Intelligent control method and system for rapid temperature change box

Abstract

The present application relates to the technical field of control, and especially relates to a kind of fast temperature change box intelligent control method and system.The method of the present application comprises: obtaining the multi-source sensor data and historical control instruction sequence of fast temperature change box;Using the hybrid model containing convolutional neural network and long short-term memory network to extract state feature vector representing the space-time dynamics of system, comprising: using convolutional neural network to extract spatial feature from multi-source sensor data, to obtain spatial feature vector;The spatial feature vector sequence and historical control instruction sequence are input into long short-term memory network, to obtain state feature vector;State feature vector is processed to obtain control instruction sequence;The control instruction sequence is sent to fast temperature change box.The present application can realize fast and accurate temperature tracking control.

Description

Technical Field

[0001] This invention relates to the field of control technology, and in particular to an intelligent control method and system for a rapid temperature change chamber. Background Technology

[0002] Rapid temperature change test chambers (rapid temperature change chambers) are key equipment for reliability testing of electronic products, automotive parts, and aerospace equipment. Their control performance directly determines the accuracy and efficiency of the test results. Current mainstream control schemes generally adopt traditional PID control technology, which has advantages such as simple structure and ease of implementation. However, it still has significant drawbacks in the complex thermal systems of test chambers, characterized by strong nonlinearity and multivariable coupling. The parameters of traditional PID controllers need to be manually tuned for specific operating conditions. When load characteristics change, the control performance deteriorates sharply, manifesting as large temperature overshoot, long settling time, and even continuous oscillation.

[0003] To address this, existing technologies have developed methods that apply algorithms such as neural networks and deep learning to the field of temperature control. These methods learn system dynamics through a data-driven approach, avoiding complex mechanistic modeling and achieving precise temperature control. However, these methods typically treat multi-source heterogeneous sensor data as isolated time points or simple stacks, failing to perform collaborative analysis and deep fusion. They also neglect the correlation between the spatial distribution characteristics of the temperature field and the multi-timescale features of system dynamics, making it difficult to accurately model the complex thermal processes of the test chamber, thus reducing the accuracy of temperature control. Summary of the Invention

[0004] To address the technical problem of neglecting the correlation between the spatial distribution characteristics of the temperature field and the multi-timescale features of system dynamics, which makes it difficult to accurately model the complex thermal processes of the test chamber, this invention provides solutions in the following aspects.

[0005] In a first aspect, the present invention provides an intelligent control method for a rapid temperature change chamber, comprising: acquiring multi-source sensor data and historical control command sequences of the rapid temperature change chamber; and extracting state feature vectors characterizing the spatiotemporal dynamics of the system using a hybrid model comprising a convolutional neural network and a long short-term memory network, comprising: extracting spatial features from the multi-source sensor data using a convolutional neural network to obtain spatial feature vectors; and inputting the spatial feature vector sequence and the historical control command sequence into the long short-term memory network to obtain state feature vectors; wherein the spatial feature vector sequence is composed of multiple spatial feature vectors. The state feature vector is processed to obtain a control command sequence; the control command sequence is then sent to the rapid temperature change box.

[0006] Furthermore, before extracting the state feature vector, the method also includes performing multivariate singular spectrum analysis on the multi-source sensor data.

[0007] Furthermore, multivariate singular spectrum analysis is performed on the multi-source sensor data, including: constructing a trajectory matrix based on the multi-source sensor data; sequentially performing singular value decomposition, grouping, and reconstruction on the trajectory matrix to obtain denoised multi-source sensor data.

[0008] Furthermore, the convolutional neural network includes multiple CNN layers and pooling layers, and the CNN layers use the ReLU activation function; the long short-term memory network is a two-layer long short-term memory network.

[0009] Further, the state feature vector is processed to obtain a control instruction sequence, including: inputting the state feature vector into a preset Actor network to obtain multiple action values; and mapping the multiple action values ​​to corresponding control instructions according to a preset mapping rule to obtain a control instruction sequence composed of control instructions.

[0010] Furthermore, the Actor network is obtained through offline training using a deep deterministic policy gradient algorithm.

[0011] Further, the state feature vector is processed to obtain a control command sequence, including: updating the simplified prediction model parameters using the state feature vector, and solving for the control commands using a model predictive control algorithm.

[0012] Furthermore, the method also includes: collecting system response data during the control process; calculating control performance evaluation indicators based on the system response data; and using the control performance evaluation indicators and system response data to perform online transfer learning fine-tuning of the hybrid model.

[0013] Furthermore, the multi-source sensor data includes at least temperature data, first pressure data, and second pressure data; acquiring the multi-source sensor data of the rapid temperature change chamber includes: acquiring temperature data at multiple spatial locations inside the rapid temperature change chamber through a temperature sensor array; and acquiring first pressure data and second pressure data of the refrigeration system of the rapid temperature change chamber through pressure sensors.

[0014] In a second aspect, the present invention provides an intelligent control system for a rapid temperature change chamber, used to implement the intelligent control method for a rapid temperature change chamber described in the first aspect; the system includes a rapid temperature change chamber and a multi-source sensor array, an edge computing unit, and an actuator drive module connected in sequence; the multi-source sensor array is used to collect multi-source sensor data from the rapid temperature change chamber; the edge computing module is used to acquire the multi-source sensor data and historical control command sequences, extract state feature vectors from the multi-source sensor data, and process the state feature vectors to obtain a control command sequence; the actuator drive module is used to convert the control commands in the control command sequence into drive signals to control the operation of the actuator of the rapid temperature change chamber.

[0015] Compared to existing technologies, the advantages of this invention are as follows: By utilizing a hybrid model incorporating convolutional neural networks and long short-term memory networks, deep fusion learning of spatial correlation and temporal evolution characteristics of multi-source sensor data is achieved, enabling a comprehensive and in-depth understanding of the system state, thereby realizing accurate modeling of complex thermal processes in the test chamber; by employing multivariate singular spectrum analysis to adaptively decompose and denoise the original multi-source sensor data, noise interference is effectively suppressed, thus avoiding incorrect correlations learned by the hybrid model and further improving the accuracy of modeling; compared to PID control, which relies on manual experience for tuning, this invention constructs a fully data-driven intelligent controller, which can achieve fast, accurate, and robust temperature tracking control without relying on precise physical mechanism models and complex manual parameter tuning, fundamentally solving the multivariable coupling problem and optimizing system energy efficiency while improving control accuracy. Attached Figure Description

[0016] Figure 1 This is a flowchart of the intelligent control method for a rapid temperature change chamber in an embodiment of the present invention; Figure 2 The structural frame of the intelligent control system for the rapid temperature change chamber in this embodiment of the invention. Figure 1 ; Figure 3 The structural frame of the intelligent control system for the rapid temperature change chamber in this embodiment of the invention. Figure 2 . Detailed Implementation

[0017] The technical solutions of 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, not all, of the embodiments of the present invention. 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.

[0018] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart of the intelligent control method for the rapid temperature change chamber in an embodiment of the present invention.

[0020] In a first aspect, the present invention provides an intelligent control method for a rapid temperature change chamber. Specifically, as... Figure 1 As shown, the method of the present invention includes the following steps.

[0021] S1. Acquire multi-source sensor data and historical control command sequences of the rapid temperature change chamber.

[0022] In this embodiment, the multi-source sensor data includes at least temperature data, first pressure data, and second pressure data. Additionally, it may include operating current data, opening degree data, and operating frequency data.

[0023] Specifically, temperature data from multiple locations within the rapid temperature change chamber can be acquired via a temperature sensor array; first and second pressure data of the refrigeration system can be acquired via a pressure sensor; operating current data of the compressor, fan, and heater can be acquired via a current sensor; opening data of the electronic expansion valve can be acquired via a position sensor; operating frequency data of the variable frequency compressor and variable frequency fan can be read via the EtherCAT bus; and historical control command sequences can be read from the historical buffer.

[0024] Furthermore, the multi-source sensor data undergoes preprocessing such as engineering unit conversion and timestamp alignment, and is then packaged into a 15-dimensional multimodal raw data vector X_raw(t): X_raw(t)=[T1(t)~T9(t),P_high(t),P_low(t),I_comp(t),I_fan(t),I_heater(t),Valve(t),Freq_comp(t),Freq_fan(t)] In the formula, X_raw(t) is the multimodal raw data vector at time t / the t-th control cycle, and T1(t)~T9(t) are the 1st to 9th spatial positions inside the rapid temperature change chamber at time t (corresponding to...). Figure 3 The values ​​of temperature measurement points 1 to 9 are given, P_high(t) is the first pressure of the refrigeration system at time t, P_low(t) is the second pressure of the refrigeration system at time t, I_comp(t) is the operating current of the compressor at time t, I_fan(t) is the operating current of the fan at time t, I_heater(t) is the operating current of the heater at time t, Valve(t) is the opening degree of the electronic expansion valve at time t, Freq_comp(t) is the operating frequency of the variable frequency compressor at time t, and Freq_fan(t) is the operating frequency of the variable frequency fan at time t.

[0025] In existing technologies, raw data is typically input directly into the model. However, raw data may contain noise, disturbances, or anomalous data, causing the model to learn incorrect associations, severely impacting the controller's generalization ability and robustness. To address this issue, and to resolve the problem of incorrect model learning caused by directly inputting raw data, the method of this invention further includes performing multivariate singular spectrum analysis on the multivariate sensor data after obtaining the multi-source sensor data.

[0026] Specifically, a trajectory matrix is ​​constructed based on multi-source sensor data. Specifically, key coupling variables are selected from the multimodal raw data vector X_raw(t); for each time series / key coupling variable, the latest N=500 data points are obtained, and a Hankel matrix H_i is constructed for each time series with a window length L=100. These 11 H_i matrices are stacked vertically to form a block Hankel trajectory matrix H_block.

[0027] In this embodiment, 11 time series data points, including 9 temperature data points [T1~T9] and 2 pressure data points [P_high, P_low], are selected as key coupling variables. By constructing a trajectory matrix, subsequent decomposition can simultaneously consider the interrelationships between all variables.

[0028] Furthermore, the trajectory matrix is ​​sequentially subjected to singular value decomposition, grouping, and reconstruction to obtain denoised multi-source sensor data.

[0029] Specifically, singular value decomposition is performed on the trajectory matrix H_block to obtain the decomposition result. In one embodiment, the decomposition result includes a singular value spectrum σ1≥σ2≥…≥σ6...≥σn and the corresponding left and right singular vectors. The singular value spectrum includes multiple singular values, where σn in the above equation is the nth singular value.

[0030] Furthermore, based on the distribution of singular values ​​and the characteristics of singular vectors, the decomposition results are grouped to obtain grouped results. In this embodiment, the grouped results include shared trend components, coupled oscillation components, and random noise components. Specifically, the components corresponding to the first two singular values ​​(i.e., σ1 and σ2) are classified as shared trend components, the components corresponding to the third to sixth singular values ​​(i.e., σ3, σ4, and σ6) are classified as coupled oscillation components, and the components corresponding to the remaining singular values ​​are considered irrelevant random noise components and discarded. Among them, the shared trend component represents the slow change main line followed by all temperature and pressure data, i.e., the set temperature change program; the coupled oscillation component has a consistent phase and frequency in different variable sequences, corresponding to the overall system oscillation caused by physical sources such as the periodic operation of the compressor and refrigerant pulsation.

[0031] Furthermore, the shared trend component and the coupled oscillation component are reconstructed by diagonal averaging to obtain the denoised signal vector X_clean(t): X_clean(t)=[T1_clean(t)~T9_clean(t),P_high_clean(t),P_low_clean(t)] In the formula, X_clean(t) is the signal component at time t after denoising, that is, the multi-source sensor data after denoising; T1_clean(t) to T9clean(t) are the temperatures of the 1st to 9th spatial locations inside the rapid temperature change chamber at time t after denoising; P_high_clean(t) is the first pressure of the refrigeration system at time t after denoising; and P_low_clean(t) is the second pressure of the refrigeration system at time t after denoising.

[0032] In one embodiment, the multivariate singular spectrum analysis process can be implemented on the edge computing unit using a highly optimized linear algebra library, with a single calculation latency consistently within 5ms.

[0033] By performing multivariate singular spectrum analysis on the raw data, we can extract pure and effective components / coupled dynamics that reflect the physical essence of the system from the messy raw data. This solves the problem that directly using the raw data leads to the hybrid model learning incorrect correlations, which seriously affects the generalization ability and robustness of the controller, thereby improving the accuracy of subsequent temperature control.

[0034] S2. A hybrid model containing convolutional neural networks and long short-term memory networks is used to extract state feature vectors that represent the spatiotemporal dynamics of the system.

[0035] Specifically, in each control cycle, the denoised multi-source sensor data are arranged and stacked according to the physical location and type of the sensors to obtain a multimodal spatiotemporal feature map. Specifically, the denoised nine temperature channels T1_clean~T9_clean at the same time are arranged into a 3x3 two-dimensional grid according to their physical spatial location (three layers: top, middle, and bottom, front, back, left, and right); at the same time, the denoised two pressure channels P_high_clean and P_low_clean are treated as two independent channels; then these 11 channels (i.e., nine temperature channels and two pressure channels) are stacked in the depth direction to form a multimodal spatiotemporal feature map; the dimension of this multimodal spatiotemporal feature map is [3x3x11].

[0036] Furthermore, spatial features are extracted from the multimodal spatiotemporal feature map using a convolutional neural network (CNN) to obtain a spatial feature vector. In this embodiment, the CNN includes multiple CNN layers and pooling layers. Specifically, the multimodal spatiotemporal feature map is sequentially input into a first CNN layer, a second CNN layer, and a 2x2 max-pooling layer. Finally, the pooled feature tensor is flattened and mapped into a 64-dimensional spatial feature vector S(t). The first CNN layer uses 16 3x3 convolutional kernels, the second CNN layer uses 32 3x3 convolutional kernels, and both the first and second CNN layers use the ReLU activation function.

[0037] By using convolutional neural networks to process multi-source sensor data, the spatial distribution pattern of the temperature field at the current moment and its spatial feature vectors related to the system pressure state were accurately extracted.

[0038] Furthermore, the spatial feature vector sequence, composed of multiple spatial feature vectors, and the historical control command sequence are input into a long short-term memory (LSTM) network to obtain a state feature vector. This LSM network is a two-layer LSM network, with each layer having 128 hidden units. Specifically, the spatial feature vector sequence S(t), S(t-1), ..., S(t-9), composed of the spatial feature vectors of the current and nine consecutive past control cycles, is concatenated with the corresponding historical control command sequence U(t-1) to form an enhanced temporal feature sequence. This temporal feature sequence is then input into the two-layer LSM network. Through the gating mechanism of the LSM network, the evolutionary rules of the multimodal spatial features are learned and memorized, and finally, the hidden state of the last time step is output as a 32-dimensional state feature vector H(t).

[0039] By utilizing a hybrid architecture of convolutional neural networks and long short-term memory networks, the nonlinear dynamic characteristics of the system are automatically learned from denoised multi-source spatiotemporal data. Deep fusion learning of spatial correlation and temporal evolution characteristics is carried out, achieving accurate modeling of complex thermal processes in the test chamber.

[0040] S3. Process the state feature vector to obtain the control command sequence.

[0041] Specifically, the state feature vector is input into the pre-trained Actor network to obtain four continuous action values ​​A(t) normalized in the range [-1, 1]: A(t) = [a1, a2, a3, a4] In the formula, a1 is the operating value of the compressor, a2 is the operating value of the electronic expansion valve, a3 is the operating value of the heater, and a4 is the operating value of the fan.

[0042] Furthermore, according to the preset mapping rules, these action values ​​are mapped to actual physical commands, i.e., control commands, thereby obtaining a control command sequence U(t) composed of these control commands: U(t) = [Freq, Valve, Heat, Fan] In the formula, Freq is the target frequency of the compressor, Valve is the opening degree of the electronic expansion valve, Heat is the duty cycle of the heater, and Fan is the target speed of the fan.

[0043] In one embodiment, the Actor network can be obtained through offline training using a depth-deterministic row policy gradient algorithm.

[0044] The specific training process and data are as follows: Training Environment and Data: Training is conducted in a simulation environment containing a high-fidelity physical model, or using a large historical dataset. The training data includes at least the multimodal raw data vector X_raw(t), the control command sequence U(t), and the reward function R(t). The reward function includes rewards and penalties for temperature tracking error and control stability, thereby quantifying control performance.

[0045] In one embodiment, the reward function R(t) can be defined as: R(t)=-w1×(T_set(t)-T_actual(t))^2-w2×||U(t)-U(t-1)||^2 In the formula, T_set(t) is the target set temperature at time t, T_actual(t) is the actual temperature at time t, U(t) is the control command sequence at time t, U(t-1) is the control command sequence at time t-1, ||| is the L2 norm, and w1 and w2 are preset weight coefficients, both of which are positive numbers.

[0046] It should be noted that during the deployment and operation phase, the target set temperature T_set(t) is a known prerequisite; the state feature vector H(t) internalizes the optimal response strategy to different target temperature change trends through learning from historical data. Therefore, in actual control, the system only needs to generate the optimal action to track the target set temperature T_set(t) based on the current state feature vector H(t), without explicitly inputting the target set temperature T_set(t) into the network at each step.

[0047] Furthermore, the Deep Deterministic Policy Gradient (DDPG) algorithm is used for end-to-end offline training: Actor Network: The input is the state feature vector H(t), and the output is the action value A(t).

[0048] Critic network: The input is the state feature vector H(t) and the action value A(t), and the output is the expected long-term cumulative reward (Q value) of the state-action pair.

[0049] Training loop: Collect empirical data (H(t), A(t), R(t), H(t+1)) and store it in the empirical replay buffer. Using randomly sampled data, the Critic network learns to accurately evaluate the Q-value, while the Actor network updates its parameters θ in the direction that increases the Q-value. This process is iterated until the policy converges.

[0050] The formation of the association between the state feature vector H(t) and the action value: Through the above training, the parameters θ of the Actor network are optimized to a stable state. At this time, the mapping relationship from H(t) to A(t) represented by the network f_actor(H(t);θ) is the strategy learned from historical best experience under a given reward function, which can map the current state feature vector H(t) to the action value A(t) that is most likely to obtain high long-term rewards (i.e., achieve accurate, stable, and efficient temperature tracking).

[0051] In one embodiment, the mapping rule can be: Freq=30+(a1+1) / 2×(120-30)Hz; Valve = (a² + 1) / 2 × 100%; Heat = (a³ + 1) / 2 × 100%; Fan = 30 + (a4 + 1) / 2 × (100 - 30)%.

[0052] By replacing the traditional PID controller with an Actor network, globally optimal control commands for multiple actuators such as compressors, electronic expansion valves, heaters, and fans can be generated synchronously and collaboratively, fundamentally solving the multivariable coupling problem. This improves control accuracy while optimizing system energy efficiency. It can achieve fast, accurate, and stable temperature tracking across the entire operating range without relying on precise physical mechanism models and complex manual parameter tuning.

[0053] In another embodiment, state feature vectors can be used to update the parameters of a pre-built simplified prediction model, and then the optimal control command sequence can be solved using a model predictive control algorithm. Those skilled in the art can configure the simplified prediction model according to actual needs; no limitations are imposed here. Furthermore, the model predictive control algorithm is prior art and will not be described in detail here.

[0054] S4. Send a sequence of control commands to the rapid temperature change box.

[0055] Specifically, the control command sequence is sent to each actuator in real time via the analog output module and the EtherCAT bus. Simultaneously, the control command sequence is stored in a history buffer for use in the next control cycle. The actuators include compressors, electronic expansion valves, heaters, and fans.

[0056] In one embodiment, the method of the present invention further includes: acquiring system response data during the control process; the system response data includes measured temperature, set temperature, real-time power consumption data of the actuator, multi-source sensor data, state feature vector, and control command sequence.

[0057] Furthermore, control performance evaluation indicators are calculated based on system response data. These indicators include the root mean square error (RMSE) of temperature, overshoot (OS), and energy consumption per unit temperature change. The overshoot can be obtained by subtracting the target steady-state temperature from the peak temperature reached during the temperature change, and then dividing by the set temperature. Energy consumption per unit temperature change is obtained by dividing the energy consumption E by the temperature change ΔT. The energy consumption can be calculated using power consumption data.

[0058] Every 24 hours of cumulative operation, the hybrid model is fine-tuned through online transfer learning using control performance evaluation metrics and system response data to slowly adapt to equipment aging. Furthermore, deep optimization can be triggered periodically: historical data from several months is retrieved, and the Sparrow Search Algorithm (SpSA) is used to globally optimize the weights of the Actor network, seeking better control strategies. The optimized Actor network, after testing, can be seamlessly updated via the cloud.

[0059] In addition, the decomposition parameters of multivariate singular spectrum analysis can be updated periodically to adapt to changes in system characteristics.

[0060] Figure 2 The structural frame of the intelligent control system for the rapid temperature change chamber in this embodiment of the invention. Figure 1 .

[0061] In a second aspect, the present invention provides an intelligent control system for a rapid temperature change chamber, used to implement the intelligent control method for a rapid temperature change chamber described in the first aspect. Specifically, as... Figure 2 and Figure 3 As shown, the system of this invention includes a rapid temperature change chamber, a multi-source sensor array, an edge computing unit, an actuator drive module, and a cloud management platform. The multi-source sensor array is used to collect the operating status data of the rapid temperature change chamber, i.e., multi-source sensor data. Figure 3 In the diagram, A is the temperature acquisition processor, B is the edge computing unit, C is the system status actuator, and D is the rapid temperature change chamber.

[0062] In this embodiment, the working volume of the rapid temperature change chamber is 1 cubic meter. It adopts a double-layer metal plate polyurethane insulation structure and includes an insulated chamber body, a refrigeration system, a heating system, and an airflow circulation system. The refrigeration system uses a binary cascade refrigeration circuit. The high-temperature stage uses a variable frequency scroll compressor (rated power 5.5kW, refrigerant R404a), and the low-temperature stage uses a variable frequency scroll compressor (rated power 3kW, refrigerant R23). The throttling component is an electronic expansion valve driven by a stepper motor. The heating system consists of two independently controlled nickel-chromium alloy resistance wire heaters with a total power of 8kW, controlled by a solid-state relay for phase angle. The airflow circulation system uses a rear centrifugal fan driven by a frequency converter, allowing for stepless speed adjustment within the range of 300 to 3000 rpm.

[0063] In one embodiment, a multi-source sensor array is used to collect operational status data of the rapid temperature change chamber, i.e., multi-source sensor data. Specifically, the multi-source sensor array includes a distributed temperature sensor / temperature sensor array, a pressure sensor, a current sensor, and a position sensor.

[0064] In one embodiment, a temperature sensor array is used to collect temperature data from multiple spatial locations inside the rapid temperature change chamber. Specifically, nine Class A PT100 platinum resistance sensors are used, and their installation positions can refer to the temperature sampling point layout in GB / T5170.2-2008. Specifically, six sensors are fixed at the front and rear geometric key points of the upper, middle, and lower layers of the working space of the rapid temperature change chamber to construct the spatial temperature field; two sensors are respectively installed at the centrifugal fan outlet and the return air grid at the bottom of the chamber to monitor airflow energy; and one sensor is attached to the evaporator fin surface to monitor the core heat exchange status. All temperature data are connected via a four-wire connection, with a sampling frequency of 10Hz, and the system uncertainty is better than ±0.3℃.

[0065] Pressure sensors are used to collect the first and second pressure data of the refrigeration system. Specifically, two high-precision piezoresistive pressure transmitters are used, installed after the discharge port of the high-temperature stage compressor (range 0-3.5MPa) and before the suction port of the low-temperature stage compressor (range 0.1-1.5MPa), respectively. Current sensors are used to collect the operating current data of the compressor, fan, and heater. Specifically, three closed-loop Hall current sensors are used, connected in series in the main power supply circuits of the compressor, circulating fan, and heater, respectively, to collect the current data of each actuator. Position sensors are used to collect the opening data of the electronic expansion valve. Specifically, a 1000-line incremental encoder is installed in the electronic expansion valve driver to collect the valve core opening position (0-500 steps).

[0066] The edge computing unit is an industrial-grade embedded computer, equipped with an Intel i7 processor, 16GB DDR4 memory, and a synchronous data acquisition system, a real-time control output system, a real-time industrial network, and an algorithm runtime environment. The synchronous data acquisition system includes a 16-bit 8-channel RTD input module and a 16-channel 4-20mA analog input module, with all channels achieving microsecond-level synchronization via the IEEE 1588 precision time protocol. The real-time control output system includes a 4-channel analog output module (±10V) and a high-speed digital output module. The real-time industrial network is equipped with an EtherCAT master card, forming a hard real-time control loop with a ≤1ms cycle with the compressor inverter, fan inverter, and electronic expansion valve driver. The algorithm runtime environment runs on a Linux real-time kernel, executing containerized algorithm modules, including a data acquisition module, a multi-source singular spectrum analysis module, a deep learning module, and an adaptive control module connected in sequence.

[0067] In one embodiment, the data acquisition module is used to synchronously acquire / obtain multi-source sensor data at the beginning of each control cycle (10Hz) and perform preprocessing; the multivariate singular spectrum analysis module is used to decompose and denoise the multi-source sensor data, extract effective signal components, and obtain denoised multi-source sensor data; the deep learning model module is used to extract state feature vectors representing the spatiotemporal dynamic characteristics of the system from the denoised multi-source sensor data; the deep learning module includes a hybrid model composed of convolutional neural networks and long short-term memory networks; and the adaptive control module is used to generate a sequence of control commands based on the state feature vectors.

[0068] In one embodiment, the multivariate singular spectrum analysis module specifically includes a trajectory matrix construction unit, a singular value decomposition unit, a component grouping unit, and a signal reconstruction unit connected in sequence. Specifically, the trajectory matrix construction unit constructs a Hankel trajectory matrix using multi-source sensor data; the singular value decomposition unit performs singular value decomposition on the trajectory matrix; the component grouping unit groups the components according to the singular value spectrum and vector characteristics to obtain the grouping results; and the signal reconstruction unit reconstructs the shared trend components and coupled oscillation components in the grouping results.

[0069] In one embodiment, the deep learning model module includes a feature map construction unit, a convolutional neural network unit, a long short-term memory network unit, and a feature fusion unit connected in sequence. The feature map construction unit is used to construct a multimodal spatiotemporal feature map from multi-source sensor data; the convolutional neural network unit includes multiple CNN layers and pooling layers to extract spatial feature vectors from the multimodal spatiotemporal feature map; the long short-term memory network unit contains multiple LSTM layers to model temporal dynamics; and the feature fusion unit is used to fuse spatial and temporal features to generate a state feature vector.

[0070] In one embodiment, the adaptive control module provides two control strategies: a deep reinforcement learning control strategy and a hybrid predictive control strategy. The deep reinforcement learning control strategy includes an Actor network trained offline, which directly maps state feature vectors to action values, and finally maps the action values ​​to control commands according to a mapping rule. The hybrid predictive control strategy includes a simplified predictive model and a rolling optimizer, which adjusts model parameters online based on state feature vectors and solves the optimization problem to obtain the optimal sequence of control commands.

[0071] The actuator drive module is used to receive control command sequences and convert the control commands in the control command sequence into drive signals, and control the actuator of the rapid temperature change box to operate according to the drive signals; the cloud management platform is used to store historical data, manage model versions, and perform remote monitoring and strategy optimization.

[0072] In the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise explicitly specified. Furthermore, the division of steps in the above method is only for clarity of description; in implementation, it can be combined into one step or some steps can be split into multiple steps, as long as they include the same logical relationship. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0073] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A method for intelligent control of a rapid temperature change chamber, characterized in that, include: Acquire multi-source sensor data and historical control command sequences from the rapid temperature change chamber; A hybrid model incorporating convolutional neural networks and long short-term memory networks is used to extract state feature vectors representing the spatiotemporal dynamics of the system. This includes: using a convolutional neural network to extract spatial features from the multi-source sensor data to obtain spatial feature vectors; and... The spatial feature vector sequence and the historical control instruction sequence are input into a long short-term memory network to obtain a state feature vector; the spatial feature vector sequence is composed of multiple spatial feature vectors. The state feature vector is processed to obtain a control command sequence; Send the control command sequence to the rapid temperature change box.

2. The intelligent control method for a rapid temperature change chamber according to claim 1, characterized in that, Before extracting the state feature vector, the method further includes performing multivariate singular spectrum analysis on the multi-source sensor data.

3. The intelligent control method for a rapid temperature change chamber according to claim 2, characterized in that, Multivariate singular spectrum analysis is performed on multi-source sensor data, including: Construct a trajectory matrix based on multi-source sensor data; The trajectory matrix is ​​sequentially subjected to singular value decomposition, grouping, and reconstruction to obtain denoised multi-source sensor data.

4. The intelligent control method for a rapid temperature change chamber according to claim 1, characterized in that, The convolutional neural network includes multiple CNN layers and pooling layers, with the CNN layers employing the ReLU activation function; the long short-term memory network is a two-layer long short-term memory network.

5. The intelligent control method for a rapid temperature change chamber according to claim 1, characterized in that, The state feature vector is processed to obtain a control command sequence, including: The state feature vector is input into a preset Actor network to obtain multiple action values; According to the preset mapping rules, multiple action values ​​are mapped to corresponding control instructions to obtain a control instruction sequence composed of control instructions.

6. The intelligent control method for a rapid temperature change chamber according to claim 5, characterized in that, The Actor network is obtained through offline training using a deep deterministic policy gradient algorithm.

7. The intelligent control method for a rapid temperature change chamber according to claim 1, characterized in that, The state feature vector is processed to obtain a control command sequence, including: The state feature vector is used to update the simplified prediction model parameters, and the control command sequence is solved by the model predictive control algorithm.

8. The intelligent control method for a rapid temperature change chamber according to claim 1, characterized in that, The method further includes: Collect system response data during the control process; Calculate control performance evaluation indicators based on system response data; The hybrid model is fine-tuned through online transfer learning using control performance evaluation metrics and system response data.

9. The intelligent control method for a rapid temperature change chamber according to claim 1, characterized in that, Multi-source sensor data includes at least temperature data, first pressure data, and second pressure data; acquiring multi-source sensor data from the rapid temperature change chamber includes: Temperature data from multiple spatial locations inside the rapid temperature change chamber are acquired using a temperature sensor array. The first and second pressure data of the refrigeration system of the rapid temperature change chamber are obtained by a pressure sensor.

10. A rapid temperature change chamber intelligent control system, characterized in that, The system is used to implement the intelligent control method for a rapid temperature change chamber as described in any one of claims 1 to 9; the system includes a rapid temperature change chamber and a multi-source sensor array, an edge computing unit, and an actuator drive module connected in sequence; the multi-source sensor array is used to collect multi-source sensor data from the rapid temperature change chamber. The edge computing module is used to acquire the multi-source sensor data and historical control command sequence, extract state feature vectors from the multi-source sensor data, and process the state feature vectors to obtain the control command sequence. The actuator drive module is used to convert control commands in the control command sequence into drive signals to control the operation of the actuator of the rapid temperature change box.

Images