according to the application environment; the specific debugging techniques are well-known in the field and will not be elaborated upon here.

[0098] Specific Implementation Example: Semiconductor Laser Array Aging Test Platform In this embodiment, the present invention is applied to a four-channel semiconductor laser array aging test system. This system requires simultaneous temperature control of four lasers with different powers. Because the heat generated by the lasers changes non-linearly over time during the aging process, traditional fixed-parameter PID controllers are insufficient to meet the long-term, high-precision steady-state requirements.

[0099] Experimental environment and parameter settings: Controlled object: Four TEC1-12706 cooling chips, each carrying a different type of laser.

[0100] Target temperature settings: Channels 1 to 4 are set to 20℃, 25℃, 35℃, and 45℃ respectively.

[0101] System time consumption verification: According to the formula At an ESP32-S3 clock speed of 240MHz, the single-channel Res-NNPID calculation time is approximately 35ms, and the signal drive time is approximately 10ms. The total time for all four channels is 180ms. (1000ms) 180ms, the system meets the real-time constraint.

[0102] Algorithm configuration: residual weight is set to 0.3, and hidden layer nodes are 8.

[0103] Experimental verification and comparative analysis: The experiment compares the Res-NNPID adaptive control of the present invention with the traditional fixed-parameter PID (tuned by the Ziegler-Nichols method).

[0104] Comparison of step response and startup performance: Taking channel 3 (target temperature 35℃) as an example, the system starts from an ambient temperature of 25℃.

[0105] See attached document Figure 6 Experimental results show that the conventional PID reaches steady state in 145 seconds, with an overshoot of about 1.5°C; while the present invention accelerates the convergence in the early stage of online learning through residual jump connection, reaching steady state in only 82 seconds with virtually no overshoot, and improving the adjustment speed by about 43%.

[0106] Verification of multi-path independence and steady-state accuracy: See attached document Figure 7 -Appendix Figure 10 The four subplots in the figure show the independent closed-loop control results executed by the system for four controlled objects with different target values ​​(20℃, 25℃, 35℃, 45℃) within the same time period.

[0107] Experimental results show that although the four channels share the same main control unit's hardware resources, thanks to the multi-task parallel scheduling mechanism (S120-S140) and the Res-NNPID independent weight maintenance strategy of this invention, the temperature curves of each channel can independently converge to their respective target setpoints. When one channel experiences adjustment or thermal fluctuations, the steady-state accuracy of the other three channels is not affected by cross-channel thermal coupling or computational delay, and the steady-state error remains within ±0.015℃. This demonstrates that this system can effectively solve the resource contention and parameter coupling problems commonly found in traditional multi-channel temperature control.

[0108] Disturbance immunity test: When the system reaches the 500th second, the driving current of the laser in channel 3 is simulated to be increased (i.e., a sudden thermal load disturbance is introduced).

[0109] See attached document Figure 11 Conventional PID controllers experience a temperature drop of 1.2°C under disturbance, with a recovery time exceeding 60 seconds; the Res-NNPID controller of this invention updates online. and The parameters control the drop temperature to within 0.4℃ and restore it to the target value within 25 seconds, demonstrating stronger adaptive disturbance rejection capabilities. Conclusion: The above experimental data comparison proves that: Effectiveness of residual structure: Formula It ensures the stability of gradient propagation in deep networks and solves the problem of easy divergence in conventional NN-PID training on embedded systems.

[0110] High precision and adaptability: Combining a high-resolution sensing unit (0.0001℃) with online parameter compensation, the system has higher thermal stability than traditional temperature control systems under complex aging test environments.

Claims

1. A four-channel thermoelectric temperature control system based on residual neural network adaptive PID, characterized in that, include: The main control unit is used to run the real-time operating system and perform multi-task scheduling; The temperature sensing unit is used to acquire real-time temperature data of the controlled object corresponding to four independent channels and feed it back to the main control unit. The Res-NNPID control algorithm module runs in the main control unit and is used to construct input features based on the real-time temperature data and the preset target temperature data, and generate PID parameters corresponding to each channel through a neural network with residual jump connections. A power drive unit is used to drive a thermoelectric cooling chip according to a PWM signal generated by the main control unit based on the PID parameters, thereby adjusting the temperature of the controlled object. A heat dissipation control unit is used to perform dynamic heat dissipation when the thermoelectric cooler is working. The human-computer interaction unit is used to set the target temperature data and display the system operating status; The remote monitoring unit is used for remote transmission of data and reception of commands from each channel.

2. The four-channel thermoelectric temperature control system based on residual neural network adaptive PID according to claim 1, characterized in that, The main control unit uses a dual-core processor and runs a real-time operating system. The real-time operating system executes the control tasks corresponding to the Res-NNPID control algorithm module, the data acquisition tasks corresponding to the temperature sensing unit, and the communication tasks corresponding to the remote monitoring unit through a multi-task parallel scheduling mechanism.

3. The four-channel thermoelectric temperature control system based on residual neural network adaptive PID according to claim 1, characterized in that, It also includes a power management module, which uses a low-dropout linear regulator to power the temperature sensing unit; the temperature sensing unit includes four independent temperature sensors, each with a preset temperature measurement accuracy.

4. The four-channel thermoelectric temperature control system based on residual neural network adaptive PID according to claim 1, characterized in that, The PWM signal extension chip receives instructions from the main control unit via the communication bus and outputs a PWM signal to the H-bridge circuit to control the direction and magnitude of the operating current of the thermoelectric cooler.

5. The four-channel thermoelectric temperature control system based on residual neural network adaptive PID according to claim 1, characterized in that, The Res-NNPID control algorithm module maintains independent neural network weight parameters for each of the four channels; the input layer of the neural network receives four feature quantities: current temperature, target temperature, temperature change rate, and temperature deviation of the previous cycle; the neural network includes two hidden layers, and each hidden layer introduces residual jump connections.

6. The four-channel thermoelectric temperature control system based on residual neural network adaptive PID according to claim 3, characterized in that, The Res-NNPID control algorithm module includes: The confidence assessment module is used to calculate the sampling confidence factor based on the temperature deviation between the preset temperature measurement accuracy and the current time. The Res-NNPID control algorithm module is also used to adjust the learning rate of the neural network and the residual weights of the residual jump connections in real time according to the sampling confidence factor, so as to shrink the learning step size in the temperature control steady-state perturbation region.

7. The four-channel thermoelectric temperature control system based on residual neural network adaptive PID according to claim 5, characterized in that, The output layer of the neural network uses a hyperbolic tangent activation function to output three normalized adjustment values. The PID parameters include proportional parameters, integral parameters, and derivative parameters. The proportional parameters, integral parameters, and derivative parameters are obtained by superimposing the normalized adjustment value and the adjustment range on their respective preset benchmark values, and are restricted to non-negative values.

8. The four-channel thermoelectric temperature control system based on residual neural network adaptive PID according to claim 5, characterized in that, The Res-NNPID control algorithm module uses the backpropagation algorithm to update the network weights online. During the update process, gradient pruning is performed on the output layer error term, and amplitude limiting is performed on the updated weights. When an invalid weight value occurs, an automatic reset operation is performed.

9. The four-channel thermoelectric temperature control system based on residual neural network adaptive PID according to claim 1, characterized in that, When calculating the control quantity, the main control unit performs saturation limiting processing on the integral term and limits the final control quantity within a preset value range. A positive control value corresponds to the heating mode, a negative control value corresponds to the cooling mode, and a zero control value corresponds to the thermoelectric cooling element stopping operation.

10. The four-channel thermoelectric temperature control system based on residual neural network adaptive PID according to claim 1, characterized in that, The human-computer interaction unit uses the LVGL graphics library to drive a color LCD display screen to present the interface; the remote monitoring unit uses the MQTT protocol to serialize the data of each channel into a JSON format string and upload it to the cloud platform, and supports receiving the target temperature setpoint and channel start / stop commands remotely issued by the cloud platform.

Images