A hydrogen storage magnesium powder preparation process parameter cooperative self-adaptive control system

Abstract

The application discloses a kind of hydrogen storage magnesium powder preparation process parameter synergic adaptive control system, specifically relates to material preparation control field, including data acquisition module, first logic judging module, second logic judging module, synergic fusion module and adaptive execution module.Data acquisition module obtains process parameter and state data in real time;First logic judging module generates first intermediate decision based on case reasoning;Second logic judging module generates second intermediate decision based on fuzzy cognitive map;Synergic fusion module carries out nonlinear synergic fusion to two decisions by tensor product and neural network, generates synergic control strategy;Adaptive execution module adjusts execution mechanism parameter accordingly dynamically.The present application fuses historical experience and causal reasoning, without presetting fixed threshold value, can adapt to complex working conditions such as cold welding and parameter drift, significantly improves hydrogen storage magnesium powder yield, cycle life and system robustness.

Description

Technical Field

[0001] This invention relates to the field of material preparation control, and more specifically, to a collaborative adaptive control system for process parameters in the preparation of hydrogen storage magnesium powder. Background Technology

[0002] With the continued growth of global demand for clean energy, hydrogen energy, as a high-energy-density, zero-carbon-emission secondary energy source, has attracted widespread attention. In the storage and transportation of hydrogen energy, solid-state hydrogen storage materials have become a research hotspot due to their advantages such as high volumetric hydrogen storage density and good safety. Among them, magnesium-based hydrogen storage materials (especially magnesium powder for hydrogen storage) are considered highly promising solid-state hydrogen storage media due to their large hydrogen storage capacity, abundant resources, and low cost. However, the preparation process of magnesium powder for hydrogen storage (such as ball milling and hydrogenation processes) involves multiple process parameters (including ball milling speed, ball-to-material ratio, ball milling time, hydrogenation temperature, hydrogenation pressure, hydrogen flow rate, and catalyst addition) and various state data (such as spindle torque, vibration spectrum, and acoustic emission signals). These parameters exhibit complex nonlinear coupling and time-varying characteristics. For example, the energy input during ball milling directly affects the grain size and defect density of magnesium powder, thus affecting the kinetic performance of the subsequent hydrogenation reaction. The hydrogenation reaction itself is also affected by the interaction of multiple factors such as temperature, pressure, and flow rate. Therefore, how to achieve efficient and precise control of process parameters in the preparation of magnesium hydride powder has become a key technical bottleneck in improving the yield, purity and cycle life of magnesium hydride.

[0003] Currently, the control of the magnesium powder preparation process for hydrogen storage mainly relies on traditional PID control, fuzzy control, or a single expert system. Traditional PID control strategies are suitable for linear, time-invariant systems, but the magnesium powder preparation process for hydrogen storage is characterized by strong nonlinearity, multivariable coupling, and operating condition drift. PID controllers often require frequent manual parameter tuning and are difficult to adapt to uncertainties such as magnesium powder particle size variations, catalyst batch differences, and sensor drift. While fuzzy control can handle some nonlinear problems, its rule base construction heavily relies on expert experience. Furthermore, when facing sudden abnormal conditions such as cold welding and agglomeration, the fixed membership function and rule table cannot be dynamically adjusted, easily leading to response lag or control instability. In recent years, some studies have attempted to use neural networks or reinforcement learning methods for parameter optimization, but these methods typically require large amounts of labeled data for offline training, resulting in limited model generalization ability and a lack of explicit expression of the process mechanism. Under conditions of insufficient training data coverage, unpredictable decision biases can easily arise.

[0004] However, existing technologies still have significant shortcomings in practical applications. First, a single control logic cannot simultaneously utilize historical successes and the causal topological relationships between parameters. When the historical case library lacks cases sufficiently similar to the current operating conditions, case-based reasoning methods will produce significant decision-making errors. Furthermore, when the process mechanism undergoes slow drift (such as changes in raw material properties), the causal model based on fuzzy cognitive maps will also experience mismatch, causing the reasoning results to deviate from the optimal control direction. Second, most control systems use linear weighting or logic gates for multi-source decision fusion, which cannot effectively capture the directional consistency, conflict degree, and high-order interaction characteristics between two decisions. This easily leads to simple averaging effects when decisions are contradictory, thus missing the optimal strategy for compromise or enhancement. In addition, existing systems typically use fixed command priorities and threshold triggering mechanisms at the execution level, lacking the ability to adaptively respond to the dynamic rate of change of parameters. This not only causes frequent operation of the actuators and increased mechanical wear, but also easily leads to command conflicts or overshoot when operating conditions fluctuate drastically. In summary, there is an urgent need for a collaborative adaptive control system for process parameters that can integrate multiple reasoning mechanisms, achieve deep interactive decision-making, and be entirely data-driven, in order to solve control challenges such as strong coupling of multiple variables, time-varying operating conditions, and frequent occurrence of abnormal operating conditions in the preparation of hydrogen storage magnesium powder. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, the present invention provides a collaborative adaptive control system for the process parameters of hydrogen storage magnesium powder preparation, which solves the problems mentioned in the background art through the following scheme.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a collaborative adaptive control system for hydrogen storage magnesium powder preparation process parameters, comprising: a data acquisition module, used to acquire process parameters and status data in real time during the preparation of hydrogen storage magnesium powder;

[0007] The first logic judgment module has its input end connected to the data acquisition module. It uses a case-based reasoning method to dynamically reason about the state data and generate a first intermediate decision.

[0008] The second logic judgment module has its input end connected to the data acquisition module. It uses a fuzzy cognitive map-based method to dynamically reason about the process parameters and generate a second intermediate decision. The first logic rule and the second logic rule have different reasoning mechanisms.

[0009] The collaborative fusion module, whose input terminals are respectively connected to the first logic judgment module and the second logic judgment module, is used to perform nonlinear collaborative fusion of the first intermediate decision and the second intermediate decision to generate a collaborative control strategy.

[0010] An adaptive execution module, whose input is connected to the collaborative fusion module and whose output is connected to each execution mechanism, is used to dynamically adjust the operating parameters of each execution mechanism according to the collaborative control strategy.

[0011] Preferably, the first logical judgment module generates the first intermediate decision in the following way:

[0012] Real-time status data is matched with cases in the historical case library for similarity. The similarity is calculated based on the weighted sum of Euclidean distance and dynamic time curvature distance. The three historical cases with the highest similarity are selected, and the parameter adjustment amounts corresponding to each case are weighted and averaged according to similarity to obtain the first intermediate decision.

[0013] Preferably, the second logic judgment module generates the second intermediate decision in the following way:

[0014] A fuzzy cognitive graph is constructed with process parameters as nodes and causal relationships between parameters as edges. Nodes in the graph are activated by real-time status data. The activation value of each node is calculated iteratively using a recurrent neural network until convergence. The converged node activation value is mapped to the adjustment trend of each process parameter as a second intermediate decision.

[0015] Preferably, the collaborative fusion module performs nonlinear collaborative fusion using the following steps:

[0016] The first and second intermediate decisions are encoded into vector form, and their tensor product is calculated to generate a joint feature space. Then, a single hidden layer neural network is used to map the joint feature space into a cooperative control strategy. The activation function of the hidden layer nodes of the neural network is the hyperbolic tangent function.

[0017] Preferably, the adaptive execution module includes:

[0018] An execution strategy parser is used to decompose the coordinated control strategy into sub-instructions corresponding to each execution mechanism;

[0019] A dynamic priority scheduler dynamically adjusts the execution order of sub-instructions based on the current rate of change of the process parameters corresponding to each sub-instruction.

[0020] An instruction smoother uses a first-order lag filter to smooth the sub-instruction values ​​at adjacent time steps.

[0021] Preferably, it also includes a knowledge update module, which is connected to the output of the collaborative fusion module. This module is used to store the current process parameters, status data and corresponding collaborative control strategies as new cases in the historical case library, and periodically delete old cases in the case library whose timeliness exceeds the set time window.

[0022] Preferably, after acquiring process parameters and status data, the data acquisition module first performs timestamp alignment and missing value repair. The missing value repair adopts a spline interpolation method based on adjacent time data. The repaired data is then sent to the first logic judgment module and the second logic judgment module respectively.

[0023] Preferably, during the execution of the collaborative control strategy, the adaptive execution module monitors the response deviation of each actuator in real time and feeds back the deviation to the collaborative fusion module, which then incrementally corrects the current collaborative control strategy accordingly.

[0024] The technical effects and advantages of this invention are as follows:

[0025] 1. This invention employs both a first logic module based on case-based reasoning and a second logic module based on fuzzy cognitive graphs, which independently make decisions using historical success experiences and causal topological relationships between parameters, respectively. Since the two modules have different data sources and reasoning mechanisms, they can compensate for each other when historical cases are missing or causal relationships drift, avoiding the blind spots of a single logic in complex nonlinear scenarios and significantly improving the control system's adaptability to multivariate coupling and time-varying characteristics during the preparation of hydrogen storage magnesium powder.

[0026] 2. By calculating the tensor product of two intermediate decision vectors to generate a joint feature space, and then mapping it to a collaborative control strategy via a single hidden layer neural network, deep interaction and nonlinear fusion between decisions are achieved. Compared with traditional weighted averaging or logic gate combination, this method can dynamically capture the directional consistency and conflict degree of the two decisions, and automatically generate the optimal strategy with compromise or enhancement. Under typical abnormal working conditions such as magnesium powder cold welding, the yield is increased by about 18% and the cycle life is increased by about 35%.

[0027] 3. The entire control process does not rely on any fixed comparison thresholds: the top three similarity matches are directly selected, the iterative convergence of the fuzzy cognitive graph depends on the minimum value of the change, and the dynamic priority scheduling is based on the real-time rate of change. The system is completely data-driven and adaptively adjusts to the operating conditions, avoiding the tediousness of repeatedly tuning thresholds required by traditional PID or fuzzy controllers. At the same time, it can automatically cope with uncertainties such as material batch changes and sensor drift, and has stronger robustness and generalization ability.

[0028] 4. The adaptive execution module incorporates a dynamic priority scheduler and instruction smoother: it dynamically adjusts the execution order of sub-instructions based on the current rate of change of each process parameter, prioritizing the handling of the most unstable parameters; simultaneously, it employs a first-order hysteresis filter to smooth instruction abrupt changes, reducing wear on the actuator. Furthermore, the actuator's response deviation is fed back to the collaborative fusion module in real time for incremental correction, forming a closed loop of perception-decision-execution-correction, ensuring control accuracy and long-term stability. Attached Figure Description

[0029] Figure 1 This is a diagram showing the overall module connection relationship of the system of the present invention.

[0030] Figure 2 A first intermediate decision flowchart is generated for the first logic judgment module of the present invention.

[0031] Figure 3 A second intermediate decision logic diagram is generated for the second logic judgment module of the present invention.

[0032] Figure 4 A schematic diagram of the collaborative control strategy generated for the collaborative fusion module of the present invention.

[0033] Figure 5 This is a schematic diagram of the internal structure of the adaptive execution module of the present invention. Detailed Implementation

[0034] 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 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.

[0035] As attached Figure 1 Appendix Figure 2 and attached Figure 5 The illustrated collaborative adaptive control system for hydrogen storage magnesium powder preparation process parameters includes: a data acquisition module, a first logic judgment module, a second logic judgment module, a collaborative fusion module, an adaptive execution module, and an execution mechanism.

[0036] The specific implementation method is as follows: The data acquisition module is the system's perception layer, responsible for acquiring all measurable variables in the hydrogen storage magnesium powder preparation process in real time. In this embodiment, the preparation process of hydrogen storage magnesium powder includes a ball milling process and a hydrogenation process. In the ball milling process, magnesium powder and catalyst are loaded into a ball mill jar at a certain ball-to-material ratio, and the ball mill operates at a set speed. In the hydrogenation process, the ball-milled magnesium powder reacts with hydrogen gas in a hydrogenation reactor to generate magnesium hydride.

[0037] Specifically, the data acquisition module collects process parameters through the following sensors:

[0038] (1) Speed ​​sensor: A Hall effect speed sensor or incremental encoder is used and installed on the main shaft of the ball mill. The sensor outputs N pulses per revolution, and the ball mill speed is calculated by measuring the number of pulses per unit time, in revolutions per minute (r / min). The acquisition frequency is 10Hz, that is, the speed value is updated every 0.1 seconds. The typical range of ball mill speed is 100 to 600 r / min, and in this embodiment, it is initially set to 400 r / min.

[0039] (2) Weighing sensors: Strain gauge weighing sensors are used, installed below the feed port of the ball mill jar and below the catalyst feeder, respectively. The weighing sensor range is 0-10 kg, with an accuracy of ±0.1 g. The ball-to-material ratio is calculated in real time by measuring the mass of magnesium powder before ball milling, the total mass of the balls (stainless steel balls or zirconia balls), and the amount of catalyst added. The ball-to-material ratio is defined as the ratio of the total mass of the balls to the mass of magnesium powder, with a typical range of 10:1 to 40:1. In this embodiment, the initial ball-to-material ratio is 20:1, and the amount of catalyst added is 5 wt% of the mass of magnesium powder.

[0040] (3) Timer: The timer module built into the PLC or an independent high-precision timer is used to accumulate the ball mill running time. The timer starts when the ball mill starts and stops when it stops, and supports multiple accumulations. The typical range of ball milling time is 2 to 24 hours, and in this embodiment, it is initially set to 8 hours.

[0041] (4) Thermocouple: A K-type or T-type thermocouple is used and installed inside the hydrogenation reactor, with an insertion depth of 1 / 3 of the reactor's inner diameter, to accurately measure the temperature at the center of the reaction. The thermocouple measurement range is -50 to 600℃, with an accuracy of ±0.5℃. The typical range of hydrogenation reaction temperature is 200 to 400℃, and in this embodiment, it is initially set to 350℃.

[0042] (5) Pressure transmitter: A piezoresistive or capacitive pressure transmitter is used and installed on the gas inlet pipe at the top of the hydrogenation reactor. The measurement range is 0 to 10 MPa, with an accuracy of ±0.1%FS. The typical range of hydrogenation pressure is 1 to 5 MPa, and in this embodiment, it is initially set to 3 MPa.

[0043] (6) Mass flow meter: A thermal or Coriolis mass flow meter is used and installed on the hydrogen inlet pipeline. The measurement range is 0 to 50 L / min, and the accuracy is ±1%FS. The typical range of hydrogen flow rate is 2 to 20 L / min, and in this embodiment, it is initially set to 10 L / min.

[0044] In addition, the data acquisition module also collects the following status data for subsequent logical judgments:

[0045] (7) Spindle torque: Measured by installing a torque sensor (strain gauge or magnetoelastic type) on the ball mill spindle. The torque value directly reflects the energy input and material viscosity changes during the ball milling process. The torque will abnormally increase when magnesium powder undergoes cold welding. The typical torque range is 0–500 N·m, with a sampling frequency of 100 Hz.

[0046] (8) Vibration spectrum: Acquired by a triaxial accelerometer (piezoelectric IEPE type) installed on the outer shell of the ball mill jar or bearing housing. The acceleration signal is transformed by FFT to obtain the vibration spectrum. The change in the amplitude of the dominant frequency of the vibration spectrum can be used to determine the motion state of the ball mill media and the uniformity of material distribution. The acquisition frequency is 10kHz, and the spectrum is calculated once every 1 second.

[0047] (9) Acoustic emission signal: Acoustic emission sensor (resonant frequency 150kHz) was attached to the outer wall of the hydrogenation reactor. The acoustic emission signal can capture the high-frequency elastic waves generated by the breakup and aggregation of magnesium hydride particles and the collapse of hydrogen bubbles. This signal is very sensitive to the microscopic processes of the hydrogenation reaction and can serve as an early indicator of the reaction process.

[0048] The data acquisition module also integrates a data preprocessing unit. After acquiring all the raw data, the preprocessing unit first timestamps the data from each sensor. Since different sensors have different sampling frequencies (e.g., 100Hz for the torque sensor, 1Hz for the temperature sensor), the preprocessing unit resamples all data to 1Hz and aligns the time axis using linear interpolation. Secondly, for missing values ​​caused by momentary sensor malfunctions or communication interference, spline interpolation is used for repair: three valid data points before and after the missing point are taken, a cubic spline function is constructed, and the estimated value of the missing point is calculated. The repaired data is then sent to the first and second logic judgment modules respectively.

[0049] The first logic judgment module uses a case-based reasoning method to dynamically reason about the state data from the data acquisition module and generate the first intermediate decision.

[0050] Case-based reasoning is an artificial intelligence method based on experiential memory. Its core idea is to find the case most similar to the current situation from historical successful cases and reuse the solution from that case. In this embodiment, the historical case library is stored in non-volatile memory, and each case contains three parts:

[0051] (1) Operating condition description, i.e., the process parameters and status data at that time; (2) Control measures taken, i.e., the adjustment amount of the actuator; (3) Control effect score, calculated based on indicators such as subsequent reaction efficiency and product purity.

[0052] The specific implementation steps are as follows:

[0053] Step S201: Receive real-time status data vector from the data acquisition module This vector contains nine dimensions: spindle torque, vibration spectrum dominant frequency amplitude, acoustic emission root mean square value, hydrogenation temperature, hydrogenation pressure, hydrogen flow rate, ball mill speed, ball-to-material ratio, and ball milling time. Among them, the vibration spectrum dominant frequency amplitude is a feature extracted from the original vibration signal: first, a fast Fourier transform is performed on the acceleration signal to extract the frequency component with the largest amplitude in the range of 0–1000Hz and its amplitude.

[0054] Step S202: Convert the current state data vector The state vector of each case in the historical case library Similarity calculation is performed. This embodiment uses a weighted sum approach, combining two metrics: Euclidean distance and dynamic time curvature distance.

[0055] First, calculate the Euclidean distance:

[0056]

[0057] in These are the weighting coefficients for each dimension. The weights are set based on prior knowledge of the degree to which each parameter affects hydrogen storage performance: spindle torque and temperature have the highest weights (set to 0.2), followed by vibration and acoustic emission (each 0.15), with the remaining parameters each having a weight of 0.05. These weights can be dynamically adjusted through the knowledge update module during system operation and are not subject to preset thresholds.

[0058] Next, calculate the dynamic time bending distance. Because the preparation process of magnesium hydrogen storage powder has obvious temporal characteristics (such as changes in state at the beginning, middle, and end of ball milling), simply comparing instantaneous values ​​may not capture the similarity of the dynamic process. Dynamic Time Warping Distance (DTW) allows for non-linear alignment between two time series, thereby comparing the similarity of the changing trends. In this embodiment, the 60 sampling points before the current moment (i.e., the past 60 seconds) are used to construct a short time series, and the DTW distance is calculated with the corresponding segments in the case library. The standard DTW algorithm is used, allowing a path slope of 0.5 to 2.

[0059] Final similarity:

[0060]

[0061] This formula ensures that the smaller the Euclidean distance and DTW distance, the higher the similarity. Note that no fixed comparison threshold is used in this similarity calculation (such as "only accept cases with similarity greater than 0.8"), but instead the three cases with the highest similarity are directly selected, thus maintaining the adaptive property.

[0062] Step S203: Select the three historical cases with the highest similarity, denoted as Case A, B, and C, with similarity scores of respectively. , , Each case record includes a corresponding parameter adjustment vector. Includes ball mill speed adjustment Hydrogenation temperature adjustment amount Hydrogen flow rate adjustment Cooling water valve adjustment amount , stirring speed adjustment amount wait.

[0063] Step S204: Calculate the parameter adjustment amount to be used based on the similarity weighted average:

[0064]

[0065] Step S205: ... This serves as the output of the first intermediate decision. The data structure of the first intermediate decision is a vector containing the adjustment direction and magnitude of each parameter. If the adjustment directions of the three most similar cases are inconsistent (e.g., case A suggests increasing the temperature, case B suggests decreasing the temperature), the weighted average may yield a small value close to zero. This indicates a contradiction in historical experience, and the confidence level of the decision is low.

[0066] The second logic judgment module employs a fuzzy cognitive graph-based method to dynamically reason about process parameters from the data acquisition module, generating a second intermediate decision. A fuzzy cognitive graph is a knowledge representation model combining fuzzy logic and neural networks, using weighted directed graphs to describe the causal relationships between concepts in the system.

[0067] Step S301: Construct a fuzzy cognitive graph. In this embodiment, the set of nodes in the cognitive graph... These represent: ball mill speed, ball-to-material ratio, ball milling time, hydrogenation temperature, hydrogenation pressure, hydrogen flow rate, and catalyst dosage, respectively. Directed edges between nodes represent causal relationships, and each edge is assigned a weight. value range Positive weights indicate positive correlation, and negative weights indicate negative correlation.

[0068] The initial causal weights are set based on the physicochemical mechanisms and expert experience in the field of hydrogen storage magnesium powder, for example:

[0069] Ball mill speed → Ball mill temperature: +0.8

[0070] Ball milling temperature → hydrogenation reaction rate: +0.6

[0071] Hydrogenation temperature → Hydrogenation pressure: +0.7

[0072] Hydrogen flow rate → Hydrogenation pressure: +0.5

[0073] Hydrogenation pressure → Hydrogenation reaction rate: +0.4

[0074] Ball milling time → Catalyst dispersion: +0.6

[0075] Catalyst addition amount → hydrogenation reaction rate: +0.7

[0076] Hydrogenation temperature → Reactor safety risk: +0.5

[0077] Furthermore, there are negative causal edges:

[0078] Hydrogen flow rate → Hydrogenation temperature: -0.3

[0079] Ball mill speed → Magnesium powder particle size: -0.7

[0080] These weights will be optimized as knowledge is updated during system operation, but the initial values ​​are sufficient to ensure the system works normally.

[0081] Step S302: Activate nodes in the graph using real-time status data. Each node Activation value at time t It is determined by two parts: the external input and the weighted influence of neighboring nodes. In this embodiment, the latest process parameters obtained by the data acquisition module are normalized to... The interval serves as the initial activation value for the node. The normalization method uses a minimum-maximum transform:

[0082]

[0083] in and These are the minimum and maximum values ​​of the parameter in the historical records, and do not depend on a fixed preset threshold.

[0084] Step S303: Iteratively calculate the activation value of each node using a recurrent neural network until convergence. The iterative formula is:

[0085]

[0086] in It is the causal weight from node j to node i. These are the self-loop weights of the nodes, typically 0 or 1. The activation function chosen is the Sigmoid function.

[0087]

[0088] in For the gain parameter, this embodiment takes... The Sigmoid function maps node activation values ​​to... The interval conforms to the fuzzy logic characteristics of fuzzy cognitive graphs.

[0089] The iterative process starts from t=0 and calculates... ,Then And so on. When the sum of the absolute values ​​of the changes in the activation values ​​of all nodes is less than a minimum value (such as...), When the iteration count reaches a certain threshold, convergence is considered achieved. Since no preset iteration threshold is required, this convergence criterion is adaptive to the dynamic process. In actual testing of this embodiment, convergence typically occurs after 20 to 50 iterations.

[0090] Step S304: Map the converged node activation values ​​to the adjustment trends of each process parameter. Specifically, for each node i, calculate its converged activation value. With initial activation value The difference:

[0091]

[0092] like If the value is >0, it indicates that the parameter is trending upwards under the current causal environment and should be adjusted in the opposite direction (i.e., the setting value should be reduced to counteract the trend); if If the value is less than 0, then a positive adjustment (increasing the setpoint) should be made. The adjustment amount and... Proportional, the proportionality coefficient is obtained through offline experimental calibration: for example, the temperature adjustment coefficient is... ,Right now =+0.1 corresponds to a decrease .

[0093] The adjustment amounts of each parameter are combined into a vector, which is the second intermediate decision. Its data structure is the same as that of the first intermediate decision.

[0094] The core task of the collaborative fusion module is to nonlinearly and collaboratively fuse the first intermediate decision (experience-oriented) and the second intermediate decision (causality-oriented) to generate the final collaborative control strategy. This embodiment uses a combination of tensor product and neural network to extract the interactive features between the two decision vectors.

[0095] Step S401: The first intermediate decision vector Second intermediate decision vector The parameters are encoded as vectors. The two vectors can have different dimensions: the first intermediate decision includes 5 parameters (rotation speed, temperature, flow rate, valve position, and stirring), and the second intermediate decision includes 7 parameters (all process parameters). To unify the dimensions, this embodiment sets the adjustment amount of parameters not included in the first intermediate decision (such as ball-to-material ratio, ball milling time, and catalyst addition) to 0, expanding it to 7 dimensions. Therefore, both vectors are 7-dimensional real vectors, denoted as... and .

[0096] Step S402: Calculate the tensor product of the two vectors to generate a joint feature space. The tensor product is defined as:

[0097]

[0098] in It is a 7×7 matrix with 49 elements. The tensor product can capture... and The pairwise product relationships between the components represent the interactive information between the two decisions. For example, if the first decision suggests increasing the temperature (+2℃) while the second decision suggests decreasing the temperature (-3℃), the product term will be negative, resulting in a canceling effect during fusion; if the two decisions are in the same direction, the product term will be a large positive value, strengthening the adjustment direction.

[0099] Step S403: Combine the feature space Expanded into a 49-dimensional vector Then it is fed into a single-hidden-layer neural network for nonlinear mapping. The input layer of the neural network has 49 nodes, corresponding to... The layer has 49 elements. The number of hidden nodes is set to 20, and the activation function is the hyperbolic tangent function.

[0100]

[0101] The hyperbolic tangent function has a range of (-1, 1) and can output both positive and negative values, making it suitable for representing the direction and magnitude of adjustment. The output layer has 7 nodes, corresponding to the cooperative control strategy vector. The value range of each output node is (-1, 1), which is converted into the actual physical adjustment amount through inverse normalization (e.g., -1 corresponds to the maximum reverse adjustment, and +1 corresponds to the maximum forward adjustment).

[0102] To ensure that the single-hidden-layer neural network in the collaborative fusion module can accurately map the joint feature space to the collaborative control policy, it needs to be trained offline before system deployment. This embodiment provides a reproducible training method, as follows:

[0103] The input to the training samples is the first intermediate decision vector from the historical running records. Second intermediate decision vector The output label represents the corresponding collaborative control strategy. .

[0104] The labeling method for output tags adopts a combination of expert review and effect verification:

[0105] Periods with excellent control performance were selected from historical databases. The quantitative criteria for excellent control performance were: magnesium hydride yield >90%, magnesium hydride purity >99%, and capacity retention rate >80% after 10 consecutive hydrogen adsorption / desorption cycles.

[0106] Three senior engineers in the field independently labeled the optimal control strategies to be adopted based on actual operating conditions within the selected time period. During the labeling process, the engineers needed to comprehensively consider the equipment safety boundaries (such as temperature <400°C, pressure <5MPa) and process economics (such as unit energy consumption).

[0107] The average of the annotations from the three engineers was used as the final label for the sample. Samples with significant disagreement among the engineers (standard deviation exceeding 20% ​​of the mean) were removed to ensure the quality of the training data.

[0108] The obtained label vector Each component is normalized to its normal value through a linear transformation. , serving as a supervisory signal for the neural network.

[0109] The neural network structure is as described in the main text: 49 nodes in the input layer, 20 nodes in the hidden layer (with tanh activation function), and 7 nodes in the output layer (without activation function, directly outputting linear values).

[0110] The training configuration is as follows:

[0111] Loss function: Mean squared error (MSE) is used, and the calculation formula is as follows:

[0112]

[0113] in For batch size, For the network's predicted output, For tags.

[0114] Optimizer: The Adam optimizer is used, with an initial learning rate of 0.001.

[0115] Training parameters: Batch size is set to 32, and the total number of training epochs is 2000. An early stopping mechanism is adopted, which terminates training early when the validation set loss function no longer decreases after 50 consecutive training epochs.

[0116] Data partitioning: The dataset is randomly divided into training set, validation set and test set in a ratio of 70%, 15% and 15% respectively.

[0117] After the above training, the weight matrix of the neural network , and bias vector , It is identified and fixed in the system. When running online, only forward propagation calculations are performed, and network weights are not updated to ensure real-time control.

[0118] The output layer of the neural network outputs a 7-dimensional vector. To obtain the actual physical adjustment amounts, a fixed linear inverse normalization method is used. The mapping range and correspondence of each physical quantity are preset according to the process safety boundary and equipment capacity, as shown in Table 1 below.

[0119] Table 1:

[0120] This fixed inverse normalization mapping ensures determinism and interpretability between the network output and physical control commands, decoupling it from the network's own learning capabilities. It is important to emphasize that the neural network's structure and parameters are trained before system deployment; runtime computation involves only forward propagation and does not involve online training, thus guaranteeing real-time performance.

[0121] Step S404: Output the cooperative control strategy, namely:

[0122]

[0123] The adaptive execution module receives the cooperative control strategy and translates it into instructions executable by the actuator. This module contains three sub-units:

[0124] (1) Execution strategy parser: decomposes each component in the cooperative control strategy vector into sub-instructions of the corresponding execution mechanism. The mapping relationship is as follows:

[0125] →The frequency adjustment value (Hz) of the ball mill variable frequency motor is sent to the frequency converter via analog output or Modbus communication.

[0126] →The power adjustment value (kW) of the electric heater and the opening adjustment value (%) of the cooling water valve work together to achieve temperature control.

[0127] → The setpoint adjustment (L / min) of the hydrogen mass flow controller is sent via a 4-20mA current loop.

[0128] →Further fine-tuning of the cooling water valve opening.

[0129] → Agitator speed adjustment value (r / min).

[0130] (2) Dynamic priority scheduler: Since the execution time of different sub-instructions is different (for example, it takes several seconds to stabilize the ball mill speed, while adjusting the flow valve only takes milliseconds), and there are mutual influences between the actuators (for example, heating will increase the pressure), the scheduler dynamically adjusts the execution order according to the current rate of change of the process parameters corresponding to each sub-instruction.

[0131] Specific method: Calculate the current rate of change of each process parameter in real time, such as the rate of temperature change.

[0132]

[0133] The larger the absolute value of the rate of change, the more unstable the parameter is, and the higher the priority it should be. Sub-instructions are sorted in descending order of the absolute value of the rate of change of their corresponding parameters and executed sequentially. For example, if the current temperature change rate is as high as 5℃ / s, while the speed change rate is only 1r / min / s, the temperature adjustment instruction is executed first, followed by the speed adjustment. This dynamic sorting avoids potential coupling conflicts that may occur when multiple instructions are executed simultaneously, and it does not require a preset fixed priority; it is entirely determined by the current operating conditions.

[0134] (3) Instruction smoother: To avoid mechanical wear and system oscillation caused by frequent actuator movements, a first-order hysteresis filter is used to smooth the sub-instruction values ​​at adjacent time points. The smoothing formula is:

[0135]

[0136] in, This represents the original subinstruction value output by the parser at the current moment. This is the value of the sub-instruction actually sent in the previous moment. For filter coefficients (0 < <1). The value is dynamically adjusted according to the type of actuator: for fast-response actuators (such as flow valves). A larger value (0.8) is chosen to enable commands to be tracked quickly; this is suitable for mechanisms with high inertia (such as heaters). A smaller value (0.3) is used to avoid drastic fluctuations. This smoother does not rely on any preset "maximum rate of change" threshold, but achieves natural signal smoothing through recursive filtering.

[0137] During the execution of the collaborative control strategy, the adaptive execution module monitors the response deviations of each actuator in real time and feeds back the deviations to the collaborative fusion module. In this embodiment, the collaborative fusion module uses a proportional correction method to incrementally correct the current collaborative control strategy.

[0138] Let the current time t be the collaborative control strategy vector output by the collaborative fusion module. Its dimension is 7, and each component corresponds to the adjustment amount of different actuators (such as temperature, speed, etc.). Simultaneously, the adaptive execution module collects actual response values ​​through built-in or external feedback sensors of each actuator and calculates the response deviation vector. ,in This represents the difference between the set value and the actual value of the i-th actuator (set value minus actual value).

[0139] Revised Cooperative Control Strategy The calculation is as follows:

[0140]

[0141] in This is the correction coefficient vector (calibrated through offline experiments). This indicates element-wise multiplication.

[0142] The revised strategy is then sent back to the strategy parser for execution. This proportional adjustment method does not change the nonlinear fusion mechanism within the collaborative fusion module; it only achieves linear compensation, forming a cascaded closed loop of "coarse adjustment + fine adjustment."

[0143] The actuators include: a ball mill variable frequency motor, a hydrogen mass flow controller, an electric heater, a cooling water valve, and a stirrer. They operate independently based on sub-instructions output by the adaptive execution module. Simultaneously, each actuator is equipped with internal or external feedback sensors (such as encoders, current transformers, position sensors, etc.) to monitor actual response values ​​(actual speed, actual opening degree, actual power, etc.) in real time, and feeds back the response deviation (set value - actual value) to the collaborative fusion module. The collaborative fusion module uses this deviation to incrementally correct the current collaborative control strategy, forming a closed-loop control.

[0144] As attached Figure 1 Appendix Figure 2 Appendix Figure 3 and attached Figure 4 The illustrated collaborative adaptive control system for hydrogen storage magnesium powder preparation process parameters further includes the following optional components:

[0145] Knowledge Update Module: This module connects to the output of the collaborative fusion module. After each control decision is completed, the system automatically records the current process parameters, status data, first intermediate decision, second intermediate decision, collaborative control strategy, and subsequent control effects (based on indicators such as magnesium hydride yield, purity, and hydrogen absorption / desorption cycle stability). This record is stored as a new case in the historical case library. Simultaneously, to control the size of the case library and ensure timeliness, the knowledge update module periodically checks the timestamp of each case and deletes old cases whose timeliness exceeds a set time window. The time window is a configurable parameter; in this embodiment, it is set to 90 days, meaning only cases from the most recent 90 days are retained. This time window can be dynamically adjusted based on the stability of the system's operating environment, but to avoid conflicting with the "no preset threshold" principle, the initial value of this window can be set by the user rather than preset by the algorithm.

[0146] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0147] The following specific application example further illustrates the beneficial effects of the present invention:

[0148] In a hydrogen storage magnesium powder preparation experiment, after 6 hours of ball milling, sensor data showed that the spindle torque gradually increased from the initial 150 N·m to 210 N·m, the amplitude of the dominant frequency of the vibration spectrum decreased by about 30%, and abnormal high-frequency spikes appeared in the acoustic emission signal. This data was sent to the first logic judgment module. After matching with the historical case database, three similar cases were found, all corresponding to the condition of "localized cold welding of magnesium powder," and the corresponding solution was to "reduce the ball mill speed by 20 r / min and suspend the cooling cycle to increase the temperature and break up the cold weld." Therefore, the first intermediate decision recommended: speed -20 r / min, temperature +10℃.

[0149] Meanwhile, the second logic judgment module receives the process parameters: current rotation speed 400 r / min, temperature 65°C, and ball-to-material ratio 20:1. After iterative calculation using the fuzzy cognitive graph, it predicts that if the current parameters are maintained, the temperature will rise to 80°C within 15 minutes, exceeding the safety limit (this limit is specified by the process safety regulations, not a preset threshold of the algorithm, but exists as a constraint in the actual system). Therefore, the second intermediate decision suggests: reduce the temperature by 5°C and increase the rotation speed by 10 r / min to enhance heat dissipation.

[0150] The two intermediate decisions were clearly contradictory (the first suggestion was to increase the temperature, the second was to decrease it). The collaborative fusion module performed a tensor product calculation on the two vectors to obtain a joint feature space, which was then mapped by a neural network. The output collaborative control strategy was: rotation speed -15 r / min, temperature +2°C. This strategy compromised between the two decisions, avoiding both the safety risks of rapid temperature increases and the abandonment of the opportunity to break the cold weld by complete cooling. After final execution, the torque recovered to 160 N·m within 30 minutes, the cold weld was eliminated, and the temperature only rose to a maximum of 72°C, which is within the safe range. The magnesium hydride yield was improved by approximately 18% compared to traditional single logic control, and the hydrogen absorption / desorption cycle life was improved by approximately 35%.

[0151] This example fully demonstrates that the dual-logic parallel reasoning and nonlinear collaborative fusion architecture adopted in this invention solves the control problem caused by multi-physics field coupling and nonlinear time-varying characteristics in the preparation of hydrogen storage magnesium powder.

[0152] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A collaborative adaptive control system for process parameters in the preparation of hydrogen-storage magnesium powder, characterized in that, include: The data acquisition module is used to acquire process parameters and status data in real time during the preparation of hydrogen storage magnesium powder; The first logic judgment module has its input end connected to the data acquisition module. It uses a case-based reasoning method to dynamically reason about the state data and generate a first intermediate decision. The second logic judgment module has its input end connected to the data acquisition module. It uses a fuzzy cognitive map-based method to dynamically reason about the process parameters and generate a second intermediate decision. The first logic rule and the second logic rule have different reasoning mechanisms. The collaborative fusion module, whose input terminals are respectively connected to the first logic judgment module and the second logic judgment module, is used to perform nonlinear collaborative fusion of the first intermediate decision and the second intermediate decision to generate a collaborative control strategy. An adaptive execution module, whose input is connected to the collaborative fusion module and whose output is connected to each execution mechanism, is used to dynamically adjust the operating parameters of each execution mechanism according to the collaborative control strategy.

2. The collaborative adaptive control system for hydrogen storage magnesium powder preparation process parameters according to claim 1, characterized in that: The first logical judgment module generates the first intermediate decision in the following way: Real-time status data is matched with cases in the historical case library for similarity. The similarity is calculated based on the weighted sum of Euclidean distance and dynamic time curvature distance. The three historical cases with the highest similarity are selected, and the parameter adjustment amounts corresponding to each case are weighted and averaged according to similarity to obtain the first intermediate decision.

3. The collaborative adaptive control system for hydrogen storage magnesium powder preparation process parameters according to claim 1, characterized in that: The second logical judgment module generates the second intermediate decision in the following way: A fuzzy cognitive graph is constructed with process parameters as nodes and causal relationships between parameters as edges. Nodes in the graph are activated by real-time status data. The activation value of each node is calculated iteratively using a recurrent neural network until convergence. The converged node activation value is mapped to the adjustment trend of each process parameter as a second intermediate decision.

4. The collaborative adaptive control system for hydrogen storage magnesium powder preparation process parameters according to claim 1, characterized in that: The collaborative fusion module performs nonlinear collaborative fusion using the following steps: The first and second intermediate decisions are encoded into vector form, and their tensor product is calculated to generate a joint feature space. Then, a single hidden layer neural network is used to map the joint feature space into a cooperative control strategy. The activation function of the hidden layer nodes of the neural network is the hyperbolic tangent function.

5. The collaborative adaptive control system for hydrogen storage magnesium powder preparation process parameters according to claim 1, characterized in that: The adaptive execution module includes: An execution strategy parser is used to decompose the coordinated control strategy into sub-instructions corresponding to each execution mechanism; A dynamic priority scheduler dynamically adjusts the execution order of sub-instructions based on the current rate of change of the process parameters corresponding to each sub-instruction. An instruction smoother uses a first-order lag filter to smooth the sub-instruction values ​​at adjacent time steps.

6. The collaborative adaptive control system for hydrogen storage magnesium powder preparation process parameters according to claim 1, characterized in that: It also includes a knowledge update module, which is connected to the output of the collaborative fusion module. This module stores the current process parameters, status data, and corresponding collaborative control strategies as new cases in the historical case library and periodically deletes old cases in the case library that have exceeded the set time window.

7. The collaborative adaptive control system for hydrogen storage magnesium powder preparation process parameters according to claim 1, characterized in that: After acquiring process parameters and status data, the data acquisition module first performs timestamp alignment and missing value repair. The missing value repair adopts a spline interpolation method based on adjacent time data. The repaired data is then sent to the first logic judgment module and the second logic judgment module respectively.

8. The collaborative adaptive control system for hydrogen storage magnesium powder preparation process parameters according to claim 1, characterized in that: During the execution of the collaborative control strategy, the adaptive execution module monitors the response deviation of each actuator in real time and feeds back the deviation to the collaborative fusion module, which then incrementally corrects the current collaborative control strategy accordingly.

Images