An axial determination and extrusion quality monitoring data processing method
Patent Information
- Application Number
- CN202611056534.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-16
- Publication Date
- 2026-10-20
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the field of extrusion processing monitoring data processing technology, and in particular to a method for axial determination and extrusion quality monitoring data processing. Background Technology
[0002] Extrusion is one of the core processes in metal forming, widely used in the mass production of automotive parts, aerospace structural components, and building profiles. During extrusion, the axial alignment of the extrusion bar directly determines the workpiece's form and position tolerances, mechanical properties, and other core quality indicators. Currently, the extrusion industry is undergoing a transformation towards intelligent manufacturing, with increasing demands for real-time processing, efficient storage, quality traceability, and intelligent prediction of process monitoring data. Data processing capabilities have become one of the core factors restricting the improvement of extrusion quality and efficiency.
[0003] Currently, the mainstream extrusion monitoring data processing solutions in the industry are mainly divided into two categories. The first category is the stand-alone monitoring solution, which works by deploying a small number of pressure and displacement sensors on a single extrusion machine. After collecting core operating parameters, the data is stored in the local industrial control unit of the machine. It only supports local data viewing and simple threshold alarms. This solution is widely used in scenarios with small processing units and single, scattered extrusion machines. Its advantages are low deployment threshold, low initial investment, and simple operation and maintenance. However, this solution has obvious drawbacks: axial status determination relies only on single-path displacement or total extrusion pressure thresholds, resulting in low accuracy. Axial off-center load fault detection is delayed, which can easily lead to batches of non-conforming products. The collected monitoring data is not categorized by processing steps and is not associated with axial status labels. The data is scattered and has poor correlation, which cannot support cross-batch quality analysis and intelligent modeling requirements. Quality traceability requires manual verification of scattered paper records and local data, and a single traceability can take several hours, resulting in extremely low efficiency.
[0004] The second type is a centralized relational database storage solution. Its working principle involves collecting and uploading monitoring data from multiple extrusion machines to a central server, storing it in a relational database, and periodically generating quality statistical reports to support batch production data analysis. This solution is often used in production management systems for medium to large-scale extrusion production lines. Its advantage lies in achieving centralized management of data from multiple machines, enabling global pass rate and output statistical analysis. However, this solution also has significant drawbacks: relational databases have low compression ratios for high-throughput time-series monitoring data, consume large amounts of storage space, and have slow response times for multi-dimensional combined searches, making it unable to support efficient and rapid quality traceability; axial state data is not linked to the monitoring data of the corresponding process, making it impossible to establish a mapping relationship between axial off-center load and quality defects. Quality identification still relies on post-process manual inspection, failing to achieve real-time quality prediction during processing. Defective products are only discovered after production, resulting in significant waste of raw materials and production capacity.
[0005] As the precision requirements of extruded products continue to increase, the existing two types of technical solutions can no longer simultaneously meet the multiple needs of high-accuracy axial determination, efficient time-series data management, real-time quality prediction, and rapid quality traceability. Developing a dedicated monitoring data processing method adapted to extrusion processing scenarios has become an urgent problem for the industry. Summary of the Invention
[0006] The present invention proposes an axial determination and extrusion quality monitoring data processing method to solve the problems mentioned in the prior art, such as low accuracy of axial determination, low data utilization, delayed quality identification, and low traceability efficiency in existing extrusion monitoring data processing.
[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for axial determination and extrusion quality monitoring data processing, comprising the following steps: S1. Deploy corresponding sensors at preset acquisition points on the extrusion processing equipment to collect multi-source time-series monitoring data of the entire extrusion processing process. The preset acquisition points include at least the two ends of the extrusion rod in the axial direction, the circumferential direction of the extrusion die cavity, the main shaft of the equipment, and the inner wall of the die cavity. The multi-source time-series monitoring data includes at least the displacement data of the two ends of the extrusion rod in the axial direction, the extrusion force data distributed at multiple points in the circumferential direction of the die cavity, the triaxial vibration data of the main shaft of the equipment, the temperature data of the inner wall of the die cavity, and the millisecond-level timestamp data corresponding to each sampling point and synchronized with the PLC clock of the equipment. After the extrusion process is started, the axial state of the current processing is determined in real time based on the axial displacement data at both ends and the extrusion force data distributed at multiple points corresponding to the same timestamp, and the result of axial normal or axial off-center load is obtained. The entire determination response delay is less than 50ms. S2. Connect the PLC process trigger signal of the extrusion processing equipment, and slice the collected continuous multi-source time-series monitoring data according to the extrusion process nodes to obtain time-series data slices corresponding to each processing stage. After the axial state determination result is used as a tag to associate with the time-series data slice of the corresponding processing stage, it is stored in a columnar time-series database that conforms to the G06F16 classification standard. S3. Retrieve time-series data slices stored in the time-series database that simultaneously carry axial status labels and historical quality inspection results of the corresponding workpieces to construct a training dataset and train the extrusion quality recognition model. The historical quality inspection results include workpiece form and position tolerance detection results and mechanical property detection results, which are automatically synchronized to the time-series database by the quality inspection equipment and bound to the unique identifier of the workpiece. S4. Real-time access to multi-source time-series monitoring data of the extrusion process via industrial bus, first complete the real-time axial state determination, then input the real-time time-series data of the current processing stage into the trained extrusion quality recognition model, output the corresponding extrusion quality level, and after associating the real-time axial state determination result, real-time multi-source time-series monitoring data and extrusion quality level with the unique identifier of the currently processed workpiece, store them in the time-series database.
[0008] Preferably, the specific process for determining the axial state in step S1 is as follows: extract the displacement data collected by two axial displacement sensors at the same time stamp, calculate the absolute difference between the two sets of displacement data, and simultaneously extract the extrusion pressure data collected by all pressure sensors in the circumferential direction of the mold cavity at the same time stamp, calculate the maximum eccentricity difference between the maximum and minimum extrusion pressure values. When the displacement difference is greater than a preset displacement threshold and the maximum eccentricity difference is greater than a preset pressure threshold, it is determined to be an axial off-center load state. If either of the two conditions is not met, it is determined to be an axial normal state. The preset displacement threshold is adaptively configured according to the axial form and position tolerance requirements of the currently processed workpiece, and the preset pressure threshold is pre-configured according to the rated mold cavity bearing pressure value of the current extrusion process.
[0009] Preferably, the specific rules for slicing according to the extrusion process nodes in step S2 are as follows: the entire extrusion process is divided into four process nodes: feeding, pre-extrusion, holding pressure, and unloading. These correspond to five segmented trigger points output by the PLC: feeding arrival signal, pre-extrusion start signal, holding pressure trigger signal, unloading start signal, and unloading arrival signal. The multi-source timing monitoring data between two adjacent trigger points is the timing data slice of the corresponding process. The four timing data slices correspond to the feeding stage, pre-extrusion stage, holding pressure stage, and unloading stage, respectively. There is no data overlap or data omission between the slices, and each slice is simultaneously associated with the preset process parameters of the corresponding process.
[0010] Preferably, in step S2, when storing the time-series data slices after associating tags into the time-series database, a three-level composite primary key index is constructed using the workpiece unique identifier, timestamp, and axial status tag. The workpiece unique identifier is a globally unique code of the workpiece obtained by scanning with RFID or QR code before extrusion processing. The timestamp adopts the Unix millisecond-level timestamp format. The axial status tag adopts a numerical code where 0 represents axial normal and 1 represents axial off-center load. The time-series database establishes inverted indexes for the three primary key dimensions, supports multi-condition retrieval based on time, workpiece, and axial state dimensions, has a retrieval response latency of less than 100ms, and automatically uses a lossless compression algorithm to compress data during storage, with a compression ratio of no less than 8:1.
[0011] Preferably, before constructing the training dataset in step S3, a data preprocessing step is also included: using the Isolation Forest algorithm to remove abnormal data collected in the time series data slices. The anomaly detection threshold of the Isolation Forest algorithm is set to 0.7. Sampling points with anomaly scores higher than 0.7 are judged as abnormal data and removed. The data gaps after removal are filled with the mean of three adjacent sampling points in the same dimension. Then, the values of all feature dimensions are normalized to the maximum and minimum values to uniformly map the value range of each feature to the interval between 0 and 1, thereby eliminating the interference of the numerical magnitude difference of different feature dimensions on model training.
[0012] Preferably, the extrusion quality recognition model trained in step S3 is a lightweight temporal convolutional network. The network structure of the lightweight temporal convolutional network includes two depthwise separable temporal convolutional layers, one global average pooling layer, and one fully connected classification layer. The dilation coefficients of the two convolutional layers are 1 and 2, respectively. The receptive field covers the time length of a complete temporal data slice. The total number of model parameters is controlled within 1M, which supports deployment and operation on edge devices. The model takes as input a multi-dimensional feature vector corresponding to a slice of time-series data and outputs three levels of extrusion quality classification: qualified, pending re-inspection, and unqualified. Qualified means that the workpiece's form and position tolerances and mechanical properties meet the process requirements. Pending re-inspection means that there are suspected defects that require manual verification. Unqualified means that there are obvious defects and the workpiece is directly judged as scrap.
[0013] Preferably, after the real-time axial status determination is completed in step S4, if the determination result is axial off-center load, an axial off-center load early warning signal is immediately generated. The early warning signal includes the early warning level, the off-center load occurrence time, the displacement difference, and the extrusion force eccentricity difference. It is sent to the PLC control terminal of the extrusion processing equipment via the Modbus protocol to trigger the equipment's deceleration buffer or emergency stop logic. At the same time, the corresponding real-time multi-source time-series monitoring data is marked as high-priority data and stored in an independent high-priority storage partition of the time-series database. The retrieval priority is higher than that of ordinary data, and the data retention time is 3 times that of ordinary data.
[0014] Preferably, the method further includes a quality traceability response step: receiving an externally input quality traceability request, wherein the input channels for the quality traceability request include API requests initiated by the production management system, HTTP requests initiated by the quality inspection end, and local requests initiated by the equipment's human-machine interface; parsing the search keywords carried in the traceability request, wherein the search keywords include the workpiece's unique identifier or processing time range; retrieving the corresponding full-process axial state determination results, full-stage multi-source time-series monitoring data, and extrusion quality level from the time-series database; automatically generating a visualized time-series curve report and summary information; supporting export to CSV or PDF format; and having a traceability response latency of less than 200ms.
[0015] Preferably, it also includes a model iteration update step: according to a preset period, the newly added time series data slices with manual review results in the time series database are retrieved to form an incremental training set. The preset period is triggered every 7 days or every 1,000 newly added time series data slices with manual review results. The incremental training set is sampled according to the ratio of qualified, pending review, and unqualified data in 10:1:1 to avoid class imbalance problem. The learning rate during incremental fine-tuning is set to 1 / 10 of the initial training learning rate. Only the parameters of the fully connected layer and the last convolutional layer are updated. The deployment and replacement are completed only when the accuracy of the updated model is verified to be higher than that of the original model on the test set; otherwise, the original model is kept running.
[0016] Preferably, the data distribution and storage steps are also included: the real-time generated axial status judgment results, extrusion quality grades, and associated multi-source time-series monitoring data are encrypted using the national cryptographic SM4 symmetric encryption algorithm, and then the content is filtered according to the permissions of different ports before being synchronously distributed to the visualization display interfaces of the equipment end, production management end, and quality inspection end. The equipment end only displays the real-time axial status, early warning information, and the quality grade of the currently processed workpiece; the production management end displays the quality statistics data and axial off-center load occurrence rate of the entire batch; and the quality inspection end displays the full-process monitoring data and quality grade of a single workpiece. The encrypted full data is synchronously backed up and stored to a distributed cloud storage node, using a three-replica backup mechanism, achieving a data reliability of 99.999%, and supporting cross-regional retrieval and access.
[0017] Compared with existing technologies, the beneficial effects of this invention are: This invention achieves high-accuracy real-time identification of axial state by using a dual-threshold method of axial displacement difference at both ends and multi-point extrusion pressure eccentricity difference. The judgment response speed meets the requirements of industrial real-time control and can quickly trigger early warning or equipment protection logic in the early stage of off-center load. It effectively solves the defects of existing technologies, such as high rate of missed judgment and false judgment in single parameter judgment and the lag in axial off-center load detection, which easily leads to batches of non-conforming products.
[0018] This invention achieves structured and associated storage of monitoring data by using axial status labels associated with extrusion process nodes and a three-level joint primary key index time-series database. It features high data storage compression ratio, fast multi-dimensional retrieval response speed, and can quickly respond to cross-dimensional quality traceability requests. It effectively solves the defects of existing technologies, such as scattered and poorly correlated monitoring data, high storage costs, and long time consumption and low efficiency in quality traceability.
[0019] This invention achieves real-time extrusion quality level identification during processing by using a lightweight temporal convolutional network trained with axial state labels and historical quality inspection data. It can predict quality defects in advance during processing and quickly intercept non-conforming products without waiting for subsequent manual inspection. This effectively solves the defects of existing technologies that rely on post-processing inspection for quality identification, have a high rate of missed detection, and easily lead to waste of raw materials and production capacity.
[0020] The overall solution of this invention is lightweight and highly adaptable. The lightweight quality recognition model can be deployed on low-computing-power hardware such as edge industrial control equipment and programmable logic controllers, without the need for high investment in computing power hardware. It is suitable for deployment on single, scattered extrusion equipment in small processing entities, as well as for deployment in clusters of multiple equipment in large-scale extrusion production lines of medium and large processing enterprises. It can be widely used in various extrusion processing scenarios such as building profile processing, cold extrusion of automotive metal parts, and hot extrusion of aerospace structural parts. The judgment threshold and model parameters can be flexibly adjusted according to the processing requirements of different products. It can adapt to various extrusion processing scenarios without changing the core technical architecture, and has high industry promotion value. Attached Figure Description
[0021] Figure 1 This is a flowchart of the overall process for axial determination and extrusion quality monitoring data processing proposed in this invention; Figure 2 This is a flowchart of the axial state determination and early warning response of the present invention; Figure 3 This is a flowchart of the multi-source monitoring data preprocessing and storage process of the present invention; Figure 4 This is a flowchart of the training and iterative update process for the extrusion quality identification model of the present invention; Figure 5 This is a flowchart of the quality traceability response and data distribution and storage process of this invention. Detailed Implementation
[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0023] Reference Figures 1 to 5 This invention discloses a method for axial determination and extrusion quality monitoring data processing, comprising the following steps: S1. Deploy corresponding sensors at the preset acquisition points of the extrusion processing equipment. The preset acquisition points specifically include one laser displacement sensor at each of the left and right ends of the extrusion rod axis, four piezoelectric pressure sensors evenly distributed around the circumference of the extrusion die cavity, one triaxial MEMS vibration sensor at the end of the main shaft of the equipment, and one K-type thermocouple temperature sensor on the inner wall of the die cavity. All sensors support the IEEE 1588 PTP clock synchronization protocol, and the sampling frequency is uniformly configured to 1kHz. The synchronization error with the PLC clock of the extrusion processing equipment is controlled within 1ms. The collection of multi-source time-series monitoring data for the entire extrusion process includes displacement data at both ends of the extrusion rod along the axis, extrusion pressure data distributed at four points around the die cavity, X / Y / Z three-axis vibration data of the equipment spindle, temperature data of the inner wall of the die cavity, and 13-bit millisecond-level Unix timestamp data corresponding to each sampling point.
[0024] After the extrusion process is started, the axial determination logic is called every 1ms. Based on the axial displacement data at both ends of the axis corresponding to the same timestamp and the extrusion pressure data distributed at multiple points, the axial state of the current processing is determined, and the result of axial normal or axial off-center load is obtained. The entire determination response delay is less than 50ms, which fully meets the requirements of industrial real-time control.
[0025] S2. The PLC control system of the extrusion processing equipment is connected via Modbus TCP protocol to obtain the process trigger signals output by the PLC in real time. The collected continuous multi-source time-series monitoring data is sliced according to the extrusion process nodes to obtain time-series data slices corresponding to each processing stage. The axial state determination result is used as a numerical label to associate with the time-series data slice of the corresponding processing stage and then stored in the InfluxDB columnar time-series database that conforms to the G06F16 classification standard. During storage, the Snappy lossless compression algorithm is automatically used to reduce the storage space occupation.
[0026] S3. Retrieve time-series data slices stored in the time-series database that simultaneously carry axial status tags and corresponding workpiece historical quality inspection results to construct a training dataset. The historical quality inspection results include the coordinate measuring machine test results of workpiece form and position tolerances and the tensile test results of mechanical properties. After the quality inspection equipment completes the inspection, it is automatically synchronized to the time-series database via the MQTT protocol and bound to the workpiece's unique RFID code to ensure accurate data association.
[0027] When training the extrusion quality recognition model, a stratified sampling method is used to divide the training set, validation set, and test set into a ratio of 7:2:1 to avoid the imbalance of class distribution affecting the model performance.
[0028] S4. Real-time access to multi-source time-series monitoring data of the extrusion process via industrial Ethernet bus. First, real-time axial state determination is completed. Then, the real-time time-series data of the current processing stage is preprocessed and input into the trained extrusion quality recognition model. The corresponding extrusion quality level is output. The real-time axial state determination result, real-time multi-source time-series monitoring data and extrusion quality level are associated with the unique RFID code of the currently processed workpiece and stored in the default storage partition of the time-series database to meet the needs of subsequent data retrieval and model iteration.
[0029] In this invention, the specific process for determining the axial state in step S1 is as follows: Displacement data collected by two axial laser displacement sensors at the same time point are extracted, and the absolute difference between the two sets of displacement data is calculated. Simultaneously, extrusion pressure data collected by four pressure sensors in the circumferential direction of the mold cavity at the same time point are extracted, and the maximum eccentricity difference between the maximum and minimum extrusion pressure values is calculated. When the displacement difference is greater than a preset displacement threshold and the maximum eccentricity difference is greater than a preset pressure threshold, the state is determined to be axially off-center. If either of these two conditions is not met, the state is determined to be axially normal. The specific determination logic is implemented through the following formula: ; ; in The difference in displacement between the two ends along the axis. This represents the real-time displacement value acquired by the axial first-end displacement sensor. The values are real-time displacement values collected by the axial second-end displacement sensor, all in millimeters. This represents the difference in eccentricity of the extrusion pressure in the mold cavity. This represents the maximum extrusion pressure collected by the circumferential pressure sensors of all mold cavities at the same time. This represents the minimum extrusion pressure collected by all circumferential pressure sensors in the mold cavities at the same time, with the unit being megapascals (MPA).
[0030] when and If the load is abnormal, it is determined to be an axial off-center load; otherwise, it is determined to be normal axial load. The preset displacement threshold is adaptively configured based on the axial form and position tolerance requirements of the currently processed workpiece, and its value is set to 0.8 times the maximum allowable axial tolerance of the workpiece. The preset pressure threshold is configured based on the rated cavity pressure of the current extrusion process, and is set to 5% of the rated extrusion pressure. The code snippet for the core decision logic is as follows: # Core logic for axial direction determination def axial_judge(d1: float, d2: float, p_list: list, d_th: float, p_th: float) ->int: delta_d = abs(d1 - d2) p_max = max(p_list) p_min = min(p_list) delta_p = p_max - p_min return 1 if (delta_d>d_th and delta_p>p_th) else 0 In this invention, the specific rules for slicing according to the extrusion process nodes in step S2 are as follows: the entire extrusion process is divided into four process nodes: feeding, pre-extrusion, holding pressure, and unloading. These correspond to five trigger points output by the PLC: feeding arrival signal, pre-extrusion start signal, holding pressure trigger signal, unloading start signal, and unloading arrival signal. The feeding arrival signal is triggered by a high-level output from the photoelectric sensor at the feeding station when it is blocked by the workpiece. The pre-extrusion start signal is triggered by the encoder feedback signal when the extrusion rod advances to the position contacting the workpiece. The holding pressure trigger signal is triggered by the pressure sensor signal when the mold cavity pressure reaches the preset holding pressure. The unloading start signal is triggered by the signal indicating the end of the holding pressure timer. The unloading arrival signal is triggered by the limit switch signal when the extrusion rod returns to its initial position. The multi-source timing monitoring data between two adjacent trigger points constitutes the timing data slice for the corresponding process. The four timing data slices correspond to the feeding stage, pre-extrusion stage, holding pressure stage, and unloading stage, respectively. There is no data overlap or omission between the slices. Each slice is also associated with the preset process parameters of the corresponding process, including pre-extrusion speed, holding pressure, and holding time.
[0031] In this invention, when storing the time-series data slices after associating tags into the time-series database in step S2, a three-level composite primary key index is constructed using the workpiece unique identifier, timestamp, and axial status tag. The workpiece unique identifier is a 12-bit globally unique code obtained by scanning with an RFID reader at the loading station before extrusion processing. The timestamp uses a 13-bit Unix millisecond-level timestamp format. The axial status tag uses an integer code where 0 represents axial normal and 1 represents axial off-center load. The time-series database establishes inverted TAG indexes for each of the three primary key dimensions, supporting multi-condition combined retrieval based on time, workpiece, and axial status dimensions. For example, retrieving all workpiece data with axial off-center load in May 2024 has a retrieval response latency of less than 100ms. When storing data, the Snappy lossless compression algorithm is automatically used for compression, with a compression ratio of no less than 8:1, meaning that 1GB of original monitoring data occupies no more than 125MB of storage space after compression, significantly reducing storage costs.
[0032] In this invention, before constructing the training dataset in step S3, a data preprocessing step is included: an isolated forest algorithm is used to remove abnormal data from the time-series data slices. The anomaly detection threshold for the isolated forest algorithm is set to 0.7; sampling points with anomaly scores higher than 0.7 are considered abnormal data and are removed. The anomaly score calculation logic of the isolated forest is implemented through the following formula: ; in For sample points Abnormal scores, For the subsampling size of each tree in an isolated forest, For sample points Path length in an isolated tree for Mean path length in all isolated trees for The average path length of a sample, when The data was identified as abnormal.
[0033] The data gaps after removal are filled by the mean of three adjacent valid sampling points in the same dimension, i.e., the missing sampling points. The value is ,in The values of the two valid sampling points before the empty defect, The first valid sample point value after the null defect. Then, the values of all feature dimensions are normalized using the maximum and minimum values. The normalization formula is: ;in These are the original eigenvalues. This is the minimum value of the feature in the training set. The maximum value of this feature in the training set is used to map the value range of each feature to the interval between 0 and 1, thereby eliminating the interference of numerical magnitude differences between different feature dimensions on model training.
[0034] In this invention, the extrusion quality recognition model trained in step S3 is a lightweight temporal convolutional network. The network structure of this lightweight temporal convolutional network sequentially includes two depthwise separable temporal convolutional layers, one global average pooling layer, and one fully connected classification layer. The first depthwise separable temporal convolutional layer has a kernel size of 3, a dilation coefficient of 1, and 16 output channels. The second depthwise separable temporal convolutional layer has a kernel size of 3, a dilation coefficient of 2, and 32 output channels. The receptive fields of the two convolutional layers, when superimposed, cover the entire temporal data slice. The global average pooling layer averages the temporal features of each channel into a single value. The fully connected layer outputs three classification results, and the activation function is Softmax. The total number of model parameters is controlled within 1M, supporting deployment and operation on low-computing-power devices such as Raspberry Pi and edge PLCs. The model's input is a multi-dimensional feature vector corresponding to a time series data slice, with dimensions of [sequence length, number of features] and 9 features: d1, d2, p1, p2, p3, p4, v_x, v_y, v_z, temp, for a total of 10 features. The sequence length is dynamically adapted according to the duration of different processes.
[0035] The output is a classification result of extrusion quality into three levels: qualified, pending re-inspection, and unqualified. Qualified means that the workpiece's form and position tolerances and mechanical properties meet the process requirements. Pending re-inspection means that there are suspected defects that need to be manually verified. Unqualified means that there are obvious defects and the workpiece is directly judged as scrap.
[0036] In this invention, after the real-time axial status determination is completed in step S4, if the determination result is axial off-center load, an axial off-center load early warning signal is immediately generated. The early warning signal includes the early warning level, the off-center load occurrence time, the displacement difference, and the extrusion force eccentricity difference. It is sent to the PLC control terminal of the extrusion processing equipment via the Modbus RTU protocol. When the displacement difference exceeds 1.5 times the threshold, a level two early warning is triggered, and the emergency shutdown logic is initiated. When the displacement difference does not exceed 1.5 times the threshold, a level one early warning is triggered, and the deceleration buffer logic is initiated to avoid workpiece scrapping or equipment damage due to axial off-center load. At the same time, the corresponding real-time multi-source time-series monitoring data is marked as high-priority data and stored in an independent high-priority storage partition of the time-series database. The retrieval priority of this partition is higher than that of the ordinary storage partition, and the data retention period is 3 times that of the ordinary storage partition. The ordinary storage partition data is retained for 1 year, and the high-priority storage partition data is retained for 3 years, which facilitates the subsequent source tracing analysis of off-center load accidents.
[0037] This invention also includes a quality traceability response step: receiving externally input quality traceability requests, the input channels of which include API requests initiated by the production management system, HTTP requests initiated by the quality inspection end, and local requests initiated by the equipment's human-machine interface; parsing the search keywords carried in the traceability request, the search keywords including the workpiece's unique identifier or processing time range; retrieving the corresponding full-process axial state determination results, full-stage multi-source time-series monitoring data, and extrusion quality grades from the time-series database; automatically generating a visualized time-series curve report and summary information, supporting export to CSV or PDF format; the traceability response latency is less than 200ms; and it supports up to 100 concurrent requests, meeting the traceability needs of large-scale production lines.
[0038] The present invention also includes a model iteration update step: according to a preset period, the newly added time series data slices with manual review results in the time series database are retrieved to form an incremental training set. The preset period is triggered every 7 days or every 1,000 newly added time series data slices with manual review results. The incremental training set is sampled in a ratio of 10:1:1 for qualified, pending review and unqualified data to avoid class imbalance.
[0039] The learning rate during incremental fine-tuning is set to 1 / 10 of the initial training learning rate. Only the parameters of the fully connected layer and the last convolutional layer are updated. This ensures that the model adapts to the new process data and avoids catastrophic forgetting of the model. The updated model is deployed and replaced only when its accuracy is verified on the test set to be more than 0.5% higher than that of the original model. Otherwise, the original model is kept running to prevent model degradation from affecting the quality recognition accuracy.
[0040] This invention also includes a data distribution and storage step: the real-time generated axial state determination results, extrusion quality grades, and associated multi-source time-series monitoring data are encrypted using the national cryptographic SM4 symmetric encryption algorithm, and then the content is filtered according to the permissions of different ports before being synchronously distributed to the visualization display interfaces of the equipment end, production management end, and quality inspection end. The equipment end only displays the real-time axial state, early warning information, and the quality grade of the currently processed workpiece; the production management end displays the quality statistics data and axial off-center load occurrence rate of the entire batch; and the quality inspection end displays the full-process monitoring data and quality grade of a single workpiece. The encrypted full data is synchronously backed up and stored on distributed cloud storage nodes, using a three-replica backup mechanism. The three replicas are stored in three different geographical availability zones, supporting cross-regional retrieval and access.
[0041] Reference Figure 1This process comprehensively demonstrates the closed-loop logic of axial force determination and extrusion quality monitoring data processing. The system first collects multi-source time-series monitoring data covering multiple dimensions such as displacement, extrusion pressure, vibration, and temperature, and uses this data to determine the axial stress state during processing. Subsequently, the system accurately segments the continuous data into multiple time-series data slices according to the process cycle and binds them with axial state labels, storing them in a time-series database. Based on the accumulated data, the system retrieves historical datasets to train and construct the extrusion quality recognition model. Finally, in a real-time production environment, the system simultaneously executes axial state monitoring and product quality prediction, archiving the output evaluation level and source data in a closed loop, realizing a complete industrial-grade data processing workflow from data perception and feature deconstruction to intelligent decision-making.
[0042] Reference Figure 2 This process details the axial state assessment and anomaly intervention mechanism during the extrusion process. The core algorithm evaluates process stability by comparing the displacement deviations at both ends of the equipment's axial direction and the extreme differences in extrusion pressure at multiple points. When both the displacement difference and the pressure eccentricity difference exceed preset safety thresholds, the system immediately classifies the current state as axial off-center loading. Once this off-center loading assessment is triggered, the control center not only sends high-priority early warning intervention signals to the underlying processing equipment terminals to prevent equipment damage or batch scrap, but also marks the multi-source time-series monitoring data of the batch that triggered the alarm with high priority and performs feature sealing, preserving the most complete high-frequency data slices from the field for subsequent fault backtracking and mechanism analysis. Under normal conditions, the data smoothly flows into subsequent pipelines.
[0043] Reference Figure 3 This workflow demonstrates the rules for cleaning, normalizing, and structuring high-frequency, multi-source monitoring data for data storage. To ensure high-quality input data for the model, the preprocessing stage first introduces the Isolation Forest algorithm to intelligently identify and remove transient flypoints and abnormal noise that may be generated by the sensors, and simultaneously performs maximum and minimum value normalization to eliminate dimensional barriers from multi-source sensors. In the data reconstruction stage, the system strictly aligns with the physical processing cycle, capturing the hard-line trigger signals of the four key process nodes—feeding, pre-extrusion, holding pressure, and unloading—as time-series cutting boundaries, decomposing the long sequence data into four highly cohesive feature slices. In the data storage stage, a composite primary key index structure consisting of timestamps, unique workpiece identifiers, and axial states is established.
[0044] Reference Figure 4This process elucidates the construction of a lightweight extrusion quality recognition engine and its self-evolution mechanism throughout its lifecycle. During model training, the traditional static threshold judgment is abandoned in favor of a lightweight temporal convolutional network sensitive to capturing temporal features. The network processes multi-dimensional slice feature vectors, ultimately outputting a three-level refined quality classification surface including qualified, pending re-inspection, and unqualified. To address the changes in feature distribution caused by aging equipment and process drift in industrial settings, the system constructs a long-term incremental model update mechanism. The underlying scheduler automatically sniffs out and extracts recently supplemented and manually verified sample sets from the temporal database at set intervals. These high-quality incremental data are input into the network for fine-tuning training, ensuring that the model parameters always closely match the latest physical state of the production line.
[0045] Reference Figure 5 This process outlines the system's service response and secure storage and distribution architecture at the back end of the data lifecycle. The system internally employs a dual-track data flow routing to address diverse business needs. In reverse query operations for quality traceability, the system possesses advanced parsing capabilities, enabling rapid reconstruction of the corresponding full-process profile and quality grading report from a vast time-series database using key credentials such as input workpiece identification codes or specified time windows. In forward distribution operations for production management, the system encrypts and packages real-time core judgment conclusions and sensor source data, then distributes them via low-latency bus multicast to the display panels of the three core business terminals: production line equipment, workshop management, and quality inspection.
[0046] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for axial determination and extrusion quality monitoring data processing, characterized in that, Includes the following steps: S1. Collect multi-source time-series monitoring data of the extrusion processing equipment. The multi-source time-series monitoring data includes at least axial displacement data at both ends, multi-point distributed extrusion pressure data, equipment vibration data, processing area temperature data, and timestamp data. Based on the displacement data at both ends of the axial direction and the multi-point distributed extrusion pressure data, the axial state of the current processing is determined, and the result of determining whether the axial direction is normal or the axial direction is off-center is obtained. S2. The collected multi-source time-series monitoring data is sliced according to the extrusion process nodes to obtain time-series data slices corresponding to each processing stage. The axial state determination result is used as a tag to associate with the corresponding time-series data slice and then stored in a preset time-series database. S3. Retrieve time-series data slices with axial status labels and historical quality inspection results stored in the time-series database to construct a training dataset and train the extrusion quality recognition model. S4. Real-time access to multi-source time-series monitoring data of the extrusion process, first complete the real-time axial state determination, then input the real-time time-series data into the trained extrusion quality recognition model, output the corresponding extrusion quality level, and store the real-time axial state determination result, real-time multi-source time-series monitoring data and extrusion quality level in the time-series database after association.
2. The method for axial determination and extrusion quality monitoring data processing according to claim 1, characterized in that, The specific process for determining the axial state in step S1 is as follows: calculate the difference between the displacement data at both ends of the axial direction, and at the same time calculate the maximum eccentricity difference of the multi-point distributed extrusion pressure. When the displacement difference is greater than the preset displacement threshold and the maximum eccentricity difference is greater than the preset pressure threshold, it is determined to be an axial off-center load state; otherwise, it is determined to be an axial normal state.
3. The method for axial determination and extrusion quality monitoring data processing according to claim 1, characterized in that, The specific rules for slicing according to the extrusion process nodes in step S2 are as follows: the entire extrusion process is divided into four process nodes: feeding, pre-extrusion, holding pressure, and unloading. The trigger signal of each process node is used as the cutting point to slice the continuous multi-source time-series monitoring data, resulting in four time-series data slices corresponding to the processing stages.
4. The method for axial determination and extrusion quality monitoring data processing according to claim 1, characterized in that, In step S2, when storing the time-series data slices after associating tags into the time-series database, a composite primary key index is constructed using timestamps, unique workpiece identifiers, and axial status tags. The time-series database supports multi-condition retrieval based on time dimension, workpiece dimension, and axial status dimension.
5. The method for axial determination and extrusion quality monitoring data processing according to claim 1, characterized in that, Before constructing the training dataset in step S3, a data preprocessing step is also included: using the isolated forest algorithm to remove abnormal data collected in the time series data slices, and then performing maximum and minimum value normalization on the values of all feature dimensions to unify the feature value range.
6. The method for axial determination and extrusion quality monitoring data processing according to claim 1, characterized in that, The extrusion quality recognition model trained in step S3 is a lightweight temporal convolutional network. The input of the lightweight temporal convolutional network is the multi-dimensional feature vector corresponding to the temporal data slice, and the output is the three-level classification results of extrusion quality: qualified, pending re-inspection, and unqualified.
7. The method for axial determination and extrusion quality monitoring data processing according to claim 1, characterized in that, After the real-time axial state determination is completed in step S4, if the determination result is axial off-center load, an axial off-center load early warning signal is immediately generated and sent to the control terminal of the extrusion processing equipment. At the same time, the corresponding real-time multi-source time-series monitoring data is marked as high-priority data and stored in the time-series database.
8. The method for axial determination and extrusion quality monitoring data processing according to claim 1, characterized in that, It also includes a quality traceability response step: receiving an externally input quality traceability request, parsing the search keywords carried in the traceability request, the search keywords including the workpiece unique identifier or processing time range, retrieving the corresponding axial state judgment result, full-process multi-source time-series monitoring data and extrusion quality level from the time-series database and outputting them.
9. The method for axial determination and extrusion quality monitoring data processing according to claim 1, characterized in that, It also includes a model iteration and update step: according to a preset cycle, the newly added time series data slices with manual review results in the time series database are retrieved to form an incremental training set, and the extrusion quality identification model is incrementally fine-tuned to update the model parameters.
10. The method for axial determination and extrusion quality monitoring data processing according to claim 1, characterized in that, It also includes data distribution and storage steps: after encrypting the real-time generated axial state judgment results, extrusion quality level and related multi-source time-series monitoring data, it is synchronously distributed to the visualization display interfaces of the equipment end, production management end and quality inspection end, and backed up and stored to distributed cloud storage nodes.