High-volume-density anode carbon block pressing parameter self-adaption method and high-volume-density anode carbon block finished product

Abstract

The invention provides a high-volume-density anode carbon block pressing parameter self-adaption method and a high-volume-density anode carbon block finished product, relates to the technical field of industrial control and adjustment, and provides a closed-loop self-adaption method integrating accurate perception, deep cognition, self-adaption control and intelligent diagnosis. By constructing a physical twin residual norm capable of comprehensively reflecting the deviation property, a fuzzy zone in'control 'and'diagnosis' decisions in the prior art is solved, so that the system can intelligently select an optimal response strategy according to the magnitude and the property of the deviation. The control precision and the response speed in the whole nonlinear suppression period are improved; two-dimensional track matching is introduced, so that the diagnosis accuracy of complex faults is improved, and the risk of false alarm is reduced; finally, by realizing dynamic optimization of control parameters, the method provides a solid technical guarantee for continuously and stably producing the anode carbon block with high volume density and high uniformity.

Description

Technical Field

[0001] This invention relates to the field of industrial control and regulation technology, specifically to an adaptive method for pressing high-density anode carbon blocks and a finished high-density anode carbon block product. Background Technology

[0002] In the field of advanced industrial control systems, particularly in process manufacturing, the deep integration of data-driven theoretical models with physical equipment to achieve process optimization has become a key development direction. Technologies such as digital twins and cloud computing are widely used to construct theoretically optimal process baselines, aiming to improve the quality of the final product and the consistency of the production process. The core objective of such programmable control systems is to ensure that the real-time physical process accurately follows a globally optimized trajectory, which is crucial for the precision manufacturing of high-performance materials. However, in complex application scenarios such as anode carbon block pressing, existing technologies still face several challenges.

[0003] The system has limited ability to distinguish between controllable process deviations and actual equipment failures. Some technical solutions, such as those described in document CN119310939A, rely on historical data to preset process parameters. When actual operation deviates from the preset model, the system lacks a mechanism for in-depth analysis of the nature of the deviation, which may lead to inappropriate control adjustments or delays in fault diagnosis.

[0004] Secondly, existing methods for quantifying the deviation between actual processes and theoretical models have limitations. Traditional deviation calculation methods often only focus on the difference in instantaneous values, ignoring the evolution trend and morphological characteristics of the deviation over time, which may lead to misjudgments of the true state of the system.

[0005] Finally, traditional fixed-parameter controllers struggle to achieve globally optimal control due to the nonlinear and time-varying dynamic characteristics of the pressing process. Their control parameters are typically a compromise, unable to dynamically adapt to drastic changes in the physical properties of the controlled object throughout the entire process cycle, thus limiting control accuracy. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive method for pressing high-density anode carbon blocks and a finished high-density anode carbon block product, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] An adaptive method for pressing parameters of high-density anode carbon blocks, comprising the following steps:

[0009] S1: Real-time acquisition of instantaneous physical state characteristics that characterize the internal physical state of the anode carbon block, calculated based on the vibration response data of the pressing equipment;

[0010] S2: Obtain the optimal process trajectory baseline corresponding to the current pressing conditions, generated based on the cloud-based federated digital twin model;

[0011] S3: The physical twin residual calculation module is used to quantify the temporal deviation between the instantaneous physical state characteristics and the optimal process trajectory baseline to generate the physical twin residual norm.

[0012] S4: Under the condition that the physical twin residual norm is within a first preset range, a control adjustment signal is generated to adjust the suppression parameter so as to minimize the physical twin residual norm;

[0013] S5: When the physical twin residual norm exceeds the first preset range, a diagnostic alarm signal is generated to indicate that there is an abnormality in the pressing equipment or the anode carbon block raw material.

[0014] A high-density anode carbon block product is prepared by the high-density anode carbon block pressing parameter adaptive method.

[0015] Compared with existing technologies, the beneficial effects of this invention are: by constructing a "physical twin residual norm" capable of comprehensively and deeply quantifying the deviation between the physical process and the theoretical model. The core idea is not simply to calculate numerical differences, but to comprehensively evaluate the deviation from three dimensions—"numerical," "time-sensitive," and "morphological"—through a time-weighted and morphologically-aware fusion mechanism. Based on this, this invention further establishes a dual-modal response framework: when the residual norm is within a controllable preset range, the system determines it as a process deviation and initiates an adaptive gain-scheduled PID control logic. This control logic can dynamically adjust the controller parameters through a preset gain scheduling function based on real-time perceived characteristics representing the internal physical state of the anode carbon block (including acoustic vibration mode damping coefficients), thereby achieving precise and robust control of the nonlinear process. When the residual norm exceeds this range, the system determines it as a potential equipment or raw material anomaly and immediately suppresses the control action, instead initiating an intelligent diagnostic program based on a two-dimensional state trajectory. This program achieves precise location of the fault root cause by matching the temporal pattern of the deviation with a preset fault feature library. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall method flow of the present invention;

[0017] Figure 2 This is a schematic diagram illustrating the execution logic of steps S1 to S3 of the present invention;

[0018] Figure 3 The average energy time-frequency spectrum diagram, which reflects the dynamic characteristics of the suppression cycle, is shown.

[0019] Figure 4 This is a schematic diagram illustrating the execution logic of steps S4 to S5 of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Example 1:

[0023] Please see Figures 1 to 4 The present invention provides a technical solution:

[0024] An adaptive method for pressing parameters of high-density anode carbon blocks, applied to the pressing process of anode carbon blocks in a pressing device, includes the following steps:

[0025] S1: Real-time acquisition of instantaneous physical state characteristics that characterize the internal physical state of the anode carbon block, calculated based on the vibration response data of the pressing equipment;

[0026] S2: Obtain the optimal process trajectory baseline corresponding to the current pressing conditions, generated based on the cloud-based federated digital twin model;

[0027] S3: The physical twin residual calculation module quantifies the temporal deviation between the instantaneous physical state characteristics and the optimal process trajectory baseline to generate the physical twin residual norm.

[0028] S4: Under the condition that the physical twin residual norm is within the first preset range, a control adjustment signal is generated to adjust the suppression parameters to minimize the physical twin residual norm;

[0029] S5: When the physical twin residual norm exceeds the first preset range, a diagnostic alarm signal is generated to indicate that there is an abnormality in the pressing equipment or the anode carbon block raw material.

[0030] for Figure 1 It should be noted that the "cloud-enabled platform digital twin" corresponds to the "cloud-based federated digital twin model" in step S2. It represents the source of the theoretically optimal process trajectory baseline.

[0031] The steps for “Perception and Quantization: Integrating Physical State and Cloud Baseline” are: S1 + S3 (with S2 as input); this logical framework summarizes the input and computation stages of the system.

[0032] "Perception" represents "perceiving physical state", which corresponds to S1 acquiring instantaneous physical state characteristics based on vibration response in real time.

[0033] "Integrating physical state and cloud baseline" corresponds to S3, which quantifies the temporal deviation between the features obtained by S1 and the baseline of S2 to generate the physical twin residual norm.

[0034] The "Evaluation and Decision: Bimodal Judgment Based on Residual Norm" corresponds to the conditional judgment parts of steps S4 and S5; this logical framework is the core decision-making hub of the entire method. It executes the "IF condition" judgment, that is, it judges whether the "physical twin residual norm" generated by S3 is within the preset range. The result of this judgment determines whether the process will proceed to S4 or S5, hence it is a "bimodal judgment".

[0035] "Adaptive control: dynamically adjusting suppression parameters" corresponds to step S4; this summarizes the entire content of step S4. When the condition "the residual norm is within the first preset range" is met, the "adaptive control" task of generating control signals to "dynamically adjust suppression parameters" is executed.

[0036] "Intelligent Diagnosis: Precisely Locating the Root Cause of Faults" corresponds to step S5; this summarizes all the content of step S5. When the condition "the residual norm exceeds the first preset range" is met, the system executes the "intelligent diagnosis" task of generating a "diagnostic alarm signal," the ultimate goal of which is to "precisely locate the root cause of the fault."

[0037] for Figure 3 It should be noted that: Figure 3 It demonstrates the data basis used to determine the "preset characteristic frequency band" and its inherent physical process correlation. Figure 3In the graph, the horizontal axis represents the time of the pressing process, and the vertical axis represents the frequency of the vibration signal. The shades of gray in the graph indicate the relative energy at that time and frequency point, obtained by statistically averaging a large amount of historical data. Unlike traditional static spectra, this graph clearly shows the "high-energy event zone" marked by the number 5, where the overall energy is significantly enhanced, precisely corresponding to the main pressing and holding pressure stages where energy is concentrated during the pressing process. Among all frequencies, the horizontal stripe indicated by the number 3 has the darkest color and exhibits a granular visual feature, indicating that this frequency band not only has the highest average energy statistically but also the strongest dynamic variability (high variance). Therefore, this area, identified by the algorithm and precisely defined by the rectangle 4, is objectively and reproducibly determined as the "preset characteristic frequency band" of this invention, ensuring that the selected feature has the highest sensitivity and information carrying capacity for key process stages. By generating an "average energy time-frequency spectrum," a "standard" vibration energy distribution pattern can be identified. Any real-time vibration signal deviating from this "standard" pattern can be considered an anomaly, thereby triggering adaptive adjustments to the pressing parameters. The prominent frequency band 3 and high-energy region 5 in the diagram represent the normal energy response of the equipment at specific frequencies and time periods during a "standard" pressing process. If the energy distribution in these regions changes significantly during actual production, it may indicate an anomaly in the pressing process, requiring parameter adjustments. For example, an abnormally high energy level in a particular frequency band may indicate die wear or uneven pressing force.

[0038] Further explanation: In S1, the step of obtaining the instantaneous physical state characteristics includes: performing time-frequency analysis on the frame vibration response signal of the pressing equipment, and calculating the acoustic vibration mode damping coefficient as the instantaneous physical state characteristic based on the energy response within the preset characteristic frequency band.

[0039] Further explanation: Time-frequency analysis includes: real-time tracking of the spectral energy distribution of the frame vibration response signal, adaptively determining a dynamic sensitive frequency band with the most significant energy change, and calculating the acoustic vibration mode damping coefficient based on the energy response within the dynamic sensitive frequency band.

[0040] Further explanation: In S2, the optimal process trajectory baseline is an ideal time-series variation curve for the acoustic vibration modal damping coefficient generated by the cloud-based federated digital twin model through the aggregation of model parameter updates from multiple pressing devices without accessing the local original vibration response data.

[0041] The system architecture of this embodiment includes a federated learning server deployed in the cloud and multiple local clients deployed on edge industrial control computers. Each client trains a model locally using its own data, uploading only the updated gradients of the model parameters to the server. The server aggregates all gradients to update the global model, and then distributes the updated model to each client. This process ensures that the original vibration data does not leave the local machine, protecting data privacy.

[0042] Further explanation: In S3, the steps for generating the physical twin residual norm include: within a sliding time window, constructing two high-dimensional vectors from the time series of instantaneous physical state features and the corresponding time series of the optimal process trajectory baseline, and calculating the Euclidean distance between the two high-dimensional vectors as the instantaneous value of the physical twin residual norm.

[0043] Further explanation: The calculation of the physical twin residual norm further includes: applying time decay weights to the time series within the sliding time window, quantifying the morphological trend difference between the two high-dimensional vectors, and fusing the weighted Euclidean distance with the morphological trend difference to generate the physical twin residual norm.

[0044] The following is a detailed implementation description of the above content: This embodiment aims to solve the technical problem of how to accurately and robustly quantify the deviation between the "actual physical process" and the "theoretical optimal model" during the pressing process of anode carbon blocks. In the prior art, the use of fixed frequency band analysis cannot adapt to the dynamic scenario of drastic changes in the physical properties of materials during the pressing process, and it is easy to lose key information; at the same time, the use of simple Euclidean distance to measure time-series deviation ignores the temporal proximate nature of the deviation (i.e., recent deviations are more important than long-term deviations) and morphological trend (i.e., the trend of deviation change is more indicative than instantaneous values), which can easily lead to misjudgment of the system state. This embodiment solves the above-mentioned technical bottlenecks by introducing a dynamic sensitive frequency band tracking mechanism and a time-weighted and morphologically-aware residual calculation method.

[0045] In the calculation process of this embodiment, all configurable operating parameters are predefined and stored in a local spreadsheet file. The key parameter definitions are as follows:

[0046] The method for determining the "preset characteristic frequency band" is based on offline statistical analysis of historical production data. The specific steps are as follows: Historical data from at least 100 complete pressing cycles are extracted from the production database, ensuring the dataset covers common raw material batches and product specifications. Each cycle's data contains a complete time series of frame vibration response signals. For each vibration signal time series in the dataset, a short-time Fourier transform is performed to generate its own time-frequency spectrum. Subsequently, all generated time-frequency spectra are aligned on the time axis, and the energy value at each point on the frequency axis is arithmetically averaged to obtain an "average energy time-frequency spectrum" that reflects the statistically "standard" vibration energy distribution of the production line. On the "average energy time-frequency spectrum," the key "main pressing and holding" time periods during the pressing process are analyzed. A frequency range is identified where the energy value consistently ranks in the top 20% of the entire spectrum, and the standard deviation of energy variation also ranks in the top 20%. The identification of this range aims to find an "information golden frequency band" that combines high signal strength and high dynamic variability. Based on the above analysis, the identified frequency range is used as the "preset characteristic frequency band." In a specific embodiment of the present invention, by analyzing historical data from a selected press production line of a particular model, a preferred "preset characteristic frequency band" is determined to be 280 Hz to 450 Hz. This frequency band is fixed and written into an external configuration file for use by systems employing this scheme.

[0047] 1.1) The dynamic sensitive frequency band is denoted as DSFB: it is a frequency range that changes dynamically over time. Its core function is to identify the vibration frequency range that most sensitively reflects changes in the densification degree within the anode carbon block during the current pressing stage. During the pressing process of the anode carbon block, the material is gradually compacted from a loose powder state into a high-density solid, and its acoustic and vibration response characteristics change drastically. Traditional signal analysis methods rely on pre-set, fixed characteristic frequency bands for analysis. However, the fundamental flaw of this static method is that it cannot track the migration of the "most information-rich" frequency band caused by changes in material density and internal structure. When key vibration response characteristics drift out of the fixed observation window, the system becomes "blind," resulting in distorted calculated characteristic parameters that cannot accurately characterize the real physical process. To solve this technical bottleneck, this invention designs the parameter "dynamic sensitive frequency band" and its determination method. The core scientific idea of ​​the method for determining the dynamic sensitive frequency band originates from non-stationary signal processing and information theory. The basic principle is that the key dynamic information of a physical process often manifests as drastic changes in energy within a specific frequency band in the spectrum. Therefore, by tracking the entire spectrum in real time and identifying the frequency band with the largest energy gradient, i.e., the most significant change, it is possible to adaptively lock onto the "information hotspot region" that best reflects the changes in the system state at the current stage.

[0048] The "dynamic sensitive frequency band" is periodically calculated and determined by a frequency band optimization module according to the following steps: The computational logic unit continuously receives the raw frame vibration response signal time series from the accelerometer model interface in the simulation environment. For this time series, a short-time Fourier transform (STFT) is performed to generate a time-spectrum graph. In this step, a window function for segmentation needs to be set. The function of this window function is to extract a small segment of the signal for spectral analysis. In this embodiment, the Hanning window is selected. The consideration for choosing this window function is that its main lobe is wide and its side lobes are attenuated significantly, which can effectively reduce spectral leakage and improve the accuracy of energy estimation. Simultaneously, the length of the window function needs to be set. This parameter determines the trade-off between frequency resolution and time resolution. In this embodiment, the window length is set to 1024 sampling points. The technical consideration for this value is that it ensures sufficient frequency resolution to distinguish different vibration modes while maintaining a high time resolution to capture rapid changes during the suppression process. In addition, a frame overlap rate also needs to be set. This parameter ensures smooth transitions between frames. In this embodiment, the overlap rate is set to 50%. The calculated entire spectrum is divided along the frequency axis into several candidate sub-bands of equal width and without overlap. In this step, a sub-band width needs to be set. This parameter defines the granularity of dynamic tracking. In this embodiment, this parameter is set to 50 Hz. The technical consideration for setting this value is that it represents a technical balance between "ensuring the capture of independent physical vibration modes" and "avoiding excessive subdivision leading to a surge in computational load and statistical instability." For each candidate sub-band, its total energy in the current time frame is calculated, and its total energy in the previous time frame is obtained. Subsequently, the following calculations are performed: the energy value of the current frame is subtracted from the energy value of the previous frame to obtain the energy difference; then, this energy difference is divided by the energy value of the previous frame to obtain the energy change rate of the sub-band. Among all the energy change rates of the candidate sub-bands, the one with the largest absolute value is selected. The start and end frequencies of this candidate sub-band are defined as the "dynamically sensitive frequency band" at the current moment and output.

[0049] 1.2) The acoustic vibration modal damping coefficient is denoted as AVMDC; it represents a dimensionless parameter characterizing the degree to which vibrational energy is absorbed and dissipated when penetrating the anode carbon block medium, and its value is positively correlated with the instantaneous volume density of the carbon block. To achieve closed-loop control of the pressing process, an index capable of quantifying the internal densification degree of the anode carbon block in real time, accurately, and non-invasively is necessary. Direct density measurement is difficult to achieve in industrial settings. This invention proposes the "acoustic vibration modal damping coefficient" as an indirect but efficient characterization parameter by revealing the physical relationship between the attenuation of vibrational energy propagating in the medium and the medium density. The physical basis of this parameter lies in acoustics and materials science. When vibrational waves propagate in a porous medium (in this example, the carbon block during the pressing process), vibrational energy is dissipated due to friction between particles within the medium and the compression and viscosity effects of gas in the pores. As pressing progresses, the porosity decreases, the density increases, and the damping characteristics of the medium change significantly, leading to an increase in energy dissipation. The acoustic vibration modal damping coefficient AVMDC quantifies this physical phenomenon.

[0050] The "hyperacoustic modal damping coefficient" is determined by a feature calculation module through the following steps: First, the "dynamic sensitive frequency band" output by the frequency band optimization module in the previous step is obtained. Then, the total energy of the frame vibration response signal and the excitation source vibration signal within this dynamic sensitive frequency band is calculated in the current time frame, denoted as "response energy" and "excitation energy," respectively. To eliminate stable vibration noise interference generated by the equipment itself and unrelated to the pressing process, a preset benchmark value needs to be subtracted from the "response energy." In this step, a system-inherent background noise energy benchmark value is set. This parameter quantifies the average vibration energy level of the equipment in each frequency band under no-load (i.e., without pressing) conditions. This parameter is determined using an experimental calibration method: during the equipment installation and commissioning phase, the press is allowed to simulate a complete pressing cycle without material, and this is repeated multiple times. During this period, frame vibration signals are continuously collected, and the average energy value in each candidate sub-frequency band is calculated. These average values ​​are stored in an external configuration file as a benchmark. The corrected response energy is obtained by subtracting the corresponding "system-specific background noise energy baseline value" from the "response energy". Then, this corrected response energy is divided by the "excitation energy" to obtain a preliminary "normalized damping ratio". The "normalized damping ratio" is then mapped to the range of 0 to 1. This step requires two preset parameters: the historical minimum damping ratio and the historical maximum damping ratio. These two parameters are obtained through statistical analysis of a large number of historical production batches' "normalized damping ratio" data, taking the 1% and 99th percentiles of their statistical distribution, respectively, to enhance robustness to extreme outliers. The mapping calculation logic is as follows: subtract the "historical minimum damping ratio" from the current "normalized damping ratio"; then, subtract the "historical minimum damping ratio" from the "historical maximum damping ratio"; finally, divide the first result by the second result. The calculated value is the final "acoustic vibration modal damping coefficient". If the calculated value is less than 0, it is taken as 0; if it is greater than 1, it is taken as 1.

[0051] 1.3) The Time Decay Weighting Factor (TDF) is denoted as a value between 0 and 1, used to assign different importance to data points at different time points when calculating time series bias. Its core idea originates from the exponential smoothing method in time series analysis, where more recent data points have a greater impact on the current state. This parameter is a configurable item and is directly read from an external spreadsheet file. In this embodiment, the value of this parameter is set to 0.95. The technical consideration for setting this value is that it achieves a technical balance between "ensuring the system is sufficiently sensitive to changes in the latest state" and "avoiding drastic fluctuations caused by instantaneous disturbances from a single noise point." Through an optimization method based on backtesting of historical data, the optimal value of the Time Decay Weighting Factor is determined, and the reasonable range of this parameter is limited to between 0.08 and 0.99. If the value is below 0.08, the weight decay of historical data is too rapid, and the system is susceptible to transient noise, resulting in excessive artifacts. If the value is above 0.99, the weight decay is too slow, and the system is sluggish in responding to recent state changes. The specific determination steps are as follows: Prepare a historical dataset containing at least 50 known events, where each event is explicitly labeled as either "normal process fluctuation" or a specific "equipment failure type." The optimization objective is to minimize the "diagnosis-control confusion rate," that is, the proportion of events where the system incorrectly identifies "equipment failure" as "normal fluctuation" and outputs a control signal. Within the range of 0.08 to 0.99, a set of candidate TDF values ​​is generated with a step size of 0.01. Then, for each candidate value in this set, the following operations are performed: set the candidate value as the TDF parameter of the current system, and perform a complete simulation backtest using the complete labeled dataset, calculating and recording the corresponding diagnosis-control confusion rate under that TDF value. After completing the iteration of all candidate values, the TDF value that minimizes the DCCR is selected as the optimal configuration in this embodiment. In this embodiment, the optimal value determined through the above optimization process is 0.95. This value is written into an external configuration file for the system to load and use during actual runtime.

[0052] Furthermore, when assessing the deviation between the actual process and the ideal model, it is necessary to know not only the numerical differences but also the morphological differences. A stable deviation parallel to the ideal curve and a drastically fluctuating deviation deviating from the trend of the ideal curve have drastically different underlying physical meanings and require different control strategies. The "morphological trend difference" and "physical twin residual norm" of this invention are designed to solve this multi-dimensional assessment problem. The method for determining these parameters is theoretically based on vector space analysis and multi-attribute decision theory. It treats two time series curves as two vectors in a high-dimensional space and comprehensively assesses their differences by analyzing their length, direction, and relative position.

[0053] 1.4) The morphological trend difference is denoted as STD: it is a dimensionless parameter used to quantify the degree of difference in the local change trends of the two time series curves, "actual state" and "ideal baseline," within a sliding time window. This parameter is determined by a morphological analysis module. Specifically, it is achieved by obtaining the actual state time series vector and the ideal baseline time series vector within the sliding time window. A first-order difference operation is performed on each vector to obtain two trend vectors representing the slope changes of their respective curves. The cosine similarity between the two trend vectors is calculated. Subtracting the cosine similarity value from a constant 1 yields the "morphological trend difference." The closer this value is to 0, the more similar the local morphologies of the two curves; the closer it is to 2, the more opposite the morphologies.

[0054] This embodiment chooses first-order difference to construct the trend vector. The technical consideration is that first-order difference is physically similar to calculating the instantaneous velocity of a time series, which can most intuitively reflect the local increasing or decreasing trend of the series, and is computationally efficient, making it suitable for industrial scenarios with high real-time requirements. Cosine similarity is chosen to measure the difference in trend vectors because cosine similarity is not sensitive to the absolute magnitude of the vectors, but only focuses on their direction.

[0055] 1.5) The physical twin residual norm is denoted as PTRN: it is the final output and a single indicator that comprehensively measures the degree of deviation between the actual process and the ideal model.

[0056] Furthermore, in S3, directly calculating the Euclidean distance between two curves has two significant limitations. First, treating all data points within a time window as equally important does not reflect the physical reality of industrial processes, where the current state of the system is more influenced by recent events. Second, it only measures point-to-point numerical differences and cannot capture the similarity or difference between two curves in terms of "shape" or "trend." For example, two curves may be numerically far apart but have completely consistent trends, which should be significantly different physically from two curves with similar values ​​but completely opposite trends. Euclidean distance cannot reflect this difference. To address these limitations, this invention designs a "time-weighted and shape-aware residual fusion mechanism." The core idea of ​​this mechanism originates from multivariate statistical analysis and pattern recognition, aiming to comprehensively evaluate deviations from three dimensions: "numerical value," "timeliness," and "shape." Its internal working mechanism is as follows:

[0057] The time-series sequences of the instantaneous physical state characteristics and the corresponding time-series sequences of the optimal process trajectory baseline are constructed into two high-dimensional vectors, representing the actual state time-series vector and the ideal baseline time-series vector, respectively. This step aims to introduce temporal recency into the calculation of Euclidean distance. Specifically, the actual state time-series vector and the ideal baseline time-series vector within the sliding time window are obtained. A weight vector of the same size as the window is generated, with its elements increasing exponentially from far to near in time, the rate of increase controlled by the "Time Decay Weighting Factor (TDF)". The difference between the two time-series vectors is calculated to obtain a deviation vector. The deviation vector is multiplied element-wise by the weight vector to obtain a weighted deviation vector. The L2 norm of the weighted deviation vector, i.e., the square root of the sum of squares of its elements, is calculated to obtain the "time-weighted Euclidean distance" denoted as TWED.

[0058] For morphological difference quantification: aiming to independently assess the dynamic similarity of two curves, this step directly calls the defined "morphological analysis module" to calculate and output "morphological trend difference STD".

[0059] Obtain the "Time-Weighted Euclidean Distance (TWED)" and the "Structure Trend Difference (STD)". Since the dimensions of TWED are related to the original data, while STD is dimensionless, dimensional consistency verification is required. In this embodiment, the value of TWED is mapped to the interval of 0 to 1 using the Min-Max normalization method based on the statistical distribution of historical data. The normalized TWED is multiplied by a numerical bias fusion coefficient, and the STD is multiplied by another morphological bias fusion coefficient. The two products are then added together. The two fusion coefficients are read from an external configuration file, and their sum is 1, used to adjust the contribution of numerical and morphological biases to the final result. The sum is output as the final "Physical Twin Residual Norm (PTRN)".

[0060] In this step, it is necessary to set the numerical deviation fusion coefficient and the morphological deviation fusion coefficient. These two parameters adjust the relative importance of the "time-weighted Euclidean distance" and the "morphological trend difference" in the final residual norm calculation. In this embodiment, the numerical deviation fusion coefficient is set to 0.6, and the morphological deviation fusion coefficient is set to 0.4. The technical consideration for this value is that, in the scenario of pressing anode carbon blocks, achieving numerical accuracy remains the primary goal, but the stability of the process (morphological similarity) is also crucial. This weighting allocation, while ensuring the dominance of numerical deviation, gives morphological deviation a significant influence; this ratio performs optimally on the key performance indicator of "diagnosis-control confusion rate."

[0061] The above fusion mechanism offers the following benefits: by introducing time weights, the system becomes more sensitive to recent, minor deviations that foreshadow changes in state. By fusing differences in morphological trends, the system can distinguish between "benign deviations" (including those with different values ​​but a good trend) and "malicious deviations" (including those with similar values ​​but diverging trends), effectively avoiding false alarms caused by instantaneous numerical disturbances.

[0062] The overall calculation process of the method of the present invention is implemented through data exchange in a simulation environment. The specific steps are as follows: The calculation logic unit of the program reads the external spreadsheet file and loads all preset configuration parameters, including but not limited to candidate sub-band division, system inherent background noise energy benchmark value, linear mapping function parameters, sliding window size, time attenuation weight factor, and fusion coefficient, etc.

[0063] The computational logic unit cyclically acquires the original frame vibration response signal and excitation source vibration signal from the sensor model interface of the simulation environment. Then, it executes all the computational model steps of the "frequency band optimization module" and the "feature calculation module", and outputs the calculated and normalized "acoustic vibration modal damping coefficient AVMDC" as a state variable to the subsequent processing module.

[0064] The computational logic unit sends the current work order information to the cloud-based federated digital twin model interface in the simulation environment through a simulated network interface. After receiving the information, the twin model interface returns a time-series data sequence containing the ideal value of the "acoustic vibration modal damping coefficient AVMDC" as the "optimal process trajectory baseline".

[0065] The computational logic unit initiates the "Physical Twin Residual Calculation Module." This module processes the corresponding outputs as inputs, following the computational model steps of "weighted distance calculation," "morphological difference quantification," and "residual fusion." Finally, the calculated "Physical Twin Residual Norm PTRN" is output as a numerical value to the decision engine model interface in the simulation environment for subsequent control and diagnostic decisions.

[0066] In this embodiment, the spreadsheet file manages parameters modularly through multiple worksheets. When the program script starts, it first reads all the contents of this spreadsheet file and loads them into a memory data structure. Subsequent calculation modules will directly query and obtain the required configuration parameters from this memory structure. The sliding window size is one of the core operating parameters determining the performance of the residual calculation module in this invention. To objectively determine its optimal value, this invention provides a systematic optimization method based on historical data backtesting. The reasonable range of this parameter is closely related to the time scale of the specific process applied. In the anode carbon block pressing scenario of this embodiment, its reasonable range is limited to between 20 and 100 sampling points. If the window size is less than 20 sampling points, the information contained is insufficient to stably determine the morphological trend and is easily affected by noise from a single data point; if the window size is greater than 100 sampling points, the average system latency increases, which may lead to untimely response to rapidly occurring equipment anomalies. The process for determining the optimal value within this range is as follows:

[0067] The same labeled historical dataset containing known "normal process fluctuations" and "equipment failure types" events was used when determining the "time decay weighting factor." The optimization objective was also set to minimize the "diagnosis-control confusion rate." A set of candidate "sliding window size" values ​​was generated within a range of 20 to 100 sampling points, with a step size of 5 sampling points. Then, for each candidate value in this set, the following operations were performed: the candidate value was set as the "sliding window size" parameter of the current system, while keeping other parameters such as TDF at their determined optimal values. A complete simulation backtest was performed using the full labeled dataset, and the diagnosis-control confusion rate corresponding to this window size was calculated and recorded. After traversing all candidate values, the "sliding window size" value that minimizes the diagnosis-control confusion rate was selected as the optimal configuration in this embodiment. In this embodiment, the optimal value determined through the above optimization process is 50 sampling points. This value is finally fixed and written to a worksheet in a spreadsheet file for the system to load and use during actual operation.

[0068] Furthermore, it should be further explained that the core output of steps S1 to S3 of this invention is the Physical Twin Residual Norm (PTRN). This parameter is normalized, and its value range is limited to the interval between 0 and 1. As a comprehensive and quantitative indicator, it is used to characterize the degree of deviation between the physical pressing process and the ideal process trajectory defined by the cloud-based digital twin model.

[0069] As the output value of the Physical Twin Residual Norm (PTRN) approaches 0, the characteristic, reflected in the time-series curve of the Acoustic-Vibration Modal Damping Coefficient (AVMDC), shows a high degree of consistency with the optimal process trajectory baseline in both numerical magnitude and trend. Technically, this means that the actual anode carbon block pressing process is in a highly stable and optimally expected state, and the densification process of the material fully conforms to the model prediction, requiring no intervention.

[0070] When the output value of the Physical Twin Residual Norm (PTRN) approaches 1, it indicates a significant deviation between the physical system and the ideal model. This deviation may stem from one or two aspects: first, a large numerical difference, i.e., an increase in the Time-Weighted Euclidean Distance (TWED); second, a severe divergence in dynamic trends, i.e., an increase in the Morphological Trend Difference (STD). Technically, this indicates an unexpected disturbance or anomaly in the pressing process, such as sudden changes in raw material batch characteristics, unstable hydraulic system pressure, or abnormal wear of equipment mechanical components. Therefore, a PTRN value approaching 1 is a powerful signal to trigger subsequent diagnostics and adaptive control, and its magnitude directly relates to the necessity and intensity of intervention.

[0071] The value of the physical twin residual norm PTRN is determined by multiple key input parameters and internal states. The following analysis examines their influence relationships to demonstrate the logical completeness of the algorithm design.

[0072] When other conditions remain unchanged, within the sliding time window, the increase in the point-to-point numerical difference between the actually measured "Acoustic and Vibration Modal Damping Coefficient (AVMDC)" time series and the "Optimal Process Trajectory Baseline" will lead to a monotonically increasing Physical Twin Residual Norm (PTRN). This is a clear positive correlation. According to the technical solution of this invention, the "Acoustic and Vibration Modal Damping Coefficient (AVMDC)" is a direct quantitative representation of the internal physical state of the anode carbon block. The "Optimal Process Trajectory Baseline" defines the ideal evolution path of this physical state on the time axis. Therefore, the numerical difference between the two directly reflects the gap between physical reality and theoretical optimality. The residual calculation module of this invention quantifies this difference by calculating the "Time Weighted Euclidean Distance (TWED)". The larger the difference between AVMDC and the baseline, the larger the calculated TWED, which, through weighted fusion, leads to a corresponding increase in the final PTRN value. This positive correlation design accurately maps the deviation magnitude of the physical world to the numerical magnitude of the residual index.

[0073] When other conditions remain constant, within the sliding time window, an increasing deviation between the trend of the AVMDC time series and the trend of the optimal process trajectory baseline will lead to a monotonically increasing physical twin residual norm (PTRN). This is also a clear positive correlation. A deviation in trend (including a baseline requiring AVMDC to increase over time, while the actual measured AVMDC decreases) indicates a more serious systemic problem than a simple numerical deviation. This invention quantifies this trend difference by calculating the cosine similarity of the first-order difference vectors and converting it into a "morphological trend difference (STD)". When the trends of the two curves are completely consistent, the cosine similarity is 1 and the STD is 0; when the trends are completely opposite, the cosine similarity is -1 and the STD is 2 before normalization. An increase in STD, through weighted fusion, also leads to an increase in the final PTRN value. This design allows the algorithm to focus not only on "whether the state is correct" but also on "whether the state change is correct," improving the early warning capability for potential faults.

[0074] The magnitude of the "Time Decay Weighting Factor (TDF)" positively modulates the sensitivity of the Physical Twin Residual Norm (PTRN) to recent deviations. That is, when deviation events of the same magnitude occur, a higher TDF value will cause PTRN to rise and fall more quickly and dramatically. TDF is a core internal parameter in the calculation of "Time Weighted Euclidean Distance (TWED)". Based on its role in the algorithm, a TDF value close to 1 means that when calculating the weighted distance, the nearest data point within the sliding window is given a high weight, while the weight of more distant data points decays rapidly. This aligns with the real-world logic of industrial control: the current state of a system and its short-term future predictions are primarily determined by its recent historical behavior. Therefore, when a deviation has just occurred, a higher TDF amplifies the impact of this new deviation, causing TWED and PTRN to rise rapidly; when the deviation disappears and the system returns to normal, a higher TDF also rapidly reduces the weight of older deviation data, causing PTRN to quickly fall back to normal levels. This parameter design gives the invention adjustable time sensitivity, enabling it to adapt to industrial processes with different dynamic response characteristics.

[0075] Further explanation: In S4, the control adjustment signal is used to drive the proportional-integral-derivative controller to dynamically modify the pressing rate or holding time parameters, thereby minimizing the physical twin residual norm.

[0076] Further explanation: The proportional, integral, and derivative gain parameters of the proportional-integral-derivative controller are dynamically adjusted based on the current values ​​of the instantaneous physical state characteristics that characterize the internal physical state of the anode carbon block obtained in step S1, through a preset gain scheduling function.

[0077] Further explanation: In S5, when the physical twin residual norm exceeds the first preset range, a diagnostic alarm signal is generated. The steps include:

[0078] The first step is to suppress the generation of the control adjustment signal;

[0079] The second step is to match the temporal pattern of the physical twin residual norm with a preset fault feature library to generate the diagnostic alarm signal containing the fault root cause category.

[0080] To further explain, the matching step in the second step further includes:

[0081] Using the time-weighted Euclidean distance (TWED) and morphological trend difference (STD) that constitute the physical twin residual norm as a two-dimensional state coordinate, an abnormal state trajectory is constructed in the preset fault state space.

[0082] The similarity between the abnormal state trajectory and the reference fault trajectory stored in the fault feature library is calculated to determine the fault root cause category.

[0083] The following is a detailed description of the implementation of the above content: In the calculation process of this embodiment, all newly added configurable running parameters are stored together with the preceding parameters in a local spreadsheet file.

[0084] The first preset range refers to dividing the range of the physical twin residual norm value into a normal area, a control area, and a diagnostic area by a first control threshold and a second diagnostic threshold, respectively.

[0085] 2.1) The conditions for the first preset range include: denoted as Th. c1 The second diagnostic threshold is denoted as Th. d2 These two dimensionless parameters define three behavioral intervals for the PTRN value range: the "normal region" (PTRN...). <Th c1 ), "Control Zone" (Th c1 ≤PTRN <Th d2 ) and "diagnostic zone" (PTRN≥Th d2 First control threshold Th c1 The second diagnostic threshold Th d2 As a configurable item, it is read directly from an external spreadsheet file. The value is determined based on statistical analysis of a large amount of historical normal and abnormal operating data. In this embodiment, Th c1 It was set to 0.15, a value corresponding to the 95th percentile of the PTRN distribution under normal production fluctuations. Th d2It was set to 0.50, which corresponds to the 5th percentile of the PTRN distribution under confirmed fault conditions. This setting strikes a technical balance between "avoiding an overreaction to normal fluctuations" and "ensuring a timely response to real faults."

[0086] It should be noted that: regarding the first control threshold Th c1 With the second diagnostic threshold Th d2 The determination method is obtained through a reproducible statistical calibration process, with the following specific steps: At least one thousand complete, fault-free pressing cycle data points are extracted from the historical production database to form a "normal operating condition dataset." Simultaneously, all pressing cycle data points with confirmed and labeled fault types are extracted to form an "abnormal operating condition dataset." The raw vibration data of each pressing cycle in the above two datasets are input into the calculation process of steps S1 to S3 of this invention to calculate the complete time series of the corresponding Physical Twin Residual Norm (PTRN) for each cycle. For the "normal operating condition dataset," all numerical points in all PTRN time series are summarized, and its probability density distribution histogram is plotted. Similarly, the same operation is performed for the "abnormal operating condition dataset" to generate the PTRN probability density distribution under abnormal operating conditions. Based on the above statistical distribution, quantile calculations are performed.

[0087] To determine the first control threshold Th c1 On the PTRN distribution of the "normal operating condition dataset", its 95th percentile is calculated. The logic of this calculation is as follows: sort all PTRN values ​​of normal operating conditions from smallest to largest, and find the value located at the 95th percentile of the total number after sorting. In this embodiment, the first control threshold Th c1 It was determined to be 0.15. The first control threshold Th c1 Its function is to define a boundary so that 95% of normal system fluctuations will not trigger the control logic, thereby ensuring the stability of the system.

[0088] To determine the second diagnostic threshold Th d2 On the PTRN distribution of the "abnormal working condition dataset", its 5th percentile is calculated. This calculation logic is similar to the former, but it finds the value located at the 5th percentile in the abnormal dataset. In this embodiment, the second diagnostic threshold Th... d2 It was determined to be 0.50. The second diagnostic threshold, Th... d2 Its function is to set a high-confidence anomaly trigger line, ensuring that the diagnostic program will only be activated when the PTRN value enters the typical fault area, effectively avoiding false alarms.

[0089] 2.2) The gain scheduling function is denoted as GSF: it represents a preset mapping relationship, and its core function is to dynamically provide a set of optimal gain parameters for the PID controller based on the real-time "Audio-Vibration Modal Damping Coefficient (AVMDC)" value, including proportional gain Kp, integral gain Ki, and derivative gain Kd. In this embodiment, this function is implemented in the form of a piecewise linear lookup table, stored in a dedicated worksheet of an external spreadsheet file. Each row of this table contains four values: an interval endpoint value of AVMDC, and the corresponding Kp, Ki, and Kd values. The generation of this lookup table originates from a series of offline system identification experiments: different AVMDC values ​​correspond to different suppression stages. By injecting a step signal and analyzing the system response, the optimal PID gain for that stage is determined using engineering tuning methods such as Ziegler-Nichols. Connecting the tuning results of multiple stages constitutes the gain scheduling function. Specifically, for nonlinear processes like anode carbon block pressing, where the characteristics (system stiffness, damping) of the controlled object change drastically over time, no single fixed set of gain parameters can achieve optimal control throughout the entire operating range. This leads to sluggish performance of traditional PID controllers in some stages and overshoot and oscillations in others. To address this fundamental technical bottleneck, this invention designs a "gain scheduling function," designed to enable the controller to "learn" to adaptively adjust its behavior based on the system state. The core scientific idea behind the parameter determination method originates from the "gain scheduling control" paradigm in modern control theory. The basic principle of this paradigm is to divide the controlled system into several operating points or operating regions, where the system can be approximated as a linear system. By individually tuning a set of optimal controller parameters for each region and associating these parameters with a "scheduling variable" that characterizes the current operating region of the system, adaptive compensation for the system's nonlinear characteristics can be achieved. In this invention, the "acoustic vibration modal damping coefficient (AVMDC)" is selected as the ideal scheduling variable because it accurately reflects the degree of material densification. The generation of the "gain scheduling function" is a system identification process completed through offline experimental calibration. The standard operating procedure is as follows:

[0090] The experimental objective is to establish the mapping relationship between the "Acoustic Vibration Modal Damping Coefficient (AVMDC)" and the optimal PID gain parameters (Kp, Ki, Kd). The experimental object is an anode carbon block pressing device in standard operating condition. The entire value range of AVMDC is divided into several non-overlapping continuous sub-intervals. In this step, a working area division number is set. This parameter determines the fineness of gain scheduling. In this embodiment, the value of this parameter is set to 10, that is, the value range of AVMDC is divided into ten sub-intervals with a width of 0.1. The technical consideration for setting this value is to achieve a technical balance between "ensuring that the dynamic characteristics of the system in each area are approximately linear" and "avoiding excessive experimental calibration workload". For each AVMDC sub-interval, the following operations are repeated: start a normal pressing cycle; continuously monitor the real-time value of AVMDC. When AVMDC enters the current target sub-interval, a preset step response test signal is injected into the hydraulic control system of the pressing device through software. In this embodiment, the target pressure setpoint of the main pressure cylinder is instantly increased by 5%. The timing response curve of AVMDC is fully recorded from the start of the injected signal to the system's re-stabilization.

[0091] For the step response curves acquired for each AVMDC sub-interval in the previous step, an engineering tuning method is applied to calculate the optimal PID gain. This embodiment uses the response curve method within the Ziegler-Nichols tuning method. The calculation logic is as follows: three characteristic parameters are measured on the response curve: the system's pure delay time, rise time, and maximum slope; then, based on a series of pure arithmetic operations defined by this tuning method for these three characteristic parameters, the recommended values ​​for the proportional gain Kp, integral gain Ki, and derivative gain Kd corresponding to the operating region are calculated respectively. All sub-intervals and their corresponding optimal (Kp, Ki, Kd) triples are organized into a lookup table. This lookup table is the entity of the "gain scheduling function" defined in this invention. Before actual deployment, simulation or actual operation verifies that when using this function for scheduling control, the overall system performance (including convergence speed and overshoot) is significantly improved compared to using any set of fixed gains. This lookup table is the entity of the "gain scheduling function" defined in this invention. For example, in a specific embodiment of the present invention, for the "acoustic vibration modal damping coefficient AVMDC" value being in the sub-range of 0.3 to 0.4, by performing the above-mentioned step response test and Ziegler-Nichols tuning method, the final determined set of preferred gain parameters is as follows: the proportional gain Kp is set to 2.5, the integral gain Ki is set to 1.2, and the derivative gain Kd is set to 0.8. This set of parameters ensures that in the intermediate stage where the material has been initially compacted but not yet fully densified, the control system can achieve a fast response and control effect without significant overshoot.

[0092] 2.3) The reference fault trajectory is denoted as RFT: it is a trajectory represented by an ordered set of points stored in a two-dimensional "fault state space". Each reference fault trajectory uniquely and typically represents a known fault mode (including "hydraulic pump leakage" and "excessive moisture in raw materials"). This parameter is generated by an offline model training module and stored in a fault feature library. The generation steps are as follows: collect a large amount of historical data of labeled fault types. For each fault type, extract the time-weighted Euclidean distance (TWED) and morphological trend difference (STD) time-series data for a period of time before and after its occurrence. Aggregate these two-dimensional time-series data of similar faults into a most representative central trajectory using a sequence averaging algorithm such as "dynamic time warping centroid averaging", which serves as the "reference fault trajectory" for that fault type.

[0093] Specifically, accurate fault diagnosis relies on having "fingerprints" or "templates" that accurately describe various fault modes. Traditional fault database construction methods often depend on manually set thresholds or simple one-dimensional waveform templates. This approach is sensitive to noise and cannot capture the complex dynamic characteristics of fault evolution. Especially when different faults exhibit similar symptoms, misdiagnosis is highly likely. To address the robustness and accuracy issues of diagnosis, this invention proposes a machine learning-based "reference fault trajectory" generation method, aiming to learn and extract the most essential dynamic "genes" of each fault mode from high-dimensional data. Its core idea originates from pattern recognition and time series data mining in machine learning, prioritizing "sequence clustering" and "sequence averaging" techniques. The basic principle is that although the details of each occurrence of the same type of fault may differ, their evolutionary trajectories in a certain abstract feature space share a common pattern. By performing a special "averaging" operation on a large amount of trajectory data of the same type of fault, the "centroid" of this trajectory cluster, i.e., the "reference fault trajectory," can be found. It represents the most stable and typical spatiotemporal evolution paradigm of this type of fault.

[0094] The offline model training module operates as follows: The model aims to generate a unique reference fault trajectory for each known fault type that needs to be identified by the system (including "hydraulic pump leakage," "excessive moisture in raw materials," and "loose frame bolts"). The training dataset is derived from the accumulation of long-term operating data from multiple pressing devices, containing at least hundreds of fault events manually labeled by domain experts. Each event record includes complete high-frequency vibration data for a period of time before and after the fault occurred, along with a clear fault category label.

[0095] For each labeled fault event in the training dataset, perform the following operations: extract the original vibration signal within a preset time window (2 seconds before and 3 seconds after) before and after the fault occurrence point. Input this signal segment into the complete calculation process described in steps S1 to S3 of this invention to obtain two corresponding time series: "Time Weighted Euclidean Distance (TWED)" and "Morphological Trend Difference (STD)". Combine these two one-dimensional time series into a two-dimensional "abnormal state trajectory", which is an ordered sequence composed of (TWED, STD) coordinate points.

[0096] For each fault category, the following operations are performed: All "abnormal state trajectory" samples belonging to that category in the dataset are collected into a set. The Dynamic Time Warping-Barycenter-Averaging (DBA) algorithm is applied to iteratively average all trajectories in the set to calculate the "centroid trajectory" for that category. This embodiment chooses the DBA algorithm because it uses dynamic time warping to intelligently align trajectories of different lengths with temporal scaling before averaging. This ensures that the averaging process preserves key morphological features of the trajectory, rather than "blurring" them due to temporal misalignment. The calculation logic of the DBA algorithm is as follows: A trajectory in the set is randomly selected as the initial "centroid trajectory"; then, an iterative loop is entered. In each iteration, all other trajectories in the set are dynamically time-warped and aligned with the current "centroid trajectory," and each point of the "centroid trajectory" is updated according to the alignment path; this iterative process is repeated until the "centroid trajectory" no longer changes significantly, or the preset number of iterations is reached. The centroid trajectory of each fault category obtained after final convergence, along with its category label, is stored in the system's fault feature library as a "reference fault trajectory" for matching during online diagnosis. In a specific embodiment of the invention, by training on a large number of fault samples labeled "hydraulic pump internal leakage," the final generated reference fault trajectory is determined to be an ordered sequence of two-dimensional coordinate points containing five key time nodes: the first point (0.1, 0.8), the second point (0.2, 0.9), the third point (0.4, 0.95), the fourth point (0.7, 0.85), and the fifth point (0.6, 0.6). This trajectory clearly depicts in the two-dimensional fault state space the typical pattern of the co-evolution of the time-weighted Euclidean distance (TWED) of the system deviation and the morphological trend difference (STD) when this type of fault occurs. That is, both the magnitude of the deviation and the morphological trend difference increase rapidly at first, reach a peak, and then slowly decrease while the morphological difference recovers rapidly.

[0097] 2.4) The trajectory distance metric is denoted as TDM: it represents a quantitative index used to calculate the similarity between two two-dimensional trajectories; the two two-dimensional trajectories represent the currently observed abnormal state trajectory and the reference fault trajectory in the database; this embodiment uses the Dynamic-Time-Warping (DTW) algorithm as the trajectory distance metric. Its core idea originates from the field of speech recognition, allowing non-linear distortion of the time axis to find the optimal alignment path between two sequences. The calculation steps are as follows: construct a distance matrix, where each element represents the Euclidean distance between the two trajectories at the corresponding time point. Based on the dynamic programming concept, find a path from the lower left corner to the upper right corner of the matrix with the minimum cumulative distance. The cumulative distance value along this path is used as the final "trajectory distance metric" output.

[0098] The innovative mechanism of step S4 in this embodiment lies in adaptive gain-scheduled PID control: the anode carbon block pressing process is a typical nonlinear, time-varying system. In the initial stage of pressing, the material is loose and the system inertia is small, requiring the controller to have a strong proportional action for rapid response; in the final stage of pressing, the material is dense and the system is rigid, and an excessively large proportional action can easily lead to overshoot and oscillation. At this time, it is necessary to reduce Kp and appropriately increase the derivative action, Kd, ​​to predictively suppress overshoot. A fixed-gain PID controller cannot simultaneously meet these two contradictory needs, and its parameters can only be selected as a compromise, resulting in the inability to achieve optimal control effect throughout the entire pressing cycle. To solve this problem, this invention designs an "adaptive gain-scheduled PID control" mechanism. This mechanism directly links the controller's performance to the physical state of the system. Its internal working mechanism is as follows: In each control cycle, the control module obtains the PTRN output from step S3 as the error input e(t). At the same time, it obtains the latest "acoustic vibration modal damping coefficient AVMDC" value output from step S1. Next, using the AVMDC value as an index, the "Gain Scheduling Function GSF" lookup table is queried to obtain the optimal Kp, Ki, Kd values ​​for the current state. Finally, it uses this dynamically acquired set of gain parameters to perform standard PID calculations (including the calculation and summation of proportional, integral, and derivative terms) on the error input e(t) to generate the final control adjustment.

[0099] This mechanism enables the PID controller to "sense" the real-time physical state of the controlled object, the anode carbon block, and adjust its behavior accordingly. This allows the control system to operate with near-optimal dynamic performance throughout the entire nonlinear process, achieving faster convergence, smaller overshoot, and higher control accuracy than traditional PID controllers, ultimately minimizing the PTRN more efficiently.

[0100] The innovative mechanism of step S5 in this embodiment lies in two-dimensional fault state space trajectory diagnosis: Traditional fault diagnosis relies on pattern matching of a one-dimensional time series of a single indicator (including PTRN). The drawback of this method is that it loses the intrinsic structural information of the deviation. For example, two completely different fault precursors, "large numerical deviation but normal morphology" and "small numerical deviation but drastically abnormal morphology," may exhibit similar patterns on the one-dimensional PTRN sequence, leading to confusion and misjudgment. Therefore, this invention designs a "two-dimensional fault state space trajectory diagnosis" mechanism. This mechanism elevates the diagnostic problem from one-dimensional pattern matching to two-dimensional trajectory classification. Its working mechanism is as follows: When PTRN triggers the diagnostic logic, the system backtracks and extracts the time series data of two components, "Time Weighted Euclidean Distance (TWED)" and "Morphological Trend Difference (STD)," from a period before the anomaly occurred. This pair of data (TWED, STD) is regarded as the coordinates of a point moving in a two-dimensional "fault state space," thereby constructing an "abnormal state trajectory." Subsequently, the diagnostic module invokes the "Track Distance Metric (TDM)" algorithm to calculate the distance between the newly generated trajectory and each "Reference Fault Trajectory (RFT)" stored in the fault feature library. Finally, the fault root cause category corresponding to the reference fault trajectory with the smallest distance is selected as the final diagnostic result. This two-dimensional trajectory analysis method utilizes both the "magnitude" information of the deviation represented by the TWED axis and the "property" information represented by the STD axis, resulting in an information dimension far exceeding that of a one-dimensional sequence. This enables the diagnostic system to capture the unique "fingerprint" trajectories of different faults in the state space, improving the accuracy of diagnosis and the ability to distinguish similar fault modes, achieving a technological leap from "anomaly detection" to "precise root cause diagnosis."

[0101] The decision-making and execution process of the method described in this invention is implemented through data exchange in a simulation environment, and its specific steps are as follows:

[0102] The program's computational logic unit reads an external spreadsheet file and loads all preset configuration parameters, including but not limited to the first control threshold Th. c1 Second diagnostic threshold Th d2 A complete gain scheduling function lookup table and a fault feature library containing all reference fault trajectories. The computational logic unit continuously receives the PTRN values ​​output by the "Physical Twin Residual Calculation Module" in step S3.

[0103] If the current PTRN value is greater than or equal to Th c1 And less than Th d2 Then the "Adaptive PID Control Module" is activated. The Adaptive PID Control Module includes:

[0104] Obtain the latest AVMDC value → Query the gain scheduling function GSF lookup table based on the AVMDC value to determine the current Kp, Ki, Kd → Perform PID calculation using dynamic gain with PTRN as the error → Encode the calculated control adjustment into a control word representing the percentage of pressure rate fine-tuning or the milliseconds of pressure holding time fine-tuning, and output it to the PLC model interface in the simulation environment.

[0105] If the current PTRN value is greater than or equal to Th d2 If this is the case, the "Fault Diagnosis Module" will be activated. The Fault Diagnosis Module includes:

[0106] Immediately send a "control suppression" command word to the PLC model interface. This command word is configured to make the PLC ignore any subsequent adjustment signals from the adaptive PID control module. → Start the "two-dimensional fault state space trajectory diagnosis" process, construct the abnormal state trajectory, and perform trajectory distance measurement (TDM) calculation with the reference fault trajectory in the library. → Encode the fault root cause category ID with the highest similarity to the matched one into a diagnostic alarm signal and output it to the human-machine interface (HMI) model interface in the simulation environment to display the alarm information.

[0107] If the current PTRN value is less than Th c1 If the system is in monitoring mode, it will not generate any control or diagnostic signals.

[0108] Furthermore, the following detailed implementation description is provided regarding the above content: In the technical solution of this invention, the physical twin residual norm PTRN is the core state indicator driving all subsequent decision-making and control behaviors. Its numerical change trend has a clear technical meaning:

[0109] As the value of the Physical Twin Residual Norm (PTRN) approaches zero, it indicates that the actual operating state of the monitored physical pressing equipment closely matches the optimal theoretical baseline for the current operating conditions deduced by the cloud-based swarm intelligence model. This directly reflects that the production process is in an efficient, stable, and best-practice-compliant ideal state.

[0110] As the physical twin residual norm (PTRN) deviates from zero and gradually increases, the deviation between the actual behavior of the physical device and the theoretical optimal baseline is widening. This trend indicates the emergence of process deviations or potential anomalies, and its magnitude directly quantifies the severity of the deviation.

[0111] The physical twin residual norm PTRN serves as the error input e(t) to the PID controller. When other parameters remain constant, an increase in the physical twin residual norm PTRN will lead to a monotonically increasing absolute value of the control adjustment signal calculated by the PID controller. This is a positive correlation.

[0112] The acoustic vibration modal damping coefficient (AVMDC) indirectly affects the control adjustment signal through the gain scheduling function (GSF). This relationship is non-linear. When the AVMDC value is low, it indicates loose material, and the GSF outputs a larger set of gain parameters, making the controller's response more aggressive to the physical twin residual norm (PTRN) error. Conversely, when the AVMDC value is high, it indicates dense material, and the GSF outputs a smaller set of gain parameters, resulting in a smoother controller response.

[0113] The positive correlation between the physical twin residual norm (PTRN) and the control output is a fundamental principle of closed-loop negative feedback control systems, ensuring that larger deviations trigger stronger corrective actions. The nonlinear relationship between the acoustic vibration modal damping coefficient (AVMDC) and the PID gain is one of the core innovations of this invention, accurately mapping the physical reality of the anode carbon block pressing process: in the initial pressing stage, large molding errors need to be quickly eliminated, hence high gain is used; in the final pressing stage, the system rigidity is high, and to avoid overshoot and equipment damage, low gain must be used for fine-tuning. This design allows the control law to dynamically match the ever-changing dynamic characteristics of the controlled object.

[0114] Time-weighted Euclidean Distance (TWED) and Morphological Trend Difference (STD): These two parameters together constitute the coordinates of points in the two-dimensional fault state space. Their individual numerical increases or decreases do not have a simple positive or negative correlation with the final diagnostic result. Their influence is reflected in the time-series trajectory formed by their combination. The final diagnostic output depends on which reference fault trajectory (RFT) in the fault feature library has the highest similarity to this trajectory. This design is based on a profound insight: a single magnitude of deviation (e.g., PTRN) is insufficient to distinguish faults of different natures. For example, "slow leakage of the hydraulic pump" and "intermittent excess moisture in raw materials" may produce similar PTRN peaks in a short period, but the former results in a smooth increase in STD, while the latter is impulsive. By upgrading the diagnostic problem from one-dimensional pattern matching to two-dimensional trajectory classification, this unique dynamic "fingerprint" encoded by both TWED and STD can be captured, thereby achieving accurate differentiation of root causes.

[0115] To quantitatively verify the advancements of the technical solution described in this invention compared to existing technologies, a high-fidelity digital model environment was constructed. This environment includes a device model capable of accurately simulating the nonlinear kinetic characteristics during the pressing of anode carbon blocks. All interactions are accomplished through data exchange between model interfaces and changes in internal state variables, thereby ensuring the objectivity and reproducibility of the experiment.

[0116] This experiment aims to compare the performance of the present invention (invention group) with two representative prior art technologies (control group A and control group B) under the same conditions. Control group A uses a fixed-gain PID controller for process regulation. Control group B uses a one-dimensional PTRN time series template matching method for fault diagnosis. Three typical scenarios were designed: Scenario 1 is the standard operating condition, used to evaluate control performance; Scenario 2 and Scenario 3 are two different typical fault operating conditions, used to evaluate the accuracy of diagnosis, as shown in Table 1 below.

[0117] Table 1: Performance Comparison Experimental Data of the Invention and Existing Technologies in Different Scenarios

[0118] Parameter name Scenario 1: Standard Operating Conditions (Experiment 1) Scenario 1: Standard Operating Conditions (Experiment 2) Scenario 2: Fault A - Hydraulic Leakage (Experiment 1) Scenario 2: Fault A - Hydraulic Leakage (Experiment 2) Scenario 3: Fault B - Raw Material Abnormality (Experiment 1) Scenario 3: Fault B - Raw Material Abnormality (Experiment 2) Experimental Groups Invention Group / Control Group A Invention Group / Control Group A Invention Group / Control Group B Invention Group / Control Group B Invention Group / Control Group B Invention Group / Control Group B Initial physical twin residual norm 0.25 / 0.25 0.26 / 0.26 0.81 / 0.81 0.83 / 0.83 0.79 / 0.79 0.80 / 0.80 Derived parameter: Process settling time (seconds) 3.2 / 5.8 3.4 / 5.9 - / - - / - - / - - / - Derived parameter: Cumulative absolute error of process control 1.8 / 3.5 1.9 / 3.6 - / - - / - - / - - / - Final diagnosis results None / None None / None Hydraulic leakage / hydraulic leak Hydraulic leakage / hydraulic leak Raw material abnormality / hydraulic leakage Raw material abnormality / hydraulic leakage

[0119] Before conducting data analysis, the two derived parameters introduced in the table for quantifying control performance are defined:

[0120] The process settling time is calculated by monitoring the time series of the physical twin residual norm (PTRN). It measures the time elapsed from the initial issuance of the control adjustment signal until the PTRN value first enters and remains within a range centered at 0 with a preset radius of 0.05. This parameter quantifies the response speed and convergence performance of the control system; a smaller value indicates a faster system elimination of process deviations and recovery to an ideal state.

[0121] The cumulative absolute error (CREE) of process regulation is obtained by integrating the absolute value of the physical twin residual norm (PTRN) over time during the process regulation settling period. The calculation logic is as follows: in a discrete time series, the absolute value of PTRN at each sampling point is multiplied by the sampling time interval, and then all these multiplications are summed. This parameter quantifies the total deviation during the entire regulation process; the smaller the value, the higher the accuracy of the entire control process and the smaller the deviation from the ideal baseline.

[0122] Comparison Scenario 1: Under the standard operating conditions of Scenario 1, compare the data of the invention group and control group A (fixed gain PID). Taking Experiment 1 as an example, the process settling time of the invention group was 3.2 seconds, while that of control group A was 5.8 seconds. The cumulative absolute error of the process settling time of the invention group was 1.8, while that of control group A was 3.5. The settling time of the invention group was shortened by 44.8% compared to control group A. The calculation method is as follows: calculate the time difference between the two, i.e., 5.8 minus 3.2, which gives 2.6; then, divide this difference by the time of control group A, i.e., 2.6 divided by 5.8, which gives 0.448. Similarly, the cumulative absolute error of the invention group was reduced by 48.6% compared to control group A.

[0123] This set of data demonstrates that the adaptive gain scheduling PID control method in S4 of this invention has significantly faster response speed and higher control accuracy compared to the fixed gain PID control in the prior art. This directly confirms that the design of dynamically adjusting the PID gain through AVMDC can effectively overcome the nonlinear characteristics of the controlled object, thereby achieving better process control.

[0124] Comparing Scenario 2 and Scenario 3: In Scenario 2, both the invention group and control group B (one-dimensional template matching) correctly identified the "hydraulic leakage" fault. However, in Scenario 3, the invention group correctly identified "raw material abnormality," while control group B incorrectly identified it as "hydraulic leakage." In the fault set designed in this experiment, the diagnostic accuracy of the invention group was 100%, while the accuracy of control group B was only 50%. Scenario 3 is the key to this experimental design. In this scenario, the PTRN one-dimensional time series generated by the "raw material abnormality" fault is highly similar in morphology to "hydraulic leakage," thus "deceiving" the one-dimensional template matching method. However, the trajectories of these two faults in the two-dimensional fault state space (defined by TWED and STD) are significantly different. This result strongly proves that the two-dimensional fault state space trajectory diagnosis method in S5 of this invention, compared with the existing one-dimensional matching method, can utilize richer information dimensions to effectively distinguish complex faults that appear similar but have different root causes, improving the accuracy and robustness of diagnosis. This solves the fundamental technical problem of insufficient diagnostic capability of existing technologies when facing fuzzy and complex faults.

[0125] It should be noted that all calculation formulas in this application employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, min-max-normalization and Z-score standardization.

[0126] The algorithm of this invention is implemented as a Python script. Before executing the core logic, the program first executes a data loading module (e.g., using the widely used pandas library in Python) configured to read the aforementioned spreadsheet file and load its contents into the program's working memory (e.g., a DataFrame data structure). Subsequent algorithm steps will directly query and retrieve the required configuration parameters from this in-memory data structure.

[0127] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An adaptive method for pressing parameters of high-density anode carbon blocks, characterized in that, The specific steps include: S1: Real-time acquisition of instantaneous physical state characteristics that characterize the internal physical state of the anode carbon block, calculated based on the vibration response data of the pressing equipment; S2: Obtain the optimal process trajectory baseline corresponding to the current pressing conditions, generated based on the cloud-based federated digital twin model; S3: The physical twin residual calculation module is used to quantify the temporal deviation between the instantaneous physical state characteristics and the optimal process trajectory baseline to generate the physical twin residual norm. S4: Under the condition that the physical twin residual norm is within a first preset range, a control adjustment signal is generated to adjust the suppression parameter so as to minimize the physical twin residual norm; S5: When the physical twin residual norm exceeds the first preset range, a diagnostic alarm signal is generated to indicate that there is an abnormality in the pressing equipment or the anode carbon block raw material.

2. The adaptive method for pressing parameters of high-density anode carbon blocks according to claim 1, characterized in that: In S1, the step of obtaining the instantaneous physical state characteristics includes: performing time-frequency analysis on the frame vibration response signal of the pressing equipment, and calculating the acoustic vibration mode damping coefficient as the instantaneous physical state characteristics based on the energy response within a preset characteristic frequency band. The time-frequency analysis includes: tracking the spectral energy distribution of the frame vibration response signal in real time, adaptively determining a dynamic sensitive frequency band with the most significant energy change, and calculating the acoustic vibration mode damping coefficient based on the energy response within the dynamic sensitive frequency band.

3. The adaptive method for pressing parameters of high-density anode carbon blocks according to claim 2, characterized in that: In S2, the optimal process trajectory baseline is an ideal time-series variation curve of the acoustic vibration mode damping coefficient generated by the cloud-based federated digital twin model through the aggregation of model parameter updates from multiple pressing devices without accessing the local original vibration response data.

4. The adaptive method for pressing parameters of high-density anode carbon blocks according to claim 3, characterized in that: In S3, the step of generating the physical twin residual norm includes: within a sliding time window, constructing two high-dimensional vectors from the time series sequence of the instantaneous physical state features and the corresponding time series sequence of the optimal process trajectory baseline, and calculating the Euclidean distance between the two high-dimensional vectors as the instantaneous value of the physical twin residual norm; The calculation of the physical twin residual norm further includes: applying a time decay weight to the time series within the sliding time window, quantifying the morphological trend difference between the two high-dimensional vectors, and fusing the weighted Euclidean distance with the morphological trend difference to generate the physical twin residual norm.

5. The adaptive method for pressing parameters of high-density anode carbon blocks according to claim 4, characterized in that: The first preset range refers to dividing the range of the physical twin residual norm value into a normal area, a control area, and a diagnostic area by a first control threshold and a second diagnostic threshold, respectively. In S4, the control adjustment signal is used to drive the proportional-integral-derivative controller to dynamically modify the pressing rate or holding time parameters, thereby minimizing the physical twin residual norm. The proportional, integral, and derivative gain parameters of the proportional-integral-derivative controller are dynamically adjusted based on the current values ​​of the instantaneous physical state characteristics that characterize the internal physical state of the anode carbon block, obtained in step S1, through a preset gain scheduling function.

6. The adaptive method for pressing parameters of high-density anode carbon blocks according to claim 5, characterized in that: In S5, when the physical twin residual norm exceeds the first preset range, a diagnostic alarm signal is generated. The steps include: First, suppressing the generation of the control adjustment signal; The second step is to match the temporal pattern of the physical twin residual norm with a preset fault feature library to generate the diagnostic alarm signal containing the fault root cause category.

7. The adaptive method for pressing parameters of high-density anode carbon blocks according to claim 6, characterized in that: The second step, the matching process, further includes: The time-weighted Euclidean distance and morphological trend difference constituting the physical twin residual norm are used as a two-dimensional state coordinate to construct an abnormal state trajectory in a preset fault state space; the abnormal state trajectory is compared with the reference fault trajectory stored in the fault feature library to determine the fault root cause category.

8. A high-density anode carbon block product, characterized in that: The high-density anode carbon block product is prepared by the adaptive method for pressing parameters of high-density anode carbon blocks as described in any one of claims 1-7.

Images