Roller brush production self-adaptive pressure regulation method and system based on multi-sensor fusion

Abstract

The application discloses a kind of based on multi-sensor fusion's drum sticky brush production self-adapting pressure regulating method and system, belong to industrial automation control field.For the problem that existing feedback regulation exists lag, difficult to deal with the quick change of process parameter, the application is by distributed sensor network synchronous acquisition drum speed, brush hardness, viscose viscosity and environmental temperature and so on multi-source parameter, constructs process parameter pressure demand mapping database and trains time series prediction model, and carries out advance prediction to future pressure demand;Predicted value and real-time feedback value are weighted fusion, and fusion weight is dynamically adjusted based on fuzzy reasoning according to working condition, generates composite control instruction to drive fast response actuator to implement pressure compensation in advance.The application changes pressure regulation from passive feedback to active prediction, effectively suppresses pressure fluctuation, and improves coating uniformity and product quality consistency.

Description

Technical Field

[0001] This application belongs to the field of industrial automation control, specifically relating to an adaptive pressure regulation method and system for roller adhesive brush production based on multi-sensor fusion. Background Technology

[0002] In the production of roller adhesive brushes, pressure control is a core process to ensure uniform adhesive coating and bonding strength. In traditional production models, pressure setting typically relies on manual experience or fixed process parameters, lacking the ability to respond in real-time to numerous dynamic factors during production. With increasing production speeds and more stringent process requirements, static pressure control methods are no longer sufficient to meet the demands for high-quality, highly consistent production.

[0003] Several technical solutions involving pressure regulation have been proposed. For example, some solutions deploy temperature and deformation sensors at key locations on the equipment to collect temperature and deformation signals in real time during production, and then dynamically adjust the pressure based on feedback from these sensors. Other solutions introduce self-learning models on top of feedback regulation, attempting to optimize pressure control parameters through analysis of historical production data. In summary, these existing solutions generally adopt a "sensing-feedback-regulation" technical approach—that is, first detecting the current state through sensors, and then making post-event corrections to the pressure based on the detection results.

[0004] However, the aforementioned existing technical solutions share a common limitation: their pressure regulation is essentially a "deviation-driven" ex-post compensation mechanism. Whether it's feedback regulation based on temperature and deformation sensors or parameter optimization incorporating self-learning models, the response and compensation only occur after a pressure deviation has occurred. When production speed is high or process parameters change frequently, this inherent regulation lag prevents the system from predicting and compensating for pressure fluctuations before they are transmitted to the output, resulting in periodic fluctuations in product quality. Therefore, there is an urgent need for a predictive pressure regulation method that can detect upstream process parameter changes in advance and proactively implement compensation before pressure fluctuations occur. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an adaptive pressure adjustment method and system for roller adhesive brush production based on multi-sensor fusion, which can effectively solve the problems mentioned in the background technology.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides an adaptive pressure adjustment method for roller adhesive brush production based on multi-sensor fusion, the method comprising the following steps: Step S1: Multi-source process parameters during the production of roller adhesive brush are synchronously collected through a distributed sensor network. The collected raw data is preprocessed to generate a time-aligned standardized process parameter sequence. Step S2: Construct a process parameter-pressure demand mapping relationship database. The process parameter-pressure demand mapping relationship database stores mapping records between multiple sets of process parameter combinations and their corresponding pressure setpoints. The historical mapping records in the process parameter-pressure demand mapping relationship database are used for training and parameter tuning of the time series prediction model. Step S3: Input the standardized process parameter sequence into the time-series prediction model. The time-series prediction model predicts the pressure demand at future moments based on the historical process parameter sequence and outputs the predicted pressure value. The predicted pressure value output by the time-series prediction model is weighted and fused with the feedback pressure value collected by the real-time pressure sensor to generate a composite control command. The feedforward weight and feedback weight used in the weighted fusion are dynamically adjusted according to the current working condition based on fuzzy inference rules. Step S4: The composite control command is converted into a drive signal for the fast-response pressure actuator, which then performs the pressure regulation action.

[0007] On the other hand, the present invention provides an adaptive pressure regulation system for roller adhesive brush production based on multi-sensor fusion, comprising: The multi-source process parameter acquisition module synchronously acquires multi-source process parameters during the production of roller adhesive brushes through a distributed sensor network. It performs preprocessing operations such as filtering, noise reduction, and normalization on the acquired raw data, and finally generates a standardized process parameter vector with unified format and time sequence alignment. The process parameter-pressure demand mapping database stores multidimensional mapping relationships between process parameters such as roller speed, brush body hardness, adhesive viscosity, and ambient temperature and optimal pressure setpoints. The historical mapping records in the process parameter-pressure demand mapping database are also used for training and parameter tuning of time series prediction models. The time-series prediction model takes the standardized process parameter vector as input, predicts the pressure demand at future moments based on the historical process parameter sequence, and outputs the predicted pressure value. The feedforward-feedback weighted fusion control module weights and fuses the predicted pressure value output by the time-series prediction model with the feedback pressure value collected by the real-time pressure sensor to generate a composite control command. The feedforward weight and feedback weight used in the weighted fusion are dynamically adjusted according to the current working conditions based on fuzzy inference rules. The fast-response pressure actuator receives the composite control command, converts the composite control command into a drive signal for the fast-response pressure actuator, and performs a pressure regulation action.

[0008] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention constructs a full-link active predictive control mechanism from front-end process parameter sensing to back-end pressure execution. By synchronously collecting multi-source process parameters through a sensor network and fusing them with a time-series prediction model, pressure regulation can sense upstream change trends in advance and complete prediction and compensation before pressure fluctuations are transmitted to the output end. This fundamentally overcomes the inherent response lag defect of traditional feedback regulation and significantly improves the stability and anti-interference capability of the production process.

[0009] 2. This invention utilizes a weighted fusion control architecture of feedforward prediction and real-time feedback, combined with a dynamic weight adjustment strategy based on fuzzy inference. This enables the control system to adaptively optimize the contribution ratio of feedforward and feedback according to changes in the current operating conditions, taking into account both the forward-looking advantages of predictive control and the steady-state accuracy of feedback control. Even under conditions of frequent fluctuations in process parameters, it can still achieve precise and stable pressure regulation, effectively ensuring the uniformity of adhesive coating and the consistency of bonding strength.

[0010] 3. This invention enables the system to continuously absorb new production data and adapt to process evolution through the online update mechanism of the process parameter pressure demand mapping database and the incremental training strategy of the time series prediction model. It continuously optimizes prediction accuracy and control performance without forgetting historical experience, and has good long-term operational adaptability and self-evolution characteristics. It is suitable for continuous production scenarios of roller adhesive brushes with multiple varieties and changing working conditions. Attached Figure Description

[0011] Figure 1 A schematic diagram of the overall technology for an adaptive pressure regulation method in roller adhesive brush production based on multi-sensor fusion; Figure 2 A schematic diagram illustrating the core principle of time series prediction model and feedforward-feedback weighted fusion control; Figure 3 This is a flowchart illustrating the logic of the multi-source process parameter acquisition and preprocessing stage. Figure 4 Flowchart for the execution of control commands for a pressure actuator to provide rapid response. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1 To be continued Figure 4 Specific embodiments are provided to further illustrate the present invention in detail.

[0013] Firstly, this invention provides an adaptive pressure regulation method for roller adhesive brush production based on multi-sensor fusion. This method achieves accurate prediction and adaptive regulation of pressure during roller adhesive brush production through the coordinated operation of a multi-source process parameter acquisition module, a process parameter-pressure demand mapping database, a time-series prediction model, a feedforward-feedback weighted fusion control module, and a fast-response pressure actuator. The overall technical architecture, internal structure of each module, data processing flow, and control strategy of this method are described in detail below.

[0014] In the adaptive pressure regulation method for roller adhesive brush production based on multi-sensor fusion, the overall technical architecture comprises four core functional layers: the process parameter sensing layer, the data processing and analysis layer, the predictive control decision layer, and the execution drive layer. The process parameter sensing layer is responsible for real-time acquisition of upstream process parameters from various stages of the roller adhesive brush production line. The data processing and analysis layer preprocesses and extracts features from the raw acquired data. The predictive control decision layer generates control commands based on a time-series prediction model and a weighted fusion control strategy. The execution drive layer receives the control commands and drives the pressure actuator to complete the specific pressure regulation action. The four functional layers interact through standardized data interfaces and communication protocols, forming a complete information and control flow closed loop, implemented according to the following steps.

[0015] The first step, S1, involves multi-source process parameter acquisition and preprocessing. Step S1 is the initial stage of the process parameter sensing layer and data processing and analysis layer within the overall technical architecture. The main task of this step is to synchronously acquire key process parameters during the production of the roller adhesive brush using multiple distributed sensors. The acquired raw data undergoes preprocessing operations such as filtering, noise reduction, and normalization to ultimately generate a standardized process parameter vector with a unified format and time sequence alignment. This provides data input for subsequent mapping relationship queries and time sequence prediction. This is implemented through the following sub-steps.

[0016] Step S101: Synchronous acquisition of multi-source process parameters. In the process parameter sensing layer, the multi-source process parameter acquisition module is responsible for the synchronous acquisition of various key process parameters through a distributed sensor network. The sensor network specifically includes the following four types of sensors.

[0017] The drum speed sensor is installed on the main shaft of the drum drive motor. It uses a non-contact photoelectric encoder or a magnetoelectric speed sensor to measure the actual speed of the drum. The drum speed sensor outputs a pulse signal or analog voltage signal that is proportional to the speed.

[0018] The brush body material hardness testing device is deployed at the entrance of the brush body conveying channel. It uses the ultrasonic hardness testing principle or a mechanical hardness measuring probe to perform online testing of the hardness of the brush body material. The test results reflect the elastic modulus and surface hardness characteristics of the current batch of brushes.

[0019] The adhesive viscosity meter is installed in the bypass position of the adhesive supply pipeline. It uses a rotary viscometer or a vibratory viscometer to measure the viscosity of the adhesive fluid in real time. The measurement results reflect the flow resistance and coating performance of the adhesive.

[0020] Ambient temperature sensors are installed in typical locations in the production line's working area. Platinum resistance temperature sensors or thermocouple sensors are selected to measure the ambient temperature, and the measurement results reflect the impact of environmental conditions on the process.

[0021] To ensure the effectiveness of subsequent data fusion, each sensor employs a synchronous sampling mechanism in the multi-source process parameter acquisition module to guarantee the temporal consistency of multi-source data. The synchronous sampling mechanism triggers data acquisition operations from all sensors using a unified clock reference signal, ensuring that data collected simultaneously, such as roller speed, brush material hardness, adhesive viscosity, and ambient temperature, remain strictly aligned in the time dimension.

[0022] Step S102: Dynamically adjust the sampling frequency. The sampling frequency is dynamically adjusted according to the operating speed of the production line. When the production line operates at a high speed, the sampling frequency is automatically increased to ensure the temporal resolution of the data, so that rapidly changing process details can be fully captured; when the production line operates at a low speed, the sampling frequency is appropriately reduced to reduce the amount of data processing.

[0023] The dynamic sampling frequency adjustment strategy is achieved by monitoring the rate of change of the drum speed signal in real time. When the rate of change of speed exceeds a preset threshold, it is determined that the current working condition is in a transition phase, and the sampling frequency is automatically increased to the highest level to capture rapidly changing process parameters; when the rate of change of speed is lower than the preset threshold, it is determined that the current working condition is in a stable phase, and the sampling frequency is adjusted to the normal level to maintain the continuity and stability of data acquisition.

[0024] Step S103: Raw data preprocessing. The collected raw data is filtered, denoised and normalized by the preprocessing unit inside the multi-source process parameter acquisition module.

[0025] The filtering and noise reduction process employs either an adaptive Kalman filter or a moving average filter. An adaptive Kalman filter automatically adjusts its parameters based on the statistical characteristics of the acquired signal, thereby achieving noise suppression. A moving average filter, by setting a sliding window length, smooths the original signal to eliminate random noise and spike interference.

[0026] Normalization maps the filtered and denoised raw data to a uniform standard range. Normalization methods include min-max normalization or Z-score standardization. Min-max normalization linearly maps the data to the [0,1] interval, while Z-score standardization converts the data into a standard normal distribution with a mean of 0 and a variance of 1.

[0027] The preprocessed data is integrated into a process parameter vector in a unified format. The dimension of the process parameter vector is consistent with the number of types of process parameters collected. Each dimension in the process parameter vector corresponds to the normalized value of a process parameter.

[0028] In summary, step S1 completes the acquisition, synchronization, dynamic adjustment, and preprocessing of multi-source process parameters. Through this series of processes, the original sensor signals from different physical dimensions and time bases are converted into standardized process parameter vectors with unified format, time alignment, and consistent dimensions. This process parameter vector serves as the standard input format for subsequent mapping relationship queries and time-series prediction model inferences, ensuring the uniformity of the data foundation throughout the control system and providing data input for the query and matching of process parameter-pressure demand mapping relationships in step S2.

[0029] The next step, S2, involves constructing a process parameter-pressure demand mapping database. This database, built within the data processing and analysis layer, stores the multidimensional mapping relationships between process parameters such as roller speed, brush hardness, adhesive viscosity, and ambient temperature, and the optimal pressure setpoint. This provides a data source for training subsequent time-series prediction models. This step is implemented through the following sub-steps.

[0030] Step S201: Database data structure organization. The process parameter-pressure demand mapping database is organized using a relational data structure. The main table in the process parameter-pressure demand mapping database records the association records of each process parameter combination and its corresponding pressure setpoint in historical production batches. Each record contains a feature vector of the process parameter combination and its corresponding pressure setpoint.

[0031] The feature vector of the process parameter combination consists of four dimensions: roller speed, brush body hardness, adhesive viscosity, and ambient temperature. The pressure setting value is the optimal pressure value verified in actual production under the conditions of this set of process parameters.

[0032] The main table structure includes the following fields: Record ID field, using an auto-incrementing integer type as the primary key; Roller speed field, using a floating-point type to store the speed value; Brush body hardness field, using a floating-point type to store the hardness value; Adhesive viscosity field, using a floating-point type to store the viscosity value; Ambient temperature field, using a floating-point type to store the temperature value; Pressure setpoint field, using a floating-point type to store the optimal pressure value corresponding to this set of process parameters; and Production timestamp field, using a time type to record the generation time of each data entry, used to support data organization and querying in chronological order.

[0033] Step S202: Hash Index Fast Retrieval Mechanism. The process parameter-pressure demand mapping database uses a hash index-based fast query mechanism to quickly retrieve historical mapping records. The hash index uses the feature vector of process parameter combinations as the key. The key is constructed by concatenating floating-point numbers from four dimensions—drum speed, brush hardness, adhesive viscosity, and ambient temperature—in a fixed order into a string. The concatenated string is then hashed using the MD5 algorithm to generate a 128-bit hash value as the index key.

[0034] The process parameter-pressure demand mapping database uses a hash function to map index keys to determine the physical location of records in the storage medium, thus completing query operations in constant time complexity. The hash index data structure uses chaining to handle hash collisions. When multiple different combinations of process parameters are mapped to the same storage location through the hash function, a linked list is maintained at that storage location. The linked list stores pointers to all records mapped to that location. During a query, the storage location is first located, and then the linked list is traversed to match the precise feature vector.

[0035] Step S203: Multi-level index structure construction. A multi-level index structure is constructed for the process parameter-pressure demand mapping database to support range queries and fuzzy queries. The multi-level index structure uses a B+ tree as the underlying storage structure. Each node of the B+ tree index stores a key value and a pointer to the data record.

[0036] The multi-level index structure is constructed as follows: the first-level index is partitioned based on the roller speed value, dividing the roller speed into multiple intervals according to the value range, with each interval corresponding to a partition; the second-level index is sorted within each partition of the first-level index based on the brush body hardness value; the third-level index is sorted based on the adhesive viscosity value based on the second-level index; and the fourth-level index is sorted based on the ambient temperature value based on the third-level index.

[0037] The multi-level index structure enables the process parameter-pressure requirement mapping database to efficiently handle multi-dimensional condition combination queries. For example, when the query conditions include both the roller speed range and the brush body hardness range, the process parameter-pressure requirement mapping database first locates the partition corresponding to the speed through the first-level index, and then locates the record range corresponding to the hardness within the partition through the second-level index.

[0038] Step S204: Online update mechanism. The process parameter-pressure demand mapping database has online update capabilities to adapt to the dynamic adjustment needs of the production process. The online update mechanism supports incremental writing of newly generated process parameter-pressure pairing data.

[0039] The specific process for incremental writes is as follows: When new process parameter-pressure pairing data is generated, the data is first inserted as a new record into the main table; then, a hash index key is calculated based on the process parameter combination feature vector of the new record, and the physical address of the record is inserted into the hash bucket corresponding to the hash index; next, the record is inserted into the B+ tree partition corresponding to the first-level index according to the drum speed value, and then the second-level, third-level, and fourth-level B+ tree indexes are updated sequentially. The above index update process is executed under the protection of database transactions to ensure consistency between the main table records and each index.

[0040] During the update process, a transaction processing mechanism is used to ensure data consistency and integrity. The transaction processing mechanism is implemented as follows: when an incremental write operation begins, the process parameter-pressure demand mapping database starts a transaction, inserts a new record into the main table, and then updates the hash index, the first-level B+ tree index, the second-level B+ tree index, the third-level B+ tree index, and the fourth-level B+ tree index in sequence. After all operations are completed successfully, the transaction is committed, and the update results are persisted to the disk storage medium.

[0041] The transaction processing mechanism comprises four aspects: atomicity guarantee, consistency constraint, isolation control, and durability guarantee. Atomicity guarantee is achieved through transaction log recording; when a failure occurs during the update process, the transaction log is used to perform a rollback operation to restore the state before the update. Consistency constraint is achieved through primary key uniqueness constraint and data type constraint. Isolation control uses the read committed isolation level to prevent interference between concurrent update operations. Durability guarantee is achieved by synchronously writing the transaction log to non-volatile storage media.

[0042] Step S205: Construction and partitioning of the training dataset. Historical mapping records in the process parameter-pressure demand mapping relationship database are used simultaneously for training and parameter tuning of the time series prediction model. The historical mapping records are organized in chronological order to form the training dataset.

[0043] The training dataset is constructed as follows: historical mapping records are extracted from the process parameter-pressure demand mapping database in ascending order of the production timestamp field. Each historical mapping record contains a feature vector of process parameter combinations at the sampling time and the corresponding pressure setpoint. All extracted historical mapping records are arranged in the extraction order to form the original time series dataset.

[0044] The original time series dataset is segmented according to a fixed time window length to generate a training sample set for the time series prediction model. Specifically, the time window length is set to T, where T is a positive integer, representing that each training sample contains a feature vector of process parameter combinations from T consecutive sampling times. For an original time series dataset of length N, from the 1st sampling time to the NTth sampling time, the feature vectors of process parameter combinations from T consecutive sampling times are extracted as the input feature sequence for the training sample. The pressure setpoint at the (T+1)th sampling time is used as the target label for that training sample.

[0045] A total of NT training samples are generated using the sliding window method described above. In the training sample set constructed using the above segmentation method, the input feature sequence dimension of each training sample is T multiplied by 4, where 4 corresponds to the four process parameter dimensions: roller speed, brush body hardness, adhesive viscosity, and ambient temperature. The output target label is a scalar pressure value.

[0046] The training dataset is divided into training and validation sets in a 7:3 ratio. The partitioning method is as follows: the first 70% of the samples in the chronologically ordered training set are assigned to the training set, and the remaining 30% are assigned to the validation set. This ensures that both the training and validation sets maintain the chronological order of the time series data, preventing future information from being leaked into the past. The training set is used for learning and optimizing the parameters of the time series prediction model, while the validation set is used for evaluating the performance of the time series prediction model and tuning hyperparameters.

[0047] Regarding step S3 in the method of this embodiment, the time-series prediction and feedforward-feedback weighted fusion control is a core component of the predictive control decision layer in the overall technical architecture. It utilizes a time-series prediction model to predict future demand pressure and generates composite control commands through feedforward-feedback weighted fusion. This step is specifically implemented through the following sub-steps.

[0048] Step S301: Select the architecture of the time series prediction model. The time series prediction model is constructed using a long short-term memory network or a Transformer architecture.

[0049] Long Short-Term Memory (LSTM) networks consist of three core gating structures: an input gate, a forget gate, and an output gate. The input gate controls the extent to which current input information enters the cell state. The forget gate controls the extent to which the cell state information from the previous time step is forgotten. The output gate controls the extent to which the current cell state information is output to the hidden layer.

[0050] Cellular states in long short-term memory networks maintain long-term memory capabilities over time. The transmission and modification of information in cellular states are finely controlled through gating mechanisms, enabling the capture of long-range dependencies in historical process parameter sequences.

[0051] The Transformer architecture employs a multi-head self-attention mechanism to model the relationships between different positions in the input sequence. This mechanism maps the input sequence to multiple distinct attention subspaces, each learning a different attention weight pattern. The attention outputs from these subspaces are then fused through a linear transformation to obtain the final attention representation. This multi-head self-attention mechanism captures the correlations between different process parameters, identifying the parameter combinations that have the greatest impact on pressure requirements.

[0052] Step S302: Configure the input and output of the time series prediction model. The input of the time series prediction model is a sequence of process parameters from several past sampling periods. The length of the input window received by the time series prediction model is set according to the variation period of the process parameters and the prediction accuracy requirements. The historical data period covered by the input window is two to three times that of the prediction time domain, ensuring that the time series prediction model can learn sufficient contextual information.

[0053] The output of the time-series predictive model is the predicted pressure demand for one or more future control cycles. The length of the output time domain is set according to the response speed of the actuator and the dynamic characteristics of the system. The output time domain covers the pure time delay of the actuator, achieving effective predictive compensation.

[0054] The time-series prediction model receives the standardized process parameter sequence generated in step S1 as input, and outputs the predicted pressure value after inference and calculation by the time-series prediction model. The predicted pressure value reflects the target pressure value required in the future period under the current trend of process parameter changes.

[0055] Step S303, training of the time series prediction model. The training of the time series prediction model adopts the rolling prediction method, using historically accumulated process parameter-pressure paired data as training samples. The training data comes from the training sample set constructed in step S2.

[0056] Mean squared error (MSE) is used as the loss function to measure the deviation between the predicted values ​​and the actual observed values ​​of the time series forecasting model. The expression for the MSE loss function is:

[0057] in, This represents the value of the loss function. Indicates the number of training samples. Indicates the index number of the training sample. Indicates the first The actual stress observation value of each training sample is the target label. This indicates that the time series prediction model is for the first... The predicted stress value for each training sample.

[0058] The gradient descent algorithm is used to optimize network weights. The gradient descent algorithm is implemented using an adaptive moment estimator (IME). The IEM automatically adjusts the learning rate parameter based on the historical gradient information of each parameter, enabling the time series prediction model to converge quickly in the early stage of training and stably converge in the later stage of training.

[0059] The time series prediction model undergoes incremental training at preset intervals to adapt to changes in production processes and the addition of new operating conditions. During incremental training, newly acquired data samples are used to fine-tune and update the parameters of the time series prediction model, while retaining the general feature representation capabilities learned by the time series prediction model from historical data.

[0060] Incremental training is divided into a parameter fine-tuning phase and a catastrophic forgetting prevention phase. In the parameter fine-tuning phase, newly acquired data samples are used to update the parameters of the fully connected layers of the time-series prediction model. Since the fully connected layer parameters have a small dimensionality, the update computation is manageable and can be performed on edge computing nodes. The catastrophic forgetting prevention phase employs a regularization strategy to limit the deviation of the time-series prediction model parameters from their original values. The expression for the regularization loss function is:

[0061] in, This represents the total loss value after adding the regularization term. This represents the mean squared error loss function value. Represents the regularization coefficient, which is used in specific implementations. The value is set to a range of 0.001 to 0.1, and is selected based on performance on the validation set. It is determined when the validation set loss no longer decreases over multiple consecutive training epochs. The value of ; This represents the parameter vector of the time-series prediction model during the current training process. This represents the time-series prediction model parameter vector saved at the start of incremental training. This represents the square of the Euclidean distance between the current time series prediction model parameter vector and the time series prediction model parameter vector saved at the start of incremental training.

[0062] The trained time-series prediction model is deployed on edge computing nodes, and the inference latency is controlled within one-tenth of the control cycle, meeting the real-time requirements.

[0063] Step S304: Real-time pressure feedback signal acquisition. The feedforward-feedback weighted fusion control module weights and fuses the predicted pressure value output by the time-series prediction model with the feedback pressure value acquired by the real-time pressure sensor to generate a composite control command.

[0064] A real-time pressure sensor is installed at the outlet position of the contact area between the roller and the brush body. A high-precision pressure transmitter is used to measure the actual pressure. The real-time pressure sensor outputs a standard current signal or digital signal proportional to the pressure. The feedback pressure value collected by the real-time pressure sensor is amplified, filtered, and converted from analog to digital by the signal conditioning circuit before being input to the feedforward-feedback weighted fusion control module.

[0065] Step S305, Adaptive weighted fusion control strategy: The feedforward-feedback weighted fusion control module dynamically calculates the values ​​of feedforward weight and feedback weight according to the current operating conditions, and the calculation of fusion weight adopts an adaptive adjustment strategy.

[0066] The value of the feedforward weight depends on the prediction accuracy assessment results of the time-series predictive model and the degree of fluctuation of the process parameters. The prediction accuracy assessment results are obtained by statistically comparing the deviations between the historical predictions of the time-series predictive model and the actual observed values. When the prediction accuracy is high, the feedforward weight is increased to leverage the look-ahead advantage of predictive control. When the prediction accuracy is low, the feedforward weight is decreased to reduce the adverse effects of prediction errors on the control effect.

[0067] The degree of fluctuation in process parameters is obtained by calculating the variance or rate of change of the process parameter series. When the process parameters are in a steady state, the feedforward weight is increased to maintain system stability using predictive control. When the process parameters undergo abrupt changes, the feedforward weight is decreased to prevent predictions based on historical trends from becoming invalid.

[0068] The value of the feedback weight depends on the magnitude of the real-time pressure deviation. The real-time pressure deviation is the difference between the pressure setpoint and the feedback pressure value. When the pressure deviation is large, the feedback weight is increased to quickly eliminate the deviation. When the pressure deviation is small, the feedback weight is decreased to maintain system stability.

[0069] The formula for calculating the composite control target value is:

[0070] in, Indicates the composite control target value. Indicates the feedforward weights. This represents the predicted stress value output by the time-series forecasting model. Indicates the feedback weight. This represents the feedback pressure value acquired by the real-time pressure sensor. Feedforward weights. With feedback weights The sum of them equals 1.

[0071] Step S306: Adaptive weighting coefficient adjustment based on fuzzy inference. The adaptive adjustment strategy of the weighting coefficients in the feedforward-feedback weighted fusion control module is implemented based on fuzzy inference rules.

[0072] The input variables for the fuzzy inference rule include prediction error, pressure change rate, and process parameter volatility. Prediction error is the difference between the predicted value and the actual observed value from the time-series prediction model. The pressure change rate is the magnitude of pressure change per unit time, expressed in kPa / s. Process parameter volatility is the statistical dispersion of the process parameter series. The output variables of the fuzzy inference rule are the feedforward weight correction and the feedback weight correction, where the correction is the incremental adjustment based on the current weights.

[0073] The fuzzy inference system is implemented using the Mamdani fuzzy inference system. The fuzzification module of the Mamdani fuzzy inference system converts precise input variable values ​​into fuzzy set membership degrees. A Gaussian membership function is used. The membership function parameters for each fuzzy subset of the input variables are set as follows: The membership function of the prediction error adopts a Gaussian form, with mean values ​​of -0.5, -0.2, 0, 0.2, and 0.5, and variance of 0.1, corresponding to five fuzzy subsets: very low, low, medium, high, and very high.

[0074] The membership function of the pressure change rate is Gaussian, with means of -2, -1, 0, 1, and 2, and variances of 0.5, corresponding to five fuzzy subsets: very low, low, medium, high, and very high.

[0075] The fluctuation of process parameters takes values ​​in the range [0, 1]. The membership function adopts a triangle, with the vertex position of the low level being (0, 0, 0.3), the medium level being (0.2, 0.5, 0.8), and the high level being (0.7, 1, 1).

[0076] The fuzzified membership vectors are input into a fuzzy rule base for fuzzy inference. Each rule in the rule base consists of an antecedent and a consequent. The antecedent defines the fuzzy conditions of the input variables, and the consequent defines the fuzzy values ​​of the output variables. The fuzzy inference process uses a maximum-minimum composition operation. The activation strength of a rule is determined by the minimum membership value of the antecedent, and the output fuzzy set of the rule is obtained by multiplying the membership function of the consequent by the activation strength.

[0077] The fuzzy inference rule base contains 125 rules, covering all combinations of fuzzy subsets of input variables. The rules are summarized as follows: When the prediction error is large and negative, it indicates that the predicted value is lower than the actual value. In this case, the feedforward weight needs to be increased to enhance the proportion of the prediction effect, while the feedback weight needs to be decreased. When the prediction error is large and positive, it indicates that the predicted value is higher than the actual value. In this case, the feedforward weight needs to be decreased to reduce the impact of the prediction error. When the pressure change rate is large and positive, it indicates that the pressure is in a rapid upward trend. In this case, the feedback weight needs to be increased to suppress the pressure increase. When the pressure change rate is large and negative, it indicates that the pressure is in a rapid downward trend. In this case, the feedback weight also needs to be increased. When the process parameter fluctuation is high, the reliability of the prediction result decreases. In this case, the feedforward weight decreases and the feedback weight increases. When the process parameter fluctuation is low, the reliability of the prediction result is high. In this case, the feedforward weight increases and the feedback weight decreases.

[0078] The above rules are superimposed under different combinations of input variables to jointly determine the output values ​​of the feedforward weight correction and the feedback weight correction. For example, when the prediction error is large negative, the pressure change rate is large positive, and the process parameter fluctuation is low, the feedforward weight correction is large positive, and the feedback weight correction is large negative; when the prediction error is large positive, the pressure change rate is large negative, and the process parameter fluctuation is high, the feedforward weight correction is large negative, and the feedback weight correction is large positive; when the prediction error is zero, the pressure change rate is zero, and the process parameter fluctuation is medium, the feedforward weight correction is zero, and the feedback weight correction is zero.

[0079] The output obtained from fuzzy inference is a fuzzy correction value, which is converted into a precise control parameter correction value by the defuzzification module. Defuzzification uses the centroid method, calculating the center of the area under the membership curve of the output fuzzy set as the precise output value. The defuzzified correction value is added to the current weight value to obtain the updated fusion weight. Upper and lower bound constraints are set during the weight update process, and feedforward weights are used. and feedback weights All values ​​are constrained within the interval [0.1, 0.9].

[0080] The time-series prediction model takes historical process parameter sequences as input and outputs predicted pressure demand values ​​for future moments after training. The feedforward-feedback weighted fusion control module dynamically weights and fuses the predicted values ​​with real-time feedback values ​​through fuzzy inference to generate a composite control command. This command is output to the fast-response pressure actuator in step S4, driving it to complete the pressure regulation action.

[0081] Finally, in step S4, the drive layer control command conversion and pressure regulation are performed. Step S4 receives the composite control command generated in step S3, converts the composite control command into a drive signal for the actuator, and completes the final pressure regulation action. Step S4 is implemented through the following sub-steps.

[0082] Step S401, Actuator selection and working principle: The fast-response pressure actuator uses a proportional valve or servo drive mechanism to control the pressure regulating valve.

[0083] The proportional valve employs an electro-hydraulic proportional control principle, controlling the valve spool opening via an input current signal to regulate the flow rate of hydraulic oil or pneumatic media, thereby achieving continuous pressure regulation. The proportional valve's response time is controlled within milliseconds, meeting the time-cycle requirements of predictive regulation. The input current signal range for the proportional valve is 4mA to 20mA, corresponding to a linear change in valve spool opening from 0% to 100%.

[0084] The servo actuator uses a servo motor to drive a ball screw or crank-connecting rod mechanism to achieve position control of the pressure regulating valve. The servo actuator employs a closed-loop position control mode, achieving position output through encoder feedback. The position output resolution reaches the micrometer level, meeting the accuracy requirements of pressure control. The rated torque of the servo motor and the lead of the ball screw are selected according to actual load requirements. The encoder uses incremental or absolute encoders, with 2500 to 10000 pulses per revolution.

[0085] Step S402, control command conversion and transmission: After receiving the composite control command, the actuator converts it into the corresponding valve opening command or servo motor target position command. The composite control command is represented by the required pressure target value.

[0086] The control conversion module inside the actuator calculates the required valve opening or motor position based on the target pressure value and the current pressure-flow characteristic curve. The pressure-flow characteristic curve is obtained through experimental calibration. Specifically, with the inlet pressure of the pressure regulating valve kept constant, the outlet pressure values ​​corresponding to different valve openings are recorded, and a valve opening-outlet pressure curve is plotted. A one-to-one correspondence between the target pressure value and the valve opening is established based on the valve opening-outlet pressure curve.

[0087] The control conversion module uses either lookup table interpolation or analytical calculation to map pressure targets to valve openings or motor positions. The lookup table interpolation method pre-establishes a correspondence between pressure target values ​​and valve openings, calculating the valve opening corresponding to the current pressure target value through linear interpolation. The analytical calculation method uses polynomial fitting to the pressure-flow characteristic curve, substituting the pressure target value into the polynomial to directly calculate the valve opening or motor position.

[0088] The calculated valve opening command or servo motor target position command is sent to the drive controller at the execution layer via a high-speed fieldbus. The high-speed fieldbus is implemented using industrial Ethernet or EtherCAT protocols, with communication cycles controlled in the sub-millisecond range to ensure timely transmission of control commands.

[0089] After receiving the valve opening command or the servo motor target position command, the drive controller drives the proportional solenoid of the proportional valve or the servo driver of the servo motor to complete the valve opening adjustment or the motor position adjustment, thereby controlling the pressure of the roller brush contact area.

[0090] Step S403: The pressure setpoint and predicted values ​​are complementaryly integrated. There is a complementary relationship between the predicted pressure value output by the time-series predictive model and the initial pressure demand reference value obtained from the process parameter-pressure demand mapping database. The initial pressure demand reference value reflects the optimal pressure setting under steady-state conditions for the current process parameter combination, providing a benchmark reference for predictive control. The predicted pressure value reflects the dynamic pressure demand in future periods under the changing trends of process parameters, providing forward-looking adjustments for predictive control. The integrated use of these two pressure values ​​enables the control system to possess both steady-state accuracy and dynamic response capability.

[0091] Step S4 responds to the composite control command of step S3 through a proportional valve or servo actuator, converting the pressure target value into a drive signal for valve opening or motor position, which is then transmitted to the drive controller via a high-speed fieldbus to complete the pressure regulation action.

[0092] To facilitate understanding, the method of this embodiment is described below in conjunction with a specific application scenario. The roller-applied brush production line includes a brush loading station, an adhesive coating station, a roller pressing station, and a curing and drying station. In the roller pressing station, the contact pressure between the brush and the roller is a key parameter affecting the uniformity of adhesive coating and the bonding strength. Excessive pressure can cause the adhesive to be squeezed out and the brush to deform, while insufficient pressure can lead to weak adhesion and incomplete bonding. The method of this embodiment, through the prediction and adaptive adjustment of the pressure at the roller pressing station, can alleviate the response lag problem of traditional feedback adjustment methods and improve the stability and consistency of adhesive coating quality.

[0093] During actual operation, the multi-source process parameter acquisition module continuously collects process parameter data such as roller speed, brush material hardness, adhesive viscosity, and ambient temperature. The data acquisition frequency is dynamically adjusted according to the production line's operating status. When a change in process parameters is detected, the time-series prediction model performs predictive inference based on the latest process parameter sequence, calculating the predicted pressure demand value for several future control cycles. The feedforward-feedback weighted fusion control module calculates the adaptive weighting coefficients of the feedforward and feedback weights based on the prediction accuracy evaluation results, real-time pressure deviation, and process parameter fluctuations. It then weights and fuses the predicted pressure value with the feedback pressure value to generate a composite control command. This composite control command is sent to the fast-response pressure actuator via a high-speed fieldbus, driving a proportional valve or servo actuator to pre-adjust the pressure. This pre-adjustment action is completed before the impact of process parameter changes reaches the pressure sensor, suppressing pressure fluctuations and transforming pressure regulation from passive feedback to active prediction.

[0094] In summary, this embodiment transforms pressure regulation from passive feedback to active prediction by introducing a process parameter prediction mechanism. Upstream process parameters such as roller speed, brush material hardness, adhesive viscosity, and ambient temperature are used as predictive inputs. A time-series prediction model is used to establish a mapping relationship between process parameters and pressure demand, allowing for early prediction of pressure change trends and proactive compensation. The feedforward-feedback weighted fusion control architecture combines the advantages of predictive and feedback control. The fusion weights are adaptively adjusted according to operating conditions, using prediction to suppress pressure fluctuations in advance and ensuring regulation accuracy through feedback. The configuration of a fast-response pressure actuator ensures timely execution of predictive regulation commands. The entire control system forms a complete prediction-decision-execution closed loop, improving the stability of the roller adhesive brush production process and the consistency of product quality.

[0095] On the other hand, the adaptive pressure regulation system for roller adhesive brush production based on multi-sensor fusion disclosed in this invention includes: The multi-source process parameter acquisition module synchronously acquires multi-source process parameters during the production of roller adhesive brushes through a distributed sensor network. It performs preprocessing operations such as filtering, noise reduction, and normalization on the acquired raw data, and finally generates a standardized process parameter vector with unified format and time sequence alignment. The process parameter-pressure demand mapping database stores the multi-dimensional mapping relationship between process parameters such as roller speed, brush body hardness, adhesive viscosity, and ambient temperature and the optimal pressure setpoint. The historical mapping records in the process parameter-pressure demand mapping database are also used for training and parameter tuning of time series prediction models. The time-series forecasting model takes a standardized process parameter vector as input, predicts the pressure demand at future moments based on historical process parameter sequences, and outputs the predicted pressure value. The feedforward-feedback weighted fusion control module weights and fuses the predicted pressure value output by the time series prediction model with the feedback pressure value collected by the real-time pressure sensor to generate a composite control command. The feedforward weight and feedback weight used in the weighted fusion are dynamically adjusted according to the current working conditions based on fuzzy inference rules. The fast-response pressure actuator receives a composite control command, converts it into a drive signal, and performs pressure regulation.

[0096] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0097] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for adaptive pressure adjustment in roller adhesive brush production based on multi-sensor fusion, characterized in that, include: Step S1: Multi-source process parameters during the production of roller adhesive brush are synchronously collected through a distributed sensor network. The collected raw data is preprocessed to generate a time-aligned standardized process parameter sequence. Step S2: Construct a process parameter-pressure demand mapping relationship database. The process parameter-pressure demand mapping relationship database stores mapping records between multiple sets of process parameter combinations and their corresponding pressure setpoints. The historical mapping records in the process parameter-pressure demand mapping relationship database are used for training and parameter tuning of the time series prediction model. Step S3: Input the standardized process parameter sequence into the time-series prediction model. The time-series prediction model predicts the pressure demand at future moments based on the historical process parameter sequence and outputs the predicted pressure value. The predicted pressure value output by the time-series prediction model is weighted and fused with the feedback pressure value collected by the real-time pressure sensor to generate a composite control command. The feedforward weight and feedback weight used in the weighted fusion are dynamically adjusted according to the current working condition based on fuzzy inference rules. Step S4: The composite control command is converted into a drive signal for the fast-response pressure actuator, which then performs the pressure regulation action.

2. The adaptive pressure adjustment method for roller adhesive brush production based on multi-sensor fusion according to claim 1, characterized in that, The synchronous acquisition of multi-source process parameters in step S1 specifically includes: The sensor network includes a roller speed sensor, a brush body material hardness detection device, an adhesive viscosity meter, and an ambient temperature sensor, which are used to collect four process parameters: roller speed, brush body material hardness, adhesive viscosity, and ambient temperature, respectively. Each sensor employs a synchronous sampling mechanism, triggering data acquisition operations of all sensors through a unified clock reference signal, ensuring that the acquired data on roller speed, brush material hardness, adhesive viscosity, and ambient temperature remain strictly aligned in the time dimension.

3. The adaptive pressure adjustment method for roller adhesive brush production based on multi-sensor fusion according to claim 1, characterized in that, In step S1, the sampling frequency is dynamically adjusted according to the operating speed of the production line: the rate of change of the roller speed signal is monitored in real time, and when the rate of change exceeds a preset threshold, the sampling frequency is increased to the highest level; when the rate of change is lower than the preset threshold, the sampling frequency is adjusted to the normal level.

4. The adaptive pressure adjustment method for roller adhesive brush production based on multi-sensor fusion according to claim 1, characterized in that, The process parameter-pressure requirement mapping database has an online update mechanism: When new process parameter-pressure pairing data is generated, the data is inserted into the main table as a new record; the hash index key is calculated based on the process parameter combination feature vector of the new record, and the physical address of the record is inserted into the hash bucket corresponding to the hash index; the record is inserted into the partition corresponding to the first-level B+ tree index according to the drum speed value, and the second-level B+ tree index, the third-level B+ tree index, and the fourth-level B+ tree index are updated in sequence. The online update mechanism adopts a transaction processing mechanism, which includes atomicity guarantee, consistency constraint, isolation control and durability guarantee. The atomicity guarantee is achieved through transaction log recording. When a failure occurs during the update process, the transaction log is used to perform a rollback operation to restore the state before the update.

5. The adaptive pressure adjustment method for roller adhesive brush production based on multi-sensor fusion according to claim 1, characterized in that, The process parameter-pressure requirement mapping database is organized using a relational data structure. The main table records in the process parameter-pressure requirement mapping database contain feature vectors of process parameter combinations and corresponding pressure setpoints. The feature vectors consist of values ​​in four dimensions: roller speed, brush hardness, adhesive viscosity, and ambient temperature. The process parameter-pressure requirement mapping database uses a hash index to achieve fast retrieval. The hash index uses the feature vector of the process parameter combination as the key value, and uses a hash function to map the index key to the physical location recorded in the storage medium. The process parameter-pressure requirement mapping database also constructs a multi-level B+ tree index structure to support range queries and fuzzy queries. The first-level index is partitioned based on the roller speed value. The second-level index is sorted within each partition of the first-level index based on the brush body hardness value. The third-level index is sorted based on the adhesive viscosity value based on the second-level index. The fourth-level index is sorted based on the ambient temperature value based on the third-level index.

6. The adaptive pressure adjustment method for roller adhesive brush production based on multi-sensor fusion according to claim 1, characterized in that, The time series prediction model uses historical mapping records extracted from the process parameter-pressure demand mapping relationship database to construct a training sample set. The original time series dataset is divided into samples according to a fixed time window length. Each training sample contains a process parameter combination feature vector of T consecutive sampling times as the input feature sequence of the training sample. The pressure set value at the (T+1)th sampling time is used as the target label of the training sample. The training sample set is divided into a training set and a validation set according to time order. The first 70% of the samples in the training sample set arranged in time order are assigned to the training set, and the remaining 30% of the samples are assigned to the validation set. The training set is used for learning and optimizing the parameters of the time series prediction model, and the validation set is used for evaluating the performance of the time series prediction model and adjusting the hyperparameters.

7. The adaptive pressure adjustment method for roller adhesive brush production based on multi-sensor fusion according to claim 1, characterized in that, The time series prediction model is incrementally trained every preset period. The incremental training includes a parameter fine-tuning stage and a catastrophic forgetting prevention stage. In the parameter fine-tuning stage, the fully connected layer parameters of the time series prediction model are updated using newly collected data samples. In the catastrophic forgetting prevention stage, a regularization strategy is used to limit the degree to which the parameters of the time series prediction model deviate from their original values.

8. The adaptive pressure adjustment method for roller adhesive brush production based on multi-sensor fusion according to claim 1, characterized in that, The feedforward and feedback weights used in the weighted fusion are dynamically adjusted based on fuzzy inference rules according to the current operating conditions, specifically as follows: The input variables of the fuzzy inference rule include prediction error, pressure change rate, and process parameter fluctuation. The prediction error is the difference between the predicted value of the time series prediction model and the actual observed value. The pressure change rate is the magnitude of the change in pressure value per unit time. The process parameter fluctuation is the statistical dispersion of the process parameter sequence. The output variables of the fuzzy inference rule are feedforward weight correction and feedback weight correction. The membership function of the prediction error adopts a Gaussian membership function, with mean values ​​of -0.5, -0.2, 0, 0.2, and 0.5, and variances of 0.1, corresponding to five fuzzy subsets: very low, low, medium, high, and very high. The membership function of the pressure change rate adopts a Gaussian membership function, with mean values ​​of -2, -1, 0, 1, and 2, and variances of 0.5, corresponding to five fuzzy subsets: very low, low, medium, high, and very high. The fluctuation of the process parameter takes values ​​in the interval [0, 1], and the membership function of the fluctuation of the process parameter adopts the triangular membership function. The fuzzy correction value obtained by fuzzy inference is converted into an accurate weight correction value by the defuzzification module. The defuzzification adopts the centroid method and calculates the center of the area under the membership curve of the output fuzzy set as the accurate output value. The weight correction value after defuzzification is added to the current fusion weight value to obtain the updated fusion weight. Both the feedforward weight and the feedback weight are constrained within the interval [0.1, 0.9].

9. The adaptive pressure adjustment method for roller adhesive brush production based on multi-sensor fusion according to claim 1, characterized in that, The predicted pressure value output by the time-series prediction model is complemented and fused with the initial pressure demand reference value obtained by querying the process parameter-pressure demand mapping database. The initial pressure demand reference value reflects the pressure setting of the current process parameter combination under steady-state conditions, providing a benchmark reference for predictive control. The predicted pressure value reflects the dynamic pressure demand in the future period under the process parameter change trend.

10. An adaptive pressure regulation system for roller adhesive brush production based on multi-sensor fusion, characterized in that, The system comprising the method according to any one of claims 1 to 9, wherein the system includes: The multi-source process parameter acquisition module synchronously acquires multi-source process parameters during the production of roller adhesive brushes through a distributed sensor network. It performs preprocessing operations such as filtering, noise reduction, and normalization on the acquired raw data, and finally generates a standardized process parameter vector with unified format and time sequence alignment. The process parameter-pressure demand mapping database stores multidimensional mapping relationships between process parameters such as roller speed, brush body hardness, adhesive viscosity, and ambient temperature and optimal pressure setpoints. The historical mapping records in the process parameter-pressure demand mapping database are also used for training and parameter tuning of time series prediction models. The time-series prediction model takes the standardized process parameter vector as input, predicts the pressure demand at future moments based on the historical process parameter sequence, and outputs the predicted pressure value. The feedforward-feedback weighted fusion control module weights and fuses the predicted pressure value output by the time-series prediction model with the feedback pressure value collected by the real-time pressure sensor to generate a composite control command. The feedforward weight and feedback weight used in the weighted fusion are dynamically adjusted according to the current working conditions based on fuzzy inference rules. The fast-response pressure actuator receives the composite control command, converts the composite control command into a drive signal for the fast-response pressure actuator, and performs a pressure regulation action.

Images