Online detection system for welding quality

Abstract

The invention relates to the technical field of welding quality detection, and discloses an online detection system for welding quality. A welding area monitoring device of the system is arranged around a working area of welding equipment and used for capturing an optical radiation signal and a heat distribution signal in the welding process; the multi-source data acquisition device is connected with the welding area monitoring device and synchronously acquires welding current fluctuation data, arc voltage change data and wire feeding speed data; the real-time feature extraction device receives the multi-source data and constructs a welding process dynamic feature set; the dynamic analysis device generates welding quality evaluation parameters according to the dynamic feature set; and the output control device generates a welding process parameter adjusting instruction based on the welding quality evaluation parameter. The system realizes real-time detection and dynamic control of welding quality, improves welding production efficiency and quality stability, and is suitable for various industrial welding scenes.

Description

Technical Field

[0001] This invention relates to the field of welding quality inspection technology, specifically to an online inspection system for welding quality. Background Technology

[0002] In modern industrial production, welding, as a key material joining process, is widely used in various fields such as machinery manufacturing, petrochemicals, aerospace, and construction engineering. Its quality directly affects the structural strength, safety performance, and service life of products. With the increasing demands for precision and reliability in industrial production, the requirements for welding quality inspection are becoming increasingly stringent. Traditional welding quality inspection methods are mostly offline, meaning that after welding is completed, non-destructive testing techniques such as radiographic testing, ultrasonic testing, and magnetic particle testing are used to assess the quality of the weld joint. This method has significant limitations. First, the inspection process must be carried out after welding is completed. If welding defects are found, the completed welded parts must be reworked or scrapped, which not only increases production costs but also extends the production cycle. This is especially true for large structural components or continuous production scenarios, where the economic losses and efficiency impact of offline inspection are more significant. Second, offline inspection cannot reflect the dynamic changes during the welding process in real time. It is difficult to capture instantaneous issues such as current fluctuations, voltage anomalies, and unstable wire feed speeds that occur during welding. These instantaneous issues are often key factors leading to welding defects, and relying solely on post-production inspection is insufficient to fundamentally control welding quality.

[0003] With the development of automated welding technology, some production scenarios have begun to explore simple online monitoring methods, such as monitoring only a single parameter like welding current or voltage. However, these monitoring methods suffer from a lack of comprehensive monitoring dimensions, failing to fully reflect the complex state of the welding process. Welding is a complex process involving the coupling of multiple physical fields, including optics, thermodynamics, and electricity. Relying solely on single-parameter monitoring makes it difficult to accurately determine key information such as the state of the weld pool and the temperature distribution of the heat-affected zone, thus hindering the timely identification of defects such as porosity, cracks, and lack of fusion that may occur during welding. Furthermore, some existing online monitoring systems lack effective data integration and analysis capabilities. The multi-source data collected is often scattered and cannot form effective feature correlations, resulting in a lack of systematic and accurate assessment of welding quality, making it difficult to meet the real-time control requirements of high-precision industrial production. Summary of the Invention

[0004] The purpose of this invention is to provide an online inspection system for welding quality, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides an online inspection system for welding quality, the system comprising: a welding area monitoring device, a multi-source data acquisition device, a real-time feature extraction device, a dynamic analysis device, and an output control device.

[0006] A welding area monitoring device is configured around the working area of ​​the welding equipment to capture optical radiation signals and thermal distribution signals during the welding process. A multi-source data acquisition device is connected to the welding area monitoring device to synchronously acquire welding current fluctuation data, arc voltage change data, and wire feed speed data. A real-time feature extraction device receives the current fluctuation data, voltage change data, and wire feed speed data output by the multi-source data acquisition device, and receives the optical radiation signals and thermal distribution signals output by the welding area monitoring device to construct a dynamic feature set of the welding process. A dynamic analysis device generates welding quality assessment parameters based on the dynamic feature set provided by the real-time feature extraction device. An output control device generates welding process parameter adjustment instructions based on the welding quality assessment parameters provided by the dynamic analysis device.

[0007] Preferably, the welding area monitoring device includes a high-resolution spectral acquisition unit and an infrared thermal imaging unit; the high-resolution spectral acquisition unit is installed at a specific distance behind the welding torch to capture spectral radiation intensity data of the weld pool area at a fixed sampling frequency; the infrared thermal imaging unit is set directly above the welding area to acquire temperature field distribution data of the weld heat-affected zone in an asynchronous sampling manner; the multi-source data acquisition device includes a current Hall sensor, a voltage differential probe, and an encoder speed measurement module; the current Hall sensor is clamped at the welding power supply output end, the voltage differential probe is connected in parallel at both ends of the welding torch, and the encoder speed measurement module is installed on the drive shaft of the wire feeding mechanism.

[0008] Preferably, the real-time feature extraction device includes a signal preprocessing unit and a feature fusion unit; the signal preprocessing unit performs sliding window normalization processing on the spectral radiation intensity data output by the high-resolution spectral acquisition unit and performs spatial interpolation compensation processing on the temperature field distribution data output by the infrared thermal imaging unit; the feature fusion unit aligns the processed spectral radiation intensity data and temperature field distribution data according to the time series, and simultaneously receives current fluctuation data collected by the current Hall sensor, voltage change data collected by the voltage differential probe, and wire feeding speed data collected by the encoder speed measurement module, and generates a multi-dimensional feature vector containing time-series correlation.

[0009] Preferably, the real-time feature extraction device further includes a feature database update unit; the feature database update unit receives the multi-dimensional feature vector generated by the feature fusion unit, extracts the feature value change trend of several consecutive sampling periods, and constructs a real-time feature library for the welding process; the real-time feature library is classified and stored according to the welding material type and thickness specification, and each storage entry contains a feature vector timestamp, corresponding welding process parameters, and environmental humidity data.

[0010] Preferably, the dynamic analysis device includes a defect probability calculation unit and a quality level determination unit; the defect probability calculation unit extracts the feature vector of the current welding cycle from the real-time feature library and calculates the deviation value between it and the feature vector of historical qualified samples; the quality level determination unit presets a first deviation threshold and a second deviation threshold, outputs a qualified determination signal when the deviation value is less than the first deviation threshold, outputs a suspicious determination signal when the deviation value is between the first deviation threshold and the second deviation threshold, and outputs a defect determination signal when the deviation value is greater than the second deviation threshold.

[0011] Preferably, the dynamic analysis device further includes a defect location unit; when the defect location unit receives the defect determination signal, it simultaneously retrieves the temperature field distribution data collected by the infrared thermal imaging unit and the spectral radiation intensity data collected by the high-resolution spectral acquisition unit, and determines the specific location coordinates of the welding defect by comparing the spatiotemporal overlap between the abnormal temperature gradient distribution area and the abrupt change in spectral characteristics.

[0012] Preferably, the output control device includes a parameter adjustment unit and an alarm triggering unit; the parameter adjustment unit receives the suspicious judgment signal output by the quality grade judgment unit, and generates the welding current correction amount, arc voltage compensation amount and wire feed speed adjustment amount according to the optimal process parameter data of the same material specification in the real-time feature library; the alarm triggering unit receives the defect judgment signal output by the quality grade judgment unit, and generates an emergency stop command containing defect type code and location information in combination with the specific location coordinates provided by the defect location unit.

[0013] Preferably, the system also includes a historical data comparison device; the historical data comparison device extracts the feature vector sequence of the most recent welding cycles from the real-time feature library, calculates the dynamic time warping distance between it and the feature vector sequence under standard process parameters; when the distance exceeds the preset process stability threshold, it sends a process parameter optimization request to the parameter adjustment unit.

[0014] Preferably, the output control device further includes a report generation unit; the report generation unit receives the judgment signal output by the quality grade judgment unit, the position coordinate data provided by the defect location unit, and the dynamic time regularization distance data calculated by the historical data comparison device, and generates a test report that includes welding quality statistical indicators, defect distribution map and process stability index.

[0015] Preferably, the system also includes a calibration device; the calibration device is connected to the welding area monitoring device and the multi-source data acquisition device, periodically triggers the standard test block welding process, collects feature vector data under standard welding conditions, updates the standard sample data in the real-time feature library, and sends a calibration completion mark to the report generation unit.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This online welding quality inspection system, by setting up welding area monitoring devices, multi-source data acquisition devices, real-time feature extraction devices, dynamic analysis devices, and output control devices, constructs a complete real-time welding quality inspection and control system, effectively solving the limitations of traditional welding quality inspection methods. From a data acquisition perspective, the welding area monitoring device is positioned around the working area of ​​the welding equipment, enabling it to specifically capture optical radiation signals and thermal distribution signals during the welding process. Simultaneously, a multi-source data acquisition device is connected to the welding area monitoring device, synchronously acquiring welding current fluctuation data, arc voltage change data, and wire feed speed data. This achieves comprehensive acquisition of optical, thermal, and electrical data across multiple dimensions during the welding process. This multi-source synchronous data acquisition method breaks through the limitations of traditional single-parameter monitoring, fully reflecting the dynamic changes during the welding process. It provides rich and comprehensive basic data for subsequent welding quality assessment, avoiding misjudgments or omissions due to insufficient data dimensions. In terms of data processing and feature construction, the real-time feature extraction device can simultaneously receive current fluctuation data, voltage change data, wire feed speed data from multi-source data acquisition devices, and optical radiation signals and heat distribution signals from welding area monitoring devices. Based on this data, it constructs a dynamic feature set of the welding process. This process integrates and correlates multi-source heterogeneous data, transforming scattered raw data into a set that reflects the essential characteristics of the welding process. This makes the data information more targeted and effective, accurately reflecting key information such as the state of the molten pool and changes in the heat-affected zone during the welding process, providing a reliable feature basis for subsequent quality assessment by dynamic analysis devices. The dynamic analysis device generates welding quality assessment parameters based on the dynamic feature set provided by the real-time feature extraction device. These parameters reflect the current quality status of the welding process in real time. Compared to the lagging quality assessment in traditional offline detection, this device enables real-time judgment of welding quality, allowing for timely detection of anomalies during the welding process and preventing further expansion of defects. The output control device generates welding process parameter adjustment commands based on the welding quality assessment parameters provided by the dynamic analysis device. This allows for timely adjustments to the welding process parameters when welding quality anomalies are detected, bringing the welding process back to normal, reducing rework or scrap due to quality issues, lowering production costs, and shortening the production cycle.

[0017] The entire system establishes a tight collaborative working mechanism among its various components, forming a closed-loop real-time control process from data acquisition, feature extraction, quality analysis to parameter adjustment. This ensures that welding quality is continuously monitored and dynamically adjusted throughout the welding process. This process design is not only applicable to conventional welding production scenarios but also addresses the quality inspection needs of different welding equipment and processes. It possesses strong adaptability and versatility, providing stable and reliable quality assurance for various industrial welding production processes. This drives the transformation of welding production from traditional post-production inspection to real-time control, improving the overall efficiency and quality stability of welding production. Attached Figure Description

[0018] Figure 1 This is a timing diagram of the online welding quality inspection system described in this invention. Figure 2 A detailed flowchart for a welding area monitoring and multi-source data acquisition device; Figure 3 A flowchart for updating the feature database of a real-time feature extraction device. Detailed Implementation

[0019] 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.

[0020] Please see Figure 1This invention provides an online inspection system for welding quality. The system integrates multiple sensors and analysis modules to achieve real-time monitoring and quality assessment of the welding process. The core components of the system include a welding area monitoring device, a multi-source data acquisition device, a real-time feature extraction device, a dynamic analysis device, and an output control device. The welding area monitoring device is positioned around the working area of ​​the welding equipment and primarily captures optical radiation and heat distribution signals generated during the welding process. These signals reflect the dynamic changes in the weld pool and heat-affected zone. The multi-source data acquisition device is connected to the welding area monitoring device and synchronously acquires welding current fluctuation data, arc voltage change data, and wire feed speed data to ensure the temporal consistency of the multi-source information. The real-time feature extraction device receives current, voltage, and wire feed speed data from the multi-source data acquisition device, as well as optical radiation and heat distribution signals from the welding area monitoring device. It constructs a dynamic feature set of the welding process using data fusion technology, which includes temporal correlation information of various physical parameters. Based on the dynamic feature set provided by the real-time feature extraction device, the dynamic analysis device performs pattern recognition and deviation calculation to generate welding quality assessment parameters, such as defect probability and quality level. The output control device generates corresponding welding process parameter adjustment instructions or control signals based on the quality assessment parameters output by the dynamic analysis device, thereby achieving closed-loop optimization of the welding process.

[0021] Example 1: See Figure 2The welding area monitoring device includes a high-resolution spectral acquisition unit and an infrared thermal imaging unit. The high-resolution spectral acquisition unit is installed at a specific distance behind the welding torch, typically set within the range of 100 mm to 200 mm to optimize signal acquisition. The unit employs a grating beam splitting principle in conjunction with a CCD sensor array, with a fixed sampling frequency of 1000 Hz, covering the spectral band from 300 nm to 1100 nm. It can continuously monitor changes in the spectral radiation intensity of the weld pool area, and the data is transmitted through shielded optical fiber to reduce interference. The infrared thermal imaging unit is positioned approximately 500 mm directly above the welding area, using a microbolometer as the core sensor. It operates in an asynchronous sampling mode with a sampling frequency of 30 Hz, acquiring temperature field distribution data of the weld heat-affected zone. The temperature resolution is designed to be 0.1 degrees Celsius. The unit incorporates a non-uniformity correction algorithm to compensate for sensor drift in real time, ensuring spatial consistency of the temperature field data. The multi-source data acquisition device connects these monitoring units and includes a current Hall sensor, a voltage differential probe, and an encoder speed measurement module. The current Hall sensor adopts a closed-loop design, clamped on the cable at the output end of the welding power supply, with a measurement range covering 0 to 500 amperes and an accuracy of ±1%. The output analog signal is transmitted after AD conversion. The voltage differential probe adopts high-voltage isolation technology and is connected in parallel to the wires at both ends of the welding gun, with a measurement range of 0 to 50 volts and a bandwidth of 1 MHz, capable of capturing microsecond-level voltage fluctuations. The encoder speed measurement module is an incremental encoder, directly mounted on the surface of the drive shaft of the wire feeding mechanism, generating 1000 pulses per revolution, which are converted into wire feeding speed values ​​through a counter circuit with an accuracy of ±0.1 m / min.

[0022] The real-time feature extraction device receives the data stream from the aforementioned device. Its signal preprocessing unit performs sliding window normalization processing on the spectral radiance intensity data output by the high-resolution spectral acquisition unit. The window size can be configured to be 100 consecutive sampling periods. During processing, the mean and standard deviation of the data within the window are calculated, and z-score normalization is performed to eliminate baseline fluctuations caused by ambient light. At the same time, spatial interpolation compensation processing is performed on the temperature field distribution data output by the infrared thermal imaging unit. A bilinear interpolation algorithm is used to fill in the spatial missing points caused by asynchronous sampling. The interpolation is based on the weighted average of neighboring pixels to ensure that the temperature field is continuous and uninterrupted. The feature fusion unit integrates preprocessed multi-source data, aligns spectral radiation intensity data and temperature field distribution data according to time series, and uses linear interpolation technology to synchronize asynchronous data to a unified time grid of 1000 Hz. At the same time, it incorporates current fluctuation data collected by current Hall sensor, voltage change data collected by voltage differential probe, and wire feeding speed data collected by encoder speed measurement module to generate a multi-dimensional feature vector. This vector includes feature dimensions such as mean spectral intensity, temperature gradient amplitude, current variance, voltage peak value, and wire feeding acceleration. Each dimension represents the physical parameter state at a sampling time. The vector structure is designed as a dynamic array, which supports real-time expansion to adapt to the needs of different welding processes.

[0023] The implementation of the high-resolution spectral acquisition unit emphasizes environmental adaptability. The unit's casing employs a water-cooling design to prevent high-temperature damage, and filters are added to the optical path to suppress stray light. The installation position is fine-tuned using a mechanical bracket to ensure the optical axis is aligned with the center of the molten pool. During data acquisition, the CCD sensor's exposure time is automatically adjusted to avoid signal saturation. The raw spectral data undergoes dark current correction before being transmitted to the preprocessing unit. The infrared thermal imaging unit's deployment considers field-of-view coverage, with a 30° × 20° angle of view covering the entire heat-affected zone. Sampling asynchronicity is controlled by an internal clock, and the timestamp is synchronized with the system clock. Emissivity compensation is performed before outputting temperature field data, dynamically adjusted based on a material emissivity database. The installation of the current Hall sensor requires the cable to be fully embedded in the magnetic ring to reduce external magnetic field interference. A low-pass filter should be installed at the signal output end, with the cutoff frequency set to 10 kHz to smooth high-frequency noise. The voltage differential probe should be connected using a high-voltage probe cable, with the parallel connection point selected at the input terminal of the welding gun. The probe grounding terminal should be independently isolated to prevent ground loop interference. The encoder speed measurement module should be installed to ensure shaft alignment. The encoder signal should be quadrupled to improve resolution, and the speed value should be updated every 10 milliseconds.

[0024] The algorithm of the signal preprocessing unit runs on an embedded digital signal processor. The sliding window normalization process adopts a real-time pipeline architecture with a window sliding step size of 1 sampling period. The dynamic range of the normalized data is compressed to the [-1,1] interval, which is convenient for subsequent fusion. The spatial interpolation compensation process is performed on each frame of the infrared thermal imaging data. The interpolation kernel size is 3×3 pixels, and the weight is calculated based on the reciprocal of the distance. The resolution of the temperature field grid after processing is increased to 1.5 times that of the original data. The feature fusion unit is implemented based on a multi-core processor. The time series alignment module first allocates a buffer for each data source, with a buffer depth of 1000 samples. During alignment, linear interpolation is used to interpolate low-frequency data (such as 30 Hz temperature data) to a 1000 Hz time point, with the interpolation error controlled within 0.1%. The feature vector generation module extracts the features of the aligned data. For example, the mean spectral intensity is obtained by averaging all wavelengths within the calculation window, the temperature gradient magnitude is obtained by convolving the temperature field image with the Sobel operator, the current variance is calculated based on 100 sampling points, the voltage peak detection uses the local maximum algorithm, and the wire feeding acceleration is obtained by differentiating continuous velocity values. The generated feature vector is appended with a timestamp and sequence number and stored in a circular memory buffer.

[0025] The data flow management of the welding area monitoring device adopts a master-slave structure. The high-resolution spectral acquisition unit and the infrared thermal imaging unit are connected to the real-time feature extraction device via a gigabit Ethernet interface. The data transmission protocol is UDP with timestamp verification. The sensor signals from the multi-source data acquisition device are connected to the analog input module via shielded cables, with a uniform sampling rate of 100 kHz, which is then downsampled to 1000 Hz after digital filtering. The signal preprocessing unit's resource configuration includes a dedicated memory pool for sliding window calculations, the window size of which can be adjusted via software parameters to adapt to different welding speeds. Spatial interpolation compensation processing is accelerated using a GPU, with interpolation parameters pre-stored in a lookup table and dynamically loaded according to the temperature field resolution. The multi-dimensional feature vector output format of the feature fusion unit is defined as a structure array, with each vector containing 20 feature values. The vector sequence is directly transmitted to the dynamic analysis device via a DMA channel at a 1-millisecond interval to ensure real-time performance.

[0026] The calibration of the welding area monitoring device is completed through periodic calibration. The high-resolution spectral acquisition unit uses a standard light source to calibrate wavelength and intensity, and the infrared thermal imaging unit uses a blackbody source to calibrate temperature accuracy. The sensors of the multi-source data acquisition device undergo zero-point calibration every shift. The current Hall sensor is verified using a standard current source, the voltage differential probe is adjusted using a standard voltage source, and the encoder speed measurement module is calibrated using a constant speed wheel. The software module of the real-time feature extraction device adopts a modular design. The signal preprocessing unit and the feature fusion unit run as independent threads with the highest thread priority to avoid data loss. The feature vector database uses high-speed solid-state drives for storage, and the index is built according to timestamps and process parameters to support fast queries. Key sensors such as the current Hall sensor and the voltage differential probe are configured with dual-path backup. The data acquisition device has a built-in self-diagnostic function, automatically switching to the backup channel in case of anomalies. The feature fusion algorithm of the real-time feature extraction device introduces a consistency check. If the data source timestamp deviation exceeds the tolerance, a re-alignment process is triggered. The influence of the welding environment is suppressed through shielding and filtering. For example, a light shield is added to the spectral acquisition unit to reduce ambient light interference, and an air blowing device is used to keep the lens of the infrared thermal imaging unit clean. The dimension selection of the feature vector is based on process knowledge. The initial setting includes 10 core features, and additional features can be dynamically added according to the type of welding material. For example, the intensity of the chromium element feature line is increased when welding stainless steel.

[0027] Example 2: See Figure 3 The feature database update unit receives multi-dimensional feature vectors generated by the feature fusion unit. These vectors contain dynamic parameters such as the mean spectral intensity, temperature gradient, and current variance. The update unit extracts the trend of feature value changes from continuous sampling periods, such as calculating the moving average and standard deviation of the feature vectors within every five sampling periods to form a trend sequence. The real-time feature library for the welding process establishes a classification index according to material type and thickness specification. Material types include common welding materials such as low-carbon steel, stainless steel, and aluminum alloy. Thickness specifications are recorded in segments from 1 mm for thin plates to 20 mm for thick plates. Each database entry stores a feature vector array, a high-precision timestamp, the corresponding welding process parameter set (including current setting value, voltage setting value, and wire feed speed setting value), and environmental humidity data. Humidity data is provided in real time by a digital humidity sensor installed in the welding chamber. The data update mechanism is to perform a write operation once after each welding cycle (approximately 100 milliseconds) is completed, while automatically archiving historical data that has exceeded the retention period.

[0028] The defect probability calculation unit of the dynamic analysis device retrieves the feature vector of the current welding cycle from the real-time feature library. The retrieval conditions match the material type and thickness specifications, and calculate the deviation value between the feature vector and the feature vector of the historical qualified samples. The historical qualified sample library is established through the initial training phase. During training, multiple defect-free test blocks are welded under standard process parameters, their feature vectors are extracted and marked as qualified samples. The deviation calculation adopts the Mahalanobis distance algorithm. This algorithm considers the covariance structure of the feature vector and effectively eliminates the influence of dimensions and correlation interference between features. During the calculation, a reference sample set of the same specification is first extracted from the qualified sample library, and the Mahalanobis distance value between the current feature vector and the center of the sample set is calculated. The larger the value, the farther it deviates from the normal welding state. The quality grade judgment unit presets two deviation thresholds. The first deviation threshold is set based on the statistical process control principle, taking the position of twice the standard deviation of the mean of the qualified sample distribution. The second deviation threshold is taken at the position of three times the standard deviation. The judgment logic is as follows: when the real-time calculated deviation value is less than the first threshold, a high-level signal is output as a qualified judgment signal; when the deviation value is between the first threshold and the second threshold, a medium-level signal is output as a suspicious judgment signal; when the deviation value exceeds the second threshold, a low-level signal is output as a defect judgment signal. All judgment signals are appended with a timestamp and welding cycle number.

[0029] The feature database update unit is deployed on an industrial server. The database adopts a time-series database structure. Each record contains a feature vector (a 20-dimensional floating-point array), a millisecond-level timestamp, a material type code (e.g., CS304 represents 304 stainless steel), a thickness value (in millimeters), an array of process parameters (floating-point values ​​for current, voltage, and wire feed speed), and an environmental humidity value. The database index is built according to the combination of material type and thickness, supporting millisecond-level query response. Update operations ensure data consistency through a transaction mechanism, while background tasks periodically clean up archived data older than three months. The historical qualified sample library is initialized through an offline training process. During training, 50 test blocks are continuously welded using standard welding parameters. 1000 feature vectors are extracted during the welding process of each test block. After non-destructive testing confirms that there are no defects, the samples are stored in the sample library. The sample library is continuously optimized, and new samples are added after each batch of new qualified weldments is completed. The defect probability calculation unit is integrated on a high-performance computing card. It triggers calculation immediately upon receiving a new feature vector, retrieving all samples with the same material-thickness combination from the qualified sample library. It calculates the sample mean and covariance matrix, incorporating regularization into the Mahalanobis distance formula to prevent matrix singularities. The deviation value is output as a floating-point number, theoretically ranging from 0 to positive infinity, but typically between 0 and 5 under normal operating conditions. The threshold management of the quality grade judgment unit is dynamically configurable. The first deviation threshold is set to 2.0 (normalized units) by default, and the second threshold is set to 3.0. Thresholds can be fine-tuned via a human-machine interface to adapt to different production line requirements. Judgment signals are transmitted through a digital output module: a qualified signal corresponds to a 24V high level, a suspicious signal to a 12V medium level, and a defect signal to a 0V low level. The signals are then transmitted to the output control device.

[0030] The real-time feature library's data structure supports fast time-series queries. Each feature vector entry is associated with metadata including welding torch number, operator ID, ambient temperature, and other auxiliary information. The database employs a partitioned storage strategy, physically separating data for different material specifications to improve query efficiency. Maintenance of the qualified sample library includes sample quality monitoring, regular checks of sample age distribution, and automatic removal of old samples exceeding a certain timeframe to prevent misjudgments due to process drift. During deviation calculation, feature vectors are pre-standardized using the mean and standard deviation of the qualified sample set to ensure consistency across different feature dimensions. The dynamic analysis device's operating cycle is strictly synchronized with the welding cycle, executing a complete analysis process every 100 milliseconds. The defect probability calculation unit is equipped with a caching mechanism, using the previous valid result for a downgraded run when a qualified sample library query fails. The quality level judgment unit's output signal undergoes anti-jitter processing, confirming a defect signal only after three consecutive cycles of defect identification to avoid false alarms due to momentary interference. A self-check process is executed upon startup to verify the feature database connection status and whether the number of qualified sample library samples meets the minimum requirements (e.g., 1000 samples per specification). If the self-check fails, the analysis process is prohibited from starting. The selection of samples for the qualified sample library incorporates diversity constraints to ensure coverage of different equipment states and environmental conditions. The deviation calculation algorithm has been optimized, employing an incremental calculation method to reduce repetitive matrix operations. Adjustments to the judgment threshold are recorded in the audit log, and any modifications require authorized operator two-factor authentication.

[0031] Example 3: The defect location unit starts immediately upon receiving a defect determination signal from the quality grade determination unit. This signal is a low-level digital pulse. The defect location unit synchronously retrieves the temperature field distribution data cache collected by the infrared thermal imaging unit in the most recent period, and at the same time reads the spectral radiation intensity data of the corresponding period from the high-resolution spectral acquisition unit. The temperature field distribution data is stored in the form of an image sequence. Each frame of the image contains a temperature value matrix of 512×640 pixels, and the spectral data is the light intensity curve of each pixel in the 300-1100 nanometer band. The defect localization unit locates defects by comparing the spatiotemporal overlap between regions with abnormal temperature gradient distributions and regions with abrupt changes in spectral features. Identification of abnormal temperature gradient distribution regions begins by calculating the temperature gradient field for each frame of the thermal image. A two-dimensional convolution operation using the Sobel operator is performed to obtain the gradient magnitude of each pixel. A gradient magnitude threshold is set, and consecutive pixel regions exceeding the threshold are marked as abnormal areas. Detection of abrupt changes in spectral features focuses on the intensity changes of emission lines of specific metal elements (such as the characteristic peak of iron at 538 nm). The first-order difference of the spectral intensity of each pixel in consecutive frames is calculated, and abrupt changes are marked when the difference exceeds three times the standard deviation of the normal fluctuation range. Spatiotemporal overlap analysis maps these two types of abnormal regions to the same coordinate system. The positions of the temperature field image and the spectral pixels are registered using a pre-calibrated transformation matrix. The area ratio of the overlapping region is calculated. When the overlap exceeds a preset threshold (e.g., 80%), the location is determined to be a defect point, and its planar coordinates relative to the welding start point are output.

[0032] The defect localization unit is implemented based on a graphics processing unit (GPU) platform. Temperature gradient calculation utilizes parallel threads to process all pixels simultaneously, and the gradient magnitude threshold is dynamically adjusted according to the material type; for example, it is set to 15℃ / mm for low-carbon steel welding. Spectral abrupt change detection employs a sliding window difference method with a window size of 5 sampling points, and the abrupt change threshold is adaptively updated every 5 minutes. Coordinate registration uses a hand-eye calibration method to pre-establish the transformation relationship between thermal image pixel coordinates and spectral pixel coordinates, and the overlap calculation uses a pixel-level traversal algorithm. The parameter adjustment unit receives a suspicious judgment signal (medium-level 12V signal) output by the quality grade judgment unit. Upon triggering this signal, the unit immediately queries the real-time feature library for the optimal process parameters of the same material specifications. The query conditions include material code, thickness value, and ambient humidity range. It obtains the historical optimal parameter set (e.g., current 210A, voltage 26.5V, wire feed speed 6.2m / min) and compares it with the current actual parameters, generating welding current correction, arc voltage compensation, and wire feed speed adjustment. The correction is calculated using an incremental PID algorithm, with the proportional coefficient set according to the welding type. The integral time constant is set to 10 control cycles, and the derivative action is limited to low-speed change scenarios. The alarm trigger unit activates upon receiving the defect judgment signal. Simultaneously, it reads the defect location coordinates (X, Y millimeters) provided by the defect location unit and generates an emergency stop command conforming to the industrial bus protocol. The command frame includes the defect type code (e.g., C for crack, P for porosity), location coordinates, timestamp, and checksum, while simultaneously activating the alarm's audible and visual output.

[0033] In the temperature gradient calculation of the defect localization unit, the gradient magnitude G of each pixel is determined by the partial derivatives of that point in different directions: Where T represents the temperature value at a point in the temperature field, and x and y represent the horizontal and vertical coordinates in the image coordinate system, respectively. This calculation is performed in parallel on the GPU, with each thread responsible for the gradient calculation of one pixel. Spectral mutation detection calculates the change in spectral intensity ΔI between adjacent frames for each pixel, and a mutation is marked when |ΔI|>3σ (σ is the standard deviation of the pixel's intensity in the most recent 100 frames).

[0034] The synchronization mechanism between the defect location unit and data acquisition adopts hardware triggering. The defect judgment signal is directly connected to the trigger pin of the FPGA, ensuring that the location delay is less than 2 milliseconds. The temperature field data buffer depth is the most recent 5 seconds of data (150 frames), and the spectral data buffer is the most recent 5000 spectral curves. The output interface of the parameter adjustment unit is an analog output module. The current correction is converted into a 4-20mA signal and sent to the welding power supply, the voltage compensation is converted into a 0-10V signal and sent to the arc controller, and the wire feed speed adjustment is output to the servo driver through a pulse signal. The alarm trigger unit has multiple interlocking functions. The emergency stop command is sent to the welding power supply, the robot controller, and the wire feed mechanism simultaneously to ensure that the system immediately enters a safe state. The defect location unit adopts a dual buffer structure to ensure data integrity. The parameter adjustment algorithm incorporates anti-saturation processing to prevent integral windup. The alarm trigger unit is equipped with a watchdog timer to monitor the communication status. All calculation modules have redundancy design. In the event of a main processor failure, the backup processor immediately takes over. The coordinate positioning results are smoothed and filtered, and the average position confirmed by three consecutive frames is taken as the final output to avoid instantaneous jumps. The output of parameter adjustment has a gradual characteristic, with each adjustment not exceeding 5% of the set value, to prevent drastic fluctuations in the process.

[0035] Example 4: The historical data comparison device periodically extracts feature vector sequences from the real-time feature library for several recent welding cycles. Typically, 100 consecutive cycles correspond to approximately 10 seconds of welding process. The extracted sequences include multi-dimensional feature vectors for each cycle, such as mean spectral intensity, temperature gradient, and current variance. These vectors are arranged in chronological order to form time series data. The historical data comparison device calculates the dynamic time warping distance between this sequence and the feature vector sequence stored under standard process parameters. The standard sequence is derived from the reference sequence of the same material specification under optimal process conditions in the historical qualified sample library. The calculation of the dynamic time warping distance uses a dynamic programming algorithm to align the length differences between the two sequences. The total distance value is obtained by accumulating the Euclidean distance between corresponding points. This distance value reflects the degree of morphological difference between the current welding process and the standard process. When the calculated distance exceeds the preset process stability threshold, the historical data comparison device sends a process parameter optimization request to the parameter adjustment unit. The request includes suggestions on the direction and magnitude of the distance deviation, such as suggesting increasing the current or adjusting the voltage. The threshold setting is based on long-term operational data statistics and is usually taken as the 95th percentile value of the distance distribution of the standard sequence.

[0036] The report generation unit receives judgment signals (qualified, suspicious, or defective) from the quality grade judgment unit, defect location coordinate data provided by the defect location unit, and dynamic time-normalized distance data calculated by the historical data comparison device. These data flow into the data buffer of the report generation unit in real time. The report generation unit generates an inspection report at predetermined time intervals (e.g., hourly) or when triggered by an event (e.g., defect occurrence). The report content includes welding quality statistics such as pass rate, number of defects, and average deviation. The defect distribution map is generated by overlaying the defect location coordinates onto the welding path diagram. The process stability index is calculated from the reciprocal of the dynamic time-normalized distance and normalized to the range of 0-100. The report format supports XML and PDF. The XML format is used for data exchange between systems, and the PDF format is used for manual viewing. The report file is stored in a network shared directory and automatically sent to the production management system. The historical data comparison device is implemented based on edge computing nodes. These nodes are equipped with multi-core processors and high-speed memory. Every 10 seconds, they query the feature vector sequences of the most recent 100 welding cycles from a real-time feature database. The query filtering conditions include material type, thickness specification, and equipment number to ensure sequence comparability. Feature vector sequences under standard process parameters are pre-loaded from the historical database, with a fixed sequence length of 100 cycles. The calculation of dynamic time warping distance uses the open-source DTW library. Algorithm parameters such as step size constraints are set to the symmetric P0.1 specification. The calculated distance value is standardized and divided by the sequence length to eliminate the influence of the number of cycles. The process stability threshold is dynamically adjusted according to the material type. For example, it is set to 5.0 for low-carbon steel welding and 6.0 for stainless steel welding. Threshold management can be modified through the configuration interface. When the distance exceeds the standard, an optimization request is sent in the form of a digital signal, which includes a deviation code and a recommended adjustment amount.

[0037] The report generation unit is deployed on a server and adopts a modular design. The data receiving module listens to the judgment signals and coordinate data in the message queue. The dynamic time-normalized distance data is updated every 10 seconds. The report generation trigger can be set to time-driven or event-driven. In time-driven mode, it executes on the hour, while in event-driven mode, it starts immediately when the defect judgment signal is issued. The welding quality statistics in the report include the proportion of qualified cycles to the total number of cycles within the calculation period, the cumulative occurrence of various defect types, and the defect distribution map, which is drawn using vector graphics. The welding path is a straight line or curve generated based on the actual trajectory. Defect locations are marked with red dots. The process stability index is calculated using the formula max(0, 100 - 10 * DTW). distanceTo ensure the index remains within a reasonable range, a data integrity check is performed before report output, and missing data is filled with interpolation. For example, consider a stainless steel welding process with material type coded SS304, thickness 3 mm, and welding speed 1 m / min. A historical data comparison device extracts the feature vector sequence of the most recent 100 welding cycles at time point T1. The sequence data is shown in Table 1, which displays the simplified feature vector values ​​and calculated DTW distance for some cycles. The standard sequence comes from the optimal process sample for SS304-3mm in the historical database. The DTW distance is calculated by aligning the current sequence and the standard sequence to obtain the cumulative distance. The process stability threshold is set to 6.0.

[0038] Table 1: Calculation Table of Dynamic Time Warping Distance In this example, the DTW distance calculation result is 7.2, exceeding the threshold of 6.0. The historical data comparison device sends an optimization request to the parameter adjustment unit. The request signal contains the code "INC_CURRENT" suggesting that the current be increased from the current 210A to 215A. At the same time, the report generation unit triggers report generation at time T1. It collects 95 qualified signals, 3 suspicious signals, and 2 defect signals within this cycle. The defect location coordinates are (102.5, 35.2) and (103.8, 36.1), respectively. The dynamic time warp distance is 7.2, and the process stability index is calculated to be 28. The report generation unit integrates these data to generate a PDF report. The report header includes the time range, equipment ID, and operator information, and the main body lists statistical tables and defect distribution maps.

[0039] The data extraction process of the historical data comparison device uses SQL queries to query the real-time feature library. The query statement specifies the time range and label conditions. After the sequence data is loaded into memory, it undergoes preprocessing, including outlier removal and smoothing filtering. When calculating the DTW distance, global constraints are used to limit the slope of the alignment path to improve calculation efficiency. The threshold comparison module performs a check every 10 seconds. Optimization requests are sent to the parameter adjustment unit via the industrial Ethernet protocol. The request message includes a timestamp, distance value, threshold, and suggested action. The report generation unit uses a subscription model for data collection. Judgment signal and coordinate data are received asynchronously through a message middleware. Dynamic time-normalized distance data is pushed by the historical data comparison device. When generating the report, the statistical index calculation uses a sliding window algorithm to avoid memory overflow. The generation of the defect distribution map calls the graphics library API to map the coordinates onto the welding drawing. The process stability index is displayed in real time in the report summary. The output management of the report generation unit includes file naming rules and storage paths. PDF reports are named in the format "Welding_Report_YYYYMMDD_HHMMSS.pdf", stored in a specified directory and uploaded to the cloud platform. XML reports contain structured data that is easy for other systems to parse. The report content also includes trend charts such as the DTW distance over time curve to help users monitor process drift. The system log records the events and data source status of each report generation, which is convenient for auditing and troubleshooting.

[0040] Example 5: The calibration device is integrated as an independent hardware module in the system cabinet. It establishes a communication connection with the welding area monitoring device and the multi-source data acquisition device through the industrial bus. Its core function is to periodically trigger the welding process of the standard test block to complete the system self-calibration. The calibration cycle can be configured as a fixed time interval (e.g., every 8 hours) or based on the cumulative welding meters (e.g., every 500 meters of weld). The specifications of the standard test block are consistent with the materials processed on the current production line, including the same material type, thickness and surface treatment. The test block is installed on a special fixture at the calibration station and is automatically picked up, placed and positioned at the welding position by the robotic arm. When the calibration cycle is triggered, the calibration device first sends a command to the welding control system to set the welding parameters to preset standard values ​​(such as current 200A, voltage 25V, wire feed speed 5m / min). Then, the standard welding process is started. During this period, the high-resolution spectral acquisition unit and infrared thermal imaging unit of the welding area monitoring device work synchronously to capture the spectral radiation intensity data and temperature field distribution data of the welding area of ​​the standard test block at a normal sampling frequency. The multi-source data acquisition device simultaneously records the real-time data of current, voltage and wire feed speed. These data are transmitted to the real-time feature extraction device through the data interface of the calibration device.

[0041] The standard welding process lasts approximately 30 seconds, completing a calibration weld of about 100 mm in length. All data collected during the welding process is labeled as "standard sample data" and temporarily stored in a buffer. The data processing unit of the calibration device performs quality checks on these raw data, including signal integrity, noise level, and data synchronization. Qualified data packets are sent to the real-time feature library update module. The real-time feature library's standard sample data updates employ a version control mechanism. Newly acquired standard sample data, after feature extraction, is compared with existing standard samples in the library. If the statistical distribution of the feature vector changes significantly (e.g., the mean shift exceeds 5%), the new sample replaces the old one; otherwise, the original standard sample is retained, and the calibration data is recorded as a historical version. Upon calibration completion, the calibration device sends a digital signal as a calibration completion marker to the report generation unit. This marker data packet contains a calibration timestamp, standard test block number, standard values ​​of welding parameters, data quality score, and checksum. The report generation unit receives this and embeds it as metadata into the header information of the test report.

[0042] The calibration device employs a modular design. Its mechanical structure includes a test block storage compartment, a six-axis robotic arm, and a vision positioning system. The storage compartment can hold 20 standard test blocks of different sizes. The robotic arm's end effector is equipped with a force-controlled gripper to ensure stable gripping. The vision positioning system uses laser ranging and CCD imaging to precisely position the test blocks on the welding platform. The calibration trigger logic is implemented by a programmable logic controller (PLC), supporting various trigger condition combinations, such as "time arrival + equipment idle" or "meter accumulation + shift change." After the calibration command is issued, the welding equipment automatically switches to calibration mode, suspending normal production. The focal length and exposure parameters of the welding area monitoring device automatically adjust to the preset calibration mode settings. During data acquisition, the calibration device monitors the operating status of each sensor. If any data anomalies are detected (such as spectral signal saturation or temperature data drift), the calibration process is interrupted and a fault code is recorded. The management of standard test blocks employs a QR code identification system. Each test block has a unique code engraved on its surface, containing information such as material, thickness, heat treatment status, and expiration date. Before the robotic arm picks up a test block, it scans the QR code to verify whether the test block specifications match the current production line requirements. Once a test block reaches its usage limit (e.g., 50 times), it is automatically discarded, and a reminder is sent to replace it with a new one. The standard sample data update algorithm for the real-time feature library adopts a gradual update strategy. New sample data does not immediately and completely replace old data; instead, the statistical parameters of the feature vector are gradually adjusted according to a time-weighted principle to avoid sudden changes in standard samples due to single calibration anomalies. The transmission of calibration completion markers uses a redundant communication protocol: the primary channel is PROFIBUS-DP, and the backup channel is Ethernet TCP, ensuring reliable delivery of marker data to the report generation unit.

[0043] Data acquisition accuracy control measures during calibration include ambient temperature compensation (welding chamber temperature controlled at 23±2℃), sensor preheating (power-on stabilization 30 minutes in advance), and electromagnetic shielding (signal lines use double-layer shielding). The fluctuation range of process parameters during standard block welding is controlled within ±1%, ensuring the repeatability of standard sample data. Statistical significance testing during feature library updates uses hypothesis testing methods, calculating the confidence interval for the difference in the mean of the feature vectors of the new and old samples. A significant change is considered to occur when the 95% confidence interval does not contain zero. The report generation unit's parsing of calibration marks includes timestamp conversion, data quality score parsing, and checksum verification. Only calibration records that pass the integrity check are displayed in the report. The calibration device's self-monitoring functions include robotic arm motion trajectory calibration, periodic automatic adjustment of the vision system's focus, and sensor zero-point drift compensation. These maintenance operations are automatically performed during calibration intervals. The weld quality after standard block welding is verified through offline flaw detection sampling. The sampling results are fed back to the calibration system to evaluate the validity of the standard sample data. The timing control of the entire calibration process is accurate to the millisecond level, ensuring strict synchronization between data acquisition and the welding process. After calibration, the system automatically switches back to production mode and generates a calibration event log to record all process parameters. The calibration markers received by the report generation unit trigger a report version update, adding a "calibrated" watermark and calibration time identifier to the report homepage. Simultaneously, the calibration data is archived to an independent database for long-term trend analysis.

[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An online inspection system for welding quality, characterized in that: The system includes a welding area monitoring device, a multi-source data acquisition device, a real-time feature extraction device, a dynamic analysis device, and an output control device; The welding area monitoring device is configured around the working area of ​​the welding equipment to capture optical radiation signals and heat distribution signals during the welding process; the multi-source data acquisition device is connected to the welding area monitoring device to synchronously acquire welding current fluctuation data, arc voltage change data and wire feed speed data. The real-time feature extraction device receives current fluctuation data, voltage change data, and wire feeding speed data output by the multi-source data acquisition device, and receives optical radiation signals and heat distribution signals output by the welding area monitoring device to construct a dynamic feature set of the welding process. The dynamic analysis device generates welding quality assessment parameters based on the dynamic feature set provided by the real-time feature extraction device; the output control device generates welding process parameter adjustment instructions based on the welding quality assessment parameters provided by the dynamic analysis device.

2. The online inspection system for welding quality according to claim 1, characterized in that: The welding area monitoring device includes a high-resolution spectral acquisition unit and an infrared thermal imaging unit. The high-resolution spectral acquisition unit is installed at a specific distance behind the welding torch to capture spectral radiation intensity data of the weld pool area at a fixed sampling frequency. The infrared thermal imaging unit is set directly above the welding area to acquire temperature field distribution data of the weld heat-affected zone in an asynchronous sampling manner. The multi-source data acquisition device includes a current Hall sensor, a voltage differential probe, and an encoder speed measurement module. The current Hall sensor is clamped at the welding power supply output end, the voltage differential probe is connected in parallel at both ends of the welding torch, and the encoder speed measurement module is installed on the drive shaft of the wire feeding mechanism.

3. The online inspection system for welding quality according to claim 2, characterized in that: The real-time feature extraction device includes a signal preprocessing unit and a feature fusion unit; The signal preprocessing unit performs sliding window normalization on the spectral radiation intensity data output by the high-resolution spectral acquisition unit and performs spatial interpolation compensation on the temperature field distribution data output by the infrared thermal imaging unit. The feature fusion unit aligns the processed spectral radiation intensity data with the temperature field distribution data according to the time series. At the same time, it receives current fluctuation data collected by the current Hall sensor, voltage change data collected by the voltage differential probe, and wire feeding speed data collected by the encoder speed measurement module to generate a multi-dimensional feature vector containing time-series correlation.

4. The online inspection system for welding quality according to claim 3, characterized in that: The real-time feature extraction device also includes a feature database update unit; the feature database update unit receives the multi-dimensional feature vector generated by the feature fusion unit, extracts the feature value change trend of several consecutive sampling periods, and constructs a real-time feature library for the welding process; the real-time feature library is classified and stored according to the welding material type and thickness specification, and each storage entry contains the feature vector timestamp, the corresponding welding process parameters, and the environmental humidity data.

5. The online inspection system for welding quality according to claim 4, characterized in that: The dynamic analysis device includes a defect probability calculation unit and a quality level determination unit; The defect probability calculation unit extracts the feature vector of the current welding cycle from the real-time feature library and calculates the deviation value between it and the feature vector of the historical qualified samples. The quality level judgment unit presets a first deviation threshold and a second deviation threshold. When the deviation value is less than the first deviation threshold, it outputs a qualified judgment signal. When the deviation value is between the first deviation threshold and the second deviation threshold, it outputs a suspicious judgment signal. When the deviation value is greater than the second deviation threshold, it outputs a defect judgment signal.

6. The online inspection system for welding quality according to claim 5, characterized in that: The dynamic analysis device also includes a defect location unit. When the defect location unit receives the defect determination signal, it simultaneously retrieves the temperature field distribution data collected by the infrared thermal imaging unit and the spectral radiation intensity data collected by the high-resolution spectral acquisition unit. By comparing the spatiotemporal overlap between the abnormal temperature gradient distribution area and the abrupt change in spectral characteristics, the specific location coordinates of the welding defect are determined.

7. The online inspection system for welding quality according to claim 6, characterized in that: The output control device includes a parameter adjustment unit and an alarm triggering unit. The parameter adjustment unit receives the suspicious judgment signal output by the quality grade judgment unit and generates the welding current correction, arc voltage compensation and wire feed speed adjustment based on the optimal process parameter data of the same material specifications in the real-time feature library. The alarm triggering unit receives the defect judgment signal output by the quality grade judgment unit and generates an emergency stop command containing defect type code and location information by combining the specific location coordinates provided by the defect location unit.

8. The online inspection system for welding quality according to claim 7, characterized in that: The system also includes a historical data comparison device; the historical data comparison device extracts the feature vector sequence of the most recent welding cycles from the real-time feature library, calculates the dynamic time warping distance between it and the feature vector sequence under standard process parameters; when the distance exceeds the preset process stability threshold, it sends a process parameter optimization request to the parameter adjustment unit.

9. The online inspection system for welding quality according to claim 8, characterized in that: The output control device also includes a report generation unit; the report generation unit receives the judgment signal output by the quality grade judgment unit, the position coordinate data provided by the defect location unit, and the dynamic time regularization distance data calculated by the historical data comparison device, and generates a test report that includes welding quality statistical indicators, defect distribution map and process stability index.

10. The online inspection system for welding quality according to claim 9, characterized in that: The system also includes a calibration device; the calibration device is connected to the welding area monitoring device and the multi-source data acquisition device, periodically triggers the standard test block welding process, collects feature vector data under standard welding conditions, updates the standard sample data in the real-time feature library, and sends a calibration completion mark to the report generation unit.

Images