Hydraulic clamp die multi-process machining process regulation and control method based on image detection

Abstract

The invention discloses a hydraulic clamp die multi-process machining process regulation and control method based on image detection. The method comprises the following steps that an image acquisition station is arranged after a die machining process to acquire image data; executing edge extraction, defect segmentation and texture analysis based on the image data, and outputting a quality index vector; inputting the quality index vector into a process condition liquid state time constant neural network, outputting a grading result and generating a corresponding process regulation and control instruction; when the grading result is in an adjusting area, feedback adjustment is conducted on cutter compensation, feeding, the rotating speed of a main shaft and clamp positioning parameters; when the grading result is a border crossing area, generating a local re-processing subtask and executing image detection again after the local re-processing subtask is completed until the detection result enters an adjusting area or a qualified area; and storing the quality index vector, the process regulation and control instruction and the process parameters in a versioning manner, and updating the process map based on a final inspection result. According to the invention, closed-loop fusion of quality detection and process regulation and control is realized, and the processing precision and stability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mold processing process control, and particularly relates to a hydraulic clamp mold multi-process processing process regulation method based on image detection. BACKGROUND

[0002] As a key component of hydraulic clamp tools, the machining precision of the hydraulic clamp mold directly affects the clamping performance and service life of the hydraulic clamp. The existing mold machining usually includes multiple processes such as rough machining, drilling, heat treatment, finishing, electrical discharge machining and polishing. The errors between different processes are easy to accumulate, and the traditional quality control mainly relies on manual detection or single-point measurement, which is difficult to realize continuous tracking and accurate quantification of dynamic errors between complex processes.

[0003] With the development of numerical control machining and image detection technology, existing methods attempt to use image recognition for defect detection and size analysis, but generally have the following shortcomings: first, multiple quality information cannot be uniformly modeled, making it difficult to form comprehensive quality characterization; second, process regulation mostly relies on fixed rules or static models, lacking self-adaptive ability to process conditions and time factors; third, when the detection results are fed back to the process level, there is a lack of fine-grained classification mechanism and closed-loop control, resulting in low rework efficiency and long-term accumulation of process optimization.

[0004] Therefore, how to provide a hydraulic clamp mold multi-process processing process regulation method based on image detection is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a hydraulic clamp mold multi-process processing process regulation method based on image detection. The present application provides a hydraulic clamp mold multi-process processing process regulation method based on image detection, which realizes dynamic perception, hierarchical judgment, fine regulation and process map adaptive update of the processing process.

[0006] A hydraulic clamp mold multi-process processing process regulation method based on image detection according to an embodiment of the present application includes the following steps:

[0007] An image acquisition station is arranged after the processing process of the hydraulic clamp mold to obtain image data of the mold processing part;

[0008] Edge extraction, defect segmentation and texture analysis are performed based on the image data to output a quality index vector;

[0009] The quality index vector is input into a process condition liquid time constant neural network, and the output hierarchical result determines that the quality index is in a qualified zone, an adjustment zone or an out-of-bound zone, and corresponding process regulation instructions are generated according to the hierarchical result;

[0010] When the grading result is the adjustment area, the adjustment process control instruction includes in-situ feedback adjustment of tool compensation, feed, spindle speed or fixture positioning parameters of the current process;

[0011] When the grading result is the out-of-bound area, the out-of-bound process control instruction includes generating a local re-machining sub-task and performing image detection again after completion, until the detection result enters the adjustment area or the qualified area;

[0012] The quality index vector, process control instruction and process parameter are versioned and stored, and the process map is updated based on the final inspection result.

[0013] Further, the machining process includes rough machining, drilling, heat treatment, finishing, electrical discharge machining and polishing of the hydraulic clamp mold.

[0014] Further, the generation step of the quality index vector includes:

[0015] Edge detection and geometric fitting are performed on the image data, and size deviation and geometric error are output;

[0016] Image segmentation is performed on the image data, and burr length and defect area are calculated by identifying burr and surface defect areas;

[0017] Frequency domain analysis is performed on the image data, texture features are extracted, and surface roughness proxy values are mapped;

[0018] Feature matching is performed on the mold images collected before and after different processes, and key point offset is calculated;

[0019] The size deviation, geometric error, burr length, defect area, roughness proxy value and key point offset are combined to form the quality index vector.

[0020] Further, the generation step of the grading result includes:

[0021] The quality index vector sequence arranged in process time sequence and corresponding process identifier, process time interval are input into the process condition liquid time constant neural network;

[0022] The process condition liquid time constant neural network includes:

[0023] The input encoding unit performs normalization processing on the quality index vector and splices the process identifier and time interval to form an input tensor;

[0024] The time constant generation unit dynamically generates the time constant and coupling weight of each liquid neuron according to the process identifier and time interval;

[0025] The liquid state update unit performs continuous time state update according to the time constant;

[0026] The cost-sensitive decoding unit outputs a confidence value based on the classification results of the misjudgment cost matrix for the qualified area, the adjustment area, or the out-of-bound area;

[0027] When the confidence value is lower than a threshold value, the control instruction is prohibited from being issued, and an image reacquisition or manual review process is triggered.

[0028] Further, the process condition liquid time constant neural network comprises:

[0029] A process embedding layer is arranged at the input end to map the process number to an embedding vector and splice the embedding vector with the quality index vector for input;

[0030] A process condition hypernetwork is configured to dynamically generate the time constant and coupling weight of the liquid neuron according to the process identifier and the time interval between processes;

[0031] A cost-sensitive decoding head is arranged at the output end in parallel with a confidence estimation branch, the preset cost matrix is used to weight the classification loss, and the classification result and the heteroscedastic confidence value are output;

[0032] A low-rank adapter weight is introduced in the hidden layer.

[0033] Further, a monotonicity constraint module is arranged between the output layer of the process condition liquid time constant neural network and the input quality index, so that the mapping relationship between the burr length and the out-of-bound probability, the key point offset and the out-of-bound probability, and the roughness proxy value and the adjustment probability is monotonically increasing, and a penalty term is applied to the samples that violate the constraint during the training process.

[0034] Further, the adjustment process control instruction comprises:

[0035] A tool radius compensation amount is calculated based on the size deviation, and the tool radius compensation amount is written into the tool compensation parameter table of the numerical control system;

[0036] A feed speed correction value is calculated based on the form and position error, and the feed speed correction value is limited within the upper and lower limits of the feed speed allowed by the machine tool;

[0037] A spindle speed correction amount is calculated based on the surface roughness proxy value, the spindle speed is set to a correction value higher than the baseline when the roughness proxy value is higher than the baseline, and the spindle speed is set to a correction value lower than the baseline when the roughness proxy value is lower than the baseline;

[0038] A jig positioning compensation amount is calculated based on the key point offset, and the jig positioning compensation amount is written into the tool setting table of the tooling;

[0039] The tool radius compensation amount, the jig positioning compensation amount, the feed speed correction value, and the spindle speed correction amount are synchronized to the next process card.

[0040] Further, the out-of-bound process regulation instruction comprises:

[0041] Generate a local re-machining subtask based on the dimensional deviation and geometric error, and set a local machining allowance and a path overlap rate in the local re-machining subtask;

[0042] Generate a local re-machining subtask based on the burr length, and set a chamfer machining path and a chamfer depth;

[0043] Generate a local re-machining subtask based on the defect area, and set a pulse width and a gap voltage of the electrical discharge machining;

[0044] Collect an image again after the local re-machining subtask is executed, regenerate a quality index vector based on the collected image again, and input the regenerated quality index vector into the process condition liquid time constant neural network for grading.

[0045] The present application has the following beneficial effects:

[0046] The multi-class image detection results are fused into a quality index vector, the comprehensive quantification of the size, geometric error, surface defect and key point offset is realized, the process condition liquid time constant neural network is introduced, the neuron characteristics can be dynamically adjusted according to the process timing and time interval, and the adaptability is stronger than that of the traditional static model.

[0047] The adjustment instruction or the out-of-bound re-machining subtask is triggered through the grading result, the process closed-loop feedback is constructed, the machining precision and stability are improved, the version storage and process map updating mechanism are adopted, the accumulation and iteration of the detection and regulation knowledge are ensured, and the long-term optimization capability is possessed. BRIEF DESCRIPTION OF DRAWINGS

[0048] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application together with the embodiments thereof, and explain the present application, and do not constitute a limitation of the present application. In the drawings:

[0049] Fig. 1 A flowchart of a hydraulic clamp mold multi-process machining process regulation method based on image detection is provided for the present application;

[0050] Fig. 2 A process condition liquid time constant neural network structure schematic diagram of a hydraulic clamp mold multi-process machining process regulation method based on image detection is provided for the present application;

[0051] Fig. 3 A quality index vector generation flowchart of a hydraulic clamp mold multi-process machining process regulation method based on image detection is provided for the present application. DETAILED DESCRIPTION

[0052] The application will be described in further detail below with reference to the drawings. These drawings are simplified schematic diagrams and only show the basic structure of the application in a schematic manner, and thus only show the components relevant to the application.

[0053] Reference Figs. 1-3 An image detection-based hydraulic clamp mold multi-process processing process control method, comprising the following steps:

[0054] An image acquisition station is arranged after the processing process of the hydraulic clamp mold to obtain image data of the mold processing part;

[0055] Edge extraction, defect segmentation, and texture analysis are performed based on the image data to output a quality index vector;

[0056] The quality index vector is input into a process condition liquid time constant neural network, and the output grading result determines that the quality index is in a qualified area, an adjustment area, or an out-of-limit area, and corresponding process control instructions are generated according to the grading result;

[0057] When the grading result is the adjustment area, the process control instructions include in-situ feedback adjustment of the tool compensation, feed, spindle speed, or fixture positioning parameters of the current process;

[0058] When the grading result is the out-of-limit area, the out-of-limit process control instructions include generating a local reprocessing subtask and performing image detection again after completion until the detection result enters the adjustment area or the qualified area;

[0059] The quality index vector, the process control instructions, and the process parameters are stored in a version, and the process map is updated based on the final inspection result.

[0060] In this embodiment, the image data includes:

[0061] Machine tool processing station camera data: industrial cameras are arranged on the processing stations of CNC machine tools, milling machines, wire cutting machines, and grinding machines to collect real-time images of mold workpieces during processing. The image resolution is 1920x1080, the frame rate is 30fps, and the global shutter is synchronized with the industrial fill light.

[0062] Auxiliary sensing image data: 3D structured light cameras are configured at the inter-process transfer positions to collect workpiece surface topography and geometric dimension data with an accuracy of 0.01mm.

[0063] Process document and CAD drawing data: mold design CAD drawings and process cards are imported through the MES system interface, including mold dimensions, tolerance bands, key surface markers, and processing process sequences.

[0064] All image data are transmitted to the image processing server through the GigE Vision interface, and the system is set to sample frequency of each process key node before and after each collection, to ensure the state comparison before and after workpiece processing.

[0065] In this embodiment, the machining process includes rough machining of the hydraulic clamp mold, drilling, heat treatment, finishing, electrical discharge machining, and polishing process.

[0066] In this embodiment, the image preprocessing includes:

[0067] Noise suppression: Gaussian filter combined with median filter is used to remove image noise caused by machine vibration and oil mist.

[0068] Illumination balancing: different station illumination imbalance is corrected by adaptive histogram equalization (CLAHE) to enhance edge features.

[0069] Workpiece region segmentation: U-Net-based deep segmentation network is used to identify the mold workpiece region, with a segmentation accuracy of more than 98%.

[0070] Key point extraction: In the workpiece segmentation area, SIFT-SURF fusion detection algorithm based on deep learning is used to extract geometric key points, including edges, corners, hole positions, and groove features.

[0071] The extraction results are stored in the form of coordinate sets and are associated with the CAD reference drawing.

[0072] In this embodiment, the quality indicator vector generation step includes:

[0073] Edge detection and geometric fitting are performed on the image data, and size deviation and geometric error are output:

[0074] Image preprocessing: distortion correction, grayscale standardization, and noise suppression are performed;

[0075] Sobel operator or Canny operator is used for edge extraction, combined with non-maximum suppression to obtain single-pixel edges;

[0076] Sub-pixel positioning (gray gravity or quadratic curve interpolation) is performed on the edge points, and least squares straight line fitting, circle fitting, or circular arc fitting is used to obtain geometric parameters;

[0077] The fitted elements are aligned with the CAD reference in coordinates, and size deviation, geometric error (at least including one of position degree, coaxiality, or flatness) are calculated, and numerical values and units are output.

[0078] Image segmentation is performed on the image data to identify burr and surface defect regions, and burr length and defect area are calculated:

[0079] Adaptive histogram equalization is performed on the original image to enhance the tiny protrusions and dark lines;

[0080] A U-Net semantic segmentation model is used to segment the image to obtain a binary mask, and the mask categories cover at least burrs and surface defects;

[0081] The area (pixel count x pixel-millimeter conversion coefficient) of each connected domain is calculated;

[0082] The maximum projection distance of the burr connected domain along the tangential direction of the hole or edge is calculated to obtain the burr length;

[0083] Connected domains with an area less than the minimum detectable threshold or a shape dispersion exceeding the limit are removed, and the effective defect area is retained;

[0084] The burr length and defect area are recorded per piece.

[0085] Frequency domain analysis is performed on the image data to extract texture features and map them to a surface roughness proxy value:

[0086] A surface texture image sub-block is obtained according to the detection area defined in the process card;

[0087] A two-dimensional fast Fourier transform (FFT) is performed on the sub-block to calculate the medium-high frequency energy ratio and the main direction consistency index;

[0088] Or a Gabor filter bank is used to extract directional texture energy;

[0089] A baseline regression model or a monotonic mapping function is used to map the frequency domain / directional features to a surface roughness proxy value;

[0090] The surface roughness proxy value and the original features used for tracing are recorded.

[0091] Feature matching is performed on mold images collected before and after different processes to calculate key point displacement:

[0092] Feature points (one of SIFT or ORB) are extracted from images before and after adjacent processes;

[0093] Feature matching is performed and RANSAC algorithm is used to remove false matches;

[0094] The displacement statistics of the matched point pairs are calculated in a unified coordinate system, and the key point displacement (at least one of the average displacement and the maximum displacement) is output;

[0095] The size deviation, geometric and position error, burr length, defect area, roughness proxy value, and key point displacement are combined to form a quality index vector.

[0096] In this embodiment, the hierarchical result generation step includes:

[0097] The quality index vector sequence arranged in the order of process time sequence and corresponding process identification and time interval between processes are input into the process condition liquid time constant neural network.

[0098] The quality index vector is a six-dimensional field, which sequentially includes size deviation, geometric error, burr length, defect area, surface roughness proxy value and key point offset;

[0099] The process identification is represented in the form of independent coding, which is composed of process number and processing type;

[0100] The time interval between processes is recorded in seconds to record the time difference between adjacent processes.

[0101] The process condition liquid time constant neural network comprises:

[0102] The input coding unit performs normalization processing on the quality index vector and splices the process identification and time interval to form an input tensor;

[0103] The input coding unit performs Z-score standardization on the quality index vector, maps the process identification into a fixed-length embedding vector, and expands the time interval into a fixed-length vector after logarithmic transformation. The three are spliced to form an input tensor. The input tensor is sent to the neural network in batches.

[0104] The time constant generation unit dynamically generates the time constant and coupling weight of each liquid neuron according to the process identification and time interval;

[0105] The time constant is obtained by superimposing a preset base value and a process condition correction amount;

[0106] The coupling weight is generated by the process condition super network to ensure that there are differences in the state evolution rate of different processes.

[0107] The liquid state update unit performs continuous time state update according to the time constant;

[0108] The liquid state update unit performs continuous time update on the neuron state according to the Euler discretization method. The update process saves the state value across time steps to reflect the cumulative effect of the previous process on the subsequent process.

[0109] The cost-sensitive decoding unit outputs the classification results of the qualified area, the adjustment area or the out-of-limit area based on the misjudgment cost matrix, and outputs the confidence value;

[0110] In the misjudgment cost matrix, the case of misjudging the "out-of-limit area" as the "qualified area" is given the highest cost, and the case of misjudging the "adjustment area" as the "qualified area" is given the second highest cost;

[0111] The decoding result is fixed as three categories: qualified area, adjustment area and out-of-limit area.

[0112] When the confidence is lower than a threshold, issuing a control instruction is prohibited, and an image reacquisition or manual review process is triggered.

[0113] Image reacquisition is reacquiring a mold processing site image by an image acquisition station;

[0114] Manual review is completed through a quality inspection terminal interface, and a quality inspector inputs a confirmation result;

[0115] The quality index vector after review confirmation replaces the original data, and the neural network is executed again to perform grading.

[0116] In the embodiment, the process condition liquid time constant neural network comprises:

[0117] A process embedding layer is arranged at the input end, the process number is mapped to an embedding vector, and the embedding vector is spliced with the quality index vector and input:

[0118] The process number is independently indexed and coded, and rough machining, drilling, heat treatment, finishing, electric discharge machining, and polishing correspond to indexes 0-5, respectively;

[0119] The process number is mapped to a 64-dimensional embedding vector through a table lookup operation;

[0120] The embedding vector is spliced with the normalized quality index vector (a six-dimensional real number vector) to form a combined input with a length of 70;

[0121] The input vector is sent to the network in batches, and the batch size is fixed at 32.

[0122] A process condition hypernetwork is configured, and the time constant and coupling weight of the liquid neuron are dynamically generated according to the process identifier and the time interval between processes:

[0123] The hypernetwork input is the process embedding vector and the time interval between processes (normalized to 32 dimensions);

[0124] The hypernetwork is composed of two fully connected networks, the first layer has 128 units, and the ReLU activation, the second layer has 256 units, and the Sigmoid activation;

[0125] The output part includes:

[0126] The time constant parameter τ of the liquid neuron, with a value range of [0.05, 1.0] seconds, controls the state update speed;

[0127] The coupling weight matrix W between the liquid neurons is sparsely constrained, with a fixed non-zero element ratio of 20%, to ensure that the weight distribution is different under different processes.

[0128] The regularization term is introduced in the training process of the super network to constrain the generated parameters to avoid unstable fluctuations.

[0129] A cost-sensitive decoding head is arranged at the output end, and a confidence estimation branch is arranged in parallel, a preset cost matrix is used to weight the hierarchical loss, and a hierarchical result and a heteroscedastic confidence value are output:

[0130] The output end includes two parts:

[0131] The cost-sensitive decoding head includes:

[0132] The input is a hidden state vector;

[0133] A two-layer fully connected structure (64 dimensions in the first layer and 3 dimensions in the second layer) is adopted to output the probability distribution corresponding to the qualified area, the adjustment area and the out-of-bound area.

[0134] A misjudgment cost matrix is introduced during training, and is set as:

[0135] The cost of out-of-bound being judged as qualified = 10, the cost of adjustment being judged as qualified = 5, and the cost of other misjudgments = 1.

[0136] The cross-entropy loss optimized by cost weighting is used to optimize the classification boundary.

[0137] The confidence estimation branch includes:

[0138] In parallel with the decoding head, a single-layer fully connected network is used to output a scalar value representing the heteroscedastic confidence, which is in the interval [0, 1] and is constrained by a Sigmoid function.

[0139] When the confidence is lower than 0.7, the system prohibits the issuance of control instructions and triggers the image reacquisition and manual review process.

[0140] A low-rank adapter weight is introduced in the hidden layer:

[0141] The hidden state dimension of the liquid neuron is fixed at 256;

[0142] A low-rank adapter (LoRA-like) structure is introduced in the hidden layer:

[0143] The original weight matrix is decomposed into A·B, where A is a 256×16 matrix and B is a 16×256 matrix.

[0144] Only the low-rank part of the parameter is updated, and the backbone weight is kept frozen;

[0145] This design improves the process condition adaptation capability while ensuring the inference efficiency.

[0146] In this embodiment, the process condition liquid time constant neural network sets a monotonicity constraint module between the output layer and the input quality indicators, so that the mapping relationship between the burr length and the out-of-bound probability, the key point offset and the out-of-bound probability, and the roughness proxy value and the adjustment probability is monotonically increasing, and a penalty term is applied to the samples that violate the constraint during the training process.

[0147] In this embodiment, the process condition liquid time constant neural network sets a monotonicity constraint module between the output layer and the input quality indicators, which is used to ensure that the prediction results meet the physical and process rules in the machining process. The design idea of this module is: when the burr length gradually increases, the out-of-bound probability output by the system should also increase; when the key point offset gradually increases, the out-of-bound probability output by the system should also increase; when the surface roughness proxy value gradually increases, the adjustment probability output by the system should also increase. Through this mechanism, the "unreasonable reverse prediction" in actual process control is avoided.

[0148] Specifically, this module adds a constraint channel for the above three key indicators inside the network, and each channel imposes a non-negative limit on the connection weight, ensuring that the larger the input value, the only way to increase the risk probability, but not to decrease. Taking the burr length as an example, the system limits the weight direction and gain amplitude, so that when the input value gradually rises from a lower level to a higher level within the continuous value range, the output of the out-of-bound probability also presents a stable upward trend. Similarly, for the key point offset and the surface roughness proxy value, the system respectively establishes a monotonically increasing mapping relationship, and dynamically maintains this trend during network reasoning.

[0149] During the training process, this module works together with the monotonicity penalty mechanism. The training data is designed as a pair of inputs, one of which has a longer burr length, a larger key point offset, or a higher roughness proxy value than the other. When the network outputs a risk probability that does not meet the monotonicity requirement during reasoning, for example, a sample with a larger input value predicts a lower risk probability, the system will treat this difference as a monotonicity violation and generate an additional penalty value. The penalty value is added to the loss function according to the preset weight coefficient and is back-propagated and optimized together with the regular classification error. The weight coefficient is usually set between zero point one and zero point five, which can effectively constrain the monotonicity without damaging the overall classification performance.

[0150] Through this design, the process condition liquid time constant neural network can ensure the consistency of the output result with the actual machining process while maintaining the prediction accuracy. For example, when the burr gradually decreases after mold polishing, the system can stably output a result of a decreasing out-of-bound probability; and when the shape and position error may increase after heat treatment, the system can output a trend of gradually increasing out-of-bound risk. This module effectively reduces the prediction abnormalities caused by random noise, local overfitting or abnormal input when applied, thereby improving the explainability and reliability of the process control instructions.

[0151] In the embodiment, the adjustment process control instruction includes:

[0152] The tool radius compensation amount is calculated based on the size deviation, and the tool radius compensation amount is written into the tool compensation parameter table of the numerical control system;

[0153] For the compensation of the size deviation, the difference between the measured size obtained by image detection and the CAD nominal size is used to distinguish the inner contour and the outer contour. When it is an inner contour such as the inside of a hole or a groove, the tool radius compensation amount is taken as the value of the diameter deviation multiplied by negative one-half; when it is an outer contour, the tool radius compensation amount is taken as the value of the outer contour size deviation multiplied by negative one-half. In this way, the case of a small hole diameter corresponds to positive compensation, and the case of an oversized contour corresponds to negative compensation. The calculated compensation amount is limited to within plus or minus 20% of the nominal radius of the tool, and is further quantized in steps of one micrometer, with the result being retained to three decimal places of millimeter value. The processed compensation amount is directly written into the radius compensation field of the tool compensation parameter table of the numerical control system, with a time stamp and a workpiece identifier, forming a traceable parameter update record.

[0154] The feed speed correction value is calculated based on the geometric and position error, and the feed speed correction value is limited within the upper and lower limits of the allowable feed speed of the machine tool;

[0155] For the correction of the feed speed, the geometric and position error in the quality index vector and the geometric and position tolerance band corresponding to the process are used to calculate the error ratio, i.e. the geometric and position error divided by the tolerance band. Based on the ratio, a correction coefficient is determined, which is defined as one minus zero point four times the error ratio, and the correction coefficient is limited to between zero point six and one. The feed speed correction value can be calculated as the baseline feed speed multiplied by the correction coefficient and then subtracted from the baseline feed speed. The corrected feed speed is limited within the upper and lower limits allowed by the machine tool, and is written into the feed parameters of the program segment. The geometric and position error, the error ratio, the correction coefficient and the corrected feed speed are recorded in the process database to ensure data reviewability.

[0156] The correction amount of the spindle speed is calculated based on the surface roughness proxy value, when the roughness proxy value is higher than the baseline, the spindle speed is set to a correction value higher than the baseline, and when the roughness proxy value is lower than the baseline, the spindle speed is set to a correction value lower than the baseline;

[0157] For the correction of the spindle speed, the roughness deviation ratio is calculated based on the surface roughness proxy value and the preset roughness baseline, that is, the roughness proxy value is subtracted from the baseline and then divided by the baseline. The correction coefficient is defined as one plus zero point two times the deviation ratio, and the value range of the correction coefficient is limited to between zero point six and one point one. Based on the correction coefficient, the spindle speed correction amount is calculated, the value of which is the baseline spindle speed multiplied by the correction coefficient and then subtracted by the baseline speed. The corrected speed is limited within the upper and lower limit range of the tool parameter table. The correction result is written into the spindle speed parameter of the program segment, and a record is formed in the process database, including the roughness proxy value, the deviation ratio, the correction coefficient, and the corrected spindle speed.

[0158] The jig positioning compensation amount is calculated based on the key point offset amount, and the jig positioning compensation amount is written into the tool setting table;

[0159] For the compensation of jig positioning, the same set of key points is extracted in the images of adjacent processes, and the pixel coordinates are converted to the machine tool coordinate system through coordinate calibration. Based on the previous and current coordinates of the key points, the centroid position is calculated and the least squares method is used to solve the plane rigid transformation to obtain the components of the translation in the X and Y directions and the angular increment around the Z axis. The compensation amount is taken as the opposite of the above components and angles to eliminate the detected offset error, and the X and Y direction compensation is limited to the range of plus or minus fifty microns, and the rotational compensation is limited to the range of plus or minus three angular minutes, and the result is quantized in steps of one micron and one angular second. The compensation value is written into the X, Y, and rotation fields corresponding to the G54 coordinate system in the tool setting table, and the time stamp and workpiece identification are recorded to ensure traceability.

[0160] After the above parameters are written in the current process, the system synchronizes the tool radius compensation amount, jig positioning compensation amount, feed speed correction value, and spindle speed correction amount to the process card of the next process. The synchronized data is recorded in the "preposition correction area", and when the next process is clamped, the numerical control program reads the preposition correction area in the opening paragraph, and performs tool compensation update, tool setting update, and preset operation of feed speed and spindle speed in order. At the same time, the system generates a unique version number for each synchronization, which is spliced from the workpiece identification, process number, and time stamp, and is used to play back the calculation input, calculation result, writing position, and version information in the quality traceability query interface, so as to realize the closed-loop control of the whole process.

[0161] In the embodiment, the cross-border process control instruction comprises:

[0162] Generate local re-machining sub-tasks based on dimensional deviation and geometric error, and set local machining allowance and path overlap rate in the local re-machining sub-tasks;

[0163] When the detection result is determined to be an out-of-bound area, the specific deviation of each index in the quality index vector is analyzed to determine the triggered re-machining type. For the case where the dimensional deviation and geometric error exceed the allowed range, a local re-machining sub-task is generated. In this sub-task, the required local machining allowance is calculated according to the ratio of the measured deviation to the tolerance band, and the machining allowance is limited to between five percent and twenty percent of the reserved allowance of the previous process. The path overlap rate is determined according to the error distribution, which is usually between twenty percent and thirty percent, to ensure that the machining path fully covers the area when it is corrected or under-cut without introducing excessive repeated cutting. The generated local tool path is converted to machine code by the post-processor and directly inserted into the task queue of the process to be executed.

[0164] Generate local re-machining sub-tasks based on burr length, and set chamfer machining path and chamfer depth;

[0165] For the out-of-bound situation triggered by the burr length exceeding the threshold, a chamfer machining sub-task is generated. This sub-task identifies the burr edge position through image segmentation results, calculates the length of the line segment that needs to be chamfered, and determines the chamfer machining path in combination with the tool diameter and safety allowance. The chamfer depth is set according to the average height of the burr, which is usually one point two times the height of the burr, to ensure complete removal of the burr. The machining path is optimized to be continuous, to reduce the tool marks caused by repeated lifting of the tool, and safety parameters such as speed and spindle speed are attached to the path after generation, to ensure stable processing.

[0166] Generate local re-machining sub-tasks based on defect area, and set pulse width and gap voltage for electrical discharge machining;

[0167] For the out-of-bound situation where the surface defect area exceeds the threshold, an electrical discharge machining sub-task is generated. The area and depth of the defect area are determined by image segmentation and gray scale estimation. The pulse width is set according to the defect area: short pulses are used when the defect area is small to reduce the secondary discharge area, and medium pulses are used when the defect area is large to improve the repair efficiency. The gap voltage is automatically adjusted according to the defect depth: when the defect is shallow, it is set to the baseline voltage plus a bias, and when the defect is deep, the voltage is increased to ensure the stability of the discharge gap. All discharge parameters are loaded into the power supply control module, and the corresponding electrical discharge machining path is generated, which strictly covers the defect area and avoids entering the qualified area.

[0168] After the local reworking subtask is completed, image acquisition is triggered again, and edge extraction, defect segmentation, and texture analysis are re-executed based on the new image data to generate a new quality indicator vector. The vector is then input into the process condition liquid time constant neural network for classification. If the result is still in the out-of-bound area, the corresponding reworking subtask will be generated again according to the new deviation result; if the classification result enters the adjustment area, the adjustment process control instruction is executed to fine-tune the parameters; if it enters the qualified area, the reworking process of the process is completed and the next process is entered. The entire process ensures that the out-of-bound situation can be gradually corrected through automatic closed-loop reworking until the processing result meets the process requirements.

[0169] Example 1

[0170] To verify the feasibility of the application in implementation, the application is applied to the batch processing production line of a certain hydraulic clamp mold. The production line includes rough machining, drilling, heat treatment, finishing, electric discharge machining, and polishing of six main processes, and an industrial camera image acquisition station is arranged at the outlet of each process.

[0171] 1. Experimental environment

[0172] Machine tool type: Mazak VCN-530C CNC vertical machining center;

[0173] Workpiece material: Cr12MoV alloy steel;

[0174] Image acquisition equipment: resolution 4096x2160 industrial camera, frame rate 30fps, with 5600K ring LED light source;

[0175] Network training data: a total of 12,000 image samples of molds processed in the past are collected, covering 6 processes.

[0176] 2. Experimental process

[0177] After each process is completed, the mold machining site image is automatically acquired, and a quality indicator vector is generated:

[0178] Dimensional deviation (mm);

[0179] Form error (mm);

[0180] Burr length (μm);

[0181] Defect area (mm 2 );

[0182] Roughness proxy value (μm);

[0183] Key point offset (μm).

[0184] The vector is inputted into the process condition liquid time constant neural network with process number and time interval between processes. The network outputs the classification result of "qualified area / adjustment area / out-of-bound area" and confidence value.

[0185] 3. Adjustment of process instance verification

[0186] In the finishing process of a batch of 20 molds, the system detects that:

[0187] Size deviation = 0.05mm;

[0188] Form error = 0.03mm;

[0189] Roughness proxy value = 0.65μm;

[0190] Key point offset = 28μm;

[0191] The network classification result is adjustment area, and the confidence value is 0.92. The system generates the following control instructions:

[0192] Tool radius compensation = 0.025mm → written into tool compensation table;

[0193] Feed speed correction value = 440mm / min (adjusted from original value 500mm / min);

[0194] Spindle speed correction = +300rpm (adjusted from 6000rpm to 6300rpm);

[0195] Clamp positioning compensation = 28μm → written into tool setting table;

[0196] After the compensation is executed, the size deviation detected by the process is reduced to 0.02mm, and the roughness proxy value is reduced to 0.52μm, successfully entering the qualified area.

[0197] 4. Out-of-bound process instance verification

[0198] In another batch of 15 molds in the electrical discharge machining process, the detection result is:

[0199] Defect area = 0.42mm 2 ;

[0200] Burr length = 130μm;

[0201] The network classification result is out-of-bound area, and the confidence value is 0.87. The system automatically generates a local re-machining subtask:

[0202] Local electrical discharge machining: pulse width is set to 120μs, gap voltage is set to 35V

[0203] Chamfering: chamfer depth is set to 0.3mm, path coverage is 100%

[0204] After reprocessing, the image is collected again, the defect area is reduced to 0.08mm 2 , the burr length is reduced to 40μm, and finally classified into the adjustment area, and after small correction, it enters the qualified area.

[0205] 5. Experimental results

[0206] In the batch processing experiment of 100 hydraulic tong molds:

[0207] The proportion of entering the qualified area for the first time is 72%;

[0208] The proportion of entering the qualified area after correction in the adjustment area is 24%;

[0209] The proportion of entering the qualified area after reprocessing is 4%;

[0210] The final inspection qualified rate is improved to 100%;

[0211] In order to further verify the advantages of the present application relative to the traditional process method, 200 workpieces of a hydraulic tong mold production line are selected and divided into experimental group and control group.

[0212] Experimental group (100 pieces): adopt the multi-process processing process regulation method based on image detection of the present application;

[0213] Control group (100 pieces): adopt the conventional process flow, only manual sampling inspection is carried out after each process, and real-time image detection and process parameter self-adaptive regulation are not performed.

[0214] Control experiment conditions:

[0215] Machine tool: Mazak VCN-530C numerical control vertical machining center;

[0216] Workpiece material: Cr12MoV alloy steel;

[0217] Process sequence: rough machining→drilling→heat treatment→finishing→electric discharge machining→polishing;

[0218] Detection method: industrial camera+process condition liquid time constant neural network are adopted in the experimental group; only manual caliper measurement and final inspection surface roughness meter are used in the control group.

[0219] Table 1 comparison of experimental results

[0220]

[0221] As shown in Table 1, in terms of processing quality, the application can timely find size deviation, shape and position error, burr and surface defect by combining image detection with process condition liquid time constant neural network, and issue compensation instructions under the "adjustment area" and generate local reprocessing subtasks under the "out-of-bound area". Therefore, the final inspection qualified rate is improved from 92.5% of the control group to 100%, and the scrap rate is completely eliminated.

[0222] In terms of processing precision, the application effectively reduces indicators such as size deviation, roughness and defect area, indicating that the method is more stable for mold surface and shape and position precision control.

[0223] In terms of processing efficiency, due to the addition of image acquisition and real-time compensation, the single piece processing time is increased by an average of 4.3%. However, due to the significant reduction in scrap rate, the unit qualified product output efficiency of the overall production line is improved by about 9%.

[0224] In terms of traceability, the application stores quality index vector, process control instructions and process parameters by versioning, and updates the process map based on the final inspection result, providing a reliable data basis for subsequent process improvement and equipment maintenance, which cannot be achieved by the control group.

[0225] Through comparative experiments, it can be proved that although the application slightly increases the single piece processing time, it significantly improves the quality consistency of mold processing and the final inspection qualified rate, and the overall benefit is obviously better than the traditional process, which has high practical value and promotion significance.

[0226] The above is only the preferred specific implementation of the application, but the protection scope of the application is not limited thereto, any skilled person in the art can make equivalent replacement or change according to the technical solution and inventive concept of the application within the technical range disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A hydraulic clamp mold multi-process machining process regulation method based on image detection, characterized in that, The method comprises the following steps: An image acquisition station is arranged after the machining process of the hydraulic clamp mold to obtain image data of the mold machining part; Based on the image data, edge extraction, defect segmentation and texture analysis are performed to output a quality index vector; The quality index vector is input into a process condition liquid time constant neural network, and the output grading result determines that the quality index is in a qualified area, an adjustment area or an out-of-bound area, and corresponding process control instructions are generated according to the grading result; When the grading result is in the adjustment area, the process control instructions include in-situ feedback adjustment of the tool compensation, feed, spindle speed or clamp positioning parameters of the current process; When the grading result is in the out-of-bound area, the out-of-bound process control instructions include generating a local re-machining subtask and performing image detection again after completion until the detection result enters the adjustment area or the qualified area; The quality index vector, process control instructions and process parameters are stored in a version, and the process map is updated based on the final inspection result.

2. The mold multi-process machining process regulation method based on image detection of hydraulic clamps according to claim 1, characterized in that, The machining process includes rough machining, drilling, heat treatment, finishing, electrical discharge machining and polishing of the hydraulic clamp mold.

3. The mold multi-process machining process regulation method based on image detection of hydraulic clamps according to claim 1, characterized in that, The generation of the quality index vector includes: Edge detection and geometric fitting are performed on the image data to output size deviation and geometric and position error; Image segmentation is performed on the image data to identify burr and surface defect areas, and calculate burr length and defect area; Frequency domain analysis is performed on the image data to extract texture features and map them into a surface roughness proxy value; Feature matching is performed on the mold images collected before and after different processes to calculate key point offset; The size deviation, geometric and position error, burr length, defect area, roughness proxy value and key point offset are combined to form the quality index vector.

4. The mold multi-process machining process regulation method based on image detection of hydraulic clamps according to claim 1, characterized in that, The generation of the grading result includes: The quality index vector sequence arranged in the order of process time sequence, corresponding process identifier and time interval between processes are input into a process condition liquid time constant neural network; The process condition liquid time constant neural network includes: An input encoding unit performs normalization processing on the quality index vector and concatenates the process identifier and the time interval to form an input tensor; A time constant generation unit dynamically generates the time constant and coupling weight of each liquid neuron according to the process identifier and the time interval; A liquid state update unit performs continuous time state update according to the time constant; A cost-sensitive decoding unit outputs the grading result of the qualified area, the adjustment area or the out-of-bound area based on a misjudgment cost matrix, and outputs a confidence value; When the confidence value is lower than a threshold value, the control instructions are prohibited from being issued, and an image reacquisition or manual review process is triggered.

5. The method of claim 1, wherein, The process condition liquid time constant neural network includes: A process embedding layer is arranged at the input end to map the process number into an embedding vector and concatenate it with the quality index vector for input; A process condition hypernetwork is configured to dynamically generate the time constant and coupling weight of the liquid neuron according to the process identifier and the time interval between processes; A cost-sensitive decoding head is arranged at the output end in parallel with a confidence estimation branch, a preset cost matrix is used to weight the grading loss, and the grading result and heteroscedastic confidence value are output; Low-rank adapter weights are introduced in the hidden layer.

6. The method of claim 1, wherein the method further comprises: The process condition liquid time constant neural network sets a monotonicity constraint module between the output layer and the input quality indicators, so that the mapping relationship between the burr length and the out-of-bound probability, the key point offset and the out-of-bound probability, and the roughness proxy value and the adjustment probability is monotonically increasing, and a penalty term is applied to the samples that violate the constraint during the training process.

7. The method of claim 1, wherein the method further comprises: The adjustment process control instruction includes: calculating the tool radius compensation amount based on the size deviation, and writing the tool radius compensation amount into the tool compensation parameter table of the numerical control system; calculating the feed speed correction value based on the form and position error, and limiting the feed speed correction value within the upper and lower limits of the feed speed allowed by the machine tool; calculating the spindle speed correction amount based on the surface roughness proxy value, setting the spindle speed to be higher than the correction value corresponding to the baseline when the roughness proxy value is higher than the baseline, and setting the spindle speed to be lower than the correction value corresponding to the baseline when the roughness proxy value is lower than the baseline; calculating the fixture positioning compensation amount based on the key point offset, and writing the fixture positioning compensation amount into the tool setting table of the tooling; synchronizing the tool radius compensation amount, the fixture positioning compensation amount, the feed speed correction value and the spindle speed correction amount to the next process process card.

8. The method of claim 1, wherein the method is characterized by: The out-of-bound process control instruction includes: generating a local re-machining subtask based on the size deviation and the form and position error, and setting the local machining allowance and the path overlap rate in the local re-machining subtask; generating a local re-machining subtask based on the burr length, and setting the chamfer machining path and the chamfer depth; generating a local re-machining subtask based on the defect area, and setting the pulse width and gap voltage of the electric discharge machining; after the local re-machining subtask is executed, the image is collected again, and the regenerated quality indicator vector is input into the process condition liquid time constant neural network for grading based on the image collected again.

Images