Machining practical training behavior detection method and system

Abstract

The invention discloses a machining practical training behavior detection method and system, and the method comprises the following steps: collecting the state data of a student in the machining practical training process, including the practical training motion track of the student, the pressure of the student on a standing position, the parameter of a practical training tool, and the parameter of a workpiece; calculating a joint angle of the trainee according to the trainee practical training action track; calculating the center-of-gravity shift of the trainee through the pressure of the trainee on the standing position; calculating track coordinates of the tool according to the parameters of the practical training tool; constructing a dynamic causal matrix W, wherein the dynamic causal matrix W represents the influence weight of the joint angle of the student, the center-of-gravity shift and the trajectory coordinate of the tool on the workpiece parameters; and judging factors influencing workpiece parameters according to the dynamic causal matrix W. The processing behavior detection device has the advantages that the processing behavior detection device is used for detecting the processing behavior in the machining practical training process to find the influence of the processing behavior on the workpiece parameters, the development of machining practical training work is facilitated, and the practical training efficiency and level are improved.

Description

Technical Field

[0001] This application relates to the field of industrial Internet of Things (IoT) technology, and in particular to a method and system for detecting machining training behavior. Background Technology

[0002] In machining training, such as filing or sawing, students often produce parts that do not meet requirements due to incorrect posture or force application, such as out-of-tolerance dimensions or incorrect shapes. During filing or sawing practice, parts are inspected by teachers or students using ordinary measuring tools like calipers and rulers. Teachers typically observe and correct students' filing or sawing postures on-site. Students have no way of knowing what posture, force application, or surface finish will produce. They also don't understand the causes of uneven surfaces or slanted saw cuts. Teachers, relying solely on the characteristics of the parts, find it difficult to accurately determine the cause of the error. Furthermore, students lack the tools or channels to learn about the condition of parts, filing and sawing knowledge and skills, and how to find the causes of errors during practical training.

[0003] In summary, existing technologies lack solutions for detecting and judging operational behaviors during machining training. Summary of the Invention

[0004] This application provides a method and system for detecting machining training behavior. Its advantage is that it can be used to detect machining behavior during machining training to discover the impact of machining behavior on workpiece parameters, which is conducive to the development of machining training and improves training efficiency and level.

[0005] The technical solution of this application is as follows:

[0006] On the one hand, this application provides a method for detecting machining training behavior, including the following steps:

[0007] S1: Collect status data of trainees during machining training, including trainee movement trajectory, pressure exerted by trainees at their positions, parameters of training tools, and parameters of workpieces;

[0008] S2: Calculate the joint angles of the trainee based on the trainee's movement trajectory; calculate the trainee's center of gravity shift based on the pressure exerted by the trainee on the standing position; calculate the trajectory coordinates of the training tool based on the parameters of the training tool;

[0009] S3: Construct a dynamic causal matrix W, which represents the influence weights of the trainee's joint angles, center of gravity offset, and tool trajectory coordinates on the workpiece parameters;

[0010] S4: Determine the factors affecting workpiece parameters based on the dynamic causal matrix W.

[0011] Furthermore, step S0 is included before step S1: constructing a machining training station.

[0012] Reflective markings were placed at the trainees' joints;

[0013] Deploy cameras at the machining training workstations;

[0014] An array of pressure sensors was installed at the trainees' standing positions;

[0015] Deploy laser displacement sensors at the machining training station;

[0016] A workpiece parameter detection device is installed at the machining training station;

[0017] In step S1, the trainee's training motion trajectory is obtained through a camera, the pressure exerted by the trainee on the standing position is obtained through a pressure sensor array, the parameters of the training tool are obtained through a laser displacement sensor, and the workpiece parameters are obtained through a laser displacement sensor and / or a parameter detection device.

[0018] Furthermore, the workpiece parameters include workpiece surface roughness, key workpiece dimensions, and workpiece shape, and the workpiece parameter detection device is a roughness meter.

[0019] Furthermore, the dynamic causal matrix W(T,ζ) is a 3×3 real matrix, with rows corresponding to posture factors, center of gravity offset factors, and tool balance factors, and columns corresponding to the first, second, and third types of defects in the workpiece. Matrix elements w ij (T,ζ) quantifies the correlation strength between the i-th type of influencing factor and the j-th type of quality defect at a certain time T and a certain working condition ζ; where:

[0020] Postural factor F1(P) represents the actual angles θ of each joint of the trainee. k Compared to the standard joint angle θ k0 The normalized joint angle deviation is obtained by the following formula:

[0021]

[0022] θ max The maximum allowable deviation angles of the joints are set for the training procedures, with k=1,2,3 corresponding to the wrist, elbow, and shoulder, respectively.

[0023] The center of gravity offset factor F2(G) is obtained by normalizing the difference between the actual coordinates (x, y) and the standard center of gravity coordinates (x0, y0) of the trainee's pressure center of gravity, and the formula is:

[0024]

[0025] d max The maximum allowable center of gravity offset distance;

[0026] The tool balance factor F3(T) is the standard deviation σ of the tool trajectory coordinates and the standard deviation σ of the maximum allowable tool jitter. max The ratio is given by the formula:

[0027]

[0028] The first type of defect, Q1, is the workpiece surface roughness deviation coefficient, which is obtained by the following formula:

[0029]

[0030] R a For the collected workpiece surface roughness, R a0 The surface roughness standard value specified in the training task sheet;

[0031] The second type of defect, Q2, is the dimensional deviation coefficient, which is obtained by the following formula:

[0032]

[0033] L represents the key dimensions of the workpiece collected, and L0 represents the dimensions marked on the drawing.

[0034] The third type of defect, Q3, is the shape deviation coefficient, obtained by the following formula:

[0035]

[0036] S is the maximum deviation between the workpiece shape and the standard shape. max To allow for the maximum shape deviation.

[0037] Furthermore, the obtained posture factors, center of gravity offset factors, tool balance factors, Type I defects, Type II defects, and Type III defects are normalized using the following formula:

[0038]

[0039] Where X represents the original data, X min X max These are the minimum and maximum values ​​of the indicator, determined by statistical analysis of 50 sets of preliminary experimental data.

[0040] Furthermore, step S3 also includes the step of initially weighting the constructed dynamic causal matrix W:

[0041] An orthogonal experiment was conducted, with the experimental factors being postural factor F1, center of gravity shift factor F2, and tool balance factor F3. Each factor was set with 3 levels, and a total of 9 sets of experiments were designed. Each set of experiments was repeated 3 times by 5 trainees, and 27 sets of valid data were collected.

[0042] For each type of quality defect Q j (j=1,2,3), establish a multiple linear regression model for influencing factors and defects:

[0043] in Let be the initial weight coefficients to be solved. This is random error;

[0044] Minimize the sum of squared errors using the least squares method:

[0045]

[0046] The initial weight coefficients are obtained by solving.

[0047] Furthermore, step S3 also includes the step of dynamic weight correction:

[0048] (1) Dynamic weight correction triggering conditions: Three types of triggering conditions are set, and dynamic weight correction will be initiated when any one of the conditions is met:

[0049] ① If the change in working condition parameters exceeds the first set threshold, the working condition parameters here include the hardness of the processed material and the amount of tool wear. When the change in any parameter exceeds the corresponding range, it is determined to be a change in working condition.

[0050] ②The cumulative deviation of influencing factors exceeds the threshold δ, i.e.

[0051] δ

[0052] T0 represents the last update time, F i0 These are the standard values ​​of the influencing factors from the last update;

[0053] ③ If the fault diagnosis accuracy is lower than the second set threshold, the diagnosis accuracy is calculated by comparing the diagnosis results output by the comparison matrix with the actual detected defects;

[0054] (2) Gradient Descent Weight Update: The gradient descent method combined with L2 regularization is used for weight update. The objective function is:

[0055]

[0056] Where N is the data window length and λ is the regularization parameter;

[0057] The weight update formula is ;

[0058] Where η is the learning rate;

[0059] Iterative updates continue until the objective function converges, with the convergence condition being...

[0060] The updated dynamic weight matrix W(T+1,ζ) is obtained.

[0061] At this point, the updated dynamic weight matrix W(T+1,ζ) is obtained and synchronized to the edge computing industrial control computer.

[0062] Furthermore, step S4 also includes the following steps:

[0063] After determining the factors affecting the workpiece parameters based on the dynamic causal matrix W, the results are then presented to the trainees.

[0064] After determining the factors affecting the workpiece parameters based on the dynamic causal matrix W, the positions of these factors are highlighted in a 3D animation for interactive display to the trainees.

[0065] Furthermore, it also includes step S5:

[0066] The trainee's operating posture and tool movement trajectory are extracted based on the trainee's practical training action trajectory, and compared with the preset 3D virtual standard operating procedure. The differences between the trainee's operating posture and tool movement trajectory and the standard operating procedure are marked and displayed to the trainee.

[0067] On another front, this application provides a machining training behavior detection system for implementing the steps in the method described above, including:

[0068] The machining training station includes: reflective markers at the trainees' joints; cameras at the machining training station; pressure sensor arrays at the trainees' standing positions; laser displacement sensors at the machining training station; and workpiece parameter detection devices at the machining training station.

[0069] The data acquisition unit collects status data of trainees during machining training at the machining training station, including trainee movement trajectory, pressure exerted by trainees at the station, parameters of training tools, and parameters of workpieces.

[0070] The preprocessing layer includes an edge computing industrial computer that calculates the joint angles of the trainee based on the trainee's training motion trajectory; calculates the trainee's center of gravity shift based on the pressure exerted by the trainee on the standing position; and calculates the trajectory coordinates of the training tool based on the parameters of the training tool.

[0071] In the cloud, a dynamic causal matrix W is constructed, which represents the influence weights of the trainee's joint angles, center of gravity offset, and tool trajectory coordinates on workpiece parameters.

[0072] And the human-computer interaction layer, including an HMI touchscreen, for visual interaction with trainees.

[0073] In summary, the beneficial effects of this application are as follows:

[0074] By integrating data from vision, pressure, displacement, and roughness sensors, a comprehensive quantitative correlation between influencing factors and quality defects can be achieved, avoiding the limitations of single data.

[0075] A dynamic matrix is ​​constructed, and the "orthogonal test calibration + real-time gradient descent correction" method is adopted. The matrix weights are dynamically updated with the training time sequence and working conditions to adapt to different scenario changes, and the diagnostic accuracy is greatly improved.

[0076] Near real-time response, combined with cloud-based distributed computing and edge computing, matrix update response speed is fast, meeting the real-time fault diagnosis needs in the training process;

[0077] Visual feedback, through HMI touch screen and 3D virtual scene, intuitively displays defect types, error areas and standard operations, to help trainees quickly correct and improve training effectiveness;

[0078] It is highly scalable, supports the expansion of new training stations, and can complete matrix adaptation by adding new orthogonal experimental data, making it widely applicable.

[0079] This invention achieves efficient acquisition and preprocessing of multi-source data through a hierarchical training and testing system. Combined with the dynamic causal matrix construction method of "orthogonal experimental calibration + gradient descent correction", it realizes the dynamic quantitative correlation between influencing factors and quality defects, solves many shortcomings of existing training and testing systems, and provides a scientific and efficient testing and guidance scheme for mechanical training, with broad application prospects. Attached Figure Description

[0080] Figure 1 This is an architecture diagram of the machining training behavior detection system according to an embodiment of this application. Detailed Implementation

[0081] The specific embodiments of this application are described in detail below with reference to the accompanying drawings.

[0082] Example 1: A method for detecting machining training behavior, implemented through a machining training behavior detection system, as follows: Figure 1 As shown, it includes a data acquisition layer, a preprocessing layer, a human-computer interaction layer, and a cloud GPU cluster.

[0083] The aforementioned method for detecting machining training behavior includes the following steps:

[0084] S1: Collect status data during the student's machining training, including the student's training movement trajectory, the pressure exerted by the student on the training position, the parameters of the training tools, and the parameters of the workpiece. This step is implemented through the data acquisition layer.

[0085] First, construct the machining training station:

[0086] Reflective markings were placed at the trainees' joints;

[0087] Deploy cameras at the machining training workstations;

[0088] An array of pressure sensors was installed at the trainees' standing positions;

[0089] Deploy laser displacement sensors at the machining training station;

[0090] A workpiece parameter detection device is installed at the machining training station;

[0091] In step S1, the trainee's training motion trajectory is obtained through a camera, the pressure exerted by the trainee on the standing position is obtained through a pressure sensor array, the parameters of the training tool are obtained through a laser displacement sensor, and the workpiece parameters are obtained through a laser displacement sensor and / or a parameter detection device.

[0092] Specifically, the cameras are deployed through a distributed camera system. Cameras are placed in front, on both sides, behind, and other key locations at the training workstation (e.g., a fitter's workstation). Reflective markers are affixed to the trainees' wrists, elbows, shoulders, feet, etc. The cameras record the trainees' movement trajectories, including their stance, actions of holding tools to perform metal processing operations (e.g., filing, sawing), body posture, and tool movement trajectories. This distributed camera system is connected to the local industrial control computer at the data acquisition layer via a GigE Vision interface.

[0093] The sensor array places pressure-sensing pads at the operator's position in the training station to record the operator's center of gravity shifts in real time. The pressure sensor array is connected to the FPGA via an SPI interface, and the FPGA preprocesses the raw signals collected by the sensors.

[0094] A laser displacement sensor is deployed at the front of the training station to detect shape errors (such as flatness and straightness) and dimensional errors of the practice workpieces. The laser displacement sensor is connected to the local industrial control computer in the data acquisition layer via Wi-Fi / Bluetooth.

[0095] The workpiece parameters include workpiece surface roughness, key workpiece dimensions, and workpiece shape. The workpiece parameter detection device is a roughness meter. The roughness meter is deployed to the side of the training station and is used to detect the roughness of the machined surface of the workpiece. The roughness meter is connected to the local industrial control computer of the data acquisition layer via a USB interface.

[0096] S2: Calculate the trainee's joint angles based on the trainee's movement trajectory; calculate the trainee's center of gravity shift based on the pressure exerted by the trainee on their stance; calculate the trajectory coordinates of the training tool based on its parameters. This step is implemented through a preprocessing layer. The preprocessing layer includes an edge computing industrial control computer. The FPGA of the data acquisition layer communicates with the edge computing industrial control computer via a PCIe x8 interface; the local industrial control computer of the data acquisition layer communicates with the edge computing industrial control computer of the preprocessing layer via Gigabit Ethernet, realizing the aggregation and secondary preprocessing of multi-source acquired data.

[0097] The edge computing industrial control computer in the preprocessing layer transmits multi-source training data to the cloud GPU cluster through dual protocols. Specifically, the OPC UA protocol is used to transmit structured sensor data (such as workpiece shape and size errors detected by laser displacement sensors, and workpiece surface roughness data detected by roughness meters), and the MQTT protocol is used to transmit unstructured compressed video streams (such as real-time footage of trainees filing surfaces). All of the above data is simultaneously uploaded to the cloud GPU cluster.

[0098] The cloud-based PyTorch DDP distributed computing framework provides parallel computing support for the 3D scene generation engine and the physics simulation engine, generating 3D virtual training scenes that are consistent with real training conditions in near real time (for example, an interactive scene simulating a trainee's filing operation, which uses the physics simulation engine to restore the material properties of the workpiece, the physical feedback of the tool operation, and the environmental parameters in the real working conditions, ensuring the consistency between the virtual scene and the real training).

[0099] The cloud platform analyzes the compressed video stream to extract the trainee's operating postures and processing methods, and compares them with the preset standard operating procedures in the 3D virtual training scene (e.g., standard filing postures and file motion paths) to verify the correctness of the trainee's operation and identify errors. The cloud platform feeds back the error information to the edge computing industrial control computer in the preprocessing layer, which then transmits it to the HMI touchscreen via a communication interface to display the error points visually (e.g., text prompts such as "filing angle deviation", image annotations of incorrect posture areas, and highlighting of standard operating paths).

[0100] S3: Construct a dynamic causal matrix W, which represents the influence weights of the trainee's joint angles, center of gravity offset, and tool trajectory coordinates on the workpiece parameters.

[0101] The dynamic causal matrix W(T,ζ) is a 3×3 real matrix, with rows corresponding to posture factors, center of gravity offset factors, and tool balance factors, and columns corresponding to the first, second, and third types of defects in the workpiece. Matrix elements w ij (T,ζ) quantifies the correlation strength between the i-th type of influencing factor and the j-th type of quality defect at a certain time T and a certain working condition ζ; where:

[0102] Postural factor F1(P) represents the actual angles θ of each joint of the trainee. k Compared to the standard joint angle θ k0 The normalized joint angle deviation is obtained by the following formula:

[0103]

[0104] θ max The maximum allowable deviation angle of the joint is set for the training procedure. k=1,2,3 correspond to the wrist, elbow, and shoulder, respectively. The value of this coefficient is in the range of [0,1]. The closer it is to 1, the greater the posture deviation.

[0105] A distributed camera system acquires image data containing reflective markers at a frequency of 50fps. Using the OpenCV algorithm, the coordinates of the reflective markers at three key joints—the wrist, elbow, and shoulder—are extracted. These coordinates are then converted into three-dimensional spatial coordinates based on camera calibration parameters, allowing the calculation of the actual angle θ of each joint. k (k=1, 2, 3 correspond to the wrist, elbow, and shoulder respectively). The standard joint angle reference is the standard joint angle θ collected by three senior fitters after completing 10 standard filing operations. k0 (Wrist 15°, elbow 90°, shoulder 30°).

[0106] The center of gravity offset factor F2(G) is obtained by normalizing the difference between the actual coordinates (x, y) and the standard center of gravity coordinates (x0, y0) of the trainee's pressure center of gravity, and the formula is:

[0107]

[0108] d max This represents the maximum allowable center of gravity offset distance.

[0109] A pressure sensor array collects pressure distribution data from the trainee at a frequency of 100 Hz. After preprocessing using an FPGA Kalman filter (process noise covariance Q = 0.01, observation noise covariance R = 0.1), the actual coordinates (x, y) of the pressure center of gravity are extracted. These coordinates are then combined with the standard center of gravity coordinates (x0, y0) (determined by the workstation design parameters) and the maximum allowable center of gravity offset distance d. max The center of gravity offset coefficient F2 is calculated using formula G, with a value range of [0,1], reflecting the severity of the center of gravity offset.

[0110] The tool balance factor F3(T) is the standard deviation σ of the tool trajectory coordinates and the standard deviation σ of the maximum allowable tool jitter. max The ratio is given by the formula:

[0111]

[0112] A laser displacement sensor acquires three-dimensional motion trajectory data of the cutting edge of tools such as files at a frequency of 200 Hz. The standard deviation σ of the trajectory coordinates (reflecting the degree of tool vibration) is calculated, and this is combined with the standard deviation σ0 of the standard tool motion trajectory and the maximum allowable standard deviation σ of the tool vibration. max (Determined by the tool's precision level), the tool balance coefficient F3(T) is obtained through formula T, with a value range of [0,1]. The larger the value, the more severe the tool vibration.

[0113] The first type of defect, Q1, is the workpiece surface roughness deviation coefficient, which is obtained by the following formula:

[0114]

[0115] R a For the collected workpiece surface roughness, R a0 The surface roughness standard value is specified in the training task description; the roughness meter collects the R value of the machined surface of the workpiece every 10 seconds. a value.

[0116] The second type of defect, Q2, is the dimensional deviation coefficient, which is obtained by the following formula:

[0117]

[0118] L represents the key dimensions of the workpiece being acquired, and L0 represents the dimensions marked on the drawing. Specifically, the key dimensions L of the workpiece are several dimensions marked on the drawing, and the laser displacement sensor acquires the key dimensions L of the workpiece.

[0119] The third type of defect, Q3, is the shape deviation coefficient, obtained by the following formula:

[0120]

[0121] S is the maximum deviation between the workpiece shape and the standard shape. max To allow for the maximum shape deviation, a laser displacement sensor collects the coordinates of feature points on the workpiece surface. After fitting a standard shape using the least squares method, the maximum deviation S between the actual shape and the standard shape is obtained.

[0122] The coefficients for all three types of defects range from [0, 1].

[0123] Since the original calculation bases for the above factors and defect indices are different, in order to eliminate the influence of dimensions, the obtained posture factors, center of gravity offset factors, tool balance factors, first type of defect, second type of defect, and third type of defect are normalized, as shown in the following formula:

[0124]

[0125] Where X represents the original data, X min Xmax These are the minimum and maximum values ​​of the indicator, determined by statistical analysis of 50 sets of preliminary experimental data.

[0126] Initial weight calibration is performed on the dynamic causal matrix W:

[0127] Orthogonal experiment was adopted: To obtain the initial correlation strength between influencing factors and quality defects, L9 (3) was used. 3 An orthogonal array was used to design a three-factor, three-level orthogonal experiment. The experimental factors were posture factor F1, center of gravity shift factor F2, and tool balance factor F3. Each factor had three levels: Level 1 (F1, F2, F3, F4, F5, F6, F7, F8, F9, F1, F1, F2, F3 ...4, F1, F2, F3, i =0.1, no effect), Level 2 (F i =0.5, moderate impact), Level 3 (F i =0.9 (strong influence), a total of 9 groups of experiments were designed. Each group of experiments was repeated 3 times by 5 trainees. 27 groups of valid data were collected and outlier judgment was performed. Judgment criteria: data deviated from the group mean ±3σ. Outlier data were removed and supplemented data were collected.

[0128] Multiple linear regression analysis: for each type of quality defect Q j (j=1,2,3), establish a multiple linear regression model for influencing factors and defects:

[0129]

[0130] in Let be the initial weight coefficients to be solved. For random errors (satisfying a normal distribution N(0,σ)) 2 );

[0131] The 27 sets of standardized influencing factor data and defect data were input into the PyTorch framework of the cloud GPU cluster, and the sum of squared errors was minimized using the least squares method:

[0132]

[0133] The initial weight coefficients are obtained by solving.

[0134] Initial weight matrix verification: The initial weight matrix obtained by solving is: ;

[0135] The regression model was validated for goodness of fit, and the coefficient of determination was calculated. RSS j For the sum of squared residuals, TSS j The total sum of squares;

[0136] Requirement of the determination coefficient R 2 ≥0.85. Satisfies R 2The engineering requirement of ≥0.85 indicates that the model fits well. The initial weight matrix is ​​synchronized to the local cache of the edge computing industrial control computer as the basis for dynamic updates.

[0137] Dynamic weight adjustment:

[0138] (1) Dynamic weight correction triggering conditions: Three types of triggering conditions are set, and dynamic weight correction will be initiated when any one of the conditions is met:

[0139] ① If the change in working condition parameters exceeds the first set threshold, which is 5% in this embodiment, the working condition parameters include the hardness of the processed material and the amount of tool wear. When the change in any parameter exceeds the corresponding range, it is determined to be a change in working condition.

[0140] ②The cumulative deviation of influencing factors exceeds the threshold δ, i.e.

[0141] δ

[0142] T0 represents the last update time, F i0 These are the standard values ​​of the influencing factors from the last update;

[0143] ③ The fault diagnosis accuracy rate is lower than the second set threshold, which is 90% in this embodiment. The diagnosis accuracy rate is calculated by comparing the diagnosis results output by the comparison matrix with the actual detected defects.

[0144] (2) Gradient Descent Weight Update: Gradient descent combined with L2 regularization is used for weight update to avoid overfitting. The objective function is:

[0145]

[0146] Where N is the data window length and λ is the regularization parameter;

[0147] The weight update formula is ;

[0148] Where η is the learning rate;

[0149] Iterative updates continue until the objective function converges, with the convergence condition being...

[0150] The updated dynamic weight matrix W(T+1,ζ) is obtained and synchronized to the edge computing industrial control computer.

[0151] At this point, the updated dynamic weight matrix W(T+1,ζ) is obtained and synchronized to the edge computing industrial control computer.

[0152] Matrix validity verification:

[0153] Three typical training scenarios were selected for verification: filing low-carbon steel, filing cast iron, and sawing round steel. Each scenario corresponded to different working parameters (hardness of low-carbon steel 170 HB, hardness of cast iron 220 HB, and hardness of round steel 190 HB) and tool types (fine-tooth flat file for filing low-carbon steel, medium-tooth flat file for filing cast iron, and high-speed steel saw blade for sawing round steel). Ten trainees were assigned to each scenario for 50 training sessions, resulting in 150 sets of valid data. During each training session, the diagnostic results of the dynamic matrix, weight update data, and response time were recorded simultaneously. After the experiment, the verification indicators were analyzed as follows:

[0154] Model diagnostic accuracy Acc: Professionals use specialized testing instruments (roughness tester, laser interferometer, geometric tolerance measuring instrument) to inspect the quality of the processed workpiece, determine the actual defects, compare the diagnostic results output by the matrix, count the number of correct diagnoses, calculate the diagnostic accuracy for each scenario, and take the average as the final result.

[0155] Weight stability (Stab): Extract 100 consecutive update data of the matrix for each scenario and calculate the weight w for each element. ij Standard deviation σ(w) ij ) and average value μ(w ij ), obtain the coefficient of variation, and take the maximum value of the coefficients of variation of all elements as the weight stability index of the scenario.

[0156] Dynamic response speed Speed: Records the total time taken from data transmission to the cloud to synchronization of the updated matrix to the edge computing industrial control computer after each weight update is triggered. The time taken for 50 updates in each scenario is statistically analyzed, and the average value is taken as the dynamic response speed for that scenario.

[0157] If the model's diagnostic accuracy (Acc) is greater than or equal to 90%, its weight stability (Stab) is less than or equal to 5%, and its dynamic response speed (Speed) is less than or equal to 0.5 seconds, then the model is considered valid.

[0158] S4: Determine the factors affecting workpiece parameters based on the dynamic causal matrix W;

[0159] After determining the factors affecting the workpiece parameters based on the dynamic causal matrix W, the results are then presented to the trainees.

[0160] After determining the factors influencing the workpiece parameters based on the dynamic causal matrix W, the locations of these factors are highlighted in a 3D animation for interactive demonstration to the trainee. This step is implemented through a human-machine interface (HMI) layer. The HMI layer includes an HMI touchscreen, and the edge computing industrial computer in the preprocessing layer is connected to the HMI touchscreen via an HDMI interface for visualizing the test results.

[0161] S5: Extract the student's operating posture and tool movement trajectory based on the student's training action trajectory, compare it with the preset 3D virtual standard operating procedure, mark the differences between the student's operating posture and tool movement trajectory and the standard operating procedure, and interact with the student.

[0162] The following section combines practical training on filing a low-carbon steel workpiece (training task: machining a 100 mm × 50 mm × 10 mm flat plate, requiring a surface roughness R). a For specific scenarios (≤1.6 μm, dimensional tolerance ±0.02 mm, flatness ≤0.01 mm), this embodiment will be further described in detail to ensure that those skilled in the art can repeat the invention based on the following description.

[0163] (I) Complete System Deployment

[0164] 1. Data Acquisition Layer Deployment

[0165] A Basler acA2500-14gm industrial camera was selected and installed in front, on the left, on the right, and behind the fitter's workstation. Highly reflective markers with a diameter of 10 mm were affixed to appropriate positions on the trainee's wrists, elbows, shoulders, and feet. The camera was connected to a local industrial computer (Intel Core i5-10400F) via the GigE Vision interface, and the Faulhaber SDK software was configured to achieve real-time image acquisition and trajectory extraction.

[0166] A 4×4 array pressure sensing pad, model Tekscan Flexiforce A201, with a measurement range of 0-100 N, an accuracy of ±0.1 N, and a sampling frequency of 100 Hz, was used. It was placed in the exerciser's standing area and connected to a Xilinx Zynq-7000 series XC7Z020-1CLG484 FPGA development board via an SPI interface. The FPGA's built-in Kalman filter algorithm (process noise covariance Q=0.01, observation noise covariance R=0.1) performed noise reduction preprocessing on the raw pressure data.

[0167] The Keyence IL-300 laser displacement sensor was selected, with a resolution of ±0.001 mm and a sampling frequency of 200Hz. It was installed in a suitable position in front of the workstation, connected to the local industrial control computer via a Wi-Fi 6 module, and configured with Keyence OP-87051 software to collect the coordinates of feature points on the workpiece surface and the three-dimensional motion trajectory of the tool (such as a file) cutting edge.

[0168] The roughness tester selected is the Mitutoyo SJ-210. Considering the spatial layout and ease of operation of the training station, it is fixed on the right side of the station. The instrument is connected to the local industrial computer via a USB 3.0 interface. It is set to automatically collect surface roughness data of the workpiece after each filing operation. To ensure accuracy, the data collection point is located in the center area of ​​the machined surface of the workpiece.

[0169] 2. Preprocessing layer deployment

[0170] The edge computing industrial PC uses the Advantech ARK-3520, equipped with an Intel Core i7-12700K processor and an NVIDIA RTX 3080 10GB graphics card to meet the requirements of parallel processing of multi-source data. The edge computing industrial PC interfaces with the FPGA module via a PCIe x8 interface to receive pre-processed center of gravity pressure data. Simultaneously, it connects to the local industrial PC in the data acquisition layer via a gigabit Ethernet interface to acquire camera video streams, laser displacement sensor data, and roughness meter detection data. On the software side, the industrial PC comes pre-installed with a Python 3.9 runtime environment, using OpenCV 4.5.5 to perform secondary noise reduction processing on the video stream; and uses the Pandas library to standardize the format of all multi-source data, uniformly converting it to CSV format, and controlling the timestamp accuracy to 1 ms to ensure the time synchronization of different data.

[0171] 3. Deployment of the Human-Computer Interaction Layer

[0172] To facilitate human-computer interaction during practical training, an industrial-grade high-definition HMI touchscreen was selected and connected to the edge computing industrial control computer via an HDMI 2.0 interface. Based on the actual training scenario, the touchscreen display interface adopts a partitioned layout design: the left area displays the real-time processing video, overlaid with the motion trajectory of reflective markers, allowing trainees to easily observe their own actions; the middle area uses red highlighting to display text prompts on defect types and dominant influencing factors, quickly reminding trainees to correct operational problems; the right area displays a 3D virtual standard operation animation for trainees' reference; and the bottom area presents a comparison bar between real-time quality indicators (covering parameters such as roughness Ra value, actual dimensions, and flatness) and standard values, allowing trainees to quickly judge the processing quality.

[0173] 4. Cloud GPU Cluster Deployment

[0174] The GPU cluster uses Huawei TaiShan 200 servers, each equipped with an NVIDIA A100 80GB high-performance GPU, an AMD EPYC 7763 processor, 256GB of RAM, and a 4TB solid-state drive. Its computing power and storage are sufficient for high-intensity scenarios. The cluster runs Ubuntu 20.04 LTS, with PyTorch 2.0.1 supporting DDP distributed training, Unreal Engine 5.1 for 3D scene creation, and PhysX 5.1 for physics simulation. Additionally, an EMQX 4.4.8 MQTT server and a FreeOpcUa 1.4.0 OPC UA server are set up to facilitate data transfer between devices.

[0175] For cluster network construction, the InfiniBand HDR high-speed interconnect solution was selected, leveraging its 200Gbps high bandwidth advantage to control the latency of distributed computing within 1ms.

[0176] (II) Implementation of Offline Initialization Phase

[0177] 1. Standard Database Construction

[0178] Three senior fitter technicians were invited to perform 10 standard filing operations. Standard data was collected during the process, including the standard joint angle θ. k0 (Wrist 15°, Elbow 90°, Shoulder 30°), the maximum allowable deviation angle θ of the joints as set in the training procedure. max =20°, standard centroid coordinates (x0 = 300 mm, y0 = 200 mm), maximum allowable centroid offset distance d max =50 mm, standard tool motion trajectory standard deviation σ0=0.002 mm, maximum allowable tool vibration standard deviation σ max =0.01 mm; Set quality standard value: R a0 =1.6 μm, L0 =100 mm, S max =0.01 mm. The above standard data is stored in a cloud-based MySQL database, and the data tables are classified according to "training scenario - standard type", such as "filing - low carbon steel - posture standard".

[0179] 2. Initial weight matrix calibration

[0180] Orthogonal experiments were conducted according to L9 (3 3 Nine experiments were designed using an orthogonal array. The experimental factors were posture factor F1, center of gravity shift factor F2, and tool balance factor F3. Each factor had three levels: level 1 (F1, F2, F3, F4, F5, F6, F7, F8, F9, F1, F1, F2, F3, F1, F2, F3, F4 ... i =0.1, no effect), Level 2 (F i=0.5, moderate impact) and level 3 (F i =0.9, strong influence).

[0181] Each experiment was repeated 3 times by 5 practitioners of different skill levels, resulting in 27 sets of valid data. Outliers were removed; the criteria for outliers was a deviation of the data from the group mean ±3σ.

[0182] The initial weights were determined by running a multiple linear regression algorithm on a cloud GPU cluster using the PyTorch framework. The input consisted of 27 sets of standardized F1, F2, F3 (influencing factors) and Q1, Q2, Q3 (quality defects) data, yielding the initial weight matrix.

[0183]

[0184] Calculate the coefficient of determination R of the regression model 2 R1 2 =0.88 (Q1), R2 2 =0.91 (Q2), R3 2 =0.89 (Q3), all satisfying R 2 The requirement of ≥0.85 will be applied to W. 0 Synchronize to the local cache of the edge computing industrial control computer.

[0185] 3. Dynamically adjust parameter configuration

[0186] In the configuration file config.ini of the edge computing industrial control computer, set the following parameters: operating condition parameter threshold of 5% (covering material hardness ±10 HB and tool wear ±0.05 mm), deviation accumulation threshold δ=0.3, learning rate η=0.005, regularization parameter λ=0.01, data window length N=100, and weight convergence threshold of 10. −6 .

[0187] (III) Implementation of the online real-time operation phase

[0188] 1. Multi-source data acquisition and transmission

[0189] After the trainee begins filing, each sensor collects data at a set frequency: the camera acquires one frame of image every 20 ms (50fps), the pressure sensor acquires pressure distribution data every 10 ms (100 Hz), the laser displacement sensor acquires trajectory and dimension data every 5 ms (200 Hz), and the roughness meter acquires roughness data every 10 s. The FPGA performs Kalman filtering on the pressure data and uploads it to the edge computing industrial control computer via the PCIe x8 interface at a period of 100 ms. The local industrial control computer compresses the camera video stream into H.265 format, converts the laser displacement sensor data into JSON format, and converts the roughness meter data into TXT format, and uploads them to the edge computing industrial control computer via Gigabit Ethernet at a period of 50 ms.

[0190] After the edge computing industrial control computer aggregates the data, it transmits the structured data (P, G, T, R) via the OPC UA protocol (transmission cycle 100 ms). a The unstructured video stream (L,S) is uploaded to the cloud via the MQTT protocol (QoS=1, transmission period 200 ms). The actual transmission latency is 85 ms, which meets the real-time requirements.

[0191] 2. Cloud-based data processing and dynamic matrix updates

[0192] The 3D scene is generated by using a cloud GPU cluster and Unreal Engine 5.1 to load a pre-configured "filing - low carbon steel" training scene template. At the same time, the PhysX 5.1 physics engine uses the tool trajectory data collected by the laser displacement sensor to recreate the dynamic process of the contact and cutting between the file and the workpiece, generating a 3D virtual scene synchronized with the real operation. The synchronization latency of the entire scene is controlled within 100 ms.

[0193] To quantify influencing factors and defects, the cloud-based YOLOv8 algorithm was used to extract the coordinates of reflective markers from the video stream and calculate the joint angle deviation. The measured deviation coefficient at a certain moment was P=0.62. Combined with the collected pressure data, the coordinates of the center of gravity (x=312 mm, y=208 mm) were extracted, and the center of gravity offset coefficient G=0.32 was obtained through algorithm calculation. Based on the tool motion trajectory data recorded by the laser displacement sensor, the trajectory standard deviation σ=0.005 mm was calculated, and the corresponding trajectory stability coefficient T=0.25 was obtained. Based on the workpiece surface roughness value Ra=2.2 μm measured by the roughness meter, the surface quality quantification index Q1=0.375 was calculated. Then, using the dimensional data L=100.03 mm and S=0.002 mm collected by the laser displacement sensor, the dimensional deviation quantification index Q2=0.15 and the form and position deviation quantification index Q3=0.2 were calculated, respectively.

[0194] The dynamic correction was triggered and executed when the measured hardness of the processed material was 180 HB, compared to the standard value of 170 HB, representing a hardness change of 5.88%, exceeding the set threshold of 5%. The system then triggered a weight update. The gradient descent algorithm was invoked from the cloud, and the most recent 100 sets of data (N=100) were input for iterative calculation. After 120 iterations, the objective function J(W) satisfied the convergence condition |J 120 -J 119 | = 8.2 × 10 −7 <10 -6 Thus, the updated dynamic matrix is ​​obtained as follows:

[0195]

[0196] The updated matrix is ​​synchronized to the edge computing industrial control computer. The update process takes 0.35 s < 0.5 s, which meets the requirements for dynamic response.

[0197] (iv) Implementation of Fault Diagnosis and Result Feedback

[0198] 1. Fault diagnosis calculation

[0199] During operation, the edge computing industrial control computer receives the dynamic weight matrix W(T,ζ) synchronously transmitted from the cloud in real time. Simultaneously, it combines this with preprocessed real-time influencing factor data F1, F2, and F3, and then uses the formula... Calculate the correlation strength S of various quality defects. j Among them, S j The value range is between [0,1]. The defect judgment threshold is set to 0.7 (this threshold was obtained through ROC curve analysis of 50 sets of preliminary experimental data; the experiment shows that when S...). j When the value is ≥0.7, the probability of an actual defect occurring exceeds 80%. The specific judgment rules are as follows:

[0200] When S j When the value is ≥0.7, it can be determined that the j-th type of process quality defect has occurred. At this time, the influencing factor with the strongest correlation is selected as the dominant factor (i.e., the weight element w corresponding to this factor). ij (Maximum value)

[0201] When 0.5≤S j When the value is less than 0.7, it indicates that the corresponding defect is in a critical state, and the system will prompt the trainee to adjust the operation method in time.

[0202] When S j When the value is less than 0.5, it indicates that the risk of this type of defect is low and no special warning is required.

[0203] The following example, taken at a specific real-time monitoring moment during the filing of low-carbon steel, illustrates the diagnostic process:

[0204] The real-time influencing factor data at a certain moment, after being collected by the sensor and preprocessed, are as follows: F1=0.62 (corresponding to an elbow angle deviation of 12°), F2=0.32 (corresponding to a center of gravity offset of 8 mm), F3=0.25 (corresponding to a tool trajectory standard deviation of 0.0025 mm).

[0205] Under the current operating conditions (the hardness of the processed material is 180 HB, which deviates from the standard value by 5.88%), the updated dynamic matrix is ​​as follows:

[0206]

[0207] Based on the updated dynamic matrix, the edge computing industrial control computer calculates the correlation strength of the three types of defects:

[0208] S1 =0.75×0.62+0.20×0.32+0.12×0.25=0.465+0.064+0.03=0.559;

[0209] S2=0.23×0.62+0.70×0.32+0.13×0.25=0.1426+0.224+0.0325=0.3991;

[0210] S3=0.17×0.62+0.25×0.32+0.72×0.25=0.1054+0.08+0.18=0.3654;

[0211] Final diagnosis results: S1=0.559 is in the range of [0.5,0.7), so the surface roughness defect is judged to be in a critical state; S2 and S3 are both less than 0.5, indicating that there is no risk of dimensional deviation or shape deviation at present; further analysis of the dominant influencing factors found that F1 has the largest weight (0.75), so it can be determined that the posture deviation is the main reason for the critical deviation of surface roughness.

[0212] 2. Visual feedback display

[0213] The HMI touchscreen displays targeted feedback according to preset interface zones. In the left video area, the position of the elbow joint is located through image recognition and highlighted with a red rectangle, while the text prompt "Elbow angle deviation 12° (standard 90°)" is superimposed on the screen. This allows the user to clearly see the specific location of the error.

[0214] In the central text prompt area, highlighted in red, it states, "The current surface roughness is approaching the critical value (Ra=2.2 μm, standard Ra≤1.6 μm), the dominant factor being filing posture deviation." Below this prompt, specific operational instructions are provided: "It is recommended to adjust the elbow posture, keeping the elbow parallel to the workpiece surface and reducing the vertical swing amplitude." This helps learners quickly find the correct correction direction.

[0215] In the 3D animation area on the right, an animation of the "standard posture for filing low-carbon steel" is loaded synchronously, with the movement trajectory of the elbow highlighted. This allows practitioners to easily observe the details of the standard movement.

[0216] The bottom indicator comparison area features an intuitive comparison bar design. The roughness comparison bar uses red to indicate the current Ra value (2.2 μm) and green to indicate the standard value (1.6 μm), allowing users to easily see the degree of deviation. The size and flatness comparison bar uses green to display the current values ​​(size 100.03 mm, flatness 0.002 mm) and is marked "Meets requirements".

[0217] It should be understood that the specific data or product types shown in this embodiment are exemplary, and those skilled in the art can use other values ​​or product types under the inventive concept of this application.

[0218] Example 2: A machining training behavior detection system, used to implement the steps in the method described in Example 1, including:

[0219] The machining training station includes: reflective markers at the trainees' joints; cameras at the machining training station; pressure sensor arrays at the trainees' standing positions; laser displacement sensors at the machining training station; and workpiece parameter detection devices at the machining training station.

[0220] The data acquisition unit collects status data of trainees during machining training at the machining training station, including trainee movement trajectory, pressure exerted by trainees at the station, parameters of training tools, and parameters of workpieces.

[0221] The preprocessing layer includes an edge computing industrial computer that calculates the joint angles of the trainee based on the trainee's training motion trajectory; calculates the trainee's center of gravity shift based on the pressure exerted by the trainee on the standing position; and calculates the trajectory coordinates of the training tool based on the parameters of the training tool.

[0222] In the cloud, a dynamic causal matrix W is constructed, which represents the influence weights of the trainee's joint angles, center of gravity offset, and tool trajectory coordinates on workpiece parameters.

[0223] And the human-computer interaction layer, including an HMI touchscreen, for visual interaction with trainees.

[0224] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the inventive concept of this application, and these all fall within the protection scope of this application.

Claims

1. A method for detecting machining training behavior, characterized in that, Includes the following steps: S1: Collect status data of trainees during machining training, including trainee movement trajectory, pressure exerted by trainees at their positions, parameters of training tools, and parameters of workpieces; S2: Calculate the joint angles of the trainee based on the trainee's movement trajectory; calculate the trainee's center of gravity shift based on the pressure exerted by the trainee on the standing position; calculate the trajectory coordinates of the training tool based on the parameters of the training tool; S3: Construct a dynamic causal matrix W, which represents the influence weights of the trainee's joint angles, center of gravity offset, and tool trajectory coordinates on the workpiece parameters; S4: Determine the factors affecting workpiece parameters based on the dynamic causal matrix W.

2. The method for detecting machining training behavior according to claim 1, characterized in that, The step S0 preceding step S1 is: constructing a machining training station. Reflective markings were placed at the trainees' joints; Deploy cameras at the machining training workstations; An array of pressure sensors was installed at the trainees' standing positions; Deploy laser displacement sensors at the machining training station; A workpiece parameter detection device is installed at the machining training station; In step S1, the trainee's training motion trajectory is obtained through a camera, the pressure exerted by the trainee on the standing position is obtained through a pressure sensor array, the parameters of the training tool are obtained through a laser displacement sensor, and the workpiece parameters are obtained through a laser displacement sensor and / or a parameter detection device.

3. The method for detecting machining training behavior according to claim 2, characterized in that, The workpiece parameters include workpiece surface roughness, key workpiece dimensions, and workpiece shape, and the workpiece parameter detection device is a roughness meter.

4. The method for detecting machining training behavior according to claim 3, characterized in that, The dynamic causal matrix W(T,ζ) is a 3×3 real matrix, with rows corresponding to posture factors, center of gravity offset factors, and tool balance factors, and columns corresponding to the first, second, and third types of defects in the workpiece. Matrix elements w ij (T,ζ) quantifies the correlation strength between the i-th type of influencing factor and the j-th type of quality defect at a certain time T and a certain working condition ζ; where: Postural factor F1(P) represents the actual angles θ of each joint of the trainee. k Compared to the standard joint angle θ k0 The normalized joint angle deviation is obtained by the following formula: θ max The maximum allowable deviation angles of the joints are set for the training procedures, with k=1,2,3 corresponding to the wrist, elbow, and shoulder, respectively. The center of gravity offset factor F2(G) is obtained by normalizing the difference between the actual coordinates (x, y) and the standard center of gravity coordinates (x0, y0) of the trainee's pressure center of gravity, and the formula is: d max The maximum allowable center of gravity offset distance; The tool balance factor F3(T) is the standard deviation σ of the tool trajectory coordinates and the standard deviation σ of the maximum allowable tool jitter. max The ratio is given by the formula: The first type of defect, Q1, is the workpiece surface roughness deviation coefficient, which is obtained by the following formula: R a For the collected workpiece surface roughness, R a0 The surface roughness standard value specified in the training task sheet; The second type of defect, Q2, is the dimensional deviation coefficient, which is obtained by the following formula: L represents the key dimensions of the workpiece collected, and L0 represents the dimensions marked on the drawing. The third type of defect, Q3, is the shape deviation coefficient, obtained by the following formula: S is the maximum deviation between the workpiece shape and the standard shape. max To allow for the maximum shape deviation.

5. The method for detecting machining training behavior according to claim 4, characterized in that, The obtained posture factors, center of gravity offset factors, tool balance factors, Type I defects, Type II defects, and Type III defects are normalized using the following formula: Where X represents the original data, X min X max These are the minimum and maximum values ​​of the indicator, determined by statistical analysis of 50 sets of preliminary experimental data.

6. The method for detecting machining training behavior according to claim 1, characterized in that, Step S3 further includes the step of initial weight calibration of the dynamic causal matrix W: An orthogonal experiment was conducted, with the experimental factors being postural factor F1, center of gravity shift factor F2, and tool balance factor F3. Each factor was set with 3 levels, and a total of 9 sets of experiments were designed. Each set of experiments was repeated 3 times by 5 trainees, and 27 sets of valid data were collected. For each type of quality defect Q j (j=1,2,3), establish a multiple linear regression model for influencing factors and defects: in Let be the initial weight coefficients to be solved. This is random error; Minimize the sum of squared errors using the least squares method: The initial weight coefficients are obtained by solving.

7. The method for detecting machining training behavior according to claim 6, characterized in that, Step S3 further includes the step of dynamic weight correction: (1) Dynamic weight correction triggering conditions: Three types of triggering conditions are set, and dynamic weight correction will be initiated when any one of the conditions is met: ① If the change in working condition parameters exceeds the first set threshold, the working condition parameters here include the hardness of the processed material and the amount of tool wear. When the change in any parameter exceeds the corresponding range, it is determined to be a change in working condition. ②The cumulative deviation of influencing factors exceeds the threshold δ, i.e. d T0 represents the last update time, F i0 These are the standard values ​​of the influencing factors from the last update; ③ If the fault diagnosis accuracy is lower than the second set threshold, the diagnosis accuracy is calculated by comparing the diagnosis results output by the comparison matrix with the actual detected defects; (2) Gradient Descent Weight Update: The gradient descent method combined with L2 regularization is used for weight update. The objective function is: Where N is the data window length and λ is the regularization parameter; The weight update formula is ; Where η is the learning rate; Iterative updates continue until the objective function converges, with the convergence condition being... The updated dynamic weight matrix W(T+1,ζ) is obtained. At this point, the updated dynamic weight matrix W(T+1,ζ) is obtained and synchronized to the edge computing industrial control computer.

8. The method for detecting machining training behavior according to claim 1, characterized in that, Step S4 also includes the following steps: After determining the factors affecting the workpiece parameters based on the dynamic causal matrix W, the results are then presented to the trainees. After determining the factors affecting the workpiece parameters based on the dynamic causal matrix W, the positions of these factors are highlighted in a 3D animation for interactive display to the trainees.

9. The method for detecting machining training behavior according to claim 1, characterized in that, It also includes step S5: The trainee's operating posture and tool movement trajectory are extracted based on the trainee's practical training action trajectory, and compared with the preset 3D virtual standard operating procedure. The differences between the trainee's operating posture and tool movement trajectory and the standard operating procedure are marked and displayed to the trainee.

10. A machining training behavior detection system, characterized in that, To implement the steps in the method as described in any one of claims 1-9, comprising: The machining training station includes: reflective markers at the trainees' joints; cameras at the machining training station; pressure sensor arrays at the trainees' standing positions; laser displacement sensors at the machining training station; and workpiece parameter detection devices at the machining training station. The data acquisition unit collects status data of trainees during machining training at the machining training station, including trainee movement trajectory, pressure exerted by trainees at the station, parameters of training tools, and parameters of workpieces. The preprocessing layer includes an edge computing industrial computer that calculates the joint angles of the trainee based on the trainee's training motion trajectory; calculates the trainee's center of gravity shift based on the pressure exerted by the trainee on the standing position; and calculates the trajectory coordinates of the training tool based on the parameters of the training tool. In the cloud, a dynamic causal matrix W is constructed, which represents the influence weights of the trainee's joint angles, center of gravity offset, and tool trajectory coordinates on workpiece parameters. And the human-computer interaction layer, including an HMI touchscreen, for visual interaction with trainees.

Images