A Multi-Source Data Fusion Processing Method for Fastener Heat Treatment

Abstract

This invention discloses a multi-source data fusion processing method for fastener heat treatment, relating to the field of data fusion, comprising five major steps: multi-source data acquisition, data preprocessing, feature extraction, fusion modeling, and result optimization. The multi-source data acquisition terminal collects process data, material data, quality inspection data, and environmental data; data preprocessing involves data cleaning and standardization; the feature extraction terminal extracts time-series feature parameters, compositional feature parameters, quality feature parameters, and interference feature parameters; fusion modeling combines prior knowledge of mechanisms with data-driven approaches to achieve feature layer and decision layer fusion; and result optimization smoothly corrects the fusion results, outputting quality assessments and process adjustment suggestions. This invention improves the accuracy and stability of fastener heat treatment quality assessment through deep multi-source data fusion, achieves adaptive optimization of process parameters, and is suitable for full-process control of heat treatment in structured industrial scenarios.

Description

Technical Field

[0001] This invention relates to the field of data fusion, specifically to a multi-source data fusion processing method for fastener heat treatment. Background Technology

[0002] In the field of fastener heat treatment production, product quality is highly correlated with process parameters. Key indicators such as hardness, strength, and deformation directly depend on the coordinated control of parameters such as heating temperature, holding time, and cooling rate. However, traditional production methods have long relied on manual experience for process control, resulting in prominent problems such as lagging quality assessment and blind parameter optimization.

[0003] Currently, industry data collection is mostly limited to a single process step, such as monitoring only the temperature of the heating furnace or the flow rate of the cooling medium. There is a lack of systematic integration of process data, such as equipment power and furnace pressure, material data, such as composition content and original hardness, quality test results, such as tensile strength and metallographic structure, and environmental interference, such as workshop temperature and humidity and power grid voltage. This makes it difficult to quickly trace the source when quality fluctuations occur.

[0004] While existing data processing methods attempt to incorporate simple statistical analysis, they lack the ability to handle noise interference and spatiotemporal misalignment in multi-source heterogeneous data, and their depth of feature correlation mining is insufficient, failing to establish a precise mapping relationship between process parameters and quality indicators. Furthermore, process parameter adjustments are often based on fixed thresholds, making it difficult to adapt to batch-to-batch material variations and environmental fluctuations, resulting in poor product consistency and significant fluctuations in yield. Therefore, developing a method that can integrate multi-source data and achieve accurate quality assessment and adaptive process optimization throughout the entire process is crucial for improving the heat treatment level of fasteners. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a rapidly deployable laser ground receiving system to address the problems raised in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multi-source data fusion processing method for fastener heat treatment, comprising:

[0007] S1. Multi-source data acquisition: The fastener to be processed is recorded as the target object. Multi-source data of the target object during the heat treatment process is collected in real time through the multi-source data acquisition terminal. The multi-source data includes process data, material data, quality inspection data and environmental data.

[0008] S2. Data Preprocessing: Multi-source data is cleaned and standardized through a data preprocessing terminal to obtain process data preprocessing parameters, material data preprocessing parameters, quality inspection data preprocessing parameters, and environmental data preprocessing parameters.

[0009] S3. Feature Extraction: Input the process data preprocessing parameters, material data preprocessing parameters, quality inspection data preprocessing parameters, and environmental data preprocessing parameters into the feature extraction terminal, and output the time series feature parameters, composition feature parameters, quality feature parameters, and interference feature parameters;

[0010] S4. Fusion Modeling: The component feature parameters, quality feature parameters and interference feature parameters are input into the fusion model through the fusion modeling terminal. The feature layer is fused by combining the mechanism prior weight and the data-driven attention mechanism. Then, the decision layer is fused through the multi-quality index sub-model, and the fusion result is output.

[0011] S5. Result Optimization: The fusion results are processed through the result optimization terminal, and the heat treatment quality assessment results of the target object are finally output. The heat treatment quality assessment results are judged, and if the preset results are not met, the process parameters are adjusted.

[0012] The technical effects and advantages of this invention are as follows:

[0013] 1. This invention integrates process, material, quality, and environmental data through multi-source data acquisition and fusion modeling, breaking through the limitations of traditional single-stage data, constructing a full-process quality mapping relationship, realizing accurate assessment of heat treatment quality, and solving the problem of difficulty in tracing the source of quality fluctuations;

[0014] 2. This invention combines mechanistic priors with a data-driven attention mechanism to dynamically optimize feature weights. Compared with fixed threshold methods, it can adapt to material batch differences and environmental fluctuations, improve product consistency, and reduce the fluctuation range of the pass rate.

[0015] 3. This invention optimizes the terminal output process adjustment suggestions through results, derives parameter adjustment amounts based on quality deviation quantification, realizes closed-loop control from data acquisition to parameter optimization, reduces reliance on manual experience, and improves production efficiency and quality stability. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a diagram illustrating the terminal connection.

[0018] Figure 2 This is a schematic diagram of the process implementation.

[0019] Figure 3 This is a diagram illustrating the data processing procedure. Detailed Implementation

[0020] 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 protection scope of the present invention.

[0021] Please see Figure 1 As shown, the present invention provides a multi-source data fusion processing method for fastener heat treatment, including a multi-source data acquisition terminal, a data preprocessing terminal, a feature extraction terminal, a fusion modeling terminal, and a result optimization terminal.

[0022] The multi-source data acquisition terminal is connected to the data preprocessing terminal to transmit the acquired process data, material data, quality inspection data, and environmental data to the data preprocessing terminal; the data preprocessing terminal is connected to the feature extraction terminal to transmit the preprocessed parameters to the feature extraction terminal; the feature extraction terminal is connected to the fusion modeling terminal to transmit the extracted time-series features, composition features, quality features, and interference feature parameters to the fusion modeling terminal; the fusion modeling terminal is connected to the result optimization terminal to transmit the fusion results to the result optimization terminal, and finally outputs the heat treatment quality assessment results, process parameter adjustment suggestions, and the entire process data trajectory.

[0023] Please see Figure 2 As shown, the present invention provides a multi-source data fusion processing method for fastener heat treatment, comprising:

[0024] S1. Multi-source data acquisition: The fastener to be processed is recorded as the target object. Multi-source data of the target object during the heat treatment process is collected in real time through the multi-source data acquisition terminal. The multi-source data includes process data, material data, quality inspection data and environmental data.

[0025] Process data includes furnace temperature, cooling medium flow rate, furnace pressure, equipment operating power, and operating time;

[0026] Material data includes the material composition and original hardness of the target object;

[0027] Quality inspection data includes hardness, tensile strength, metallographic images, and deformation after heat treatment;

[0028] Environmental data include workshop temperature, humidity, and grid voltage.

[0029] Process data refers to the data collected in real time by various online sensors during the heat treatment process of fasteners, which are related to the equipment operating status and process parameters, and directly reflect the dynamic changes in each stage of heat treatment.

[0030] In this embodiment, the data acquisition range and specific method are as follows: The furnace temperature is acquired using distributed thermocouples within the furnace. Three measuring points (a total of nine thermocouples) are arranged in each of the upper, middle, and lower layers of the furnace cavity to ensure coverage of the temperature field in different areas of the furnace. The measurement range is 20-1200℃, with an accuracy of ±1℃, and the sampling frequency is set to 1Hz. The acquired raw data includes the real-time temperature values ​​of each measuring point. For example, the upper layer thermocouples acquire 902℃, 899℃, and 903℃; the middle layer acquires 898℃, 897℃, and 899℃; and the lower layer acquires 900℃, 898℃, and 902℃. This data is transmitted to the data acquisition card via dedicated thermocouple signal lines, and then uploaded to the data server via industrial Ethernet for subsequent reflection of the temperature distribution and trends within the furnace.

[0031] The real-time flow rate of the cooling medium is collected using an electromagnetic flow meter installed at the inlet of the cooling pipe. The measurement range is 0-50 L / min, with an accuracy of ±0.5% and a sampling frequency of 0.5 Hz. Due to the influence of the equipment's power supply system, the collected raw data is subject to 50 Hz power frequency interference, exhibiting periodic fluctuations, such as alternating within the range of 25-27 L / min. The flow data is transmitted to the PLC via a 4-20mA analog signal, and then forwarded by the PLC to the data server for monitoring whether the cooling rate meets the process requirements.

[0032] The furnace internal pressure is collected using a pressure transmitter mounted on the side of the furnace body. The measurement range is -0.1 to 0.5 MPa, with an accuracy of ±0.2% and a sampling frequency of 0.5 Hz. Due to the influence of furnace sealing performance and ambient temperature changes, the raw data exhibits baseline drift. For example, if the pressure increases by 0.01 MPa per hour, the data collected at 8:00 AM will be 0.20 MPa, at 9:00 AM 0.21 MPa, and at 10:00 AM 0.22 MPa. The pressure data is transmitted to the data acquisition terminal via an RS485 bus and finally uploaded to the server to ensure that the furnace internal pressure remains stable within the process-specified range.

[0033] The operating power of the heat treatment equipment is collected using power sensors installed on the equipment's power supply line. The measurement range is 0-500kW, with an accuracy of ±1% and a sampling frequency of 1Hz. Power spikes occur during equipment start-up and shutdown; for example, the power can reach 600kW during startup, exceeding the normal operating range. Power data is transmitted wirelessly to a gateway, which then aggregates and uploads it to a server for monitoring equipment energy consumption and operating status. Operating time is recorded by the system clock for each process stage, with an accuracy of ±0.1s. For example, the start time of the heating stage is marked as t=0s; when the temperature reaches the austenitizing temperature, such as 850℃, it is marked as t=300s, indicating the start of the holding stage; the start time of the cooling stage is marked as t=900s, indicating the end of the holding stage. Operating time data serves as a time-series reference and is stored in association with parameters such as temperature, pressure, and flow rate to ensure consistency of various process data across time.

[0034] Material data refers to the original property data of fasteners before heat treatment. It reflects the inherent characteristics of the material and has a direct impact on the quality after heat treatment.

[0035] In this embodiment, the material data collection content and method are as follows: The material composition content of the fastener raw materials is detected offline using a spectrometer. The detected elements include key alloying elements such as C, Si, Mn, Cr, Ni, and Mo, with a detection accuracy of ±0.01%. For example, the material composition of a batch of bolts is: C=0.45%, Si=0.25%, Mn=1.2%, Cr=0.8%, Ni=0.3%, and Mo=0.1%. After the detection is completed, the composition data is entered into the material information management system to form a material composition database, which serves as the basis for subsequent analysis of material hardenability, hardness, and other properties.

[0036] The initial hardness of the fasteners before heat treatment was tested using a Rockwell hardness tester. Tests were performed three times at different locations on the fastener (e.g., head, shank, tail), and the average value was taken as the initial hardness value. For example, the test results for a nut were 25 HRC, 26 HRC, and 25 HRC, with an average of 25.3 HRC. The initial hardness data was recorded in the quality inspection report and uploaded to the data server for comparison with the hardness after heat treatment to evaluate the effectiveness of the heat treatment.

[0037] Quality inspection data refers to data related to product quality obtained through offline testing after fastener heat treatment, directly reflecting the final effect of the heat treatment process.

[0038] In this embodiment, the data collection content and method for quality inspection are as follows: The hardness after heat treatment is measured using the same Rockwell hardness tester as the original hardness test. Five test points (uniformly distributed along the axial direction) are selected on the heat-treated fastener, and each point is tested three times. The average value is taken as the hardness value at that location. For example, the test results for a bolt are: head 44.8 HRC, 45.2 HRC, average 45.0 HRC; middle 43.9 HRC, 44.1 HRC, average 44.0 HRC; tail 42.8 HRC, 43.2 HRC, average 43.0 HRC. The hardness data is entered into the quality inspection system to evaluate whether the hardness of the fastener meets the standard and its uniformity.

[0039] Tensile strength was tested using a universal testing machine on five fastener specimens from the same batch. The maximum load at which the specimen fractured was recorded during the test, and the tensile strength was calculated based on the specimen's cross-sectional area. For example, test results of 1000 MPa, 1020 MPa, 990 MPa, 1010 MPa, and 1030 MPa for the five specimens were used to assess whether the fastener's mechanical strength met the design requirements.

[0040] Metallographic images of the fastener cross-section were acquired using a metallographic microscope. The sample was first ground, polished, and etched before being observed under the microscope. The image resolution was 2048×1536 pixels, and the magnification was 500×. For example, the acquired images may contain martensite and pearlite structures, which are used to analyze grain size and phase composition ratios. Deformation was measured using a laser diameter gauge to measure the diameter and length of the fastener before and after heat treatment. The dimensional change was calculated as the deformation. For example, a bolt with a diameter of 10.000 mm before treatment and a diameter of 10.005 mm after treatment had a deformation of 5 μm; its length was 50.00 mm before treatment and 50.02 mm after treatment, with a deformation of 20 μm. The deformation data was used to assess whether the dimensional stability of the fastener met assembly requirements.

[0041] Environmental data refers to the environmental parameters of the heat treatment workshop. Although they do not directly participate in process control, they may indirectly affect equipment operation and process stability. In this embodiment, the environmental data collection content and method are as follows: Workshop temperature is collected using temperature and humidity sensors. The sensors are installed in corners of the workshop and near the equipment (a total of 5 measuring points). The measurement range is 0-50℃, with an accuracy of ±0.5℃ and a sampling frequency of 0.1Hz. For example, the collected data are 25℃, 26℃, 25℃, 27℃, and 25℃, used to analyze the impact of ambient temperature on the heat dissipation of the heating furnace.

[0042] Humidity and workshop temperature are collected using the same humidity sensor, with a measurement range of 10%-90%RH, an accuracy of ±3%RH, and a sampling frequency of 0.1Hz. For example, the collected data are 40%RH, 42%RH, 39%RH, 41%RH, and 40%RH, which are used to assess the impact of humidity on the electrical components of the equipment and the evaporation of the cooling medium.

[0043] The workshop's mains voltage is collected using a voltage sensor installed in the distribution cabinet. The measurement range is 380±50V, with an accuracy of ±0.5V and a sampling frequency of 0.1Hz. For example, the collected data are 380V, 379V, 381V, 380V, and 378V, which are used to monitor the impact of voltage fluctuations on the power output of the heating furnace.

[0044] S2. Data Preprocessing: Multi-source data is cleaned and standardized through a data preprocessing terminal to obtain process data preprocessing parameters, material data preprocessing parameters, quality inspection data preprocessing parameters, and environmental data preprocessing parameters.

[0045] Process data preprocessing parameters were obtained using a smoothing algorithm; material data preprocessing parameters were obtained using the range standardization method, and the calculation formula is as follows:

[0046] ,

[0047] in, The output value after standardization. It is the minimum value under this feature dimension. The maximum value under the feature dimension. These are the original data values.

[0048] The specific processing method for process data preprocessing is as follows: For heating furnace temperature preprocessing, firstly, the spatial consistency of the original temperature data of 9 measuring points is checked, and abnormal measuring points that deviate significantly from the overall trend are removed. For example, if a measuring point displays 1200℃ due to a fault, it is directly removed. Then, the arithmetic mean of the three layers of measuring points is calculated. For example, the average value of the upper layer is 901.3℃, the average value of the middle layer is 898.0℃, and the average value of the lower layer is 900.0℃. The overall average value is (901.3+898.0+900.0) / 3≈899.8℃.

[0049] Then, a sliding window mid-range filter is applied, with the window size set to 5 sampling points (i.e., 5 seconds). For example, if the original data is 899℃, 900℃, 950℃, 901℃, 900℃, after filtering, it is corrected to 899℃, 900℃, 900℃, 901℃, 900℃, filtering out instantaneous pulse interference. Finally, a smooth temperature time series curve is obtained, which is used for subsequent extraction of features such as the rate of temperature change.

[0050] Cooling medium flow rate preprocessing addresses the 50Hz power frequency interference in the original data. Fourier transform is used to convert the data to the frequency domain, identifying and removing the 50Hz and its harmonic components. Then, inverse Fourier transform is used to convert the data back to the time domain, obtaining the denoised flow rate data. For example, the data fluctuated between 25-27 L / min before denoising, and stabilized at around 26 L / min after denoising. Subsequently, the average value of a 1-minute sliding window (i.e., the average of 12 sampling points) is calculated to further smooth the data. For example, the 12 denoised data points are 26.1, 25.9, 26.0, 26.2, 25.8, 26.0, 26.1, 25.9, 26.0, 26.1, 25.9, and 26.0, with an average of 26.0 L / min, serving as the basic data for the average flow rate characteristic.

[0051] To address baseline drift, the furnace pressure preprocessing employed a linear fitting method to eliminate it. Assuming the pressure data from 8:00 to 10:00 were 0.20 MPa, 0.205 MPa, 0.21 MPa, 0.215 MPa, and 0.22 MPa, the fitted drift formula was 0.01t, where t is the number of hours. 8:00 corresponds to t=0, 9:00 corresponds to t=1, and so on. Subtracting the drift from the original data yielded corrected pressure data: 0.20 MPa, 0.195 MPa, 0.20 MPa, 0.205 MPa, and 0.21 MPa, thus eliminating the linear drift. Then, the pressure fluctuation range within 5 minutes is calculated, and the difference between the maximum and minimum values ​​is calculated. For example, the corrected pressure data for 8:00-8:05 are 0.20MPa, 0.198MPa, 0.202MPa, 0.199MPa, and 0.201MPa, with a maximum value of 0.202MPa and a minimum value of 0.198MPa, and a fluctuation range of 0.004MPa.

[0052] The equipment's operating power preprocessing sets a power threshold of 0-500kW. Peak data exceeding this threshold is corrected by replacing it with the mean of three adjacent normal sampling points. For example, if a data segment contains 480kW (t=1s), 600kW (t=2s, abnormal), and 470kW (t=3s), then the power correction at t=2s is (480+470) / 2=475kW. The corrected data is then used to calculate the power standard deviation over 30 seconds (i.e., the dispersion of the 30 sampling points). For example, if the mean of the 30 data points is 490kW and the standard deviation is 5kW, it reflects the stability of the power output.

[0053] Runtime preprocessing requires no complex steps; it only aligns with other process data using timestamps to ensure that each data point, such as temperature, pressure, and flow rate, corresponds to a unique time stamp. For example, the starting time of the heating phase, t=0s, is associated with the first temperature data, 20℃, and t=300s is associated with the starting temperature of the heat preservation phase, 850℃, forming a multi-parameter time series matrix with time as the horizontal axis.

[0054] The specific processing method for material data preprocessing is as follows: The original content of C element is 0.45%, the minimum value of this feature dimension is 0.3%, and the maximum value is 0.6%. The calculated normalized value of C element is (0.45-0.3) / (0.6-0.3)=0.5. Similarly, the normalized value of Si element is calculated to be 0.5, the normalized value of Mn element is 0.4, the normalized value of Cr element is 0.6, the normalized value of Ni element is 0.33, and the normalized value of Mo element is 0.33.

[0055] Furthermore, these standardized values ​​are combined into a component feature vector [0.5, 0.5, 0.4, 0.6, 0.33, 0.33] for subsequent fusion modeling. The original hardness preprocessing does not require standardization; only the average value of multiple tests needs to be calculated and directly stored as the initial material performance parameter for comparison with the hardness after heat treatment. For example, if the average original hardness of a batch of fasteners is 25.3 HRC and the average hardness after heat treatment is 45.0 HRC, the hardness increase can be directly calculated as 19.7 HRC.

[0056] The specific processing method for quality inspection data preprocessing is as follows: After heat treatment, the average hardness values ​​at each test point are processed, and the hardness gradient along the fastener axis is calculated using the following formula:

[0057] ,

[0058] For example, if a bolt has a head hardness of 45.0 HRC, a tail hardness of 43.0 HRC, and an axial length of 20 cm, then the hardness gradient is 0.1 HRC / cm, which reflects the rate of change of hardness along the axial direction and assesses the uniformity of hardness.

[0059] For tensile strength pretreatment, the mean and standard deviation of the tensile strength data of 5 samples were calculated. The mean was (1000+1020+990+1010+1030) / 5=1010MPa, and the standard deviation was 15MPa. The mean was used as the representative value of the tensile strength of the batch, and the standard deviation was used as the strength dispersion to evaluate the performance consistency within the batch.

[0060] The metallographic image preprocessing first involves Gaussian filtering the original image to remove noise, then using an edge detection algorithm to identify grain boundaries, randomly selecting 100 grains to calculate the equivalent diameter, and then using grayscale thresholding to segment and statistically analyze the proportion of martensite phase. These two quantitative indicators are then used as image features.

[0061] The maximum dimensional difference before and after heat treatment is recorded in the deformation preprocessing. The maximum value is taken as the peak value of deformation and compared with the standard value to evaluate dimensional stability.

[0062] The specific processing methods for environmental data preprocessing are as follows: For workshop temperature preprocessing, calculate the temperature fluctuation range within one hour. For example, if the temperature data for a certain hour is 23-27℃, with a fluctuation range of 4℃, and mark the times exceeding the normal range (15-35℃), such as a temperature of 36℃ in summer, it is marked as abnormal. Humidity preprocessing requires no additional changes; it is only recorded when the humidity exceeds the 10%-90%RH range, such as 92%RH during the rainy season, as a reference for environmental interference factors. For power grid voltage preprocessing, calculate the voltage standard deviation within 10 minutes, then divide it by the rated voltage of 380V to obtain the voltage stability coefficient, which is used to judge the power grid stability (requirement ≤2%).

[0063] S3. Feature Extraction: Input the process data preprocessing parameters, material data preprocessing parameters, quality inspection data preprocessing parameters, and environmental data preprocessing parameters into the feature extraction terminal, and output the time series feature parameters, composition feature parameters, quality feature parameters, and interference feature parameters;

[0064] The time-series characteristic parameters include the rate of temperature change, pressure fluctuation amplitude, mean flow rate, and standard deviation of power.

[0065] Component characteristic parameters include the content percentage of each element and the uniformity of the components;

[0066] Quality characteristic parameters include hardness gradient, strength dispersion, and peak deformation.

[0067] Interference characteristic parameters include temperature fluctuation range and voltage stability coefficient.

[0068] The temperature change rate is calculated every 10 seconds based on the pre-treated furnace temperature time-series curve, using the temperature values ​​from five consecutive sampling points. The formula is as follows:

[0069] ,

[0070] in, This is the temperature value at the current sampling point. This is the average temperature of the first four sampling points. The sampling time interval, The temperature change rate is used; the component uniformity is calculated by the ratio of the standard deviation to the mean of the content of each element.

[0071] Taking this embodiment as an example, the current sampling point temperature is 900.5℃, the average temperature of the first 4 sampling points is 900.1℃, the sampling time interval is 1 second, and the result is 0.4℃ / s when substituted into the formula. This feature reflects the heating rate. If it exceeds the process threshold (such as 1℃ / s), it indicates that the heating is too fast and may cause uneven tissue.

[0072] The pressure fluctuation range is calculated based on the pre-processed furnace pressure data, which is the difference between the maximum and minimum pressure values ​​within 5 minutes. Taking this embodiment as an example: in the pressure data from 8:00 to 8:05, the maximum value is 0.202 MPa, the minimum value is 0.198 MPa, and the fluctuation range is 0.004 MPa. If it exceeds 0.01 MPa, it indicates that there may be an abnormality in the furnace body seal, which needs to be checked in conjunction with the equipment log.

[0073] The average flow rate is calculated based on the pre-processed cooling medium flow rate data, using the average value of a 1-minute sliding window (12 sampling points). Taking this embodiment as an example, the average flow rate data within a certain window is 26.0 L / min. This characteristic directly reflects the stability of the cooling medium supply. If it is lower than the process requirement of 25 L / min, the pump power needs to be adjusted.

[0074] The power standard deviation is calculated based on preprocessed equipment operating power data, determining the power dispersion over 30 seconds. For example, if the mean of 30 data points is 490kW and the standard deviation is 5kW, the formula is:

[0075] ,

[0076] in, For a single power value, The mean is n=30. The smaller the standard deviation, the more stable the power output of the equipment. If it exceeds 10kW, it indicates that there are abnormal fluctuations in the power grid or the equipment.

[0077] The content ratio of each element is based on the standardized data of the material composition after pretreatment. The standardized value is used directly as a feature to reflect the proportion of each element within the allowable range of the process. For example, a standardized value of 0.5 for element C indicates that its content is at a medium level in historical batches. The holding time needs to be adjusted in combination with the austenitizing temperature. High C content requires extended holding time.

[0078] Compositional homogeneity is calculated based on the original material composition data of three samples from the same batch, using the ratio of the standard deviation to the mean, as shown in the formula:

[0079] ,

[0080] in , Substituting the values, we get approximately 2.2%. If the uniformity exceeds 5%, it indicates that the raw material components are segregated, and the heating temperature should be appropriately increased during heat treatment to promote component diffusion.

[0081] The hardness gradient is calculated based on the hardness data after pretreatment and heat treatment, using the following formula to determine the rate of change of hardness along the axial direction:

[0082] ,

[0083] The axial length is 20cm, which gives 0.1HRC / cm. The smaller the gradient, the better the uniformity of hardness. If it exceeds 0.3HRC / cm, it indicates that the cooling medium is unevenly distributed and the nozzle position needs to be adjusted.

[0084] The strength dispersion is calculated based on the tensile strength data after pretreatment. The standard deviation of 5 samples is obtained as 15 MPa. The smaller the dispersion, the more stable the batch performance. If it exceeds 30 MPa, the uniformity of the heating furnace temperature field needs to be checked.

[0085] The metallographic structure characteristics are extracted based on the preprocessed metallographic image quantization data, extracting the equivalent diameter of grains and the proportion of martensite phase. In this embodiment, 100 grains are randomly selected with an average diameter of 15μm. If the diameter exceeds 20μm, it indicates that the heat preservation time is too long, resulting in coarse grains. The proportion of grains is obtained by grayscale segmentation. If the proportion is less than 80%, it indicates that the cooling rate is insufficient and the flow rate of the cooling medium needs to be increased.

[0086] The peak deformation value is based on the preprocessed deformation data, with the maximum value of 20μm taken as the feature. If it exceeds the process threshold of 10μm, it indicates that the quenching stress distribution is uneven and the cooling path needs to be optimized.

[0087] The temperature fluctuation range is based on the pre-processed workshop temperature data. The maximum temperature difference within 1 hour is calculated to be 4℃. If it exceeds 5℃, it indicates that the workshop temperature control system is abnormal, which may lead to unstable heat dissipation of the heating furnace. It needs to be corrected in conjunction with the furnace temperature data.

[0088] The voltage stability coefficient is calculated based on preprocessed grid voltage data, as the ratio of the standard deviation of voltage over 10 minutes to the rated value. The formula is as follows:

[0089] ,

[0090] in , Substituting this, we get approximately 0.4%. If the coefficient exceeds 2%, it indicates that voltage fluctuations may affect the heating power output, and the weight of the power feature needs to be increased in the fusion model.

[0091] Finally, the extracted feature parameters are integrated into feature vectors according to categories. For example, the time series feature vector is: [0.4℃ / s, 0.004MPa, 26.0L / min, 5kW]; the composition feature vector is: [0.5, 0.5, 0.4, 0.6, 0.33, 0.33, 2.2%]; the mass feature vector is: [0.1HRC / cm, 15MPa, 15μm, 85%, 20μm]; and the interference feature vector is: [4℃, 0.4%]. These feature vectors will be used as inputs to the fusion modeling terminal, and deep fusion of multi-source data will be achieved through mechanistic prior weights and data-driven attention mechanisms.

[0092] S4. Fusion Modeling: The component feature parameters, quality feature parameters and interference feature parameters are input into the fusion model through the fusion modeling terminal. The feature layer is fused by combining the mechanism prior weight and the data-driven attention mechanism. Then, the decision layer is fused through the multi-quality index sub-model, and the fusion result is output.

[0093] The fusion model includes: feature layer fusion and decision layer fusion;

[0094] The mechanism prior weights are set based on the correlation between the austenitizing temperature and the holding time of the fastener heat treatment. The data-driven attention mechanism learns the feature importance coefficients, and the fused feature vector is the sum of the products of various feature parameters and their corresponding weights.

[0095] The multi-quality index sub-model includes a hardness uniformity sub-model, a strength stability sub-model, a microstructure conformity sub-model, and a dimensional accuracy sub-model.

[0096] The hardness uniformity sub-model takes hardness gradient features as its core input and outputs a hardness distribution uniformity score, as shown in the formula:

[0097] ,

[0098] in This represents the hardness gradient value. This is the hardness influence coefficient. It is a natural constant;

[0099] The strength stability sub-model outputs a strength consistency coefficient based on the strength dispersion characteristic, and the formula is as follows:

[0100] ,

[0101] in For strength standard deviation, This represents the maximum permissible standard deviation.

[0102] The microstructure qualification sub-model, combining grain diameter and martensite phase ratio characteristics, outputs the microstructure qualification rate, as shown in the formula:

[0103] ,

[0104] Where d is the actual grain diameter. These are the upper and lower limits of the grain diameter, respectively. The proportion of martensite phase;

[0105] The dimensional accuracy sub-model outputs the dimensional deviation level based on the peak deformation characteristics, using the following formula:

[0106] ,

[0107] in This represents the peak value of the deformation. This is the deformation threshold;

[0108] The output values ​​of each sub-model are assigned preset weights. Weighted summation, where The preset weights are 1.

[0109] Preferably, the following embodiment is used: The core of feature layer fusion is to integrate component feature parameters, quality feature parameters, and interference feature parameters in a spatial dimension. Through mechanism prior weights and data-driven attention mechanism, the contribution of key features is highlighted. First, feature parameters are input, and three types of feature parameter vectors are input: component feature vector: [0.5, 0.5, 0.4, 0.6, 0.33, 0.33, 2.2%], quality feature vector: [0.1HRC / cm, 15MPa, 15μm, 85%, 20μm], and interference feature vector: [4℃, 0.4%]. Then, mechanism prior weights are assigned. Based on the heat treatment process mechanism, prior weights are assigned to different features: In the component features, C element (0.5) and Cr element (0.6) have a significant impact on hardenability, and their weights are set to 0.2 and 0.15, respectively. The component uniformity of 2.2% is weighted at 0.1.

[0110] In terms of quality characteristics, the martensite content of 85% and the hardness gradient of 0.1 HRC / cm directly reflect the phase transformation quality, with weights set to 0.2 and 0.15, respectively.

[0111] Among the interference characteristics, the voltage stability coefficient of 0.4% has a greater impact on the heating power, and its weight is set to 0.1. The weight of the temperature fluctuation range (4℃) is set to 0.05. The total weight is 1.0 to ensure that the contribution of the characteristics can be directly compared.

[0112] Furthermore, an attention network is constructed using a data-driven attention mechanism, learning dynamic feature weights based on historical data: the input layer receives standardized feature vectors, the hidden layer uses the ReLU activation function, and the output layer generates attention weights through the Softmax function. For example, when a batch has a high C content (0.5%) and a low martensite content (85%), the network automatically increases the attention weight of the C element feature (e.g., from 0.2 to 0.25) and decreases the weight of the temperature fluctuation feature (e.g., from 0.05 to 0.03). Finally, the attention weight and the prior mechanism weight are weighted and fused together, using the following formula:

[0113] ,

[0114] Where 0.7 and 0.3 are the balance coefficients between mechanism and data. As a priori weight of the mechanism, For attention weights.

[0115] Finally, it is the feature fusion output. The weighted feature vectors are concatenated into a fusion feature matrix, which is a 14-dimensional matrix in this embodiment.

[0116] The decision-level fusion is based on the fusion feature matrix output by the feature level. Through the integrated learning of multiple quality index sub-models, the comprehensive quality evaluation result is output. First, multiple quality index sub-models are constructed, and sub-models are trained respectively for key quality indexes: The hardness uniformity sub-model takes the fusion feature matrix as the input, adopts a random forest regression model, and outputs the hardness uniformity score. For example, after inputting the features of a certain batch, the model outputs 85 points, corresponding to a hardness gradient of 0.1 HRC / cm, which is better than the threshold of 0.3 HRC / cm.

[0117] The strength stability sub-model uses support vector machine regression to output the predicted value of strength dispersion and converts it into a score. For example, for 15 MPa, the score is 90 points because 15 MPa < the threshold of 30 MPa.

[0118] The microstructure qualification sub-model is based on a convolutional neural network. With the grain diameter of 15 μm and the martensite proportion of 85% in the fusion features as the input, the microstructure qualification rate is output as 92%.

[0119] The dimensional accuracy sub-model uses linear regression to output the predicted peak value of deformation as 20 μm, which is converted into a score of 60 points because 20 μm > the threshold of 10 μm.

[0120] Furthermore, sub-model weight assignment is carried out. Weights are assigned to sub-models according to the importance of quality indexes: The weights of hardness uniformity and microstructure qualification rate are relatively high, both set to 0.3; the weight of strength stability is set to 0.25; the weight of dimensional accuracy is set to 0.15, and the total weight is 1.0, reflecting the influence degree of different indexes on the final quality. Finally, it is the comprehensive decision output. The weighted sum of the sub-model outputs is calculated as the fusion result, and 84.1 points are obtained. If the fusion result ≥ 80 points, it is determined that the heat treatment quality of this batch of fasteners is qualified.

[0121] S5. Result optimization: The fusion result is processed through the result optimization terminal, and finally the heat treatment quality evaluation result of the target object is output. The heat treatment quality evaluation result is judged. If it does not meet the preset result of 80 points, the process parameters are adjusted.

[0122] The data processing process is referred to Figure 3 as shown, specifically:

[0123] S51. Calculate the mean value for multiple consecutive fusion results;

[0124] S52. Conduct anomaly detection on the mean value to identify outliers;

[0125] S53. Replace outliers with the median of multiple adjacent results.

[0126] The process parameters are adjusted by deriving the adjustment amount based on the deviation between the fusion result and the target quality threshold. When the predicted hardness value is lower than the target hardness value, the heat preservation time for optimizing the hardness index is adjusted. The formula for calculating the adjustment amount is as follows:

[0127] ,

[0128] in, To increase the required insulation time, The target hardness value, The hardness value is predicted by the fusion model. Material coefficient, This is the temperature correction factor.

[0129] Furthermore, the specific implementation is as follows: For the instantaneous fluctuations in the fusion results, the temporal smoothing process calculates the moving average value of the continuous fusion results, such as 84.1 points, 83.9 points, 85.2 points, 84.5 points, 83.8 points, and 84.7 points, to obtain 84.4 points, thereby eliminating the result jumps caused by instantaneous sensor noise.

[0130] The anomaly detection is based on the smoothed fusion result, combined with historical data to set a deviation threshold, such as ±5 points. When the fusion result of a certain batch is 78 points, which is lower than the threshold of 80 points, it is judged as an anomaly. During correction, the median of the results of five adjacent normal batches is used for replacement. For example, if the results of adjacent batches are 84.2 points, 84.5 points, 84.0 points, 84.7 points, and 83.9 points, the median is 84.2 points, and the anomaly value of 78 points is replaced to ensure the continuity of the data trajectory.

[0131] When the fusion model predicts hardness values Below the target value When optimizing the heat preservation time for hardness index, the adjustment amount is calculated using the following formula:

[0132] ,

[0133] in, This is the material coefficient (1.2 for carbon steel). The temperature correction factor (1.1 when the actual temperature is 5°C lower than the standard temperature) is used to calculate the required increase in insulation time. It is recommended to increase the heat preservation time by about 2 minutes.

[0134] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0135] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-source data fusion processing method for fastener heat treatment, characterized in that, Comprise the following steps: S1, multi-source data acquisition: the fastener to be processed is recorded as a target object, and multi-source data in the heat treatment process of the target object is collected in real time through a multi-source data acquisition terminal, wherein the multi-source data comprises process data, material data, quality detection data and environmental data; S2, data preprocessing: the multi-source data is cleaned and standardized by a data preprocessing terminal, and process data preprocessing parameters, material data preprocessing parameters, quality detection data preprocessing parameters and environmental data preprocessing parameters are obtained respectively; S3, feature extraction: the process data preprocessing parameters, material data preprocessing parameters, quality detection data preprocessing parameters and environmental data preprocessing parameters are input into a feature extraction terminal, and time sequence feature parameters, composition feature parameters, quality feature parameters and interference feature parameters are output; S4, fusion modeling: the composition feature parameters, quality feature parameters and interference feature parameters are input into a fusion model through a fusion modeling terminal, and feature layer fusion is performed by combining mechanism prior weight and data-driven attention mechanism, and then decision layer fusion is performed through a multi-quality index sub-model, and a fusion result is output; The multi-quality index sub-model comprises a hardness uniformity sub-model, a strength stability sub-model, a structure eligibility sub-model and a size accuracy sub-model; The hardness uniformity sub-model takes hardness gradient features as core input, and outputs hardness distribution uniformity scores, and the formula is: ， wherein is a hardness gradient value, is a hardness influence coefficient, is a natural constant; The strength stability sub-model is based on strength dispersion degree features, and outputs strength consistency coefficients, and the formula is: ， wherein is the standard deviation of the intensity, is the maximum allowed standard deviation; The structure eligibility sub-model combines grain diameter and martensite phase proportion features, and outputs microscopic structure eligibility rates, and the formula is: ， wherein d is the actual grain diameter, are the upper and lower grain diameter limits, respectively, is the martensite phase fraction; The size accuracy sub-model is based on deformation amount peak value features, and outputs size deviation levels, and the formula is: ， wherein is a peak value of the deformation, is a threshold value of the deformation; each sub-model output value is weighted by a preset weight weighted summation, wherein are preset weights and the sum is 1; S5, result optimization: the fusion result is processed by a result optimization terminal, and finally a heat treatment quality evaluation result of the target object is output, and the heat treatment quality evaluation result is judged, and if it does not meet the preset result, process parameter adjustment is performed.

2. The multi-source data fusion processing method for fastener-oriented heat treatment according to claim 1, characterized in that: The process data comprises heating furnace temperature, cooling medium flow, furnace pressure, equipment running power and running time; the material data comprises material composition content and original hardness of the target object; the quality detection data comprises hardness, tensile strength, metallographic structure image and deformation amount after heat treatment; and the environmental data comprises workshop temperature, humidity and power grid voltage.

3. The multi-source data fusion processing method for fastener-oriented heat treatment according to claim 1, characterized in that: The time sequence feature parameters comprise temperature change rate, pressure fluctuation amplitude, flow mean value and power standard deviation; The composition feature parameters comprise element content proportion and composition uniformity; the quality feature parameters comprise hardness gradient, strength dispersion degree and deformation amount peak value; The interference feature parameters comprise temperature fluctuation range and voltage stability coefficient.

4. The multi-source data fusion processing method for fastener-oriented heat treatment according to claim 3, characterized in that: The temperature change rate is calculated by temperature values of 5 consecutive sampling points, and the formula is: ， wherein, Tcurrentis the temperature value of the current sampling point, Taverageis the average temperature value of the previous 4 sampling points, Tintervalis the sampling time interval, Tderivativeis the temperature change rate; the homogeneity of the composition is calculated by the ratio of the standard deviation to the average of the content of each element.

5. The multi-source data fusion processing method for fastener-oriented heat treatment according to claim 1, characterized in that: The mechanism prior weight is set based on the relationship between austenitizing temperature and holding time in fastener heat treatment, the data-driven attention mechanism learns feature importance coefficients, and the fusion feature vector is the sum of the product of each type of feature parameter and the corresponding weight.

6. The multi-source data fusion processing method for fastener-oriented heat treatment according to claim 1, characterized in that: The data processing process is specifically as follows: S51, the mean value is calculated for a plurality of fusion results in succession; S52, performing anomaly detection on the mean to identify outliers; S53, replacing the outliers with the median of the adjacent multiple results.

7. The multi-source data fusion processing method for fastener-oriented heat treatment according to claim 6, characterized in that: The anomaly detection is constructed based on the time sequence distribution characteristics of the fusion results, the mean and the standard deviation of the continuous N fusion results are calculated to determine the normal fluctuation interval, and when a fusion result exceeds the range of 1.5 times the standard deviation of the interval, it is determined to be an outlier. Wherein N is a preset sliding window size and the normal fluctuation interval is dynamically adjusted with the update of the data in the window.

8. The multi-source data fusion processing method for fastener-oriented heat treatment according to claim 1, characterized in that: The process parameter adjustment mode is to derive the adjustment amount based on the deviation of the fusion result and the target quality threshold value, and when the predicted hardness value is lower than the target hardness value, the adjustment amount is calculated as follows: ， wherein, is the increased holding time, is the target hardness value, is the hardness value predicted by the fusion model, is the material coefficient, is the temperature correction coefficient.

Images