Adaptive compensation method, system, medium and computer device for equipment changeover

Abstract

This invention discloses an adaptive compensation method, system, medium, and computer equipment for equipment changeover, belonging to the field of industrial automation technology. It includes: establishing a real-time monitoring network, setting up a real-time sensing method, continuously collecting operational data, preprocessing data using edge computing and extracting feature parameters, and identifying equipment dynamic characteristics; dynamically predicting the optimal parameter combination, setting up an adaptive calibration method, and adjusting equipment parameters in real time based on processing errors; triggering a parameter memory clearing mechanism during changeover, setting up a residual suppression method, analyzing the degree of interference and designing a compensation algorithm, and adopting a collaborative adjustment strategy after performing coupling analysis on the parameters; integrating multi-source data to construct a dynamic verification model, setting up a classification compensation method, assessing equipment status risks, and designing customized compensation strategies for different operating conditions; establishing a full-process data feedback mechanism from equipment status sensing to process verification, setting up optimization methods, and using data analysis algorithms to iteratively optimize the models and algorithms at each stage.

Description

Technical Field

[0001] This invention belongs to the field of industrial automation technology and relates to adaptive compensation methods, systems, media, and computer equipment for equipment changeover. Background Technology

[0002] Against the backdrop of rapid development in industry and intelligent manufacturing, the global manufacturing industry is undergoing digital transformation. Multi-variety, high-volume mixed-line production has become a core demand. With the widespread application of intelligent equipment such as CNC machine tools and sensors, workshop production data is characterized by large scale, high diversity, and high real-time performance, providing a foundation for data-driven optimization methods. The integration of industrial internet, Internet of Things, and artificial intelligence technologies has promoted the intelligent upgrading of production processes. Enterprises need to optimize process parameters, improve equipment energy efficiency, and dynamically control quality through data mining and analysis to meet the challenges of personalized market demands and flexible production.

[0003] However, existing technologies are still limited by the lack of real-time adaptability to the dynamic characteristics of equipment caused by frequent production changes in small-batch production, and face multiple challenges: when frequent production changes lead to high-frequency reset of equipment parameters and dynamic adjustment of physical components, such as rapid mold replacement and tool precision calibration, the existing equipment stability maintenance mechanism relies on static working condition models, resulting in a delay in dynamic balance reconstruction. Traditional parameter calibration strategies are difficult to maintain the continuity of processing accuracy during production changeover. In addition, the lack of a mechanism to suppress the residual effect of historical parameters in the control system makes it impossible for the equipment to clear process parameter memory interference in real time. Parameter coupling fluctuations during production changeover cause the process output to deviate from the preset range. In small-batch scenarios, insufficient process verification sample size and imbalance of dynamic verification mechanisms result in the first piece inspection not being able to cover the time-varying risks of equipment status. The lack of verification strategies for special production changeover conditions further reduces process reliability. Insufficient semantic understanding of equipment status leads to the inability of control compensation to adapt synchronously when parameters drift and wear accumulate, resulting in defects in production continuity and quality stability, making it difficult to meet the dynamic adaptation requirements of rapid equipment switching in small-batch, multi-variety production. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide an adaptive compensation method, system, medium, and computer equipment for equipment replacement. It accurately identifies equipment status through multi-dimensional sensing and data fusion, dynamically predicts parameters and calibrates them in real time using deep learning, suppresses residual and coupled fluctuations through two-level clearing and collaborative optimization, covers verification blind spots through data augmentation and transfer learning, and improves adaptability through closed-loop optimization of the entire process.

[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention provides an adaptive compensation method for equipment changeover, comprising: Establish a real-time monitoring network, set up real-time sensing methods, continuously collect operational data, preprocess the data using edge computing, extract feature parameters from the time domain, frequency domain, and time-frequency domain using feature extraction algorithms, and construct a multi-dimensional characterization model of equipment status to identify the dynamic characteristics of the equipment. Based on the equipment status and production changeover process requirements, the optimal parameter combination is dynamically predicted, an adaptive calibration method is set, and the equipment parameters are adjusted in real time in combination with the processing error, and a mapping relationship between parameter adjustment and equipment response is established. When changing production, a parameter memory clearing mechanism is triggered, historical parameters are initialized, a method to suppress residual effects is set, the degree of interference is analyzed through a residual effect prediction model and a compensation algorithm is designed, and a collaborative adjustment strategy is adopted after coupling analysis of the parameters. Integrate multi-source data to build a dynamic verification model, set up classification compensation methods, assess equipment status risks, collect unconventional operating condition data to train the compensation model, and design customized compensation strategies for different operating conditions. Establish a full-process data feedback mechanism from equipment status perception to process verification, set up classification compensation methods, and use data analysis algorithms to iteratively optimize the models and algorithms of each link.

[0006] Specifically, the real-time sensing method includes: A three-layer stereo sensor array layout is adopted to form a three-dimensional monitoring network; The design incorporates a dynamic sampling mechanism. Under normal operating conditions, sampling is performed synchronously at the basic sampling interval. Once a production change signal is detected, fuzzy logic is used to calculate and adjust the sampling rate. The collected vibration signal is divided into data segments, the kurtosis index is calculated, and a time-domain attention weight matrix is ​​generated using a sliding time window; The vibration signal is decomposed by wavelet packet and key frequency bands are selected by autoencoder and fully connected network; The temperature signal is subjected to S-transform and its time-frequency domain features are enhanced by 2D convolution and global average pooling; The time domain, frequency domain, and time-frequency domain features are integrated into a multi-dimensional feature vector. After being divided into sub-vectors according to sensor type, the vectors are input into the evidence generator, and the output probability allocation sequence is generated. Based on the variance and mean of each probability assignment sequence in the sliding window, the activity index is calculated and the weights are dynamically adjusted. Calculate the conflict coefficient between pieces of evidence and redistribute the weights of evidence that exceeds the preset value; The device state vector is generated by using a pre-trained state transition matrix and an extended Kalman filter.

[0007] Specifically, the adaptive calibration method includes: Construct a pre-trained model library containing product changeover data. When the initial data for a new changeover is insufficient, activate the domain adaptation mechanism to standardize the source and target domain data and calculate the MMD distance. Expand the target domain data after achieving distribution alignment through adversarial training and gradient inversion layer. The design employs a dual-memory LSTM architecture. The main memory unit processes the state feature sequence of the current operating condition, while the auxiliary memory unit stores historical knowledge. The new and old knowledge are integrated through a gating mechanism, and the probability distribution of parameter combinations is predicted using a Gaussian mixture model and the EM algorithm.

[0008] Specifically, the adaptive calibration method further includes: Construct a parameter-optimized Siamese network, learn the parameter-response mapping by comparing loss functions, and fine-tune it online periodically; The sparrow search algorithm is initialized with a population, a dynamic search radius is introduced, and the search range is adjusted based on the parameter similarity of the output of the twin network according to the parameters. The parameter combination is optimized by combining adaptive weight update and elite reverse learning strategy. A twin dynamics model is constructed based on the equipment physical model and LSTM error prediction. The model parameters are corrected through a twin network, and a parameter smoothness penalty term is introduced to perform multi-objective rolling optimization to generate a parameter optimization sequence. The equipment is divided into subsystems and a coupling matrix is ​​established to trigger cross-subsystem collaborative adjustment. The production change data is stored in the knowledge graph constructed by the graph database, and the inference of similar working condition parameters is achieved through graph neural network training.

[0009] Specifically, the method for suppressing residues includes: When a production change instruction is received, the primary clearing mechanism is activated to reset the historical process parameters to the default base values; The secondary algorithm cleanup mechanism is activated. The Transformer network is used to perform semantic parsing on the parameter sequence. Combined with the current state vector of the device and the optimization parameters of the previous batch, the residual risk score of each parameter is calculated through the attention mechanism and a residual risk score matrix is ​​generated. Risk parameters that exceed the preset threshold are screened out and the gradient zeroing algorithm is executed. A multi-dimensional feature vector is constructed by integrating real-time equipment status characteristics, parameter deviations between historical process parameters and default baseline values, and environmentally relevant parameters. The multidimensional feature vector is input into the trained XGBoost model to predict the intensity of residual effects in the short period of time after the production change, and outputs a three-dimensional prediction vector, including the influence of residual effect amplitude, phase and duration. Key influencing factors are identified by ranking features by importance, and a warning is issued for risk parameters when the predicted residual effect amplitude exceeds a preset threshold.

[0010] Specifically, the method for suppressing residues further includes: Define optimization parameters and integrate them according to production change time and batch based on historical data to generate a historical optimization parameter sequence. Use mutual information theory to calculate the coupling strength between parameters, construct a dynamically updated coupling degree matrix, and set a strong coupling threshold to define strong coupling. By learning higher-order dependencies between parameters through graph neural networks, the coupling degree matrix is ​​transformed into a weighted directed graph model and the coupling degree matrix is ​​updated periodically. Strongly coupled parameter pairs are identified and the coupling relationship is visualized in the form of a heatmap. A state space containing device status, three-dimensional prediction vectors and parameter deviations is constructed, an action space for parameter adjustment is defined, a reward function that integrates accuracy, smoothness and energy consumption is designed, a deep Q-network training optimization strategy is used, and training stability is improved through experience replay and target network mechanisms. Upon receiving the target value of the optimization parameters, the optimal action combination is calculated based on the current state, and a proportional collaborative adjustment strategy is generated for strongly coupled parameter pairs. A comprehensive error signal is formed by integrating real-time machining accuracy error, three-dimensional prediction vector and parameter deviation. The PID parameters are then dynamically adjusted by a fuzzy controller based on the comprehensive error signal and the rate of change of the comprehensive error signal. The PID parameters are updated at fixed intervals until the processing error converges to a preset range. At the same time, during the compensation process, the parameter adjustment trajectory and error data are recorded in real time and used as training samples for the reinforcement learning model.

[0011] Specifically, the classification compensation method includes: Integrate equipment state vectors, optimization parameters, and machining accuracy data to form a dynamic verification data pool; A generative adversarial network with conditional input is constructed. The virtual samples are generated by the concatenation vector of random noise, equipment state vector and process parameters. The discriminator is trained by real data, virtual samples and conditional variables. The consistency of sample distribution is verified by the maximum mean difference method. Using ResNet50 as the base network, combined with a domain adaptation mechanism and a gradient inversion layer, a validation model is generated. The enhanced virtual samples are used as input, and the output is the probability distribution of device state risk. The status features are extracted from the equipment operation data, and the weights are determined by the analytic hierarchy process. After standardization, the weighted sum is obtained to obtain a comprehensive risk index. Construct a two-dimensional risk map and set thresholds to trigger response strategies.

[0012] Specifically, the classification compensation method further includes: Collect unconventional operating condition data and perform signal processing and manual annotation to establish a database; A hybrid model is constructed using convolutional neural networks and bidirectional gated recurrent units, and the recognition accuracy is optimized through an attention layer; Based on the working condition probability distribution output by the hybrid model, set graded thresholds for working condition confirmation, secondary identification, or collection of unknown working conditions. Set up compensation strategies that include pre-compensation, correction, and optimization layers, and store them according to working conditions; The strategy association relationship is constructed using a knowledge graph, and the parameters of the compensation strategy are updated using a particle swarm optimization algorithm; Based on the identified abnormal operating conditions, the compensation model is retrieved and invoked, and the parameters are adjusted according to the equipment state vector. After calculating and standardizing the accuracy, efficiency, and reliability indicators, a comprehensive score and a compensation effectiveness score are generated. The Q-learning algorithm is used to update the parameters of the compensation strategy with the compensation parameter combination as the state, the parameter adjustment amount as the action, and the effectiveness score as the reward.

[0013] Specifically, the optimization method includes: Construct a four-layer closed-loop architecture comprising a perception layer, a decision-making layer, an execution layer, and a verification layer; The federated learning mechanism is adopted, where each edge node trains the model locally and then uploads the gradient, and the central server aggregates and updates the global model. A dynamic weighting function is constructed by integrating production efficiency, quality stability, and equipment health objectives. The weights are adjusted according to the production stage using fuzzy logic. Equipment safety restrictions and process specifications are transformed into soft constraints and processed through penalty functions. Based on the NSGA-Ⅲ algorithm, the collaborative parameters and compensation parameters are merged and encoded into a real number vector. Simulated binary crossover and polynomial mutation are used to maintain the diversity of solutions. The proportion of elite solutions retained is adaptively adjusted according to the compensation effect score. The prediction time domain is dynamically adjusted and optimized by combining the two-dimensional risk map early warning, and the optimization is performed periodically and updated based on the latest status. The system integrates LSTM architecture and hybrid model, uses transfer learning to periodically update the state prediction model, and prioritizes the replay of high-risk working condition data to update the compensation model based on reinforcement learning experience replay mechanism. Add device status, compensation strategy and verification result entities to the knowledge graph and define relation types. Update the node embedding representation through graph convolutional network to dynamically update the knowledge graph and perform similar working condition retrieval and reasoning. A general optimizer is constructed using meta-learning, and the knowledge transfer conditions are determined by calculating the differences in the distribution of device features and adjusting the optimizer using support set data. Smoothing filtering is applied to the clearing parameters and collaborative parameters. The executed value of the compensation parameter is compared with the target value and the deviation is corrected by proportional-integral control. The strongly coupled parameters are adjusted collaboratively based on the parameter coupling degree matrix. Anomaly detection is performed using the isolated forest algorithm in conjunction with risk map data. Compensation strategies or early warning logic are triggered based on the degree of anomaly, and recovery is verified through no-load operation. Establish a performance evaluation index system that includes equipment utilization, process capability index and mean time between failures, and regularly use SWOT analysis to compare with industry benchmarks and formulate evolutionary strategies. A digital twin is constructed to simulate the input device status and control parameters of the device actuators. Optimization strategies are tested, and the collaborative adjustment and compensation strategies that have passed virtual verification are migrated to the physical device.

[0014] This invention provides an adaptive compensation system for equipment changeover, comprising: a state perception module, a parameter calibration module, a residual suppression module, an operating condition compensation module, and a closed-loop optimization module; The state perception module is used to collect real-time operating data of key components of the equipment through a three-layer three-dimensional sensor array, extract and fuse features through edge computing, so as to identify the dynamic characteristics of the equipment. The parameter calibration module is used to predict the optimal parameter combination under different working conditions, adjust the parameters in real time in combination with the processing error, and establish a mapping relationship between parameter adjustment and equipment response. The residual suppression module is used to activate a hardware-level and algorithm-level dual-level parameter clearing mechanism during production change. It predicts residual effects and designs compensation algorithms through the XGBoost model, and achieves multi-parameter collaborative optimization by combining the parameter coupling degree matrix and the deep Q network. The operating condition compensation module is used to integrate multi-source data to build a dynamic verification model, use generative adversarial networks to enhance small batch samples, combine transfer learning to assess equipment status risks, identify unconventional operating conditions through a hybrid model, and call customized compensation strategies. The closed-loop optimization module is used to establish a full-process data feedback mechanism, and to iteratively optimize the models of each stage through multi-objective optimization algorithms and federated learning, and to realize cross-condition knowledge transfer and strategy verification by combining knowledge graphs and digital twins.

[0015] The beneficial effects of this invention are: A real-time monitoring network and dynamic operating condition model are constructed to achieve accurate identification of equipment status and dynamic prediction of parameters, eliminating reliance on static models, shortening dynamic balance reconstruction time, and maintaining the continuity of processing accuracy. A two-level parameter clearing and multi-parameter collaborative compensation are designed to suppress historical parameter residues and coupled fluctuations, stabilizing process output. A dynamic verification model is built using data augmentation and transfer learning, combined with customized strategies for unconventional operating conditions, covering blind spots in small-batch verification and risks associated with unconventional operating conditions. A closed-loop optimization mechanism is established throughout the entire process to iteratively optimize the model at each stage, ensuring that control compensation and equipment status are synchronously adapted. Ultimately, the dynamic adaptability of the equipment is improved, addressing deficiencies in production continuity and quality stability, and meeting the needs for rapid switching between small batches and multiple product varieties. Attached Figure Description

[0016] Figure 1Flowchart of the adaptive compensation method for equipment changeover; Figure 2 This is a flowchart of the adaptive calibration method of the present invention; Figure 3 This is a flowchart of the residue suppression method of the present invention; Figure 4 Structure diagram of the adaptive compensation system for equipment changeover. Detailed Implementation

[0017] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0018] Example 1 refer to Figures 1 to 3 As shown in the figure, this embodiment introduces an adaptive compensation method for equipment changeover, including: Establish a real-time monitoring network containing multiple types of sensors, such as vibration sensors, temperature sensors, and pressure sensors. Set up real-time sensing methods to continuously collect operating data of key components of the equipment during equipment changeover. Use edge computing nodes to preprocess the collected data and extract feature parameters that characterize the dynamic characteristics of the equipment from multiple dimensions, such as time domain, frequency domain, and time-frequency domain, using feature extraction algorithms. Employ data fusion technology to fuse the data collected by different sensors and the extracted feature parameters to construct a multi-dimensional representation model of the equipment status. Through this multi-dimensional representation model, accurately identify the current dynamic characteristics of the equipment and track the status changes of the equipment in real time during the changeover process. This provides accurate status basis for subsequent adaptive compensation, enabling control compensation to be synchronously adapted to equipment parameter drift and wear accumulation, thereby improving the accuracy of equipment status semantic understanding. A deep learning-based dynamic operating condition model is constructed to dynamically predict the optimal parameter combination of the equipment under different operating conditions based on the real-time perceived equipment status and production changeover process requirements. An adaptive calibration method is set up. During the production changeover process, when equipment parameters need to be reset and physical components need to be dynamically adjusted, an intelligent optimization algorithm is used to calculate and adjust various parameters of the equipment in real time based on the prediction results of the dynamic operating condition model and the current actual processing error. By establishing a mapping relationship between parameter adjustment and equipment dynamic response, the parameter adjustment can be quickly converged, shortening the time for dynamic balance reconstruction. This eliminates the dependence on static operating condition models and enables the equipment to optimize parameters in real time according to the actual operating conditions during the production changeover period, maintaining the continuity of processing accuracy and improving the efficiency and accuracy of parameter calibration. During production changeover, a parameter memory clearing mechanism is triggered, historical parameters are initialized, a method for suppressing residual effects is set, the degree of interference is analyzed through a residual effect prediction model, a compensation algorithm is designed to offset the impact in real time, a relationship model is established by performing coupling analysis on the parameters, a collaborative adjustment strategy is adopted, and the collaborative adjustment is carried out by comprehensively considering the coupling relationship between parameters. The strategy is also dynamically optimized based on process output deviations. This method solves the problems of missing historical parameter residual effect suppression mechanisms and process output deviations caused by parameter coupling fluctuations, clears parameter memory interference, suppresses coupling fluctuations, and ensures that the process output remains stable within a preset range. By integrating multi-source data to construct a dynamic verification model based on big data, setting up classification compensation methods, utilizing data augmentation and transfer learning techniques, and combining real-time monitoring mechanisms to assess the time-varying risks of equipment status, collecting data on unconventional operating conditions to train the compensation model, designing customized compensation strategies for different operating conditions, and automatically triggering corresponding strategies when unconventional operating conditions are detected, the model solves the problems of insufficient verification sample size in small-batch scenarios, imbalance of dynamic verification mechanisms, and lack of verification strategies for special production changeover conditions. This model covers equipment status risks and improves the reliability of processes under unconventional operating conditions. Establish a full-process data feedback mechanism from equipment status perception, parameter adjustment, production processing to process verification. Set optimization methods and use data analysis algorithms to analyze and mine the feedback data, identify problems affecting production continuity and quality stability in the production process, and iteratively optimize the models and algorithms of each link, such as the real-time dynamic characteristic perception method and the adaptive parameter calibration method, based on the analysis results. Continuously improve the system's performance and adaptability, and continuously optimize it according to the actual production situation to solve the defects in production continuity and quality stability, meet the dynamic adaptation requirements of rapid equipment switching in small-batch, multi-variety production, and improve production efficiency and product quality.

[0019] Specifically, real-time sensing methods include: Since a fixed sampling rate cannot capture transient characteristics during production changes, a three-layer three-dimensional sensor array layout is adopted, including a triaxial acceleration vibration sensor installed on the end cover of the main shaft of the equipment, temperature sensors arranged at fixed intervals along the guide rail, and a piezoresistive thin-film pressure sensor embedded in the mold mounting surface, forming a three-dimensional monitoring network. Design a dynamic sampling mechanism, and under normal operating conditions, set the basic sampling interval to [value missing]. The data acquisition card samples synchronously at a basic sampling interval. Once a production change signal is detected, it immediately triggers adaptive sampling rate control, which increases the sampling rate through fuzzy logic control. Inputs include the production change stage identifier and equipment vibration intensity, while the output is a sampling rate adjustment coefficient. The sampling rate is switched using the adjustment coefficient and the basic sampling interval. During the switching process, linear interpolation is used to fill in transitional data to avoid signal interruption. Dual redundant data transmission links are deployed, including an industrial Ethernet main link and a fiber optic ring network as a backup. A hardware automatic switching protocol enables rapid fault switching, eliminating blind spots in status identification caused by data transmission interruptions, and outputting raw time-domain data sequences, including vibration acceleration waveforms, temperature field distribution, and pressure curves. The basic sampling interval is set differently depending on the sensor type. The vibration sensor's basic sampling interval... The corresponding sampling frequency is 1kHz, used to capture high-frequency impact characteristics during production changeover. The basic sampling interval for the temperature sensor is... The corresponding sampling frequency is 1Hz, which is suitable for the slow-changing characteristics of the temperature field; the basic sampling interval of the pressure sensor. The corresponding sampling frequency is 10Hz. To balance the data volume and pressure fluctuation monitoring requirements, after the production change signal is triggered, the sampling rate adjustment coefficient ranges from 2 to 10 times. That is, the maximum sampling frequency of the vibration sensor can reach 10kHz, the maximum sampling frequency of the temperature sensor can reach 10Hz, and the maximum sampling frequency of the pressure sensor can reach 100Hz. The collected vibration signal was divided into several data segments according to the sampling points. The kurtosis index was calculated for each segment. A sliding time window was introduced to calculate the deviation of the kurtosis of each segment from the global mean. The attention weight matrix in the time domain was generated by the Softmax function. To complete time-domain processing, when the kurtosis value of a certain segment exceeds twice the standard deviation, its weight is automatically increased to 0.15, thus increasing the proportion of time periods with obvious shock characteristics in subsequent processing by 3 times. At the same time, an L2 regularization term is added to prevent overfitting; the expression is as follows:

[0020]

[0021] In the formula, For the first kurtosis index of the segment for The corresponding deviation, for The corresponding weights The mean of global kurtosis. The global kurtosis standard deviation, The L2 regularization term, representing the window length, is applied to the weight parameters of the autoencoder and fully connected network to prevent overfitting. Its value ranges from 1e-4 to 1e-2. When the amount of training data is less than 1000 sets, the upper limit of 1e-2 is used, and when the amount of training data is greater than 10000 sets, the lower limit of 1e-4 is used. The amount of intermediate data is determined by linear interpolation. The vibration signal was decomposed into 8 frequency bands using a 3-layer wavelet packet decomposition with db9 wavelet. The energy values ​​of each frequency band were input into an autoencoder. Key frequency bands were selected based on reconstruction error. Frequency bands with reconstruction errors less than 0.05 were retained. The contribution of each frequency band was then calculated using a two-layer fully connected neural network. Finally, the top 3 frequency bands with the highest contribution were retained for frequency domain analysis. The autoencoder consisted of an 8-5-3 encoder structure and a 3-5-8 decoder, with ReLU activation function. The fully connected network consisted of an input layer, a hidden layer, and an output layer, with LeakyReLU activation function. The temperature signal is subjected to S-transform, and a variable window function is used to generate a time-frequency energy matrix. Spatial features are extracted by 2D convolution, and then global average pooling and fully connected layers are used to calculate attention weights along the time axis and frequency axis, respectively, to enhance the representation ability of key features in the time-frequency domain. Features from the time domain, frequency domain, and time-frequency domain are integrated to generate a multi-dimensional feature vector. Each dimension carries an attention weight, and the vectors are divided into vibration sub-vectors, temperature sub-vectors, and pressure sub-vectors according to sensor type. These sub-vectors are then input into three independent evidence generators. Each evidence generator contains two layers of neural networks and outputs a probability assignment sequence corresponding to the proposition space. The proposition space is a set of device states, including normal, slight wear, severe wear, and abnormal vibration. For example, the probability assignment sequence output by the vibration evidence generator... This indicates that the probability of the current state being "normal" is 70%. The activity index is calculated by performing a weighted summation operation based on the variance and mean of each probability assignment sequence within the sliding window. The fusion weights of probability allocation are dynamically adjusted using an activity index. The adjusted fusion weights are obtained by multiplying the activity index by the initial fusion weights and adding the initial fusion weights. Simultaneously, an exponentially weighted moving average is used to introduce a forgetting factor, which is used as the weight of the adjusted fusion weights. This forgetting factor is then combined with the final fusion weights from the previous time step for weighted summation. Smoothing and decay corrections ensure real-time response to dynamic changes, resulting in the final fusion weights. The initial fusion weights are determined based on the analytic hierarchy process (AHP) combined with fault injection experiments. The forgetting factor is determined by those skilled in the art based on equipment changeover frequency and operating status. Through simulation using historical changeover data, the forgetting factor is determined based on the criteria of optimal equipment status identification accuracy and minimal weight fluctuation. For example, a value of 0.90~0.93 is selected when there are frequent changeovers and rapid equipment dynamic changes, while a value of 0.95~0.99 is selected when there is steady-state production and small equipment status fluctuations. For the probability allocation sequences generated by the three evidence generators, the probability allocation sequence of each evidence generator is weighted and corrected according to the fusion weight; Based on the corrected probability allocation sequence, the probabilities of the same device state are multiplied, and the product values ​​of each device state are summed to calculate the conflict coefficient between the evidence. When the conflict coefficient between any number of pieces of evidence exceeds the conflict threshold, a scenario conflict exists. For example, vibration evidence supports "abnormal vibration," while pressure evidence supports "normal," triggering a weight redistribution mechanism. This includes calculating the average conflict degree between each piece of evidence and other evidence, defining the evidence with the highest average conflict degree as high-conflict evidence, and applying an exponential decay function to high-conflict evidence to reduce the fusion weight. The decay coefficient in the exponential decay function is dynamically adjusted by the sigmoid function according to the conflict coefficient between the evidence. The expression is shown below:

[0022]

[0023]

[0024] In the formula, For the first The evidence is fused with weights after exponential decay. For the first The original fusion weights of the evidence, For evidence serial number, These correspond to three types of sensory evidence: vibration, temperature, and pressure. This is the attenuation coefficient, used to control the rate at which the weights decay with increasing conflict level. The preset conflict threshold is set at the 90th quantile of the conflict coefficient distribution under normal operating conditions. , To define the upper and lower limits of the attenuation coefficient, multiple sets of independent historical production change samples were obtained to construct a calibration dataset. Each set of samples included the evidence conflict coefficient, the original fusion weights of the three types of evidence, and the actual equipment status label. High-conflict samples with conflict coefficients greater than the conflict threshold were selected. The optimization objective was to ensure that the equipment identification status output after weight correction and evidence fusion of the high-conflict sample perfectly matched the actual operating status of the corresponding equipment. Within the attenuation coefficient traversal interval, candidate attenuation coefficients were traversed in ascending order, with a traversal step size of 1 / 50 of the interval length. All attenuation coefficient values ​​that could result in correct status identification were selected, and the attenuation coefficient with the smallest value was taken as the optimal attenuation coefficient corresponding to the high-conflict sample. The optimal attenuation coefficients of all high-conflict samples were then sorted in ascending order, and the quintile was used as the quartile. 95th percentile as The upper and lower bounds of the attenuation coefficient traversal interval are obtained from the calibration dataset. The lower bound is set to 0 to ensure that all possible optimal attenuation coefficients are covered. The upper bound is derived from the historical maximum conflict coefficient and weight suppression constraints. , To determine the maximum value of the conflict coefficient in the dataset, As the kurtosis control coefficient, for high-conflict samples in the calibration dataset, the conflict coefficients are sorted in ascending order. The 20th quantile of the conflict coefficients of high-conflict samples is used as the left boundary, and the 80th quantile as the right boundary. These two boundaries are then substituted into the sigmoid function. After simplifying the equations, the steepness control coefficient can be directly derived. The conflict coefficient between pieces of evidence. , , Evidence from vibration, temperature, and pressure sensors, respectively. This represents the device state in the device state set. For vibration evidence generator The probability of each equipment state and the conflict threshold are obtained based on a large amount of historical production change evidence conflict data. The 90th percentile of the conflict distribution under normal operating conditions is taken as the conflict threshold. For example, when the conflict coefficient is the left boundary When the sigmoid adjustment term outputs 0.25 (the attenuation coefficient is close to the lower limit), and when the conflict coefficient is at the right boundary... At that time, the sigmoid adjustment term outputs 0.75 (the attenuation coefficient is close to the upper limit). Substituting 0.25 and 0.75 into the sigmoid function... ,in ; Substitute the left boundary conditions Solving for ,Right now ; Substitute the right boundary conditions Solving for ,Right now ; Subtract the two equations to eliminate ,get ; The fused evidence is mapped to a three-dimensional state space through a pre-trained state transition matrix, and the state vector is updated using an extended Kalman filter (EKF) to generate the device state vector. The state transition matrix is ​​trained based on 1000 sets of historical production change data.

[0025] Specifically, the adaptive calibration method includes: Build includes A pre-trained model library for product changeover data, where each category contains several sets of data, each set including equipment state vectors, process parameters, and processing results. When a new changeover condition (such as the first...) occurs... When the initial data for a product category is less than 10 sets, a domain adaptation mechanism is initiated. This includes standardizing the source and target domain data, calculating the MMD distance, and when the MMD is greater than the distance threshold, adversarial training is performed. For example, a generator generates target domain style data, and a discriminator distinguishes between the source and target domains to achieve distribution alignment. During training, a gradient inversion layer is used to force the model to learn domain-invariant features. After the transfer is completed, the target domain data is expanded by generating three times the original data volume through data augmentation methods such as rotation and scaling. The source domain is the most similar historical product, and the target domain is the new production condition. The distance threshold is the critical value for judging whether the difference in data distribution between the source domain (historically similar products) and the target domain (newly produced products) is acceptable. Standard production data from different batches of the same product are selected as the benchmark dataset. The benchmark dataset is randomly divided into two groups, and the MMD distance between the two groups is calculated. This process is repeated 100 times to obtain the MMD distance distribution, and the 95th quantile of the distribution is taken as the distance threshold. The design employs a dual-memory LSTM architecture. The main memory unit processes the state feature sequence of the current operating condition, while the auxiliary memory unit stores key knowledge of the historical operating conditions. The new and old knowledge are integrated through a gating mechanism. The output layer adopts a Gaussian mixture distribution model to predict the probability density function of the parameter combination. The parameters are iteratively optimized through the EM algorithm. The system inputs historical operating condition data to predict the probability distribution of the optimal parameter combination for the future operating condition. A parameter-optimized Siamese network is constructed, and the parameter-response mapping relationship is learned by comparing the loss function. The network parameters are fine-tuned online every 10 production changes. The sparrow search algorithm is initialized with a population, a dynamic search radius is introduced, the search range is adjusted based on the parameter similarity of the Siamese network output, and the weights are updated adaptively. The inertial weights are dynamically adjusted according to the ratio of individual fitness to the average fitness of the population to enhance the global search capability. The parameter combination is optimized by combining an elite reverse learning strategy. A twin dynamics model is constructed based on the equipment physical model and LSTM error prediction. The model parameters are corrected in real time through the twin network. The objective function introduces a parameter smoothness penalty term for multi-objective rolling optimization to generate a parameter optimization sequence for future operating conditions. The equipment is divided into subsystems according to its functions, and a coupling matrix is ​​established. When the absolute value of the coupling coefficient exceeds the preset value, cross-subsystem collaborative adjustment is triggered. At the same time, the data of each production change is stored in the Neo4j graph database, including status features, optimization parameters, and processing accuracy. A knowledge graph containing equipment status, process parameters, and processing results is constructed. 1000 production change records are trained through graph neural network (GCN) to achieve optimal parameter reasoning for similar working conditions.

[0026] Specifically, methods for suppressing residues include: When a production change instruction is received, the primary clearing mechanism is activated to reset the historical process parameters to the default base values; Simultaneously, a secondary algorithm cleanup mechanism is initiated, including using the Transformer network to perform semantic parsing on the parameter sequence, combining the current device state vector and the optimized parameters of the previous batch, calculating the residual risk score of each parameter through the attention mechanism, and performing min-max normalization on the residual risk score based on the minimum and maximum original score values ​​of the corresponding parameters in all historical production change batches to generate a residual risk score matrix. Parameters with residual risk scores greater than the mandatory threshold are selected as risk parameters, and the gradient zeroing algorithm is performed on the risk parameters to eliminate their residual effects in the control logic and ensure the thoroughness of parameter initialization. Among them, complete parameter sequences and corresponding residual effect data of at least 50 successful production changes are collected, the residual risk score distribution of all parameters after production change is calculated, and the 95th quantile is taken as the mandatory threshold. By integrating real-time equipment status characteristics, parameter deviations between historical process parameters and default baseline values, and environmental parameters, a multi-dimensional feature vector is constructed. The multidimensional feature vector is input into the trained XGBoost model to predict the intensity of the residual effect in the short period of time after the production change, and outputs a three-dimensional prediction vector containing the influence of amplitude, phase and duration. Key influencing factors are identified by ranking features by importance. When the predicted residual effect amplitude exceeds the warning threshold, an automatic warning is issued for the risk parameters, and the risk parameters are highlighted. The warning threshold is the critical value for judging whether the residual historical parameters after production change will lead to excessive processing accuracy. Complete data of successful production change without compensation is obtained, and a mapping relationship between residual amplitude and processing error is established. All residual amplitude samples that cause processing error to exceed the tolerance by 50% are screened out, and the 90th quantile is taken as the warning threshold. When changing production lines, process parameters need to be adjusted to meet different needs. Process parameters that have been verified to be effective in actual production are defined as optimized parameters. Based on historical adjustment records, the optimized parameters are organized into a time-series sequence according to the changeover time and batch, thereby generating a historical optimized parameter sequence. The coupling strength between parameters is calculated using mutual information theory, and a dynamically updated coupling degree matrix is ​​constructed. At the same time, a strong coupling threshold is set, and the coupling strength exceeding the strong coupling threshold is defined as strong coupling. The strong coupling threshold is the critical value for judging whether there is a strong dependency relationship between two process parameters. The complete process parameter sequence and corresponding processing quality data for normal changeover are obtained, covering all product types and operating conditions. The mutual information of all process parameter pairs is calculated, and the distribution of mutual information values ​​of all parameter pairs is statistically analyzed. All parameter pairs that would cause the processing error to exceed the tolerance by more than 30% if adjusted independently are selected. The minimum value of the mutual information distribution of the parameter pairs is taken as the strong coupling threshold. By learning higher-order dependencies between parameters through graph neural networks, the coupling degree matrix is ​​transformed into a weighted directed graph model, where nodes are parameters and edge weights are coupling degrees. The coupling degree matrix is ​​updated periodically based on the latest production data to identify strongly coupled parameter pairs and visualize the coupling relationship in the form of a heatmap, providing a data foundation for multi-parameter collaborative optimization. A state space containing equipment status, 3D prediction vectors, and parameter deviations is constructed. An action space for parameter adjustment is defined, and a reward function for fusion accuracy, smoothness, and energy consumption is designed. A deep Q-network training optimization strategy is used, and training stability is improved through experience replay and target network mechanisms. The weights of fusion accuracy, smoothness, and energy consumption in the reward function are calibrated based on offline data. Complete production data of historical production change batches are obtained. The analytic hierarchy process (AHP) is used to determine subjective weights based on expert experience, and the entropy weight method is used to determine objective weights based on the dispersion of historical data. The weights of fusion accuracy, smoothness, and energy consumption are obtained by taking the arithmetic mean of the subjective and objective weights. Based on product requirements, ideal values ​​of parameters are set as target values ​​for optimization parameters. Upon receiving the target values ​​for optimization parameters, the network calculates the optimal action combination based on the current state and generates a proportional collaborative adjustment strategy for strongly coupled parameter pairs to ensure the stability of the process output after parameter adjustment. By integrating real-time machining accuracy error, three-dimensional prediction vector, and parameter deviation, a comprehensive error signal is formed. The fuzzy controller dynamically adjusts the PID parameters based on the comprehensive error signal and its rate of change, which then acts on the equipment actuator in real time. The PID parameters are updated at a fixed period until the machining error converges to the compliance range. During the compensation process, the parameter adjustment trajectory and error data are recorded in real time and used as training samples for the reinforcement learning model to continuously optimize the adaptability of multi-parameter collaborative optimization. The compliance range is the workpiece size tolerance of ±0.01mm.

[0027] Specifically, the compensation method is categorized as follows: Integrate equipment status vectors, optimization parameters, and processing accuracy data, access the time-series database, create indexes by production batch, and synchronize the three types of data through timestamps to ensure that different types of data are aligned in the time dimension, forming a dynamic verification data pool. A generative adversarial network with conditional input is constructed. The input is a finite sample of the current small batch of products, and the conditional variables are the equipment state vector and process parameters. The generator takes random noise and the concatenated vector of equipment state vector and process parameters as input. After multi-layer transposed convolution operation, it generates virtual samples with the same dimension as the real data. The discriminator receives real data, virtual samples and conditional variables to judge the authenticity of the data. At the same time, a specific loss function is used to train the model. After generating virtual samples, the maximum mean difference method is used to verify the consistency of the distribution of virtual samples with real data. Based on ResNet50, the first few convolutional layers are frozen to extract general features, and the later layers are replaced with fully connected layers adapted to small-batch scenarios. Using the domain adaptation mechanism, the gradient inversion layer enables the model to learn general features across product families and generate a validation model. The model takes the enhanced virtual samples as input and outputs the device state risk probability distribution. Multiple state features are extracted from equipment operation data, such as vibration spectrum entropy, temperature change rate, and current harmonic distortion rate. The weight of each feature is determined by the analytic hierarchy process. After standardizing the feature values, the weighted sum is used to obtain a comprehensive risk index. A two-dimensional risk map is constructed with comprehensive risk indicators as the vertical axis and time as the horizontal axis. The risk level is represented by the color intensity. Gradient activation mapping technology is used to map high-risk areas to equipment components, generating a heat map which is then overlaid on the equipment's three-dimensional model. At the same time, different risk indicator thresholds are set, corresponding to different warning levels. Specifically, the equipment status and fault data corresponding to production changes are obtained, and the probability distribution of comprehensive risk indicators is calculated. The 90th percentile is taken as the medium-risk threshold, and the 98th percentile is taken as the high-risk threshold. The medium-risk threshold is used to divide low-risk and medium-risk, and the high-risk threshold is used to divide medium-risk and high-risk. The comprehensive risk indicators are updated at fixed time intervals, and local real-time analysis is achieved through edge computing to ensure a short warning delay time. Corresponding response strategies, such as parameter fine-tuning or shutdown intervention, are triggered according to the warning level. Collect various unconventional working condition data, including equipment state vectors, parameter sequences, and machining accuracy data; simultaneously perform short-time Fourier transform on vibration signals to generate vibration time-frequency diagrams; extract multiple current features from current signals; manually label working condition types; and establish a database containing multiple types of features. A hybrid model is constructed using convolutional neural networks and bidirectional gated recurrent units. The model includes a CNN layer that extracts spatial features of the time-frequency graph through multi-layer convolution operations, a Bi-GRU layer that processes temporal features, an attention layer that calculates the weights of each time step, and training parameters that are set. The model is trained with a large amount of data to improve the accuracy of working condition recognition. The current vibration time-frequency diagram, current characteristics, and process parameters are input into the hybrid model, which outputs the probability distribution of each operating condition type. Different processing is performed according to the probability value. A two-level threshold is set to divide the condition into first-level confidence, second-level confidence, and third-level confidence. The probability value corresponding to the first-level confidence is relatively large, and the operating condition is confirmed at this time. The probability value corresponding to the second-level confidence is within the range of the second-level threshold, and more sampling points are added for secondary identification. The probability value corresponding to the third-level confidence is relatively small, and it is marked as an unknown operating condition and data acquisition is started. The compensation strategy is set up, including a pre-compensation layer, a correction layer and an optimization layer. The pre-compensation layer is based on modal analysis theory to establish a deformation compensation calculation model. The correction layer adopts an iterative learning control method to update the compensation parameters according to the previous processing error. The optimization layer uses the Bayesian optimization algorithm to search for the best parameter combination with a multi-objective function. The compensation strategies are categorized and stored according to the type of working condition. Each strategy includes a compensation model, parameter range, and application conditions. Knowledge graph technology is used to build the relationship between strategies, and particle swarm optimization algorithm is used to update and optimize the strategy parameters. When an unconventional operating condition is identified, the corresponding compensation model is retrieved and invoked from the strategy library. Based on the current equipment state vector, the compensation parameters are adjusted in real time. In this process, a machine learning model with fused features, such as XGBoost, is used to construct a compensation model. The model collects equipment state vectors (multi-dimensional sensing data), corresponding historical compensation parameters (effective adjustment records), and compensation effect data (accuracy, efficiency, and other indicators) under various unconventional operating conditions as inputs to output adapted compensation parameters. The calculation of indicators related to processing accuracy, production efficiency, and equipment reliability is performed. Each indicator is standardized, and the entropy weight method is used to calculate its weight. The standardized indicator values ​​are then weighted and summed to calculate a comprehensive score. The overall score is determined by focusing on the degree of improvement in the indicators before and after compensation. , At the same time, obtain the theoretically optimal comprehensive score for the corresponding scenario. ,based on , The difference and , The ratio of the differences is used to calculate the compensation effectiveness score. Based on the score, a feedback mechanism is triggered to analyze the source of error and generate an optimization scheme. Among them, the theoretical optimal comprehensive score is the comprehensive score corresponding to the ideal best state that the equipment can achieve in the current production scenario. It is obtained by multiplying the theoretical optimal values ​​of three indicators, namely processing accuracy, production efficiency and equipment reliability (e.g., all of which are 1.0) by the corresponding entropy weight method weights and summing them. The Q-learning algorithm is used to optimize the strategy parameters. The combination of compensation parameters is the state, the parameter adjustment amount is the action, and the effectiveness score is the reward. The optimization is initiated at fixed production change cycles to update the parameters of the compensation strategy.

[0028] Specifically, the optimization methods include: A four-layer closed-loop architecture is constructed, comprising a perception layer, a decision layer, an execution layer, and a verification layer. Federated learning is employed to protect data privacy. Each edge node trains its model locally and then uploads its gradient, while the central server aggregates and updates the global model. The perception layer integrates equipment state vectors, residual risk scoring matrices, and classification results of unconventional operating conditions. Edge computing nodes preprocess this data to extract key features, such as vibration entropy and thermal deformation, and push the processed data to the decision layer at fixed intervals. The decision layer integrates collaborative adjustment strategies and compensation parameters, and combines them with verification scores, including comprehensive scores and compensation effectiveness scores, to construct a multi-objective optimization decision model. The optimization strategy is updated periodically. Upon receiving optimization parameters, the execution layer simultaneously triggers a two-level clearing mechanism and fuzzy PID compensation to update the equipment parameter configuration. The verification layer collects risk map data and compensation effect scores, and feeds them back to the decision layer through a unified data interaction protocol, thus forming a complete data closed loop. By integrating three types of objectives—production efficiency, quality stability, and equipment health—a dynamic weighting function is constructed. Fuzzy logic is used to adjust the weights according to different production stages. For example, higher weight is given to efficiency objectives at the beginning of production changeover, while more emphasis is placed on quality objectives during the processing stage. Equipment safety limits and process specifications are transformed into soft constraints, and penalty functions are used to handle the constraints to ensure that the optimization process is carried out within the range that meets actual production requirements. Based on the NSGA-Ⅲ algorithm, the cooperative parameters and compensation parameters are merged and encoded into real number vectors. During the algorithm execution, simulated binary crossover and polynomial mutation operations are used to perform crossover and mutation to maintain the diversity of solutions. Based on the compensation effect score fed back by the verification layer, the retention ratio of elite solutions is adaptively adjusted. When the score is high, the proportion of elite solutions is increased to accelerate the convergence speed of the algorithm. By combining the early warning information from the risk map, the prediction time domain is dynamically adjusted and optimized. A longer prediction time domain is set under normal operating conditions, while when a high-risk operating condition is detected, the prediction time domain is shortened to improve the response speed. Each optimization only executes the result of the current cycle, and the next cycle is re-optimized based on the latest equipment status to ensure that the optimization strategy can adapt to the operating status of the equipment in real time. For the state prediction model, the existing LSTM architecture and hybrid model are integrated, and the model is updated regularly using transfer learning. After every 20 production changes, the pre-trained model is fine-tuned using new task data. During fine-tuning, the transfer coefficient is fixed at 0.2 to avoid significantly altering the generalization ability of the original model. The model's generalization ability and task adaptability are balanced by adjusting the transfer coefficient in transfer learning. Meanwhile, for the compensation model, the experience playback mechanism based on reinforcement learning prioritizes the playback of data from high-risk working conditions to improve the model's compensation ability under complex working conditions.

[0029] Three new entity types—device status, compensation strategy, and verification result—are added to the knowledge graph, and relationship types, such as cause, optimization, and verification, are defined. After each closed-loop iteration, key knowledge is extracted and the embedding representation of nodes in the graph is updated through a graph convolutional network, thereby realizing the dynamic updating of the knowledge graph. The updated knowledge graph is then used for retrieval and reasoning of similar working conditions, improving the accuracy of strategy retrieval.

[0030] A meta-learning approach is used to construct a general optimizer. By calculating the difference in feature distribution between the source device and the target device, it is determined whether the conditions for knowledge transfer are met. When the distribution difference is small, the transfer process is initiated. The general optimizer is quickly adjusted using support set data to adapt it to the operating characteristics of the new device and shorten the debugging time of the new device. Smoothing filtering is applied to the clearing parameters and coordination parameters. By setting a smoothing coefficient, the impact of parameter mutations on the equipment is reduced. The execution of compensation parameters is monitored in real time. The executed value is compared with the target value. The deviation is corrected by a proportional-integral control algorithm. Based on the parameter coupling degree matrix, the strongly coupled parameters are coordinated and adjusted to ensure the consistency and stability of parameter adjustment. By combining risk map data, the isolated forest algorithm is used to detect anomalies in equipment operation status. When a minor anomaly is detected, a compensation strategy is automatically invoked and the correction amount is increased. If a serious anomaly is detected, an early warning mechanism is immediately triggered and the corresponding early warning logic is executed. After the anomaly handling is completed, the compensation parameters are regenerated, and the equipment is verified to return to normal operation through a no-load run. Establish a comprehensive performance evaluation index system, in which efficiency indexes include equipment utilization rate, calculated by the ratio of actual operating time to planned operating time; quality indexes include process capability index, calculated based on the mean, standard deviation, and specification limits of processing data; and reliability indexes include mean time between failures, determined by statistically averaging the mean time between failures.

[0031] A full-process performance evaluation is scheduled every 50 production changes or monthly. A SWOT analysis is used to compare the process with industry benchmarks to identify strengths, weaknesses, opportunities, and threats. For identified weaknesses, corresponding iterative evolution strategies are developed, such as optimizing sensor network layout to improve data acquisition efficiency and updating validation models to enhance risk identification capabilities. The effectiveness of these strategies is then verified using actual production data. A digital twin corresponding to the physical equipment is constructed. Real-time collected equipment status and control parameters of the equipment actuators are input into the digital twin for simulation to predict equipment performance trends. New parameter optimization strategies are tested in the digital twin environment. The effectiveness of the strategies is evaluated by comparing simulation output with actual operating data. Virtually validated collaborative adjustment and compensation strategies are then migrated to the physical equipment, reducing trial-and-error costs in actual production. The parameter optimization strategies include collaborative adjustment and compensation strategies.

[0032] Example 2 Please see Figure 4 Another embodiment of the present invention provides an adaptive compensation system for equipment changeover, comprising: a state perception module, a parameter calibration module, a residual suppression module, an operating condition compensation module, and a closed-loop optimization module; The state perception module is used to collect real-time operating data of key components of the equipment through a three-layer three-dimensional sensor array. It performs time domain, frequency domain, and time-frequency domain feature extraction through edge computing, and combines improved DS evidence theory to fuse multi-source data to build a multi-dimensional representation model of equipment state, so as to achieve accurate identification of equipment dynamic characteristics and solve the problem of control compensation lag caused by insufficient semantic understanding of equipment state. The parameter calibration module is used to predict the optimal parameter combination under different working conditions based on the deep learning dynamic working condition model. It combines intelligent optimization algorithm and processing error to adjust parameters in real time, establishes the mapping relationship between parameter adjustment and equipment response, realizes rapid parameter convergence, solves the dynamic balance reconstruction delay caused by traditional static model, and maintains the continuity of processing accuracy during production changeover. The residual suppression module is used to activate a hardware-level and algorithm-level dual-level parameter clearing mechanism during production changeover. It predicts residual effects through the XGBoost model and designs a compensation algorithm. It combines the parameter coupling degree matrix and deep Q network to achieve multi-parameter collaborative optimization, solve the problem of process output deviation caused by historical parameter residues and coupling fluctuations, and ensure the thoroughness of parameter initialization and process stability. The operating condition compensation module is used to integrate multi-source data to build a dynamic verification model, use generative adversarial networks to enhance small batch samples, combine transfer learning to assess equipment status risks, identify unconventional operating conditions through a hybrid model and call customized compensation strategies to solve the problem of low process reliability caused by insufficient small batch verification samples and lack of unconventional operating condition strategies. The closed-loop optimization module is used to establish a full-process data feedback mechanism. It iteratively optimizes the models of each link through multi-objective optimization algorithms and federated learning. Combined with knowledge graphs and digital twins, it realizes cross-operating condition knowledge transfer and strategy verification, solves the defects of production continuity and quality stability, and meets the dynamic adaptation requirements of rapid equipment switching.

[0033] Example 3 An embodiment of the present invention provides a medium storing a computer program, which, when executed by a processor, implements the aforementioned adaptive compensation method for equipment replacement.

[0034] Example 4 A computer device according to an embodiment of the present invention includes a processor and the aforementioned medium. The processor executes a computer program in the medium to implement the aforementioned adaptive compensation method for equipment replacement.

[0035] In summary, this invention establishes a multi-type sensor monitoring network to collect data from key equipment components in real time. This data is preprocessed using edge computing and extracted from multiple dimensions, then fused to construct an equipment state model, accurately identifying dynamic characteristics. Based on deep learning, it dynamically predicts optimal parameters and intelligently adjusts them during production changes, combining processing errors to establish a parameter-response mapping to accelerate convergence. During production changes, it triggers a two-stage parameter clearing process, predicts and compensates for residual effects, analyzes parameter coupling, and implements collaborative optimization. It integrates multi-source data to enhance verification samples, combines transfer learning to assess risks, trains models for unconventional operating conditions, and invokes customized strategies. Through full-process data feedback, it iteratively optimizes the model algorithms at each stage, achieving adaptive compensation and improving the equipment's dynamic adaptability.

[0036] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An adaptive compensation method for equipment changeover, characterized in that, include: Establish a real-time monitoring network, set up real-time sensing methods, continuously collect operational data, preprocess the data using edge computing, extract feature parameters from the time domain, frequency domain, and time-frequency domain using feature extraction algorithms, and construct a multi-dimensional characterization model of equipment status to identify the dynamic characteristics of the equipment. Based on the equipment status and production changeover process requirements, the optimal parameter combination is dynamically predicted, an adaptive calibration method is set, and the equipment parameters are adjusted in real time in combination with the processing error, and a mapping relationship between parameter adjustment and equipment response is established. When changing production, a parameter memory clearing mechanism is triggered, historical parameters are initialized, a method to suppress residual effects is set, the degree of interference is analyzed through a residual effect prediction model and a compensation algorithm is designed, and a collaborative adjustment strategy is adopted after coupling analysis of the parameters. Integrate multi-source data to build a dynamic verification model, set up classification compensation methods, assess equipment status risks, collect unconventional operating condition data to train the compensation model, and design customized compensation strategies for different operating conditions. Establish a full-process data feedback mechanism from equipment status perception to process verification, set optimization methods, and use data analysis algorithms to iteratively optimize the models and algorithms of each link.

2. The adaptive compensation method for equipment changeover according to claim 1, characterized in that: The real-time sensing method includes: A three-layer stereo sensor array layout is adopted to form a three-dimensional monitoring network; The design incorporates a dynamic sampling mechanism. Under normal operating conditions, sampling is performed synchronously at the basic sampling interval. Once a production change signal is detected, fuzzy logic is used to calculate and adjust the sampling rate. The collected vibration signal is divided into data segments, the kurtosis index is calculated, and a time-domain attention weight matrix is ​​generated using a sliding time window; The vibration signal is decomposed by wavelet packet and key frequency bands are selected by autoencoder and fully connected network; The temperature signal is subjected to S-transform and its time-frequency domain features are enhanced by 2D convolution and global average pooling; The time domain, frequency domain, and time-frequency domain features are integrated into a multi-dimensional feature vector. After being divided into sub-vectors according to sensor type, the vectors are input into the evidence generator, and the output probability allocation sequence is generated. Based on the variance and mean of each probability assignment sequence in the sliding window, the activity index is calculated and the weights are dynamically adjusted. Calculate the conflict coefficient between pieces of evidence and redistribute the weights of evidence that exceeds the preset value; The device state vector is generated by using a pre-trained state transition matrix and an extended Kalman filter.

3. The adaptive compensation method for equipment changeover according to claim 2, characterized in that: The adaptive calibration method includes: Constructing A pre-trained model library for product changeover data. When the initial data for the new changeover is insufficient, a domain adaptation mechanism is activated to standardize the source and target domain data and calculate the MMD distance. After achieving distribution alignment through adversarial training and gradient inversion layer, the target domain data is expanded. The design employs a dual-memory LSTM architecture. The main memory unit processes the state feature sequence of the current operating condition, while the auxiliary memory unit stores historical knowledge. The new and old knowledge are integrated through a gating mechanism, and the probability distribution of parameter combinations is predicted using a Gaussian mixture model and the EM algorithm.

4. The adaptive compensation method for equipment changeover according to claim 3, characterized in that: The adaptive calibration method further includes: Construct a parameter-optimized Siamese network, learn the parameter-response mapping by comparing loss functions, and fine-tune it online periodically; The sparrow search algorithm is initialized with a population, a dynamic search radius is introduced, and the search range is adjusted based on the parameter similarity of the output of the twin network according to the parameters. The parameter combination is optimized by combining adaptive weight update and elite reverse learning strategy. A twin dynamics model is constructed based on the equipment physical model and LSTM error prediction. The model parameters are corrected through a twin network, and a parameter smoothness penalty term is introduced to perform multi-objective rolling optimization to generate a parameter optimization sequence. The equipment is divided into subsystems and a coupling matrix is ​​established to trigger cross-subsystem collaborative adjustment. The production change data is stored in the knowledge graph constructed by the graph database, and the inference of similar working condition parameters is achieved through graph neural network training.

5. The adaptive compensation method for equipment changeover according to claim 4, characterized in that: The method for inhibiting residues includes: When a production change instruction is received, the primary clearing mechanism is activated to reset the historical process parameters to the default base values; The secondary algorithm cleanup mechanism is activated. The Transformer network is used to perform semantic parsing on the parameter sequence. Combined with the current state vector of the device and the optimization parameters of the previous batch, the residual risk score of each parameter is calculated through the attention mechanism and a residual risk score matrix is ​​generated. Risk parameters that exceed the preset threshold are screened out and the gradient zeroing algorithm is executed. A multi-dimensional feature vector is constructed by integrating real-time equipment status characteristics, parameter deviations between historical process parameters and default baseline values, and environmentally relevant parameters. The multidimensional feature vector is input into the trained XGBoost model to predict the intensity of residual effects in the short period of time after the production change, and outputs a three-dimensional prediction vector, including the influence of residual effect amplitude, phase and duration. Key influencing factors are identified by ranking features by importance, and a warning is issued for risk parameters when the predicted residual effect amplitude exceeds a preset threshold.

6. The adaptive compensation method for equipment changeover according to claim 5, characterized in that: The method for inhibiting residues also includes: Define optimization parameters and integrate them according to production change time and batch based on historical data to generate a historical optimization parameter sequence. Use mutual information theory to calculate the coupling strength between parameters, construct a dynamically updated coupling degree matrix, and set a strong coupling threshold to define strong coupling. By learning higher-order dependencies between parameters through graph neural networks, the coupling degree matrix is ​​transformed into a weighted directed graph model and the coupling degree matrix is ​​updated periodically. Strongly coupled parameter pairs are identified and the coupling relationship is visualized in the form of a heatmap. A state space containing device status, three-dimensional prediction vectors and parameter deviations is constructed, an action space for parameter adjustment is defined, a reward function that integrates accuracy, smoothness and energy consumption is designed, a deep Q-network training optimization strategy is used, and training stability is improved through experience replay and target network mechanisms. Upon receiving the target value of the optimization parameters, the optimal action combination is calculated based on the current state, and a proportional collaborative adjustment strategy is generated for strongly coupled parameter pairs. A comprehensive error signal is formed by integrating real-time machining accuracy error, three-dimensional prediction vector and parameter deviation. The PID parameters are then dynamically adjusted by a fuzzy controller based on the comprehensive error signal and the rate of change of the comprehensive error signal. The PID parameters are updated at fixed intervals until the processing error converges to a preset range. At the same time, during the compensation process, the parameter adjustment trajectory and error data are recorded in real time and used as training samples for the reinforcement learning model.

7. The adaptive compensation method for equipment changeover according to claim 6, characterized in that: The classification compensation method includes: Integrate equipment state vectors, optimization parameters, and machining accuracy data to form a dynamic verification data pool; A generative adversarial network with conditional input is constructed. The virtual samples are generated by the concatenation vector of random noise, equipment state vector and process parameters. The discriminator is trained by real data, virtual samples and conditional variables. The consistency of sample distribution is verified by the maximum mean difference method. Using ResNet50 as the base network, combined with a domain adaptation mechanism and a gradient inversion layer, a validation model is generated. The enhanced virtual samples are used as input, and the output is the probability distribution of device state risk. The status features are extracted from the equipment operation data, and the weights are determined by the analytic hierarchy process. After standardization, the weighted sum is obtained to obtain a comprehensive risk index. Construct a two-dimensional risk map and set thresholds to trigger response strategies.

8. The adaptive compensation method for equipment changeover according to claim 7, characterized in that: The classification compensation method also includes: Collect unconventional operating condition data and perform signal processing and manual annotation to establish a database; A hybrid model is constructed using convolutional neural networks and bidirectional gated recurrent units, and the recognition accuracy is optimized through an attention layer; Based on the working condition probability distribution output by the hybrid model, set graded thresholds for working condition confirmation, secondary identification, or collection of unknown working conditions. Set up compensation strategies that include pre-compensation, correction, and optimization layers, and store them according to working conditions; The strategy association relationship is constructed using a knowledge graph, and the parameters of the compensation strategy are updated using a particle swarm optimization algorithm; Based on the identified abnormal operating conditions, the compensation model is retrieved and invoked, and the parameters are adjusted according to the equipment state vector. After calculating and standardizing the accuracy, efficiency, and reliability indicators, a comprehensive score and a compensation effectiveness score are generated. The Q-learning algorithm is used to update the parameters of the compensation strategy with the compensation parameter combination as the state, the parameter adjustment amount as the action, and the effectiveness score as the reward.

9. The adaptive compensation method for equipment changeover according to claim 8, characterized in that: The optimization method includes: Construct a four-layer closed-loop architecture comprising a perception layer, a decision-making layer, an execution layer, and a verification layer; The federated learning mechanism is adopted, where each edge node trains the model locally and then uploads the gradient, and the central server aggregates and updates the global model. A dynamic weighting function is constructed by integrating production efficiency, quality stability, and equipment health objectives. The weights are adjusted according to the production stage using fuzzy logic. Equipment safety restrictions and process specifications are transformed into soft constraints and processed through penalty functions. Based on the NSGA-Ⅲ algorithm, the collaborative parameters and compensation parameters are merged and encoded into a real number vector. Simulated binary crossover and polynomial mutation are used to maintain the diversity of solutions. The proportion of elite solutions retained is adaptively adjusted according to the compensation effect score. The prediction time domain is dynamically adjusted and optimized by combining the two-dimensional risk map early warning, and the optimization is performed periodically and updated based on the latest status. The system integrates LSTM architecture and hybrid model, uses transfer learning to periodically update the state prediction model, and prioritizes the replay of high-risk working condition data to update the compensation model based on reinforcement learning experience replay mechanism. Add device status, compensation strategy and verification result entities to the knowledge graph and define relation types. Update the node embedding representation through graph convolutional network to dynamically update the knowledge graph and perform similar working condition retrieval and reasoning. A general optimizer is constructed using meta-learning, and the knowledge transfer conditions are determined by calculating the differences in the distribution of device features and adjusting the optimizer using support set data. Smoothing filtering is applied to the clearing parameters and collaborative parameters. The executed value of the compensation parameter is compared with the target value and the deviation is corrected by proportional-integral control. The strongly coupled parameters are adjusted collaboratively based on the parameter coupling degree matrix. Anomaly detection is performed using the isolated forest algorithm in conjunction with risk map data. Compensation strategies or early warning logic are triggered based on the degree of anomaly, and recovery is verified through no-load operation. Establish a performance evaluation index system that includes equipment utilization, process capability index and mean time between failures, and regularly use SWOT analysis to compare with industry benchmarks and formulate evolutionary strategies. A digital twin is constructed to simulate the input device status and control parameters of the device actuators. Optimization strategies are tested, and the collaborative adjustment and compensation strategies that have passed virtual verification are migrated to the physical device.

10. An adaptive compensation system for equipment changeover, used to implement the adaptive compensation method for equipment changeover as described in any one of claims 1-9, characterized in that, include: The module includes a state perception module, a parameter calibration module, a residual suppression module, a working condition compensation module, and a closed-loop optimization module. The state perception module is used to collect real-time operating data of key components of the equipment through a three-layer three-dimensional sensor array, extract and fuse features through edge computing, so as to identify the dynamic characteristics of the equipment. The parameter calibration module is used to predict the optimal parameter combination under different working conditions, adjust the parameters in real time in combination with the processing error, and establish a mapping relationship between parameter adjustment and equipment response. The residual suppression module is used to activate a hardware-level and algorithm-level dual-level parameter clearing mechanism during production change. It predicts residual effects and designs compensation algorithms through the XGBoost model, and achieves multi-parameter collaborative optimization by combining the parameter coupling degree matrix and the deep Q network. The operating condition compensation module is used to integrate multi-source data to build a dynamic verification model, use generative adversarial networks to enhance small batch samples, combine transfer learning to assess equipment status risks, identify unconventional operating conditions through a hybrid model, and call customized compensation strategies. The closed-loop optimization module is used to establish a full-process data feedback mechanism, and to iteratively optimize the models of each stage through multi-objective optimization algorithms and federated learning, and to realize cross-condition knowledge transfer and strategy verification by combining knowledge graphs and digital twins.

11. A medium, characterized in that, The medium stores a computer program, which, when executed by a processor, implements the adaptive compensation method for equipment replacement as described in any one of claims 1-9.

12. A computer device, characterized in that, include: Memory, used to store instructions; A processor for executing the instructions to cause the device to perform operations implementing the adaptive compensation method for device turnover as described in any one of claims 1 to 9.

Images