A multi-parameter perception spraying process production line control method and system

Abstract

The application discloses a multi-parameter sensing spraying process production line control method and system, and relates to the technical field of industrial automation control. The method comprises the following steps: deploying multiple types of sensor groups to synchronously collect coating temperature, pressure, flow, atomizing air pressure and flow, spraying gun pose, environmental temperature and humidity, and workpiece surface temperature; performing dimension normalization and covariance matrix operation on the collected data to generate a multi-parameter dynamic coupling state vector; using a long short-term memory network to construct a process parameter coupling dynamics model, predicting coating quality and outputting multi-process parameter adjustment amounts; constructing a coupling compensation matrix to cross-couple and compensate the adjustment amounts, generating control signals to perform collaborative control; and detecting coating quality in real time and comparing it with the predicted value, and performing online incremental updating of the model when the error exceeds a threshold. The application realizes multi-parameter collaborative sensing and dynamic compensation, and improves coating quality consistency and long-term stability of the production line.

Description

Technical Field

[0001] This invention relates to the field of industrial automation control technology, and in particular to a multi-parameter sensing control method and system for a spraying process production line. Background Technology

[0002] Spray coating is a widely used surface treatment technology in manufacturing. The final product quality—namely, the uniformity of coating thickness, surface gloss, adhesion strength, and defect density on the workpiece surface—is influenced by the coupling of multiple physical field parameters during the spraying process. These parameters include, but are not limited to: the spray gun outlet pressure, paint volumetric flow rate, atomizing air pressure, the spatial orientation parameters of the spray gun relative to the workpiece surface (including spraying distance and spraying angle), the moving speed of the spray gun along the preset trajectory, the real-time temperature and apparent viscosity of the paint in the delivery pipeline, the ambient temperature and relative humidity at the spraying site, and the initial temperature of the workpiece surface before spraying.

[0003] In existing spray coating production line control technologies, open-loop control strategies or closed-loop control strategies based on single-parameter feedback are commonly used. In open-loop control strategies, all process parameters are set to fixed values ​​by operators based on experience before the spraying operation begins and remain unchanged throughout the entire spraying cycle. This strategy cannot cope with process disturbances caused by factors such as paint characteristic drift, environmental condition fluctuations, and equipment status changes during the spraying process, resulting in poor coating quality consistency between different workpieces in the same batch. Although closed-loop control strategies based on single-parameter feedback achieve real-time monitoring and closed-loop adjustment of a specific parameter—such as spray gun outlet pressure or paint temperature—spraying quality is essentially a function determined by the collaboration of multiple process parameters. Local optimization of a single parameter cannot guarantee that the overall process state is at its optimal operating point. On the contrary, adjusting one parameter may disrupt the original matching relationship between other parameters, leading to a deterioration in the overall spraying effect.

[0004] In recent years, a few technical solutions have attempted to introduce the concept of multi-sensor data fusion. For example, some solutions calculate spraying parameters by collecting real-time rheological property data of the coating and combining it with the local geometric features of the workpiece; others input multimodal sensor data during the spraying process into a pre-trained process diagnostic model, outputting a semantic diagnostic description of the current painting process status. However, these solutions share the following shortcomings: First, the acquisition and processing of multi-sensor data only remain at the level of parallel acquisition and independent processing. The physical coupling relationship between parameters—that is, how the change of one parameter affects the optimal value range of other parameters through physical mechanisms—is not described by a mathematical model, resulting in a lack of synergy between parameter adjustment decisions. Second, existing solutions typically use offline trained static models for process diagnosis or parameter optimization. Once deployed, the model is no longer updated and cannot adapt to the drift in process characteristics caused by coating batch changes, equipment wear, and seasonal environmental changes during long-term production line operation. Third, the feedback control of existing solutions is mostly post-diagnostic—that is, parameter adjustments are made only after spraying is completed or defects have occurred, lacking the ability to perform feedforward parameter pre-compensation for unsprayed areas based on real-time sensing data during the spraying process.

[0005] In summary, the core technical problem faced by existing spraying process production line control technology is the lack of a control method that can simultaneously perceive multiple process parameters in real time, establish a mathematical model of dynamic coupling relationship between parameters, and realize collaborative prediction and dynamic compensation of spraying parameters based on the model. This makes it difficult to meet the increasingly stringent industrial manufacturing requirements for batch consistency and long-term stability of spraying quality. Summary of the Invention

[0006] To achieve the above objectives, the present invention provides a multi-parameter sensing method for controlling a spraying process production line, comprising the following steps:

[0007] Step 1: Deploy a multi-type sensor group on the spraying actuator of the spraying process production line. The multi-type sensor group includes a paint temperature sensor, a paint pressure sensor, and a paint flow sensor installed at the inlet of the spray gun feed pipe; an atomizing air pressure sensor and an atomizing air flow sensor installed at the inlet of the spray gun atomizing air pipe; a spray gun posture sensor installed on the spray gun body; an ambient temperature sensor and an ambient relative humidity sensor installed in the spraying operation area; and a workpiece surface temperature sensor installed on the workpiece conveying device. Control all sensors in the multi-type sensor group to synchronously collect their corresponding physical quantities at a sampling frequency of 50 Hz to 200 Hz. Arrange all sensor readings obtained from each synchronous acquisition into a multi-dimensional data vector according to a preset data structure. Arrange multiple multi-dimensional data vectors obtained from multiple consecutive synchronous acquisitions in the order of acquisition time to form a multi-parameter synchronous acquisition time series dataset.

[0008] Step 2: Extract the multidimensional data vector at the current moment from the multi-parameter synchronous acquisition time series dataset. Perform dimensional normalization processing on each physical quantity value in the multidimensional data vector. The dimensional normalization processing is to subtract the moving average value of the physical quantity within a preset historical time window from the physical quantity value, and then divide it by the moving standard deviation of the physical quantity within the preset historical time window to obtain the normalized value of the physical quantity. Arrange the normalized values ​​corresponding to all physical quantities in the multidimensional data vector according to the preset data structure to form a normalized multidimensional data vector. Perform covariance matrix operation on the normalized multidimensional data vector to obtain the multi-parameter covariance matrix at the current moment. Perform tensor concatenation between the normalized multidimensional data vector and the multi-parameter covariance matrix to generate the multi-parameter dynamic coupling state vector at the current moment.

[0009] Step 3: Using the multi-parameter dynamic coupling state vector generated at multiple consecutive moments as the input sequence, and the process parameter adjustment amount of the spraying actuator within a preset time window after the current moment as the output variable, a process parameter coupling dynamic model is constructed. The process parameter coupling dynamic model adopts a long short-term memory network structure, which includes an input layer, two hidden layers, and an output layer. Each hidden layer contains 128 long short-term memory units. The multi-parameter dynamic coupling state vector generated in real time in Step 2 at the current moment is input into the trained process parameter coupling dynamic model. The process parameter coupling dynamic model outputs a sequence of predicted values ​​for the coating quality index within a preset prediction time window after the current moment. At the same time, it outputs the adjustment amount that needs to be applied to multiple process parameters of the spraying actuator to make the coating quality index in the predicted value sequence approach the preset target value. The multiple process parameters include paint pressure adjustment amount, paint flow rate adjustment amount, atomized air pressure adjustment amount, spray gun moving speed adjustment amount, and spray gun spatial pose adjustment amount.

[0010] Step 4: Construct a coupling compensation matrix. The coupling compensation matrix is ​​a lower triangular matrix with all diagonal elements equal to 1. The off-diagonal elements of the coupling compensation matrix represent the cross-coupling influence coefficients between different process parameter adjustment amounts. The cross-coupling influence coefficients are obtained by linear transformation of the hidden layer state vector in the process parameter coupling dynamics model. The paint pressure adjustment amount, paint flow rate adjustment amount, atomized air pressure adjustment amount, spray gun moving speed adjustment amount, and spray gun spatial pose adjustment amount output in Step 3 are combined into an adjustment amount vector. The adjustment amount vector is multiplied left by the coupling compensation matrix to obtain the collaborative adjustment amount vector after cross-coupling compensation. Each element in the collaborative adjustment amount vector is converted into the spray gun feed pipeline regulating valve opening control signal, the spray gun atomized air pipeline regulating valve opening control signal, the spraying actuator drive motor speed control signal, and the spraying actuator joint driver pose control signal according to the corresponding process parameter type. The above control signals are sent to the corresponding actuators to perform collaborative control.

[0011] Step 5: Real-time measured coating quality data of the sprayed area is collected by an online coating quality detection sensor deployed behind the spraying actuator. The measured coating quality data includes measured values ​​of coating thickness, coating surface gloss, and the number of coating surface defects. The measured coating quality data is compared with the predicted coating quality value of the process parameter coupled dynamics model in Step 3 at the same time, and the prediction error is calculated. When the absolute value of the prediction error exceeds a preset error threshold, the multi-parameter synchronously collected time-series data, the process parameter adjustment amount output by the process parameter coupled dynamics model, and the measured coating quality data at the current time are combined into a new training sample. Single-step incremental training is performed on the process parameter coupled dynamics model to update the network parameters of the process parameter coupled dynamics model. The updated process parameter coupled dynamics model is used for feedforward prediction in the subsequent Step 3.

[0012] Preferably, the preset data structure in step 1 is a fixed-length data format arranged in the order of paint temperature value, paint pressure value, paint flow rate value, atomizing air pressure value, atomizing air flow rate value, spray gun space three-dimensional coordinate value, spray gun space attitude angle value, ambient temperature value, ambient relative humidity value and workpiece surface temperature value.

[0013] The spray gun posture sensor includes three orthogonally arranged linear displacement sensors and three orthogonally arranged angular displacement sensors mounted on the spray gun body. The three orthogonally arranged linear displacement sensors are used to collect the X-axis, Y-axis and Z-axis coordinate values ​​of the spray gun nozzle relative to the workpiece surface in a three-dimensional spatial coordinate system. The three orthogonally arranged angular displacement sensors are used to collect the pitch angle, yaw angle and roll angle values ​​of the spray gun axis relative to the normal of the workpiece surface in a three-dimensional spatial coordinate system.

[0014] Preferably, the length of the preset historical time window in step 2 is 100 to 500 consecutive sampling times;

[0015] The moving average is calculated according to the following formula:

[0016]

[0017] in, Let N represent the moving average of the i-th physical quantity at time t, and let N represent the total number of sampling times included in the preset historical time window. This indicates that the i-th physical quantity is at time i. The collected values, Indicates the time interval between two adjacent samples;

[0018] The moving standard deviation is calculated according to the following formula:

[0019]

[0020] in, This represents the standard deviation of the movement of the i-th physical quantity at time t;

[0021] The normalized value is calculated according to the following formula:

[0022]

[0023] in, This represents the normalized value of the i-th physical quantity at time t.

[0024] Preferably, the covariance matrix operation in step 2 is performed according to the following formula:

[0025]

[0026] in, Let M represent the multi-parameter covariance matrix at time t, and M represent the number of normalized multidimensional data vectors involved in the covariance matrix operation, with values ​​ranging from 10 to 50. Indicates time The normalized multidimensional data vector, Indicates from time The mean vector of the M normalized multidimensional data vectors up to time t, where the superscript T indicates the vector transpose operation;

[0027] The tensor concatenation is performed as follows: the normalized multidimensional data vector is used as the first dimension component of the tensor, the multi-parameter covariance matrix is ​​expanded into a one-dimensional vector and used as the second dimension component of the tensor, and the first dimension component and the second dimension component are concatenated along the last dimension of the tensor to generate the multi-parameter dynamic coupling state vector.

[0028] Preferably, the training samples of the process parameter coupled dynamics model in step 3 are derived from the multi-parameter synchronous acquisition time series dataset recorded during historical spraying operations and the corresponding coating quality detection data. The coating quality detection data includes coating thickness uniformity index, coating surface gloss index and coating surface defect density index.

[0029] The coupled dynamics model of process parameters aims to minimize the mean square error between the predicted coating quality index and the actual coating quality index. It employs a backpropagation algorithm over time for parameter optimization training. The mean square error is calculated using the following formula:

[0030]

[0031] in, The mean square error is represented by T, and T represents the total number of prediction times included within the preset prediction time window. This indicates that the process parameter coupled dynamic model is at the prediction time. The output vector of predicted coating quality indicators. Indicates the time of prediction The corresponding actual coating quality index vector;

[0032] The length of the preset prediction time window is from 50 consecutive sampling times to 200 consecutive sampling times.

[0033] Preferably, the coating pressure adjustment, coating flow rate adjustment, atomized air pressure adjustment, spray gun moving speed adjustment, and spray gun spatial pose adjustment output by the process parameter coupled dynamics model in step 3 are all relative adjustment percentage values ​​relative to the actual values ​​of the corresponding process parameters at the current moment.

[0034] The spray gun spatial pose adjustment includes the percentage values ​​of the X-axis coordinate adjustment, Y-axis coordinate adjustment, Z-axis coordinate adjustment, pitch angle adjustment, yaw angle adjustment, and roll angle adjustment in the three-dimensional coordinate system relative to the workpiece surface.

[0035] Preferably, the cross-coupling influence coefficient in step 4 is obtained as follows: extract the hidden state vector of the last time step from the output state vector of the long short-term memory unit of the second hidden layer in the process parameter coupling dynamics model, input the hidden state vector into a fully connected linear layer, the output dimension of the fully connected linear layer is equal to the square of the dimension of the adjustment vector, rearrange the output of the fully connected linear layer into a square matrix, perform lower triangularization on the square matrix and set all diagonal elements to 1 to obtain the coupling compensation matrix;

[0036] The coupling compensation matrix has a dimension of 5×5. The off-diagonal element in the i-th row and j-th column of the coupling compensation matrix, where i>j, represents the cross-coupling influence coefficient of the i-th process parameter adjustment on the j-th process parameter adjustment.

[0037] Preferably, in step 4, the opening control signal of the spray gun supply pipeline regulating valve is a current signal of 4 mA to 20 mA, the opening control signal of the spray gun atomizing air pipeline regulating valve is a current signal of 4 mA to 20 mA, the speed control signal of the drive motor of the spraying actuator is a voltage signal of 0 V to 10 V, and the pose control signal of the joint actuator of the spraying actuator is a pulse width modulation signal, wherein the duty cycle of the pulse width modulation signal is 0% to 100%.

[0038] The paint pressure adjustment element and paint flow adjustment element in the coordinated adjustment vector are respectively mapped to the opening control signal of the spray gun feed pipeline regulating valve. The atomizing air pressure adjustment element in the coordinated adjustment vector is mapped to the opening control signal of the spray gun atomizing air pipeline regulating valve. The spray gun moving speed adjustment element in the coordinated adjustment vector is mapped to the speed control signal of the spraying actuator drive motor. The spray gun spatial pose adjustment element in the coordinated adjustment vector is mapped to the pose control signal of the spraying actuator joint actuator.

[0039] A multi-parameter sensing spraying process production line control system, comprising:

[0040] A multi-parameter synchronous acquisition module is used to control a group of multiple types of sensors deployed on the spraying actuator of the spraying process production line to synchronously acquire their respective physical quantities at a sampling frequency of 50 Hz to 200 Hz. The group of multiple types of sensors includes a paint temperature sensor, a paint pressure sensor, and a paint flow sensor installed at the inlet of the spray gun feed pipe; an atomizing air pressure sensor and an atomizing air flow sensor installed at the inlet of the spray gun atomizing air pipe; a spray gun posture sensor installed on the spray gun body; an ambient temperature sensor and an ambient relative humidity sensor installed in the spraying operation area; and a workpiece surface temperature sensor installed on the workpiece conveying device. The multi-parameter synchronous acquisition module arranges all sensor readings obtained in each synchronous acquisition into a multi-dimensional data vector according to a preset data structure, and arranges multiple multi-dimensional data vectors obtained from multiple consecutive synchronous acquisitions in the order of acquisition time to form a multi-parameter synchronous acquisition time series dataset.

[0041] A dynamic coupling state construction module extracts the multi-dimensional data vector at the current moment from the multi-parameter synchronous acquisition time series dataset output by the multi-parameter synchronous acquisition module, performs dimensional normalization processing on each physical quantity value in the multi-dimensional data vector to obtain a normalized multi-dimensional data vector, performs covariance matrix operation on the normalized multi-dimensional data vector to obtain the multi-parameter covariance matrix at the current moment, and concatenates the normalized multi-dimensional data vector and the multi-parameter covariance matrix using tensors to generate the multi-parameter dynamic coupling state vector at the current moment.

[0042] The feedforward prediction module incorporates a process parameter coupling dynamics model. This model employs a long short-term memory (LSTM) network structure, which includes an input layer, two hidden layers, and an output layer. Each hidden layer contains 128 LSM units. The feedforward prediction module receives the multi-parameter dynamic coupling state vector generated in real-time by the dynamic coupling state construction module. Through the process parameter coupling dynamics model, it outputs a sequence of predicted values ​​for coating quality indicators within a preset prediction time window after the current moment. Simultaneously, it outputs the adjustment amounts that need to be applied to multiple process parameters of the spraying actuator to make the coating quality indicators in the predicted value sequence approach the preset target values. These multiple process parameters include paint pressure adjustment, paint flow rate adjustment, atomized air pressure adjustment, spray gun movement speed adjustment, and spray gun spatial pose adjustment.

[0043] The collaborative control execution module receives the paint pressure adjustment, paint flow adjustment, atomized air pressure adjustment, spray gun movement speed adjustment, and spray gun spatial pose adjustment output by the feedforward prediction module. It then constructs a coupling compensation matrix, which is a lower triangular matrix with all diagonal elements equal to 1. The off-diagonal elements of the coupling compensation matrix represent the cross-coupling influence coefficients between different process parameter adjustments. The collaborative control execution module combines the paint pressure adjustment, paint flow adjustment, atomized air pressure adjustment, spray gun movement speed adjustment, and spray gun spatial pose adjustment into an adjustment vector. It then multiplies this adjustment vector by the coupling compensation matrix to obtain a collaborative adjustment vector after cross-coupling compensation. Each element in the collaborative adjustment vector is converted into a spray gun feed pipeline regulating valve opening control signal, a spray gun atomized air pipeline regulating valve opening control signal, a spraying actuator drive motor speed control signal, and a spraying actuator joint actuator pose control signal, according to the corresponding process parameter type. These control signals are then sent to the corresponding actuators.

[0044] The online evaluation and update module collects real-time measured coating quality data of the sprayed area through an online coating quality detection sensor deployed behind the spraying actuator. The measured coating quality data includes measured values ​​of coating thickness, coating surface gloss, and the number of coating surface defects. The measured coating quality data is compared with the predicted coating quality value predicted by the feedforward prediction module at the same time, and the prediction error is calculated. When the absolute value of the prediction error exceeds a preset error threshold, the multi-parameter synchronously collected time-series data, the process parameter adjustment amount output by the feedforward prediction module, and the measured coating quality data at the current time are combined into a new training sample. Single-step incremental training is performed on the process parameter coupled dynamics model built into the feedforward prediction module to update the network parameters of the process parameter coupled dynamics model.

[0045] Preferably, the dimensional normalization process in the dynamic coupling state construction module is to subtract the moving average value of the physical quantity within a preset historical time window from the physical quantity value, and then divide it by the moving standard deviation of the physical quantity within the preset historical time window to obtain the normalized value of the physical quantity. The length of the preset historical time window is from 100 consecutive sampling times to 500 consecutive sampling times.

[0046] The training samples of the process parameter coupled dynamics model built into the feedforward prediction module are derived from the multi-parameter synchronous acquisition time series dataset recorded during historical spraying operations and the corresponding coating quality detection data. The coating quality detection data includes coating thickness uniformity index, coating surface gloss index and coating surface defect density index. The process parameter coupled dynamics model takes minimizing the mean square error between the predicted coating quality index and the actual coating quality index as the training objective and uses the backpropagation algorithm over time for parameter optimization training.

[0047] The cross-coupling influence coefficient in the coordinated control execution module is obtained as follows: the hidden state vector of the last time step is extracted from the output state vector of the long short-term memory unit of the second hidden layer in the process parameter coupling dynamics model, the hidden state vector is input into a fully connected linear layer, the output dimension of the fully connected linear layer is equal to the square of the dimension of the adjustment vector, the output of the fully connected linear layer is rearranged into a square matrix, the square matrix is ​​subjected to lower triangularization and all diagonal elements are set to 1 to obtain the coupling compensation matrix.

[0048] The beneficial effects of this invention are:

[0049] 1. By using a multi-type sensor group to synchronously collect multi-dimensional physical quantity data such as paint temperature, paint pressure, paint flow rate, atomized air pressure, atomized air flow rate, spray gun spatial pose, ambient temperature, ambient relative humidity, and workpiece surface temperature at high frequency, and combining dimensional normalization processing and covariance matrix operation to construct a multi-parameter dynamic coupled state vector, the control system can simultaneously perceive the instantaneous changes of each parameter and the strength of their linear correlation, which significantly improves the characterization accuracy and real-time response of the dynamic state of the spraying process;

[0050] 2. A long short-term memory network is used to construct a coupled dynamic model of process parameters to achieve feedforward prediction of coating quality indicators such as coating thickness uniformity, coating surface gloss and coating surface defect density. Through the coupling compensation matrix, the adjustment of coating pressure, coating flow rate, atomized air pressure, spray gun movement speed and spray gun spatial pose are cross-coupled and compensated. This overcomes the problem of global effect deterioration caused by independent adjustment of single parameters in traditional methods, realizes the coordinated dynamic control of multiple process parameters, effectively improves coating thickness uniformity and surface gloss and reduces defect density.

[0051] 3. The measured values ​​of coating thickness, coating surface gloss, and the number of coating surface defects are collected in real time by online coating quality detection sensors. The measured data are compared with the model prediction values. When the absolute value of the prediction error exceeds the preset error threshold, the model is triggered to perform single-step incremental training. This enables the process parameter coupled dynamic model to continuously adapt to the drift of process characteristics caused by coating batch changes, equipment wear, and seasonal environmental changes, thus ensuring the prediction accuracy and control effect stability of the control system during long-term operation. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0054] Figure 2 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0055] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0056] Please see Figures 1-2 This invention provides a multi-parameter sensing control method and system for a spraying process production line. Its core lies in the tight coupling of five stages: high-frequency synchronous acquisition by multiple types of sensors, construction of multi-parameter dynamically coupled state vectors, feedforward prediction of a process parameter coupled dynamic model based on a long short-term memory network, multi-parameter collaborative control based on a coupling compensation matrix, and online model updating. This forms a complete closed-loop control system of "sensing-modeling-prediction-control-update". The following provides a detailed description of several key technical features involved in the technical solution of this invention.

[0057] Regarding the deployment and synchronous data acquisition of multiple sensor groups:

[0058] This invention deploys multiple types of sensor groups on the spraying actuator in a spraying process production line. The term "spraying actuator" refers to automated equipment used to perform spraying operations, including but not limited to multi-axis industrial robots and their end-effector spray gun assemblies, conveying devices for transporting workpieces, and piping systems for supplying paint and atomizing air.

[0059] The multi-type sensor group includes the following nine types of sensors:

[0060] The first type is a paint temperature sensor installed at the inlet of the spray gun's feed line, used to collect the real-time temperature value of the paint before it enters the spray gun. The feed line inlet refers to the connection point where the paint enters the spray gun body from the feed tank through the delivery pipe. This location can most accurately reflect the actual temperature state of the paint received by the spray gun.

[0061] The second type is a paint pressure sensor installed at the inlet of the spray gun's feed line, used to collect real-time pressure values ​​of the paint at the inlet of the feed line.

[0062] The third type is a paint flow sensor installed at the inlet of the spray gun's feed line, used to collect the real-time volumetric flow rate of paint at the inlet of the feed line.

[0063] The fourth type is an atomizing air pressure sensor installed at the inlet of the spray gun's atomizing air pipeline, used to collect the real-time pressure value of the atomizing air before it enters the spray gun. The atomizing air pipeline refers to the independent pipeline used to supply compressed air to the spray gun to assist in paint atomization.

[0064] The fifth type is an atomizing air flow sensor installed at the inlet of the spray gun's atomizing air pipeline, used to collect the real-time volumetric flow rate of atomizing air at the inlet of the atomizing air pipeline.

[0065] The sixth type is the spray gun pose sensor mounted on the spray gun body. This sensor specifically includes three orthogonally arranged linear displacement sensors and three orthogonally arranged angular displacement sensors. The three orthogonally arranged linear displacement sensors are used to collect the X-axis, Y-axis, and Z-axis coordinates of the spray gun nozzle relative to the workpiece surface in a three-dimensional spatial coordinate system. The three orthogonally arranged angular displacement sensors are used to collect the pitch, yaw, and roll angles of the spray gun axis relative to the normal to the workpiece surface in a three-dimensional spatial coordinate system. These six dimensions of pose data together constitute a complete pose description of the spray gun in space.

[0066] The seventh category consists of ambient temperature sensors installed in the spraying operation area to collect ambient temperature values ​​within the spraying operation area.

[0067] The eighth category consists of ambient relative humidity sensors installed in the spraying operation area to collect ambient relative humidity values ​​within the spraying operation area.

[0068] The ninth category is workpiece surface temperature sensors installed on workpiece conveying devices, used to collect the surface temperature values ​​of the workpiece to be sprayed before entering the spraying area.

[0069] All nine types of sensors are synchronously acquired using the same sampling frequency. "Synchronous acquisition" means that all sensors sample their respective physical quantities at the same time under the control of the same clock signal, ensuring that all acquired data are strictly aligned in time. The sampling frequency is 50 Hz to 200 Hz, or 50 to 200 times per second. All sensor readings obtained from each synchronous acquisition are arranged into a multi-dimensional data vector according to a preset data structure. This preset data structure is a fixed-length data format arranged in the following order: paint temperature, paint pressure, paint flow rate, atomizing air pressure, atomizing air flow rate, spray gun spatial three-dimensional coordinates, spray gun spatial attitude angle, ambient temperature, ambient relative humidity, and workpiece surface temperature. This multi-dimensional data vector contains data in 10 dimensions (the spray gun spatial three-dimensional coordinates include three sub-dimensions: X-axis, Y-axis, and Z-axis; the spray gun spatial attitude angle includes three sub-dimensions: pitch angle, yaw angle, and roll angle; therefore, the multi-dimensional data vector actually contains values ​​in 10 sub-dimensions). Multiple multidimensional data vectors obtained from consecutive synchronous acquisitions are arranged in the order of acquisition time to form a multi-parameter synchronous acquisition time-series dataset.

[0070] Regarding dimensional normalization:

[0071] Because the physical quantities in the aforementioned 10 sub-dimensions have different dimensions and orders of magnitude—for example, the dimension of coating temperature is degrees Celsius (°C), the dimension of coating pressure is megapascals (MPa), and the dimension of coating flow rate is liters per minute (L / min)—directly performing mathematical operations on these raw values ​​would cause the physical quantities with larger dimensions to dominate the calculation, masking the changes in the physical quantities with smaller dimensions. Therefore, this invention performs dimension normalization processing on the value of each physical quantity in the multidimensional data vector.

[0072] The specific method for dimensional normalization is as follows: For each physical quantity value, subtract the moving average value of the physical quantity within a preset historical time window from the physical quantity value, and then divide by the moving standard deviation of the physical quantity within the preset historical time window to obtain the normalized value of the physical quantity.

[0073] The preset historical time window has a length of 100 to 500 consecutive sampling moments. This time window is sliding, meaning that as the current moment progresses, the window continuously moves forward, always including data from the most recent N sampling moments before the current moment. The value of N can be adjusted according to the dynamic response speed of the actual production line—for production lines where process parameters change rapidly, a smaller N value (e.g., 100) can be selected to make the normalization result respond to parameter changes more quickly; for production lines where process parameters change slowly, a larger N value (e.g., 500) can be selected to smooth short-term noise fluctuations.

[0074] The moving average is calculated using the following formula:

[0075]

[0076] in, Let N represent the moving average of the i-th physical quantity at time t, and let N represent the total number of sampling times included in the preset historical time window. Indicates the time of the i-th physical quantity. The collected values, This represents the time interval between two consecutive samples. When the sampling frequency is 50 Hz... Equal to 0.02 seconds; when the sampling frequency is 200 Hz, It equals 0.005 seconds.

[0077] The moving standard deviation is calculated using the following formula:

[0078]

[0079] in, Let represent the standard deviation of the i-th physical quantity at time t.

[0080] Normalized values ​​are calculated using the following formula:

[0081]

[0082] in, This represents the normalized value of the i-th physical quantity at time t. After the above normalization process, the normalized value of each physical quantity follows a distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of differences in dimensions and orders of magnitude on subsequent mathematical operations.

[0083] Regarding the concatenation of multi-parameter covariance matrices and tensors:

[0084] The normalized values ​​corresponding to all physical quantities in the multidimensional data vector are arranged according to the preset data structure to form a normalized multidimensional data vector. The normalized multidimensional data vector is a 10-dimensional vector, where each element corresponds to the normalized value of a physical quantity.

[0085] Perform covariance matrix operations on the normalized multidimensional data vector to obtain the multi-parameter covariance matrix at the current time step. The covariance matrix operation is performed according to the following formula:

[0086]

[0087] in, Let M represent the multi-parameter covariance matrix at time t, and M represent the number of normalized multidimensional data vectors participating in the covariance matrix operation, with a value of 10 to 50. Indicates time Normalized multidimensional data vectors, Indicates from time The mean vector of the M normalized multidimensional data vectors up to time t, with the superscript T indicating the vector transpose operation.

[0088] The covariance matrix The dimension is (Because the normalized multidimensional data vector contains data in 10 sub-dimensions), each of its matrix elements This represents the linear correlation strength between the normalized value of the p-th physical quantity and the normalized value of the q-th physical quantity at time t. When When the value is positive, it indicates that the two physical quantities are changing in the same direction; when... When the value is negative, it indicates that the two physical quantities are changing in opposite directions; The larger the absolute value, the stronger the linear correlation between the two physical quantities.

[0089] The tensor concatenation is performed as follows: normalizing the multidimensional data vector... (Dimension 10) is used as the first dimension component of the tensor, and the multi-parameter covariance matrix is... (dimension is) The first and second dimensions of the tensor are then expanded into a one-dimensional vector (100 dimensions) and used as the second dimension component. The first and second dimension components are then concatenated along the last dimension of the tensor to generate a multi-parameter dynamic coupling state vector with dimensions 10 + 100 = 110. This multi-parameter dynamic coupling state vector contains both the instantaneous numerical information of each physical quantity at the current moment (provided by the normalized multi-dimensional data vector) and the linear correlation strength between the physical quantities at the current moment (provided by the multi-parameter covariance matrix).

[0090] Regarding the coupled dynamics model of process parameters:

[0091] The process parameter coupling dynamics model employs a Long Short-Term Memory (LSTM) network structure. LSTM is a special type of recurrent neural network specifically designed for processing sequential data, effectively capturing long-term dependencies in time series. In this invention, the LSTM network structure comprises one input layer, two hidden layers, and one output layer, with each hidden layer containing 128 LSTM units.

[0092] The input layer receives multi-parameter dynamically coupled state vectors generated at multiple consecutive time points as the input sequence. Specifically, at each prediction time point, the multi-parameter dynamically coupled state vectors of the current time point and multiple consecutive time points before it (such as 50 to 200 time points before the current time point) are arranged in chronological order to form the input sequence.

[0093] Two hidden layers are used for feature extraction and temporal modeling of the input sequence, respectively. The first hidden layer receives the output sequence of the input layer and extracts preliminary temporal features; the second hidden layer receives the output sequence of the first hidden layer and further extracts higher-level abstract features. Each LSTM unit contains three gating structures: a forget gate, an input gate, and an output gate, as well as a memory unit. The forget gate controls the degree to which the information from the previous time step is forgotten, the input gate controls the degree to which the input information from the current time step is written into the memory unit, and the output gate controls the degree to which the information from the memory unit is output to the next layer. The output values ​​of the three gating structures are all between 0 and 1, calculated using the sigmoid activation function.

[0094] The output layer receives the output state of the second hidden layer at the last time step and maps the 128-dimensional hidden state to the dimensions of the output variables through a fully connected linear layer. The output variables include a sequence of predicted coating quality indicators for a preset prediction time window following the current time, and adjustments to various process parameters of the spraying actuator to bring the coating quality indicators in this predicted sequence closer to preset target values. The preset prediction time window has a length of 50 to 200 consecutive sampling times.

[0095] The training samples for the coupled dynamics model of process parameters are derived from multi-parameter synchronously acquired time-series datasets recorded during historical spraying operations and corresponding coating quality inspection data. The coating quality inspection data includes coating thickness uniformity indices, coating surface gloss indices, and coating surface defect density indices. The coating thickness uniformity index is obtained by measuring the coating thickness at multiple points on the workpiece surface after spraying and calculating its standard deviation or coefficient of variation; the coating surface gloss index is obtained through gloss meter measurement; and the coating surface defect density index is obtained by counting the number of defects per unit area using a visual inspection system.

[0096] The model aims to minimize the mean squared error between the predicted and actual coating quality indicators during training, and employs the Backpropagation Through Time (BPTT) algorithm for parameter optimization. The mean squared error is calculated using the following formula:

[0097]

[0098] in, The mean square error is represented by T, and T represents the total number of prediction times included within the preset prediction time window. This indicates that the process parameter coupled dynamic model is at the prediction time. The output vector of predicted coating quality indicators. Indicates the time of prediction The corresponding actual coating quality index vector. During training, the weight and bias parameters in the network are continuously adjusted through the backpropagation algorithm, so that... It gradually decreases until it converges.

[0099] During the spraying operation, the multi-parameter dynamic coupling state vector generated in real time is input into the trained process parameter coupling dynamics model. The model outputs a sequence of predicted coating quality indicators within a preset prediction time window after the current moment, and simultaneously outputs the adjustment amounts that need to be applied to multiple process parameters of the spraying actuator to make the coating quality indicators in the predicted value sequence approach the preset target value. These multiple process parameters include paint pressure adjustment, paint flow rate adjustment, atomizing air pressure adjustment, spray gun movement speed adjustment, and spray gun spatial pose adjustment. All of these adjustment amounts are relative percentage values ​​compared to the actual values ​​of the corresponding process parameters at the current moment. For example, a paint pressure adjustment of +5% means increasing the paint pressure by 5% from the current value; the spray gun spatial pose adjustment includes percentage values ​​for the X-axis, Y-axis, Z-axis, pitch angle, yaw angle, and roll angle adjustments of the spray gun nozzle relative to the workpiece surface in a three-dimensional spatial coordinate system.

[0100] Regarding the coupling compensation matrix:

[0101] The coupling compensation matrix is ​​a lower triangular matrix with all diagonal elements equal to 1. A lower triangular matrix is ​​a square matrix in which all elements above the main diagonal are 0. The dimension of this matrix is... (Because the adjustment vector contains adjustments for 5 process parameters). The off-diagonal element in the i-th row and j-th column of this matrix, where i>j, represents the cross-coupling influence coefficient of the i-th process parameter adjustment on the j-th process parameter adjustment. For example, the element in the 2nd row and 1st column of the matrix represents the cross-coupling influence coefficient of the coating flow rate adjustment on the coating pressure adjustment—that is, when the coating flow rate is adjusted, due to the physical coupling relationship, the coating pressure also needs to be adjusted accordingly.

[0102] The cross-coupling influence coefficient is obtained as follows: The hidden state vector of the last time step is extracted from the output state vector of the LSTM unit in the second hidden layer of the process parameter coupling kinetic model. This hidden state vector has a dimension of 128 (because each hidden layer contains 128 LSTM units). This hidden state vector is then input into a fully connected linear layer whose output dimension is equal to the square of the dimension of the adjustment vector. Fully connected linear layers undergo linear transformations. The 128-dimensional hidden state vector is mapped to a 25-dimensional output vector, where for The weight matrix, This is a 25-dimensional bias vector. The 25-dimensional output of the fully connected linear layer is rearranged as follows: The matrix is ​​in the form of a square matrix. The matrix is ​​then subjected to lower triangularization, which means setting all elements above the main diagonal to 0 and setting all diagonal elements to 1 to obtain the coupling compensation matrix.

[0103] The adjustment values ​​of paint pressure, paint flow rate, atomizing air pressure, spray gun movement speed, and spray gun spatial pose output from step 3 are arranged in a fixed order to form an adjustment vector (5-dimensional). This adjustment vector is then multiplied by a coupling compensation matrix to obtain a collaborative adjustment vector (5-dimensional) after cross-coupling compensation. The physical meaning of the left multiplication operation is that for each process parameter adjustment, not only is the adjustment requirement of the parameter itself considered, but also the cross-coupling effect of other parameter adjustments on that parameter, thereby achieving collaborative adjustment among multiple parameters.

[0104] Each element in the coordinated adjustment vector is converted into a corresponding actuator control signal according to the corresponding process parameter type. Specifically: the paint pressure adjustment element and the paint flow adjustment element in the coordinated adjustment vector are mapped to the spray gun feed pipeline regulating valve opening control signal (4 mA to 20 mA current signal); the atomizing air pressure adjustment element in the coordinated adjustment vector is mapped to the spray gun atomizing air pipeline regulating valve opening control signal (4 mA to 20 mA current signal); the spray gun movement speed adjustment element in the coordinated adjustment vector is mapped to the speed control signal of the spraying actuator drive motor (0 V to 10 V voltage signal); and the spray gun spatial pose adjustment element in the coordinated adjustment vector is mapped to the pose control signal of the spraying actuator joint actuator (pulse width modulation signal, duty cycle 0% to 100%).

[0105] Regarding online model updates:

[0106] Online coating quality inspection sensors deployed behind the spraying actuator collect real-time measured data of the sprayed area's coating quality. These online sensors include, but are not limited to: laser displacement sensors or eddy current thickness gauges for measuring coating thickness, gloss sensors for measuring coating surface gloss, and machine vision systems for detecting coating surface defects. The measured coating quality data includes measured values ​​of coating thickness, coating surface gloss, and the number of coating surface defects.

[0107] The measured coating quality data is compared with the predicted coating quality values ​​generated by the coupled dynamic model of process parameters at the same time point, and the prediction error is calculated. The prediction error can be calculated separately for each coating quality indicator or as a comprehensive error. When the absolute value of the prediction error exceeds a preset error threshold, the online model update mechanism is triggered. The specific value of the preset error threshold is set according to the actual process requirements—for example, the coating thickness prediction error threshold can be set to ±5 micrometers, and the coating surface gloss prediction error threshold can be set to ±3 gloss units.

[0108] During online model updates, the current multi-parameter synchronously acquired time-series data, the process parameter adjustments output by the process parameter coupled dynamics model, and the measured coating quality data are combined into new training samples. Single-step incremental training is then performed on the process parameter coupled dynamics model. Single-step incremental training means that only the newly added training sample is used to update the model parameters once using gradients, rather than retraining using all historical data. This update method has lower computational overhead and enables real-time online updates. After the update is complete, the updated model is used for subsequent feedforward predictions.

[0109] Example

[0110] This embodiment uses an automotive body exterior panel spraying process production line as an example to provide a detailed description of the method and system of the present invention.

[0111] Application scenarios and initial conditions:

[0112] In the body painting workshop of an automobile manufacturing company, a six-axis industrial robot carrying an air spray gun is used to spray topcoat onto the outer panels of automobile bodies. The workpiece to be painted is an automobile body that has undergone pretreatment (including phosphating and electrophoretic primer coating). The outer panel material is cold-rolled steel sheet, and the surface has been polished to a surface roughness of [missing information]. to The spraying operation is carried out in a closed spray booth, and the temperature of the spray booth is set at [temperature value missing]. The relative humidity is set to The coating is a two-component polyurethane topcoat, and the coating temperature is maintained in the feed tank at [temperature value missing]. The spray gun is an air-assisted airless spray gun with a nozzle diameter of 1.3 mm. The robot sprays the outer body panel of the vehicle segment by segment according to the preset spraying path, with each segment being approximately 500 mm to 800 mm in length.

[0113] Deployment and synchronous data acquisition of multiple sensor groups:

[0114] In this embodiment, multiple types of sensor groups are deployed on the spraying actuator, as detailed below:

[0115] Install a PT100 platinum resistance temperature sensor with a range of [missing information] at the inlet of the spray gun's feed line. to The accuracy is This is used to collect coating temperature data. A piezoresistive pressure sensor with a range of 0 MPa to 5 MPa and an accuracy of [missing information] is installed at the same location. MPa, used to collect coating pressure values. An oval gear flow meter with a range of 0.1 L / min to 5.0 L / min and an accuracy of [missing information] is installed at the same location. It is used to collect paint flow rate data.

[0116] A piezoresistive pressure sensor with a range of 0 MPa to 1 MPa and an accuracy of [missing information] is installed at the inlet of the atomizing air line of the spray gun. MPa, used to collect atomized air pressure values. A thermal mass flow meter with a range of 0 L / min to 500 L / min and an accuracy of [missing information] is installed in the same location. It is used to collect the flow rate of atomized air.

[0117] A spray gun position sensor is mounted on the spray gun body. This sensor includes three orthogonally arranged linear displacement sensors and three orthogonally arranged angular displacement sensors. The three linear displacement sensors are wire-type displacement sensors with a range of 0 mm to 2000 mm and an accuracy of [insert accuracy here]. Millimeters are used to collect the X-axis, Y-axis, and Z-axis coordinates of the spray gun nozzle relative to the workpiece surface in a three-dimensional coordinate system. The three angular displacement sensors are triaxial gyroscopes within a MEMS inertial measurement unit, with a range of [missing information - likely millimeters]. The accuracy is It is used to collect the pitch angle, yaw angle and roll angle values ​​of the spray gun axis relative to the normal of the workpiece surface in a three-dimensional coordinate system.

[0118] A thermocouple temperature sensor with a range of [value missing] is installed in the spraying area. to The accuracy is It is used to collect ambient temperature values. A capacitive relative humidity sensor with a range of 0% to 100% and an accuracy of [missing information] is installed at the same location. It is used to collect relative humidity values ​​of the environment.

[0119] An infrared temperature sensor with a measuring range of [value missing] is installed on the workpiece conveying device. to The accuracy is It is used to collect the surface temperature value of the workpiece.

[0120] All the aforementioned sensors are controlled by a synchronous data acquisition card, which incorporates a high-precision clock generator that outputs a 100Hz synchronous sampling clock signal. All sensors are triggered to sample on the rising edge of the same clock signal, achieving strictly synchronous acquisition. Each synchronous acquisition generates a multi-dimensional data vector, arranged in the following order: paint temperature, paint pressure, paint flow rate, atomizing air pressure, atomizing air flow rate, three-dimensional coordinates of the spray gun (X-axis, Y-axis, Z-axis), spray gun attitude angles (pitch, yaw, roll), ambient temperature, ambient relative humidity, and workpiece surface temperature. This process is repeated 1000 times to create a multi-parameter synchronous acquisition time-series dataset.

[0121] Construction of multi-parameter dynamically coupled state vectors:

[0122] Extract the multidimensional data vector at the current time t from the multi-parameter synchronously acquired time-series dataset. This multidimensional data vector contains 10 physical quantity values: Paint temperature value (unit: ), The pressure value of the coating (unit: megapascals). This refers to the paint flow rate (unit: liters per minute). This represents the atomizing air pressure value (unit: megapascals). This refers to the atomized airflow rate (unit: liters per minute). The value is the X-axis coordinate (unit: millimeters). Y-axis coordinate value (unit: millimeters). Z-axis coordinate value (unit: millimeters). This is the pitch angle value (unit: degrees). This is the yaw angle value (unit: degrees). This is the roll angle value (unit: degrees). Ambient temperature value (unit: ), The value represents the relative humidity of the environment (unit: %). The surface temperature of the workpiece (unit: ).

[0123] Set the preset historical time window length N = 200 consecutive sampling moments, which includes 200 sampled data points prior to the current moment. The time interval between two adjacent samples... Seconds (corresponding to a 100 Hz sampling frequency).

[0124] For the i-th physical quantity, calculate its moving average value at time t. :

[0125]

[0126] Calculate moving standard deviation :

[0127]

[0128] Calculate normalized values :

[0129]

[0130] After the above calculations, the normalized multidimensional data vector is obtained. The dimension is 14.

[0131] Set M=30, and use the normalized multidimensional data vectors from the current time t and the previous 29 time points to participate in the covariance matrix operation. Calculate the mean vector. :

[0132]

[0133] Calculate the multi-parameter covariance matrix :

[0134]

[0135] The dimension is .Will Expand into a one-dimensional vector (dimension is...) ),and (Dimension 14) Concatenate along the last dimension to obtain the multi-parameter dynamic coupling state vector (dimension 14+196=210).

[0136] Construction and training of the coupled dynamic model of process parameters:

[0137] The LSTM network structure is constructed as follows: The input layer receives a sequence of multi-parameter dynamically coupled state vectors from multiple consecutive time steps, with a sequence length of 100 time steps (corresponding to a 1-second time window). The first hidden layer contains 128 LSTM units and receives the output sequence from the input layer. The second hidden layer contains 128 LSTM units and receives the output sequence from the first hidden layer. The output layer is a fully connected linear layer that maps the 128-dimensional output of the last time step of the second hidden layer to the output variable.

[0138] The output variables consist of two parts: the first part is the predicted value sequence of coating quality indicators, with a prediction time window of T=100 time points (corresponding to 1 second). At each time point, three coating quality indicators are predicted (coating thickness uniformity, coating surface gloss, and coating surface defect density). Therefore, the dimension of the predicted value sequence is... The second part consists of five process parameter adjustments (coating pressure adjustment, coating flow rate adjustment, atomizing air pressure adjustment, spray gun movement speed adjustment, and spray gun spatial pose adjustment), each of which is a scalar value. Therefore, the total dimension of the output layer is 300 + 5 = 305.

[0139] The model's training data comes from the historical production data of the spraying production line over the past three months, totaling approximately 5,000 complete spraying segments. Each training sample includes an input sequence (a multi-parameter dynamically coupled state vector at 100 consecutive time points) and corresponding label data (measured coating quality data after those 100 time points and the process parameter adjustments required to achieve the target quality). Coating quality inspection data is obtained through offline detection: after each spraying segment is completed, a coating thickness gauge is used to measure the coating thickness at 20 uniformly distributed measurement points on the workpiece surface, and its standard deviation is calculated as an index of coating thickness uniformity; a gloss meter is used to measure the 60° gloss at 10 measurement points, and the average value is taken as an index of coating surface gloss; a machine vision system is used to detect defects such as bubbles, runs, and orange peel on the coating surface, and the number of defects per unit area is counted as an index of coating surface defect density.

[0140] The backpropagation algorithm was used for training, with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs. The mean squared error loss function was:

[0141]

[0142] in For the model at the prediction time The output is a vector of predicted 3D coating quality indicators. This represents the corresponding actual coating quality index vector. After training, the mean squared error of the model on the validation set converged to below 0.02.

[0143] Feedforward prediction and coordinated regulation:

[0144] During the spraying operation, the multi-parameter dynamically coupled state vector generated in real time is input into the trained LSTM model. The model outputs a sequence of predicted coating quality indicators for the next second (100 time points) after the current time, and also outputs the adjustment amounts of 5 process parameters.

[0145] Assume the actual process parameters at the current moment are: coating pressure P = 2.5 MPa, coating flow rate Q = 0.8 L / min, and atomizing air pressure... Megapascals, spray gun movement speed v = 0.6 m / s, spray gun spatial pose (X-axis = 500 mm, Y-axis = 300 mm, Z-axis = 200 mm, pitch angle) Yaw angle Roll angle The model output adjustments are: paint pressure adjustment +3%, paint flow rate adjustment -2%, atomized air pressure adjustment +5%, spray gun movement speed adjustment +1%, and spray gun spatial pose adjustment (X-axis 0%, Y-axis 0%, Z-axis -2%, pitch angle +1%, yaw angle 0%, roll angle 0%).

[0146] The adjustment amounts of the above five process parameters are combined into an adjustment amount vector. ,in The spatial pose adjustment vector for the spray gun (containing 6 sub-dimensions).

[0147] Extract the hidden state vector from the last time step of the second hidden layer of the LSTM model. (Dimension 128), input to a fully connected linear layer, output a vector of dimension 25, rearranged as follows: The square matrix is ​​lower triangularized and its diagonals are set to 1 to obtain the coupling compensation matrix. Assuming The specific values ​​are:

[0148]

[0149] The value of 0.15 in the second row and first column indicates that the cross-coupling effect coefficient between the paint flow rate adjustment and the paint pressure adjustment is 0.15—meaning that when the paint flow rate needs to be adjusted, the paint pressure also needs to be adjusted by 15%. The value of -0.08 in the third row and first column indicates that the cross-coupling effect coefficient between the atomizing air pressure adjustment and the paint pressure adjustment is -0.08—meaning that when the atomizing air pressure increases, the paint pressure needs to be slightly reduced to maintain the atomization effect.

[0150] Multiply the adjustment vector by the coupling compensation matrix on the left to obtain the coordinated adjustment vector. :

[0151]

[0152]

[0153]

[0154] Each element in the coordinated adjustment vector is converted into an actuator control signal according to its corresponding process parameter type:

[0155] Coating pressure adjustment +3% — Current coating pressure is 2.5 MPa, target pressure after adjustment is Megapascals. The target pressure value is mapped to a current signal of 4 mA to 20 mA. Assuming the range of 0 MPa to 5 MPa corresponds to 4 mA to 20 mA, the control signal current is... Milliampere.

[0156] Paint flow rate adjustment -1.55% — Current paint flow rate is 0.8 liters per minute, target flow rate after adjustment is: The rate of increase per minute is mapped to the current control signal of the regulating valve in the feed pipeline.

[0157] Atomizing air pressure adjustment +4.22% — Current atomizing air pressure is 0.4 MPa, target pressure after adjustment is... Megapascals (MPa) are mapped to the current control signal of the atomizing air pipeline regulating valve.

[0158] Spray gun movement speed adjustment +1.21% — Current speed is 0.6 meters per second, target speed after adjustment is... Meters per second, mapped to the speed control signal (0 to 10 volts voltage signal) of the robot's drive motor.

[0159] Spray gun spatial pose adjustment amount - based on the pose adjustment amount after collaborative adjustment, it is mapped to the pulse width modulation signal of each joint actuator of the robot.

[0160] The aforementioned control signals are sent to the corresponding actuators to achieve coordinated control of multiple process parameters.

[0161] Online model updates:

[0162] During the spraying operation, online coating quality inspection sensors deployed behind the spraying actuator collect real-time measured data of the coating quality in the sprayed area. In this embodiment, a laser triangulation sensor is installed 200 mm behind the robot's end effector to measure the coating thickness immediately after spraying; a gloss sensor is installed to measure the coating surface gloss in real time; and a high-resolution industrial camera, along with image processing algorithms, is installed to detect coating surface defects in real time.

[0163] Assume that at a certain moment, the model predicts the coating thickness to be... The gloss level is 88 GU (gloss unit), and the defect density is 0.5 defects per square centimeter. The coating thickness measured by the online sensor is... The gloss level is 85 GU, and the defect density is 0.8 defects / cm². The coating thickness prediction error is... The gloss prediction error is For GU, the defect density prediction error is |0.5-0.8|=0.3 defects / square centimeter.

[0164] Set the coating thickness prediction error threshold to: The gloss prediction error threshold is 2 GU, and the defect density prediction error threshold is 0.2 defects / cm². Since at least one of the three indicators (in fact, all three) exceeds the threshold, the online model update mechanism is triggered.

[0165] The current time-series data (including sensor data from the previous 100 time steps), the model's output adjustments for five process parameters, and the measured coating quality data (coating thickness 32 μm, gloss 85 GU, defect density 0.8 defects / cm²) are combined to form a new training sample. This new sample is labeled: the model's output process parameter adjustments are used as part of the input features, and the difference between the measured and predicted coating quality is used as the correction target. This single sample is used to perform a gradient update on the LSTM model—that is, calculating the gradient of the loss function with respect to the model parameters and updating the parameters once along the gradient descent direction, with the learning rate set to 0.0001 (less than the initial training learning rate to avoid catastrophic forgetting). After the update, the updated model is used for subsequent feedforward predictions.

[0166] Comparison of effects:

[0167] To verify the technical effectiveness of this invention, a comparative experiment was conducted on the same spraying production line. The control group employed a traditional open-loop control strategy—all process parameters were set to fixed values ​​before the spraying operation started (coating pressure 2.5 MPa, coating flow rate 0.8 L / min, atomizing air pressure 0.4 MPa, spray gun moving speed 0.6 m / s, distance between spray gun and workpiece surface 200 mm, spray gun axis perpendicular to workpiece surface), and were not adjusted throughout the entire spraying cycle. The experimental group adopted the multi-parameter sensing spraying process production line control method of this invention. Each group sprayed 50 car bodies, and the various indicators of coating quality were statistically analyzed. The results are shown in the table below:

[0168] Average coating thickness (micrometers) 34.2 35.1 — Coating thickness standard deviation (micrometers) 4.8 1.2 75.0% Coating thickness uniformity (coefficient of variation, %) 14.0 3.4 75.7% Average gloss level (GU) of the coating surface 86.5 89.2 — Standard deviation of coating surface gloss (GU) 5.2 1.8 65.4% Defect density on coating surface (number of defects per square centimeter) 1.6 0.4 75.0% Batch pass rate (%) 78.0 96.0 23.1%

[0169] As shown in the table above, after adopting the method of the present invention, the standard deviation of coating thickness decreased from 4.8 micrometers to 1.2 micrometers, a reduction of 75.0%; the coating thickness uniformity (coefficient of variation) decreased from 14.0% to 3.4%, a reduction of 75.7%; the standard deviation of coating surface gloss decreased from 5.2 GU to 1.8 GU, a reduction of 65.4%; the coating surface defect density decreased from 1.6 defects / cm² to 0.4 defects / cm², a reduction of 75.0%; and the batch pass rate increased from 78.0% to 96.0%, an increase of 23.1 percentage points.

[0170] Comparative results show that the present invention significantly improves the batch consistency and overall pass rate of spraying quality through multi-parameter synchronous acquisition, dynamic coupling state construction, LSTM-based feedforward prediction, coordinated regulation based on coupling compensation matrix, and complete closed-loop control with online model updates.

[0171] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0172] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-parameter sensing method for controlling a spraying process production line, characterized in that, Includes the following steps: Step 1: Deploy a multi-type sensor group on the spraying actuator of the spraying process production line. The multi-type sensor group includes a paint temperature sensor, a paint pressure sensor, and a paint flow sensor installed at the inlet of the spray gun feed pipe; an atomizing air pressure sensor and an atomizing air flow sensor installed at the inlet of the spray gun atomizing air pipe; a spray gun posture sensor installed on the spray gun body; an ambient temperature sensor and an ambient relative humidity sensor installed in the spraying operation area; and a workpiece surface temperature sensor installed on the workpiece conveying device. Control all sensors in the multi-type sensor group to synchronously collect their corresponding physical quantities at a sampling frequency of 50 Hz to 200 Hz. Arrange all sensor readings obtained from each synchronous acquisition into a multi-dimensional data vector according to a preset data structure. Arrange multiple multi-dimensional data vectors obtained from multiple consecutive synchronous acquisitions in the order of acquisition time to form a multi-parameter synchronous acquisition time series dataset. Step 2: Extract the multidimensional data vector at the current moment from the multi-parameter synchronous acquisition time series dataset. Perform dimensional normalization processing on each physical quantity value in the multidimensional data vector. The dimensional normalization processing is to subtract the moving average value of the physical quantity within a preset historical time window from the physical quantity value, and then divide it by the moving standard deviation of the physical quantity within the preset historical time window to obtain the normalized value of the physical quantity. Arrange the normalized values ​​corresponding to all physical quantities in the multidimensional data vector according to the preset data structure to form a normalized multidimensional data vector. Perform covariance matrix operation on the normalized multidimensional data vector to obtain the multi-parameter covariance matrix at the current moment. Perform tensor concatenation between the normalized multidimensional data vector and the multi-parameter covariance matrix to generate the multi-parameter dynamic coupling state vector at the current moment. Step 3: Using the multi-parameter dynamic coupling state vector generated at multiple consecutive moments as the input sequence, and the process parameter adjustment amount of the spraying actuator within a preset time window after the current moment as the output variable, a process parameter coupling dynamic model is constructed. The process parameter coupling dynamic model adopts a long short-term memory network structure, which includes an input layer, two hidden layers, and an output layer. Each hidden layer contains 128 long short-term memory units. The multi-parameter dynamic coupling state vector generated in real time in Step 2 at the current moment is input into the trained process parameter coupling dynamic model. The process parameter coupling dynamic model outputs a sequence of predicted values ​​for the coating quality index within a preset prediction time window after the current moment. At the same time, it outputs the adjustment amount that needs to be applied to multiple process parameters of the spraying actuator to make the coating quality index in the predicted value sequence approach the preset target value. The multiple process parameters include paint pressure adjustment amount, paint flow rate adjustment amount, atomized air pressure adjustment amount, spray gun moving speed adjustment amount, and spray gun spatial pose adjustment amount. Step 4: Construct a coupling compensation matrix. The coupling compensation matrix is ​​a lower triangular matrix with all diagonal elements equal to 1. The off-diagonal elements of the coupling compensation matrix represent the cross-coupling influence coefficients between different process parameter adjustment amounts. The cross-coupling influence coefficients are obtained by linear transformation of the hidden layer state vector in the process parameter coupling dynamics model. The paint pressure adjustment amount, paint flow rate adjustment amount, atomized air pressure adjustment amount, spray gun moving speed adjustment amount, and spray gun spatial pose adjustment amount output in Step 3 are combined into an adjustment amount vector. The adjustment amount vector is multiplied left by the coupling compensation matrix to obtain the collaborative adjustment amount vector after cross-coupling compensation. Each element in the collaborative adjustment amount vector is converted into the spray gun feed pipeline regulating valve opening control signal, the spray gun atomized air pipeline regulating valve opening control signal, the spraying actuator drive motor speed control signal, and the spraying actuator joint driver pose control signal according to the corresponding process parameter type. The above control signals are sent to the corresponding actuators to perform collaborative control. Step 5: Real-time measured coating quality data of the sprayed area is collected by an online coating quality detection sensor deployed behind the spraying actuator. The measured coating quality data includes measured values ​​of coating thickness, coating surface gloss, and the number of coating surface defects. The measured coating quality data is compared with the predicted coating quality value of the process parameter coupled dynamics model in Step 3 at the same time, and the prediction error is calculated. When the absolute value of the prediction error exceeds a preset error threshold, the multi-parameter synchronously collected time-series data, the process parameter adjustment amount output by the process parameter coupled dynamics model, and the measured coating quality data at the current time are combined into a new training sample. Single-step incremental training is performed on the process parameter coupled dynamics model to update the network parameters of the process parameter coupled dynamics model. The updated process parameter coupled dynamics model is used for feedforward prediction in the subsequent Step 3.

2. The multi-parameter sensing spraying process production line control method according to claim 1, characterized in that, The preset data structure mentioned in step 1 is a fixed-length data format arranged in the order of paint temperature value, paint pressure value, paint flow rate value, atomizing air pressure value, atomizing air flow rate value, spray gun space three-dimensional coordinate value, spray gun space attitude angle value, ambient temperature value, ambient relative humidity value and workpiece surface temperature value. The spray gun posture sensor includes three orthogonally arranged linear displacement sensors and three orthogonally arranged angular displacement sensors mounted on the spray gun body. The three orthogonally arranged linear displacement sensors are used to collect the X-axis, Y-axis and Z-axis coordinate values ​​of the spray gun nozzle relative to the workpiece surface in a three-dimensional spatial coordinate system. The three orthogonally arranged angular displacement sensors are used to collect the pitch angle, yaw angle and roll angle values ​​of the spray gun axis relative to the normal of the workpiece surface in a three-dimensional spatial coordinate system.

3. The multi-parameter sensing spraying process production line control method according to claim 1, characterized in that, The length of the preset historical time window mentioned in step 2 is from 100 consecutive sampling moments to 500 consecutive sampling moments; The moving average is calculated according to the following formula: in, Let N represent the moving average of the i-th physical quantity at time t, and let N represent the total number of sampling times included in the preset historical time window. Indicates the time of the i-th physical quantity. The collected values, Indicates the time interval between two adjacent samples; The moving standard deviation is calculated according to the following formula: in, This represents the standard deviation of the movement of the i-th physical quantity at time t; The normalized value is calculated according to the following formula: in, This represents the normalized value of the i-th physical quantity at time t.

4. The multi-parameter sensing spraying process production line control method according to claim 1, characterized in that, The covariance matrix operation described in step 2 is performed according to the following formula: in, Let M represent the multi-parameter covariance matrix at time t, and M represent the number of normalized multidimensional data vectors involved in the covariance matrix operation, with values ​​ranging from 10 to 50. Indicates time The normalized multidimensional data vector, Indicates from time The mean vector of the M normalized multidimensional data vectors up to time t, where the superscript T indicates the vector transpose operation; The tensor concatenation is performed as follows: the normalized multidimensional data vector is used as the first dimension component of the tensor, the multi-parameter covariance matrix is ​​expanded into a one-dimensional vector and used as the second dimension component of the tensor, and the first dimension component and the second dimension component are concatenated along the last dimension of the tensor to generate the multi-parameter dynamic coupling state vector.

5. The multi-parameter sensing spraying process production line control method according to claim 1, characterized in that, The training samples for the process parameter coupled dynamics model in step 3 are derived from the multi-parameter synchronous acquisition time series dataset recorded during historical spraying operations and the corresponding coating quality detection data. The coating quality detection data includes coating thickness uniformity index, coating surface gloss index and coating surface defect density index. The coupled dynamics model of process parameters aims to minimize the mean square error between the predicted coating quality index and the actual coating quality index. It employs a backpropagation algorithm over time for parameter optimization training. The mean square error is calculated using the following formula: in, The mean square error is represented by T, and T represents the total number of prediction times included within the preset prediction time window. This indicates that the process parameter coupled dynamic model is at the prediction time. The output vector of predicted coating quality indicators. Indicates the time of prediction The corresponding actual coating quality index vector; The length of the preset prediction time window is from 50 consecutive sampling times to 200 consecutive sampling times.

6. The multi-parameter sensing spraying process production line control method according to claim 1, characterized in that, In step 3, the coating pressure adjustment, coating flow rate adjustment, atomizing air pressure adjustment, spray gun moving speed adjustment, and spray gun spatial pose adjustment output by the coupled dynamics model of process parameters are all relative adjustment percentage values ​​relative to the actual values ​​of the corresponding process parameters at the current moment. The spray gun spatial pose adjustment includes the percentage values ​​of the X-axis coordinate adjustment, Y-axis coordinate adjustment, Z-axis coordinate adjustment, pitch angle adjustment, yaw angle adjustment, and roll angle adjustment in the three-dimensional coordinate system relative to the workpiece surface.

7. The multi-parameter sensing spraying process production line control method according to claim 1, characterized in that, The cross-coupling influence coefficient in step 4 is obtained as follows: the hidden state vector of the last time step is extracted from the output state vector of the long short-term memory unit of the second hidden layer in the process parameter coupling dynamics model. The hidden state vector is input into a fully connected linear layer. The output dimension of the fully connected linear layer is equal to the square of the dimension of the adjustment vector. The output of the fully connected linear layer is rearranged into a square matrix. The square matrix is ​​then subjected to lower triangularization and all diagonal elements are set to 1 to obtain the coupling compensation matrix. The coupling compensation matrix has a dimension of 5×5. The off-diagonal element in the i-th row and j-th column of the coupling compensation matrix, where i>j, represents the cross-coupling influence coefficient of the i-th process parameter adjustment on the j-th process parameter adjustment.

8. The multi-parameter sensing spraying process production line control method according to claim 1, characterized in that, In step 4, the opening control signal of the spray gun supply pipeline regulating valve is a current signal of 4 mA to 20 mA; the opening control signal of the spray gun atomizing air pipeline regulating valve is a current signal of 4 mA to 20 mA; the speed control signal of the drive motor of the spraying actuator is a voltage signal of 0 V to 10 V; and the pose control signal of the joint actuator of the spraying actuator is a pulse width modulation signal with a duty cycle of 0% to 100%. The paint pressure adjustment element and paint flow adjustment element in the coordinated adjustment vector are respectively mapped to the opening control signal of the spray gun feed pipeline regulating valve. The atomizing air pressure adjustment element in the coordinated adjustment vector is mapped to the opening control signal of the spray gun atomizing air pipeline regulating valve. The spray gun moving speed adjustment element in the coordinated adjustment vector is mapped to the speed control signal of the spraying actuator drive motor. The spray gun spatial pose adjustment element in the coordinated adjustment vector is mapped to the pose control signal of the spraying actuator joint actuator.

9. A multi-parameter sensing spraying process production line control system, characterized in that, The system is used to execute the multi-parameter sensing spraying process production line control method according to any one of claims 1 to 8, including: A multi-parameter synchronous acquisition module is used to control a group of multiple types of sensors deployed on the spraying actuator of the spraying process production line to synchronously acquire their respective physical quantities at a sampling frequency of 50 Hz to 200 Hz. The group of multiple types of sensors includes a paint temperature sensor, a paint pressure sensor, and a paint flow sensor installed at the inlet of the spray gun feed pipe; an atomizing air pressure sensor and an atomizing air flow sensor installed at the inlet of the spray gun atomizing air pipe; a spray gun posture sensor installed on the spray gun body; an ambient temperature sensor and an ambient relative humidity sensor installed in the spraying operation area; and a workpiece surface temperature sensor installed on the workpiece conveying device. The multi-parameter synchronous acquisition module arranges all sensor readings obtained in each synchronous acquisition into a multi-dimensional data vector according to a preset data structure, and arranges multiple multi-dimensional data vectors obtained from multiple consecutive synchronous acquisitions in the order of acquisition time to form a multi-parameter synchronous acquisition time series dataset. A dynamic coupling state construction module extracts the multi-dimensional data vector at the current moment from the multi-parameter synchronous acquisition time series dataset output by the multi-parameter synchronous acquisition module, performs dimensional normalization processing on each physical quantity value in the multi-dimensional data vector to obtain a normalized multi-dimensional data vector, performs covariance matrix operation on the normalized multi-dimensional data vector to obtain the multi-parameter covariance matrix at the current moment, and concatenates the normalized multi-dimensional data vector and the multi-parameter covariance matrix using tensors to generate the multi-parameter dynamic coupling state vector at the current moment. The feedforward prediction module incorporates a process parameter coupling dynamics model. This model employs a long short-term memory (LSTM) network structure, which includes an input layer, two hidden layers, and an output layer. Each hidden layer contains 128 LSM units. The feedforward prediction module receives the multi-parameter dynamic coupling state vector generated in real-time by the dynamic coupling state construction module. Through the process parameter coupling dynamics model, it outputs a sequence of predicted values ​​for coating quality indicators within a preset prediction time window after the current moment. Simultaneously, it outputs the adjustment amounts that need to be applied to multiple process parameters of the spraying actuator to make the coating quality indicators in the predicted value sequence approach the preset target values. These multiple process parameters include paint pressure adjustment, paint flow rate adjustment, atomized air pressure adjustment, spray gun movement speed adjustment, and spray gun spatial pose adjustment. The collaborative control execution module receives the paint pressure adjustment, paint flow adjustment, atomized air pressure adjustment, spray gun movement speed adjustment, and spray gun spatial pose adjustment output by the feedforward prediction module. It then constructs a coupling compensation matrix, which is a lower triangular matrix with all diagonal elements equal to 1. The off-diagonal elements of the coupling compensation matrix represent the cross-coupling influence coefficients between different process parameter adjustments. The collaborative control execution module combines the paint pressure adjustment, paint flow adjustment, atomized air pressure adjustment, spray gun movement speed adjustment, and spray gun spatial pose adjustment into an adjustment vector. It then multiplies this adjustment vector by the coupling compensation matrix to obtain a collaborative adjustment vector after cross-coupling compensation. Each element in the collaborative adjustment vector is converted into a spray gun feed pipeline regulating valve opening control signal, a spray gun atomized air pipeline regulating valve opening control signal, a spraying actuator drive motor speed control signal, and a spraying actuator joint actuator pose control signal, according to the corresponding process parameter type. These control signals are then sent to the corresponding actuators. The online evaluation and update module collects real-time measured coating quality data of the sprayed area through an online coating quality detection sensor deployed behind the spraying actuator. The measured coating quality data includes measured values ​​of coating thickness, coating surface gloss, and the number of coating surface defects. The measured coating quality data is compared with the predicted coating quality value predicted by the feedforward prediction module at the same time, and the prediction error is calculated. When the absolute value of the prediction error exceeds a preset error threshold, the multi-parameter synchronously collected time-series data, the process parameter adjustment amount output by the feedforward prediction module, and the measured coating quality data at the current time are combined into a new training sample. Single-step incremental training is performed on the process parameter coupled dynamics model built into the feedforward prediction module to update the network parameters of the process parameter coupled dynamics model.

10. A multi-parameter sensing spraying process production line control system according to claim 9, characterized in that, The dimensional normalization process in the dynamic coupling state construction module is to subtract the moving average value of the physical quantity within a preset historical time window from the physical quantity value, and then divide it by the moving standard deviation of the physical quantity within the preset historical time window to obtain the normalized value of the physical quantity. The length of the preset historical time window is from 100 consecutive sampling times to 500 consecutive sampling times. The training samples of the process parameter coupled dynamics model built into the feedforward prediction module are derived from the multi-parameter synchronous acquisition time series dataset recorded during historical spraying operations and the corresponding coating quality detection data. The coating quality detection data includes coating thickness uniformity index, coating surface gloss index and coating surface defect density index. The process parameter coupled dynamics model takes minimizing the mean square error between the predicted coating quality index and the actual coating quality index as the training objective and uses the backpropagation algorithm over time for parameter optimization training. The cross-coupling influence coefficient in the coordinated control execution module is obtained as follows: the hidden state vector of the last time step is extracted from the output state vector of the long short-term memory unit of the second hidden layer in the process parameter coupling dynamics model, the hidden state vector is input into a fully connected linear layer, the output dimension of the fully connected linear layer is equal to the square of the dimension of the adjustment vector, the output of the fully connected linear layer is rearranged into a square matrix, the square matrix is ​​subjected to lower triangularization and all diagonal elements are set to 1 to obtain the coupling compensation matrix.

Images