Self-evolutionary control method for wire harness welding fume capture and tin powder recovery

Abstract

The present application relates to the field of welding environmental protection technology, and discloses a wire harness welding welding fume capturing and tin powder recycling self-evolution control method based on working condition sensing, comprising the following steps: S1, collecting multi-source working condition data and welding gun area space image; S2, data preprocessing to generate working condition feature vector; S3, inputting the working condition feature vector and the space image into a self-evolution space-time prediction model of fused 2D-CNN, 1D-CNN and BiLSTM, and outputting welding fume production prediction value and tin powder spatter space distribution prediction map; S4, driving the capturing device to perform grid level differentiated variable negative pressure capturing according to the prediction map; S5, multi-stage separation and recycling, and online detection of recovery rate and purity; S6, constructing a comprehensive reward, and using a PPO deep reinforcement learning model to cooperatively optimize the capturing and recycling parameters; S7, performing parameter updating and incrementally training the prediction model to form a self-evolution closed loop. The present application solves the problem of lack of spatial targeting of traditional uniform air extraction, and significantly improves the welding fume capturing efficiency and tin powder recycling quality.

Description

Technical Field

[0001] This invention relates to the field of welding environmental protection technology, specifically to a self-evolving control method for welding fume capture and solder powder recovery in wire harness welding based on working conditions. Background Technology

[0002] Wire harness welding is a core process in electrical connections for automotive, aerospace, and industrial equipment. It involves metallurgically connecting wires to terminals using resistance welding or laser welding. During welding, the welding wire melts at high temperatures, and localized overheating of the molten pool causes the tin and flux components to vaporize and rapidly condense, forming tin-containing welding fumes and micron-sized tin powder splatter. If these are not effectively collected, they will cause environmental pollution and waste of tin resources. Currently, wire harness welding stations are generally equipped with welding fume collection and tin powder recovery devices. Collection devices typically include negative pressure fans, ductwork, and suction ports, continuously drawing welding fumes away from the work area. Recovery devices are connected in series at the rear end of the collection duct, and common configurations include cyclone separators, filter cartridges, or electrostatic adsorption modules, used to separate and collect tin powder particles from the airflow.

[0003] However, most existing welding fume capture and control methods can only make scalar estimates of the total amount of welding fumes generated, that is, only output a single intensity value. They cannot predict the two-dimensional spatial distribution characteristics of solder powder splash in the welding area. Due to the lack of prior information on the spatial distribution of splash, the capture system can only apply uniform negative pressure to the entire welding area for extensive ventilation. It cannot concentrate capture resources on high-splash-risk areas, resulting in a lack of targeting in the spatial dimension of the capture action and limited welding fume capture efficiency. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a self-evolving control method for wire harness welding fume collection and solder powder recovery based on working condition perception. This method solves the problem that existing methods cannot predict the spatial distribution characteristics of solder powder splashes, resulting in a lack of targeted collection actions in the spatial dimension, leading to only a rough and uniform ventilation.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A condition-aware, self-evolving control method for welding fume collection and solder powder recovery in wire harness welding, characterized by comprising the following steps:

[0007] S1, collect multi-source working condition data for wire harness welding, including at least wire harness type, welding current, welding voltage, welding torch moving speed, welding wire feeding speed, ambient temperature and humidity, welding area dust concentration, and spatial images of the welding torch area simultaneously acquired by a high-speed camera.

[0008] S2 performs time synchronization, outlier removal, and normalization on the S1 data to generate a working condition feature vector.

[0009] S3, input the working condition feature vector of S2 and the spatial image of the welding torch area of ​​S1 into the self-evolving spatiotemporal prediction model; the self-evolving spatiotemporal prediction model includes a two-dimensional convolutional module, a one-dimensional convolutional neural network and a bidirectional long short-term memory network. First, the two-dimensional convolutional module extracts the spatial features of the spatial image and fuses them with the working condition feature vector. Then, the one-dimensional convolutional neural network and the bidirectional long short-term memory network process the data to output the predicted value of welding fume generation and the predicted spatial distribution map of solder powder splash; the self-evolving spatiotemporal prediction model is configured with an online update interface;

[0010] S4, based on the prediction results of S3 and the real-time position of the collection device, generates a negative pressure control command for exhaust and a guide angle adjustment command for the air inlet, driving the collection device to perform directional variable negative pressure collection;

[0011] S5 performs multi-stage separation and recovery of tin-containing solder fumes captured in S4, and detects the purity and particle size distribution of the recovered tin powder online to calculate the real-time tin powder recovery rate.

[0012] S6 constructs a comprehensive reward based on the recovery rate, purity compliance status, and capture energy consumption of S5. It calls a deep reinforcement learning model based on the near-end policy optimization algorithm, and takes the working condition feature vector of S2, the capture action of S4 in the previous control cycle, and the recovery rate and purity feedback of S5 as state inputs. It outputs the optimized ventilation negative pressure curve, guide angle sequence, and multi-stage separator working parameters.

[0013] S7 updates the optimized parameters of S6 to the collection device and the recovery module. At the same time, the working condition feature vector and spatial image collected in S1 are used as input, and the measured amount of solder fume and the spatial distribution map of solder powder splash at the corresponding time step are used as labels to form training samples. The self-evolving spatiotemporal prediction model of S3 is incrementally trained through the online update interface, and the weights are updated and returned to S3.

[0014] By adopting the above technical solution, the welding process status is digitized in real time through multi-source working condition data acquisition and working condition feature vector construction, thereby providing a unified and accurate state input for prediction models and reinforcement learning decisions. As a result, a comprehensive perception and accurate representation effect of complex wire harness welding working conditions is achieved, laying a solid data foundation for subsequent self-evolutionary control.

[0015] Preferably, in step S1, the multi-source operating condition data is collected by a welding power supply Hall sensor, a temperature and humidity probe, a laser scattering fume concentration sensor, and the high-speed camera; the multi-source operating condition data also includes the welding wire diameter and the shielding gas flow rate; the welding wire feeding speed is collected by a photoelectric encoder installed on the end of the wire feeding wheel shaft; and the data is collected using the welding controller clock as a reference for unified time synchronization.

[0016] By adopting the above technical solution, multiple sensors are used to synchronously collect working condition data from multiple dimensions such as electrical, visual, and thermodynamic, and the welding controller clock is used for unified timing, which ensures the temporal consistency and spatial correlation of heterogeneous data. At the same time, the real-time acquisition of the welding wire feeding speed ensures the accuracy of the recovery rate calculation. Therefore, the obtained data can fully reflect the instantaneous state of the welding process and its temporal evolution, which improves the accuracy and reliability of subsequent working condition feature vector construction.

[0017] Preferably, in step S2, the normalization process uses minimum-maximum normalization to map the feature values ​​of each dimension to a closed interval between zero and one. When the completeness of the key data of the working condition feature vector in S2 does not reach the preset threshold, the average optimal capture parameter based on the same wire harness model in history is called as the preset safe capture parameter. The average optimal capture parameter is the average value of the parameters of the top 20% of records of the wire harness model in the historical processing database after sorting by the comprehensive score of recovery rate and energy consumption.

[0018] By adopting the above technical solution, the use of minimum-maximum normalization to unify the features of different physical dimensions to the same scale eliminates the interference of dimensional differences on the training of the prediction model and accelerates the model convergence speed. At the same time, when the data completeness is insufficient, it automatically switches to the average optimal parameters based on the top 20% of historical processing data sorted by comprehensive score. Therefore, it not only ensures the input quality of the high-precision prediction model under normal working conditions, but also avoids the risk of unreliable control commands output by the system when the sensor is abnormal or the data is missing, thus realizing the robust operation of the system in areas with weak data.

[0019] Preferably, in step S3, the tin powder splash spatial distribution prediction map output by the self-evolving spatiotemporal prediction model provides the splash intensity value of each grid in the form of a welding area grid.

[0020] By adopting the above technical solution, since the predicted model outputs a spatial distribution prediction map of tin powder splash in the form of a welding area grid, which refines the splash intensity to each grid unit, it provides an accurate spatial quantitative basis for the subsequent collection device to realize grid-level differentiated negative pressure distribution and guide angle adjustment, thus avoiding the inefficiency of coarse uniform collection.

[0021] Preferably, in step S4, the real-time pose of the capture device includes the three-dimensional coordinates of the suction port and the pitch or yaw angle; when generating the suction negative pressure control command, the target negative pressure value of each suction port is calculated based on the spatter intensity of each grid using a grid intensity weighted average algorithm, and the guide angle of the suction port is adjusted according to the direction of the welding torch movement.

[0022] By adopting the above technical solution, the negative pressure distribution of multiple air inlets is differentiated according to the intensity value of each grid in the solder powder splash distribution prediction diagram. Combined with the advance adjustment of the guide angle in the direction of soldering gun movement, the high splash area obtains stronger capture negative pressure. The capture window can actively follow the drift direction of the solder fume cloud. Therefore, the inefficiency and lag of traditional uniform negative pressure capture are avoided, and the efficiency of solder fume capture and the conveying efficiency of solder powder into the recycling channel are significantly improved.

[0023] Preferably, in step S5, the multi-stage separation and recovery is performed in series by a cyclone separator, an electrostatic adsorption grid, and an ultrasonic vibrating screen; online detection is performed using an online laser particle size analyzer and an X-ray fluorescence elemental analyzer; the real-time tin powder recovery rate is the ratio of the mass of recovered tin powder to the theoretical mass of tin consumed in the solder wire.

[0024] By adopting the above technical solution, the tiered separation mechanism of cyclone separation, electrostatic adsorption and ultrasonic sieving is used to process tin powder particles of different sizes step by step. The purity and particle size distribution of the recovered tin powder are fed back in real time by online particle size analyzer and elemental analyzer. Therefore, while improving the recovery rate and purity of tin powder, timely and accurate quality feedback signals are provided for the reinforcement learning agent, so that the parameter adjustment of the recovery process can be closely focused on the recovery quality target.

[0025] Preferably, in S6, the comprehensive reward is constructed with tin powder recovery rate and purity compliance as positive reward items, and capture energy consumption and temperature fluctuation in the soldering area as penalty items; the deep reinforcement learning model based on the near-end strategy optimization algorithm has an action space including the negative pressure value of each air inlet, the guide angle increment, the electrostatic adsorption voltage, and the frequency of the ultrasonic vibrating screen.

[0026] By adopting the above technical solution, a multi-objective comprehensive reward function that takes into account recycling efficiency, quality, energy consumption, and thermodynamic stability is constructed. The optimal strategy is explored in a high-dimensional continuous action space containing all controllable parameters of collection and recycling using a near-end strategy optimization algorithm. Therefore, the collaborative optimization of multiple links of solder fume collection and solder powder recycling is realized. Under the premise of ensuring welding quality, it can automatically find the optimal working condition that balances high recovery rate, high purity, and low energy consumption, and make the strategy update process more stable.

[0027] Preferably, in step S7, incremental training adopts an experience replay mechanism, storing the newly generated working condition feature vector, spatial image and its corresponding measured solder fume generation and solder powder splash spatial distribution map tuple into the experience pool. When the amount of data in the experience pool reaches the batch size, a small batch gradient update is performed on the self-evolving prediction model. During the update, the old model copy is retained. If the prediction error of the updated model on the validation set increases by more than a preset threshold, the old model copy is activated.

[0028] By adopting the above technical solution, the prediction model can continuously learn from actual feedback by using the experience playback mechanism and the real-time collected working condition images and measured welding fume and spatter distribution as supervision labels for batch incremental training. By retaining copies of old models and using an automatic rollback mechanism based on verification errors, the model's adaptability to time-varying working conditions is guaranteed, and performance degradation caused by abnormal samples or over-updates is prevented, thus achieving a self-evolutionary effect of monotonic improvement of the prediction model.

[0029] Preferably, step S7 further includes: real-time monitoring of the welding current change rate; when the change rate exceeds a preset threshold, pausing the online update of the self-evolutionary prediction model and switching to operation with preset conservative trapping parameters and preset conservative recovery parameters; when the change rate is lower than the preset threshold and continues for a preset duration, resuming closed-loop self-evolutionary control; the preset conservative trapping parameters are the lowest energy consumption safe operating parameters of the wire harness model under historical normal operating conditions; the preset conservative recovery parameters are a set of static optimal recovery rate parameters.

[0030] By adopting the above technical solution, the model update is frozen and switched to the lowest energy consumption safe operation parameters and static optimal recovery rate parameter group based on the historical normal operation conditions of the welding torch in a timely manner under abnormal conditions such as short circuit and arc interruption. The self-evolution process is only restarted after the operating conditions return to stability. Therefore, it effectively prevents abnormal data from polluting the prediction model and reinforcement learning strategy, ensures the operational safety of the system under extreme conditions, and guarantees the controllability and reliability of the self-evolution process.

[0031] Preferably, in step S3, the online update interface of the self-evolving spatiotemporal prediction model is connected to the cloud platform and adopts a federated learning architecture: multiple edge models are trained locally and then uploaded to the encrypted gradient, which is aggregated in the cloud and then the updated weights are distributed, and at least three historical model versions are retained in the cloud.

[0032] By adopting the above technical solution, the federated learning architecture completes local training at the edge and only uploads encrypted gradients. After aggregating experience from multiple production lines in the cloud, the global model is distributed. Combined with a multi-version retention mechanism, the generalization ability and robustness of the model are improved by making full use of distributed production line data while protecting the privacy of enterprise data. At the same time, it supports secure version rollback, enabling the system to continuously integrate multi-source knowledge and evolve securely.

[0033] This invention provides a self-evolving control method for welding fume collection and solder powder recovery in wire harnesses based on operational conditions. It offers the following advantages:

[0034] 1. This invention employs a spatiotemporal prediction model that uses a spatial image of the soldering torch area and a feature vector of the working condition as input. The model first extracts spatial features through two-dimensional convolution, then fuses them with the feature vector of the working condition, and finally processes the data using 1DCNN and BiLSTM. This model outputs a predicted map of the amount of solder fumes generated and the spatial distribution of solder powder splashes. This achieves a dual-modal, accurate prediction of the intensity and spatial distribution of solder fumes and solder powder splashes. Compared to existing technologies that only perform scalar estimation of the total amount of solder fumes, this invention solves the problem that these technologies cannot predict the spatial distribution characteristics of solder powder splashes, resulting in a lack of specificity in the spatial dimension of the collection action, leading to only a coarse and uniform extraction method.

[0035] 2. This invention employs a technical solution that uses achieving target recovery rate and purity as positive rewards and energy consumption and temperature fluctuations as penalties to construct a comprehensive reward system. Deep reinforcement learning is used to collaboratively optimize the capture and recovery parameters in a multi-dimensional action space, achieving a globally coordinated optimization effect across multiple stages of capture and recovery. Compared to existing technologies where capture control and recovery control are adjusted independently, this invention solves the problem of difficulty in simultaneously achieving high recovery rate, high purity, and low energy consumption due to the fragmentation of processes.

[0036] 3. This invention employs a technical solution that uses real-time acquired operating condition images and measured welding fume and spatter distribution labels to perform online incremental training of the prediction model, combined with backup of the old model and automatic rollback. This achieves the technical effect of continuous self-evolution of the prediction model as the operating conditions drift, with monotonic performance improvement. Compared to the existing technology of offline training and fixed deployment of the prediction model, this invention solves the problems of lack of online adaptive capability and continuous deterioration of control effect due to operating condition drift after long-term operation. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0038] Figure 2 This is a schematic diagram of the self-evolutionary prediction model structure of the present invention;

[0039] Figure 3 This is a schematic diagram of the reinforcement learning decision-making mechanism of the present invention;

[0040] Figure 4 This is a schematic diagram of the multi-stage separation, recovery, and detection process of the present invention;

[0041] Figure 5 This is a schematic diagram of the online update architecture for federated learning according to the present invention. Detailed Implementation

[0042] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Please see the appendix Figure 1 -Appendix Figure 5 This invention provides a condition-aware, self-evolving control method for welding fume collection and solder powder recovery in wire harness welding, comprising the following:

[0044] S1, collect multi-source working condition data for wire harness welding, including at least wire harness type, welding current, welding voltage, welding torch moving speed, welding wire feeding speed, ambient temperature and humidity, welding area dust concentration, and spatial images of the welding torch area simultaneously acquired by a high-speed camera.

[0045] Furthermore, in S1, multi-source operating condition data is collected through a welding power supply Hall sensor, a temperature and humidity probe, a laser scattering fume concentration sensor, and the high-speed camera; the multi-source operating condition data also includes the welding wire diameter, shielding gas flow rate, and welding wire feeding speed; the welding wire feeding speed is collected by a photoelectric encoder installed on the end of the wire feeding wheel shaft; the data is collected using the welding controller clock as a reference for unified time synchronization;

[0046] Specifically, this system forms the perceptual foundation for all subsequent intelligent decisions. Hall effect sensors capture the transient waveforms of current and voltage output from the welding power source in real time at kilohertz frequencies. A high-speed camera, fixed diagonally above the workstation, records the three-dimensional spatial pose of the welding torch relative to the wire harness pins and the molten pool morphology at a rate of no less than 30 frames per second. Temperature and humidity probes monitor the thermodynamic boundaries affecting welding fume diffusion within the environmental chamber, while a laser scattering fume concentration sensor provides absolute quantitative values ​​of particulate matter concentration in the welding area. To construct spatiotemporally correlated samples, all sensor data is immediately timestamped by the edge computing gateway using a hardware-level timestamp based on the welding controller's internal clock upon acquisition.

[0047] In this scheme, the sampling frequency of the laser scattering dust concentration sensor is preferably 10Hz to ensure that it can capture the dynamic burst process of dust generation without causing excessive load on the data transmission at the edge.

[0048] S2 performs time synchronization, outlier removal, and normalization on the S1 data to generate a working condition feature vector.

[0049] Furthermore, in S2, the normalization process uses minimum-maximum normalization to map the feature values ​​of each dimension to a closed interval between zero and one. When the completeness of the key data of the working condition feature vector in S2 does not reach the preset threshold, the average optimal capture parameter based on the same wire harness model in history is called as the preset safe capture parameter. The average optimal capture parameter is the average value of the parameters of the top 20% of records of the wire harness model in the historical processing database after sorting by the comprehensive score of recovery rate and energy consumption.

[0050] Specifically, the original sensor stream of S1 is transformed into a structured tensor that can be directly consumed by the prediction model. The edge computing program first uses the timestamp generated by the welding controller clock reference in S1 as the alignment basis to perform linear interpolation resampling on the data streams of each heterogeneous sensor to ensure that there is a complete sensor snapshot at each logical time point. Then, it starts automatic outlier detection based on Grubbs' criterion to remove outliers caused by electromagnetic interference or optical obstruction. For the removed transient missing bits, the values ​​of the previous valid time point are used for linear filling. After cleaning, min-max normalization is performed on each continuous feature dimension, including current, voltage, temperature and humidity, dust concentration, moving speed, and wire feeding speed. The physical quantity values ​​under the original engineering units are uniformly scaled to a closed interval of zero to one to eliminate the gradient bias caused by different dimensions to the subsequent neural network training. The key data completeness is defined as the proportion of valid data points in each of the above dimensions to the total sampling points in the time window.

[0051] In this scheme, the threshold is preferably 90%. When the integrity does not meet the standard, it indicates that the sensor is in an unreliable state. The system will bypass the subsequent intelligent model and directly call the preset safety capture parameters. These parameters are derived from the historical processing database of the same wire harness model. After sorting by the comprehensive score of recovery rate and energy consumption, the average value of the top 20% of the recorded parameters is taken.

[0052] The final generated working condition feature vector is a one-dimensional vector. A real vector, where The vector represents the total number of dimensions for all working condition features. It carries information on the welding thermodynamics, kinematics, and environmental state at the current moment and is directly used as the input to the S3 prediction model.

[0053] S3, input the working condition feature vector of S2 and the spatial image of the welding torch area of ​​S1 into the self-evolving spatiotemporal prediction model; the self-evolving spatiotemporal prediction model includes a two-dimensional convolutional module, a one-dimensional convolutional neural network and a bidirectional long short-term memory network. First, the two-dimensional convolutional module extracts the spatial features of the spatial image and fuses them with the working condition feature vector. Then, the one-dimensional convolutional neural network and the bidirectional long short-term memory network process the data to output the predicted value of welding fume generation and the predicted spatial distribution map of solder powder splash; the self-evolving prediction model is configured with an online update interface;

[0054] Furthermore, in S3, the self-evolving spatiotemporal prediction model is a spatiotemporal prediction model that integrates two-dimensional convolution, one-dimensional convolutional neural networks, and bidirectional long short-term memory networks, taking the time series of working condition feature vectors and the spatial image of the welding torch area as inputs; the spatial image of the welding torch area is first processed by a two-dimensional convolution module to extract spatial features, and then concatenated with the working condition feature vectors at the corresponding time steps to form a fused feature sequence; the fused feature sequence is then processed by a 1DCNN to extract local temporal features, and stacked BiLSTMs are used to learn long-term dependencies, and finally, a fully connected layer is used to output the predicted value of welding fume generation and a two-dimensional solder powder spatter intensity distribution map; the solder powder spatter spatial distribution prediction map output by the self-evolving spatiotemporal prediction model gives the spatter intensity value of each grid in the form of a welding area grid;

[0055] In S3, the online update interface of the self-evolving spatiotemporal prediction model is connected to the cloud platform and adopts a federated learning architecture: multiple edge models are trained locally and then uploaded to the encrypted gradient, which is aggregated in the cloud and then distributed to update the weights. The cloud retains at least three historical model versions.

[0056] According to the appendix Figure 2 As shown, specifically, it serves as the core mapping bridge between perception and control. The self-evolving spatiotemporal prediction model is a spatiotemporal prediction model that integrates two-dimensional convolution, one-dimensional convolution, and bidirectional long short-term memory networks. Its inputs include:

[0057] Generated by S2, containing continuous The time series of working condition feature vectors at each time step, with dimensions of... ;

[0058] A sequence of spatial images of the welding torch area corresponding to the same time window.

[0059] The spatial image of the welding torch area is first processed by a two-dimensional convolutional neural network module with shared weights. This two-dimensional convolutional neural network module contains two layers with convolutional kernels of size [missing information]. The convolutional layer encodes each frame of the image into a fixed-length spatial feature vector. After encoding, the spatial feature vector is concatenated with the condition feature vector at the corresponding time step to form a fused feature vector. Thus, a feature vector with dimension [missing information] is obtained. The fusion feature sequence, where is the length of the spatial feature vector.

[0060] Subsequently, the fused feature sequence is input into a 1DCNN module consisting of three one-dimensional convolutional layers to extract temporally correlated features. The kernel size of the first convolutional layer is [size missing]. The stride is 1, the number of output channels is 32; the size of the second convolutional kernel is... The stride is 1, the number of output channels is 64; the kernel size of the third convolutional layer is... The stride is 1, and the number of output channels is 128. Each convolutional layer is followed by a ReLU activation function. The feature map output by the 1DCNN module has a dimension of [missing information]. Next, the feature map is fed into a two-layer stacked bidirectional long short-term memory network. The first layer, BiLSTM, has 64 hidden units, and the second layer, BiLSTM, also has 64 hidden units. The forward LSTM encodes from the past to the present, while the backward LSTM backtracks from the present to the past. The hidden states of both layers at each time step are concatenated to form a dimension. The vector.

[0061] The BiLSTM hidden state at the last time step is taken and mapped through two parallel fully connected layers: the first fully connected layer maps the 128-dimensional features into a scalar, which is used as the predicted value of the current welding fume generation, denoted as . The unit is milligrams per cubic meter; the second fully connected layer maps the 128-dimensional features into a single dimension. The matrix, i.e., the spatial distribution prediction map of tin powder splash. The prediction map divides the welding area into a grid with a physical scale of 1 cm by 1 cm. The matrix contains the first... Line number Column element values This represents the relative intensity of solder powder splashing at that location, and its value ranges within a closed interval of zero to one.

[0062] The online update interface of the self-evolving spatiotemporal prediction model is connected to a federated learning server deployed on the factory's private cloud. During the training process, the original working condition data is not directly uploaded. Instead, the model weight difference generated by the local incremental training is uploaded to the cloud after differential privacy encryption. The cloud performs federated averaging aggregation on the uploaded gradients of multiple production line edge nodes and retains at least three historical model versions so that it can be safely rolled back to the verified old version at any time.

[0063] S4, based on the prediction results of S3 and the real-time position of the collection device, generates a negative pressure control command for exhaust and a guide angle adjustment command for the air inlet, driving the collection device to perform directional variable negative pressure collection;

[0064] Furthermore, in S4, the real-time pose of the capture device includes the three-dimensional coordinates of the suction port and the pitch or yaw angle; when generating the suction negative pressure control command, the negative pressure value of each suction port is allocated according to the spatter intensity of each grid, and the guide angle of the suction port is adjusted according to the direction of the welding torch movement.

[0065] Specifically, this method transforms the spatial distribution information generated by the predictive model into precise fluid dynamics execution actions. The capture device comprises two independent three-degree-of-freedom robotic arms arranged on both sides of the welding station, each arm having a rectangular suction port at its end. The dimension generated by S3 is... The predicted spatial distribution map of solder powder splatter was first loaded into a digital twin space registered with the coordinate system of the actual soldering station. For the first... Each air intake is defined as a grid set within its current effective coverage area. Calculate the splash intensity-weighted centroid coordinates of all grids within the region, and then use the direction of the line connecting the three-dimensional coordinates of this centroid to the geometric center of the air intake to calculate the target pitch angle. and target yaw angle This generates a guide angle adjustment command and simultaneously calculates the weighted average of the splash intensity within the coverage area. Its expression is:

[0066] ;

[0067] in, For grid The splash intensity value, For grid Weighting factor for the Euclidean distance from the geometric center of the air intake. The value is the reciprocal of that distance. Based on this weighted average, the th... Target negative pressure value of each air intake Calculated using the following formula:

[0068] ;

[0069] in, For the first The target negative pressure value for each air intake is expressed in Pascals. The basic negative pressure value; This is the system's maximum safe negative pressure value; For the first Weighted average splash intensity of the area covered by each air intake; The theoretical maximum value of splash intensity is 1.0. In this scheme, Preferably 500 Pa, The preferred pressure is 2000 Pa to accommodate the pipe pressure capacity of most wire harness welding stations;

[0070] To further advance compensation, the system uses kinematic extrapolation to predict the position of the welding torch in the next two time steps based on the history of the welding torch moving speed vector in the characteristic vector of S2 working condition for the previous five time steps. This makes the response of the suction port guide angle advance the current welding point, ensuring that the collection window always covers the drift direction of the welding fume cloud.

[0071] S5 performs multi-stage separation and recovery of tin-containing solder fumes captured in S4, and detects the purity and particle size distribution of the recovered tin powder online to calculate the real-time tin powder recovery rate.

[0072] Furthermore, in S5, multi-stage separation and recovery is performed sequentially by a cyclone separator, an electrostatic adsorption grid, and an ultrasonic vibrating screen; online detection is performed using an online laser particle size analyzer and an X-ray fluorescence elemental analyzer; the real-time tin powder recovery rate is the ratio of the mass of recovered tin powder to the theoretical mass of tin consumed in the solder wire.

[0073] According to the appendix Figure 4 As shown, specifically, the high-concentration solder fumes captured by S4 first enter the cyclone separator tangentially. Under centrifugal force, coarse solder powder particles larger than 10 micrometers in diameter are thrown against the separator wall and fall into the bottom ash hopper. This stage can remove more than 70% of the particulate matter in the flue gas. The airflow carrying the remaining fine particles then passes through a set of parallel electrostatic adsorption grids with a uniform electric field generated by a high-voltage DC power supply. The electric field strength between the grids is maintained at 3 to 5 kV per centimeter, which charges fine particles with a diameter of 1 to 10 micrometers and captures them by the electrodes. The grids are vibrated at fixed intervals to allow the adsorbed layer to fall into the collection pipe. Finally, the airflow passes through an ultrasonic vibrating screen, the screen aperture of which is preferably 20 micrometers. High-frequency, low-amplitude vibration prevents solder powder bridging and blockage and completes the final fine screening. At the outlet of the main collection pipe, a bypass sampling airflow is split off and flows sequentially through an online laser particle size analyzer and an X-ray fluorescence elemental analyzer. The laser particle size analyzer, based on Mie scattering theory, outputs the D10, D50, and D90 characteristic values ​​of the particle size distribution in real time; the X-ray fluorescence elemental analyzer calculates the mass percentage purity of tin by analyzing the characteristic X-ray energy spectrum of the sample under stimulated emission, denoted as . Real-time tin powder recovery rate Calculated using the following formula:

[0074] ;

[0075] in, The real-time tin powder recovery rate is dimensionless and ranges from zero to one in a closed interval. The net mass of recovered tin powder accumulated by the electronic balance within the sampling time window, in grams; This refers to the real-time wire feeding speed, measured in meters per second. The sampling time window length is in seconds. This represents the theoretical mass density of tin in the solder wire, expressed in grams per meter. According to the diameter of the welding wire Density of welding wire material and the mass percentage of tin element Through formula The recovery rate is pre-calculated and determined. This recovery rate is directly input into the reward function of S6 as a key performance indicator.

[0076] The true values ​​required for incremental training are obtained through the following parallel measurement and computation system:

[0077] Actual value of welding fume generation: A conical element oscillating microbalance and an anemometer are installed in the main air duct of the collection device. The conical element oscillating microbalance measures the particulate matter mass concentration of the flowing air at a frequency of 1Hz; the anemometer simultaneously measures the volumetric flow rate within the air duct. The calculation unit instantaneously multiplies the mass concentration and volumetric flow rate, converts the volume of the welding area, and generates the actual value of welding fume generation in mg / m³, denoted as [value missing]. .

[0078] Realistic spatial distribution of solder powder splatter: The sequence of images of the welding area synchronously captured by the high-speed camera is transmitted to the edge computing node. The edge computing node is equipped with an image processing pipeline for performing meshed splatter intensity quantization. This quantization process includes: using the image frame before arc initiation as the background reference image; for each subsequent image frame, calculating its difference from the background reference image, and performing threshold segmentation on the difference image to obtain a binary image of candidate splatter particles; performing connected component analysis on the binary image to identify independent splatter particle regions, and extracting the centroid coordinates and area of ​​each region; dividing the welding area into a mesh format consistent with the prediction image mesh division method in S3. Grid, statistics of falling into each grid The total area of ​​all particles within the grid is calculated, and this area value is divided by the grid area for normalization. The resulting value is the measured splash intensity of that grid. Composed of the measured splash intensity values ​​of all grids, with dimension [missing information]. A gridded real splash space distribution map, denoted as The image processing pipeline is accelerated by FPGA or GPU to meet the time step alignment requirements of incremental training.

[0079] S6 constructs a comprehensive reward based on the recovery rate, purity compliance status, and capture energy consumption of S5. It calls a deep reinforcement learning model based on the near-end policy optimization algorithm, and takes the working condition feature vector of S2, the capture action of S4 in the previous control cycle, and the recovery rate and purity feedback of S5 as state inputs. It outputs the optimized ventilation negative pressure curve, guide angle sequence, and multi-stage separator working parameters.

[0080] Furthermore, in S6, the comprehensive reward is constructed with tin powder recovery rate and purity compliance as positive reward items, and capture energy consumption and temperature fluctuation in the soldering area as penalty items; the deep reinforcement learning model adopts the near-end policy optimization algorithm, and its action space includes the negative pressure value of each air intake, the guide angle increment, as well as the electrostatic adsorption voltage and the frequency of the ultrasonic vibrating screen.

[0081] According to the appendix Figure 3As shown, this is specifically the core decision-making mechanism for the entire system to achieve self-evolution. First, a comprehensive reward function is constructed. Its expression is:

[0082] ;

[0083] in, This is the total reward value; Real-time tin powder recovery rate calculated for S5; For purity compliance indicator function, when When greater than or equal to the preset target value ,otherwise The preset target value in this scheme is preferably 95%; It is the sum of the power consumption of the fan and the power consumption of the electrostatic adsorption high voltage power supply in the current control cycle, normalized to the historical maximum value. It is dimensionless and takes the value in a closed interval from zero to one. , which is the standard deviation of the welding area temperature in the characteristic vector of S2 operating condition over the most recent ten control cycles, in degrees Celsius, and is used as a thermodynamic stability penalty term; , , , These are the weighting coefficients. In this scheme, the preferred values ​​for each weighting coefficient are... , , , .

[0084] The deep reinforcement learning agent employs a proximal policy optimization algorithm. Both its policy network and value network are three-layer fully connected networks with 256 hidden units per layer. The state space vector dimension is... It is composed of three parts: the current working condition feature vector generated by S2, with dimensions of The actual capture action vector generated by S4 in the previous control cycle has a dimension of [missing information]. ,in The number of air intakes, each containing two action quantities: negative pressure value and guide angle; and a two-dimensional vector composed of the recovery rate and purity fed back by S5, with dimensions of... The action space is a one-dimensional space. , where Each dimension represents the negative pressure setting value for each air intake. Each dimension represents the incremental value of the guide angle of each air intake. Each dimension represents the electrostatic adsorption voltage value. Each dimension represents the frequency value of the ultrasonic vibrating screen.

[0085] It should be noted that when the self-evolving spatiotemporal prediction model predicts the amount of welding fumes generated and the distribution of spatter in S3, it uses a time series of working condition feature vectors (with dimensions of T×N) and a spatial image sequence containing T consecutive time steps as input to capture the spatiotemporal evolution pattern; while when the deep reinforcement learning model makes real-time decisions in S6, it only uses the working condition feature vector (with dimensions of 1×N) at the current moment as part of the state input to meet the real-time requirements of control.

[0086] The agent runs a control cycle every 0.5 seconds, outputs action vectors, and then converts them into physical setpoints for each actuator through linear mapping. These setpoints are then encapsulated into optimized time series of the ventilation negative pressure curve and working parameters of the multi-stage separator. The objective function of the near-end policy optimization algorithm introduces a pruning mechanism to limit the update range of the old and new policies, thus avoiding the catastrophic policy forgetting that may be caused by a single large-sample update in the traditional policy gradient method.

[0087] S7 updates the optimized parameters of S6 to the collection device and the recovery module. At the same time, the working condition feature vector and spatial image collected by S1 are used as input, and the measured amount of solder fume and the spatial distribution map of solder powder splash at the corresponding time step are used as labels to form training samples. The self-evolving spatiotemporal prediction model of S3 is incrementally trained through the online update interface, and the weights are updated and returned to S3.

[0088] Furthermore, in S7, incremental training adopts an experience replay mechanism, which stores the newly generated working condition feature vectors, spatial images and their corresponding measured solder fume generation and solder powder splash spatial distribution map tuples into the experience pool. When the amount of data in the experience pool reaches the batch size, a small batch gradient update is performed on the self-evolving prediction model. During the update, the old model copy is retained. If the prediction error of the updated model on the validation set increases by more than a preset threshold, the old model copy is activated.

[0089] S7 also includes: real-time monitoring of the welding current change rate; when the change rate exceeds a preset threshold, pausing the online update of the self-evolutionary prediction model and switching to the operation of preset conservative capture and recovery parameters; when the change rate is lower than the preset threshold and continues for a preset duration, resuming closed-loop self-evolutionary control; the preset conservative capture parameters are the lowest energy consumption safe operation parameters of the wire harness model under historical normal working conditions; the preset conservative recovery parameters are the static optimal recovery rate parameter set.

[0090] According to the appendix Figure 5As shown, this implements the closed-loop iteration and safe operation mechanism of the system. The continuous motion vector generated by S6 is sent to the local controller of each actuator via the OPCUA industrial communication protocol. The exhaust negative pressure servo driver, stepper motor guide driver, and high-voltage power supply respectively convert the target value into physical quantity output, completing one complete intelligent control cycle. At the same time, the working condition feature vector and spatial image collected by S1 corresponding to this decision, as well as the actual value of welding fume generation at the corresponding time step obtained by the synchronous measurement system, are used. Real spatial distribution diagram of tin powder splash This constitutes a complete training tuple, which is stored in an experience replay pool with a capacity of 100,000 records. When the accumulated new data in the experience pool reaches a batch size (preferably 512 records in this scheme), the incremental training process of the self-evolving spatiotemporal prediction model described in S3 is triggered. The training uses this batch of data as a mini-batch, with the working condition feature vector and spatial image as input, and the measured welding fume generation as the input. Spatial distribution diagram of measured tin powder splash Calculate the prediction loss function for the labels. Its expression is:

[0091] ;

[0092] in, Represents the mean square error function; The amount of welding fumes predicted by the model. This represents the actual measured amount of welding fumes generated. This is a predicted spatial distribution map of tin powder splashes based on the model, with dimensions of [dimension not specified]. , This is a spatial distribution map of actual tin powder splashes obtained through high-speed camera image analysis, with the same dimension. ; and The weighting coefficients for the two losses are, in this scheme, preferably, both. The Adam optimizer, with a learning rate much lower than that used in the initial training phase, is employed. The optimal learning rate is [missing value]. A gradient descent step is performed to avoid catastrophic forgetting. Before the update, the current model weights are fully backed up. After the update, the system quickly evaluates the root mean square error of the new model's welding fume prediction using a reserved validation set that has not been used for training in the past hour. If this error increases by more than 10% compared to the error of the old model on the validation set, the system automatically rolls back to the backup weights and clears the current batch of data, thus ensuring the monotonic improvement characteristic of the system performance.

[0093] In addition, a separate high-priority safety monitoring task runs in the system. This task continuously reads the raw value of the welding current from the Hall sensor at 1-millisecond intervals and calculates the first derivative of the current with respect to time in real time, denoted as... ,when If the power consumption exceeds a preset safety threshold (preferably 500 amperes per second), it indicates that the welding torch may be experiencing abnormal conditions such as a short circuit, arc interruption, or pin collision. At this point, both the prediction model and the reinforcement learning strategy are unreliable. The system immediately suspends the online update of the S3 prediction model and the action output of the S6 reinforcement learning, and switches to operation with preset conservative capture parameters and preset conservative recovery parameters. The preset conservative capture parameters are the lowest energy consumption safe operating parameters for the wire harness model under historical normal operating conditions; the preset conservative recovery parameters are a set of static optimal recovery rate parameters. Once the voltage drops below 500 amperes per second and remains stable within this safe range for at least 3 seconds, the system automatically resumes normal closed-loop self-evolutionary control and restarts a new decision cycle from S3.

[0094] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A self-evolving control method for welding fume collection and solder powder recovery in wire harnesses based on operational conditions, characterized in that, Includes the following steps: S1, collect multi-source working condition data for wire harness welding, including at least wire harness type, welding current, welding voltage, welding torch moving speed, welding wire feeding speed, ambient temperature and humidity, welding area dust concentration, and spatial images of the welding torch area simultaneously acquired by a high-speed camera. S2 performs time synchronization, outlier removal, and normalization on the S1 data to generate a working condition feature vector. S3, input the working condition feature vector of S2 and the spatial image of the welding torch area of ​​S1 into the self-evolving spatiotemporal prediction model; the self-evolving spatiotemporal prediction model includes a two-dimensional convolutional module, a one-dimensional convolutional neural network and a bidirectional long short-term memory network. First, the two-dimensional convolutional module extracts the spatial features of the spatial image and fuses them with the working condition feature vector. Then, the one-dimensional convolutional neural network and the bidirectional long short-term memory network process the data to output the predicted value of welding fume generation and the predicted spatial distribution map of solder powder splash; the self-evolving spatiotemporal prediction model is configured with an online update interface; S4, based on the prediction results of S3 and the real-time position of the collection device, generates a negative pressure control command for exhaust and a guide angle adjustment command for the air inlet, driving the collection device to perform directional variable negative pressure collection; S5 performs multi-stage separation and recovery of tin-containing solder fumes captured by S4, and detects the purity and particle size distribution of the recovered tin powder online to calculate the real-time tin powder recovery rate. S6 constructs a comprehensive reward based on the recovery rate, purity compliance status, and capture energy consumption of S5. It calls a deep reinforcement learning model based on the near-end policy optimization algorithm, and takes the working condition feature vector of S2, the capture action of S4 in the previous control cycle, and the recovery rate and purity feedback of S5 as state inputs. It outputs the optimized ventilation negative pressure curve, guide angle sequence, and multi-stage separator working parameters. S7 updates the optimized parameters of S6 to the collection device and the recovery module. At the same time, the working condition feature vector and spatial image collected in S1 are used as input, and the measured amount of solder fume and the spatial distribution map of solder powder splash at the corresponding time step are used as labels to form training samples. The self-evolving spatiotemporal prediction model of S3 is incrementally trained through the online update interface, and the weights are updated and returned to S3.

2. The method according to claim 1, characterized in that, In step S1, multi-source operating condition data is collected through a welding power supply Hall sensor, a temperature and humidity probe, a laser scattering fume concentration sensor, and the high-speed camera; the multi-source operating condition data also includes the welding wire diameter and the shielding gas flow rate; the welding wire feeding speed is collected by a photoelectric encoder installed on the end of the wire feeding wheel shaft; the data is collected using the welding controller clock as a reference for unified time synchronization.

3. The method according to claim 1, characterized in that, In S2, the normalization process uses minimum-maximum normalization to map the feature values ​​of each dimension to a closed interval between zero and one. When the completeness of the key data of the working condition feature vector in S2 does not reach the preset threshold, the average optimal capture parameter based on the same wire harness model in history is called as the preset safe capture parameter. The average optimal capture parameter is the average value of the parameters of the top 20% of records of the wire harness model in the historical processing database after sorting by the comprehensive score of recovery rate and energy consumption.

4. The method according to claim 1, characterized in that, In S3, the tin powder splash spatial distribution prediction map output by the self-evolving spatiotemporal prediction model gives the splash intensity value of each grid in the form of a welding area grid.

5. The method according to claim 1, characterized in that, In S4, the real-time pose of the capture device includes the three-dimensional coordinates of the suction port and the pitch or yaw angle; when generating the suction negative pressure control command, the target negative pressure value of each suction port is calculated according to the spatter intensity of each grid using a grid intensity weighted average algorithm, and the suction port guide angle is adjusted according to the welding torch movement direction.

6. The method according to claim 1, characterized in that, In step S5, the multi-stage separation and recovery is performed in series by a cyclone separator, an electrostatic adsorption grid, and an ultrasonic vibrating screen; online detection is performed using an online laser particle size analyzer and an X-ray fluorescence elemental analyzer; the real-time tin powder recovery rate is the ratio of the mass of recovered tin powder to the theoretical mass of tin consumed in the solder wire.

7. The method according to claim 1, characterized in that, In S6, the comprehensive reward is constructed with tin powder recovery rate and purity compliance as positive reward items, and capture energy consumption and temperature fluctuation in the soldering area as penalty items; the deep reinforcement learning model based on the near-end strategy optimization algorithm has an action space including the negative pressure value of each air inlet, the guide angle increment, the electrostatic adsorption voltage, and the ultrasonic vibration screen frequency.

8. The method according to claim 1, characterized in that, In S7, incremental training adopts an experience replay mechanism, which stores the newly generated working condition feature vector, spatial image and its corresponding measured welding fume generation and solder powder splash spatial distribution map tuple into the experience pool. When the amount of data in the experience pool reaches the batch size, a small batch gradient update is performed on the self-evolving prediction model. During the update, the old model copy is retained. If the prediction error of the updated model on the validation set increases by more than a preset threshold, the old model copy is activated.

9. The method according to claim 1, characterized in that, S7 further includes: real-time monitoring of the welding current change rate; when the change rate exceeds a preset threshold, pausing the online update of the self-evolutionary prediction model and switching to operation with preset conservative trapping parameters and preset conservative recovery parameters; when the change rate is lower than the preset threshold and continues for a preset duration, resuming closed-loop self-evolutionary control; the preset conservative trapping parameters are the lowest energy consumption safe operating parameters of the wire harness model under historical normal operating conditions; the preset conservative recovery parameters are a set of static optimal recovery rate parameters.

10. The method according to claim 1, characterized in that, In S3, the online update interface of the self-evolving spatiotemporal prediction model is connected to the cloud platform and adopts a federated learning architecture: multiple edge models are trained locally and then uploaded to the encrypted gradient, which is aggregated in the cloud and then distributed to update the weights. At least three historical model versions are retained in the cloud.

Images