Welding process multi-sensor data fusion processing system

Abstract

The present application relates to the technical field of welding data processing, and discloses a multi-sensor data fusion processing system for a welding process. The system comprises a multi-modal welding data acquisition module, a welding process feature image generation module, a feature preprocessing module, an environment encoding module and a process parameter mapper. The multi-modal welding data acquisition module acquires welding current, voltage, molten pool image, weld appearance, environmental temperature and humidity and equipment vibration frequency; the feature image generation module generates a feature image containing welding quality and process parameter feature vectors accordingly; the feature preprocessing module performs time domain segmentation and frequency domain conversion on the multi-modal data to generate a standard welding process parameter data set; the environment encoding module encodes the environmental temperature and equipment vibration frequency to generate an environment state encoding sequence; and the process parameter mapper matches and maps the two to output process parameter mapping characteristic values. The system realizes multi-modal data integration and process parameter dynamic matching, and ensures the stability of the welding process.

Description

Technical Field

[0001] This invention relates to the field of welding data processing technology, specifically a multi-sensor data fusion processing system for welding processes. Background Technology

[0002] In modern manufacturing, welding, as a key joining technology, is widely used in various fields such as machinery manufacturing, shipbuilding, aerospace, and petrochemicals. Its process stability is directly related to the performance of the final welded product. With the increasing demands for welding quality in industrial production, traditional methods relying on manual experience to monitor the welding process are no longer sufficient to meet the demands for high precision and consistency. Traditional welding process monitoring often only collects and analyzes single or a few parameters, such as focusing solely on changes in welding current and voltage, neglecting the impact of multi-dimensional data on welding quality, including molten pool images, weld morphology, environmental factors, and equipment vibration.

[0003] The molten pool, as the core area of ​​metal melting and solidification during welding, directly determines the internal quality and appearance of the weld in terms of its shape, size, and flow state. Without effective acquisition and analysis of molten pool images, it is difficult to judge the trend of defect formation during welding in real time. Weld morphology, as a direct reflection of welding quality, is traditionally inspected through offline sampling, which cannot provide real-time feedback during welding. This leads to the inability to adjust process parameters promptly when quality problems occur, resulting in material waste and reduced production efficiency. Meanwhile, the interference of environmental factors on the welding process is often overlooked. Fluctuations in ambient temperature affect arc stability and the solidification rate of the molten pool; high ambient humidity can easily lead to defects such as porosity in the weld; abnormal equipment vibration frequency can cause welding torch position deviation, affecting the trajectory accuracy and penetration consistency of the weld. These factors all have a significant impact on welding quality but are not included in a unified data analysis system.

[0004] Traditional data processing methods lack effective integration and preprocessing of multimodal data. Different types of data differ in format, dimensions, and variation patterns. Direct analysis can easily lead to the neglect of correlations between data points, hindering a comprehensive understanding of welding process characteristics. Regarding process parameter adjustment, traditional methods rely heavily on manual experience, making it difficult to dynamically match and optimize process parameters based on real-time multi-dimensional data and environmental conditions. This results in poor welding process stability, significant product quality fluctuations, and an inability to meet the demands of modern manufacturing for intelligent and precise welding processes. Therefore, constructing a system capable of acquiring, integrating, analyzing, and dynamically matching process parameters for multimodal welding data has become a key direction for solving current welding process quality control challenges. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-sensor data fusion processing system for welding processes to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a multi-sensor data fusion processing system for welding processes, the system comprising:

[0007] The multimodal welding data acquisition module is used to acquire multimodal welding data during the welding process. The multimodal welding data includes welding current, welding voltage, molten pool image, weld morphology, ambient temperature, ambient humidity, and equipment vibration frequency.

[0008] A welding process feature profile generation module is used to generate a welding process feature profile based on the multimodal welding data. The welding process feature profile includes a welding quality feature vector and a process parameter feature vector.

[0009] The feature preprocessing module is used to perform time-domain segmentation and frequency-domain transformation on the multimodal welding data to generate a standard welding process parameter dataset.

[0010] The environment coding module is used to perform spectral feature coding on the ambient temperature and the vibration frequency of the equipment according to the environmental state attributes, and generate an environmental state coding sequence.

[0011] A process parameter mapper is used to match and map the environmental state encoding sequence with the standard welding process parameter dataset, and output process parameter mapping feature values.

[0012] Preferably, the feature preprocessing module includes:

[0013] The time-domain segmentation unit is used to segment the welding current and the welding voltage according to the welding time window, and generate a current time segment set and a voltage time segment set;

[0014] The frequency domain transformation unit is used to perform a fast Fourier transform on the current time segment set and the voltage time segment set to extract frequency domain energy distribution features;

[0015] The process parameter integration unit is used to integrate the frequency domain energy distribution features with the gray-level co-occurrence matrix features of the molten pool image to generate the standard welding process parameter dataset.

[0016] Preferably, the environment coding module includes:

[0017] The vibration spectrum analysis unit is used to perform wavelet packet decomposition on the vibration frequency of the equipment and extract vibration energy spectrum features.

[0018] A temperature frequency domain conversion unit is used to convert the ambient temperature into a temperature change frequency spectrum;

[0019] The encoding and integration unit is used to fuse the vibration energy spectrum features and the temperature change frequency spectrum to generate the environmental state encoding sequence.

[0020] Preferably, the welding process feature profile generation module includes:

[0021] The historical feature profiling unit is used to extract historical welding process feature profiles based on a preset welding quality assessment cycle. The historical welding process feature profiles include historical welding quality feature vectors and historical process parameter feature vectors.

[0022] The real-time feature profiling unit is used to acquire the current welding process feature profiling, which includes a real-time welding quality feature vector and a real-time process parameter feature vector.

[0023] The image deviation analysis unit is used to calculate the first deviation vector between the historical welding quality feature vector and the real-time welding quality feature vector, and the second deviation vector between the historical process parameter feature vector and the real-time process parameter feature vector, and merge them to generate a welding process feature image deviation vector.

[0024] Preferably, the system further includes:

[0025] The feature database node is used to store the welding process feature dataset associated with the welding process feature profile deviation vector;

[0026] The cloud retrieval module is used to send an encrypted retrieval request to the cloud server to obtain the target feature data index that matches the deviation vector of the welding process feature profile;

[0027] The feature extraction module is used to extract a subset of target welding process feature data from the feature database node based on the target feature data index.

[0028] Preferably, the cloud retrieval module includes:

[0029] A vector encryption unit is used to perform homomorphic encryption on the deviation vector of the welding process feature profile to generate an encrypted feature vector.

[0030] A node retrieval unit is used to distribute the encrypted feature vector to distributed feature database nodes to obtain an initial feature data index set;

[0031] An outlier processing unit is used to delete outlier index values ​​from the initial feature data index set and generate the target feature data index.

[0032] Preferably, the process parameter mapper includes:

[0033] A feature grouping unit is used to group and encode the feature values ​​of the process parameters according to the type of welding material.

[0034] The dynamic matching unit is used to match the grouped and coded process parameters to the corresponding welding time intervals based on the welding stage sequence.

[0035] The parameter optimization unit is used to dynamically adjust the set thresholds of the welding current and the welding voltage based on the matching results.

[0036] Preferably, the system further includes:

[0037] The mapper training module is used to collect multiple sets of welding process sample data to train the process parameter mapper. The multiple sets of welding process sample data include sample environmental state encoding sequences and sample standard welding process parameters.

[0038] The energy efficiency assessment module is used to calculate the mapped energy efficiency characterization value of the process parameter mapper;

[0039] The parameter binding module is used to bind process parameter mappers that meet the preset mapping energy efficiency characterization value thresholds to preset welding material types.

[0040] Preferably, the feature database node includes:

[0041] The node decryption unit is used to decrypt the received encrypted feature vector and restore the welding process feature profile deviation vector.

[0042] The feature retrieval unit is used to retrieve the local feature database based on the deviation vector of the reconstructed welding process feature profile;

[0043] The index feedback unit is used to return the feature data index corresponding to the search results to the cloud search module.

[0044] Preferably, the system further includes:

[0045] The load balancing module is used to count the real-time query load of each characteristic database node;

[0046] The priority allocation unit is used to dynamically allocate the storage priority of welding process feature data according to the real-time query load.

[0047] Distributed storage units are used to distribute and store subsets of target welding process feature data to corresponding feature database nodes according to storage priority.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] By setting up a multimodal welding data acquisition module, the system can comprehensively acquire multi-dimensional data during the welding process, including welding current, welding voltage, molten pool image, weld morphology, ambient temperature, ambient humidity, and equipment vibration frequency. This breaks through the limitations of traditional welding monitoring that only focuses on a few parameters, achieving comprehensive data coverage of the welding process from core process parameters to external influencing factors. This multi-dimensional data acquisition method enables the system to capture the correlation between different factors during the welding process. For example, the interaction between changes in ambient temperature and fluctuations in welding current and voltage, and the intrinsic relationship between abnormal equipment vibration frequency and deviations in molten pool morphology and weld morphology. This provides a more comprehensive reflection of the true state of the welding process and avoids the one-sidedness of welding quality judgment caused by single data acquisition.

[0050] The welding process feature profile generation module generates a welding process feature profile based on multimodal welding data, containing welding quality feature vectors and process parameter feature vectors. This transforms scattered, multi-dimensional data into structured, visualized feature information, allowing staff to intuitively grasp the overall characteristics and quality trends of the welding process. By constructing feature vectors for welding quality-related data and process parameter data respectively, the module clearly presents the impact of process parameter changes on welding quality. This establishes a clear correlation between quality fluctuations during the welding process and the corresponding direction of process parameter adjustments, providing a clear analytical basis for subsequent process parameter optimization without relying on vague judgments based on human experience.

[0051] The feature preprocessing module performs time-domain segmentation and frequency-domain transformation on multimodal welding data to generate a standard welding process parameter dataset, effectively solving the data integration challenges caused by differences in data formats and dimensions. Time-domain segmentation divides continuous welding data into segments with specific process significance according to the time dimension, facilitating the analysis of parameter variation characteristics at different welding stages. Frequency-domain transformation can extract hidden information from the data from a frequency perspective, such as identifying the frequency components of abnormal vibrations from equipment vibration frequency data and extracting periodic fluctuation characteristics from welding current and voltage data. These processing methods enable multimodal data to be transformed into a unified standard dataset, laying a solid foundation for subsequent data analysis and process parameter matching, and improving data utilization efficiency.

[0052] The environmental coding module encodes the ambient temperature and equipment vibration frequency using spectral features based on environmental state attributes, generating an environmental state coding sequence. This transforms environmental factors that are difficult to quantify directly into structured coded information. This coding method can accurately capture the dynamic trends of ambient temperature changes and the characteristic patterns of equipment vibration frequencies. For example, it can distinguish between slow drifts and sudden changes in ambient temperature through spectral feature coding, and identify normal fluctuations and abnormal impacts in equipment vibration. This allows the impact of environmental factors on the welding process to be represented in a calculable and matchable form, avoiding the problem of insufficient analytical accuracy caused by the qualitative description of environmental factors in traditional analysis.

[0053] The process parameter mapper matches and maps the environmental state encoding sequence with a standard welding process parameter dataset, outputting process parameter mapping feature values, thus realizing a dynamic correlation between environmental states and process parameters. Through this matching mapping, the system can automatically find suitable process parameter combinations based on real-time changes in environmental states. For example, when the ambient temperature rises, the system adjusts the welding current and voltage to a reasonable range through mapping; when abnormal equipment vibration frequencies occur, it matches welding torch operating parameters that can counteract the vibration effects. This allows welding process parameters to be optimized in real time according to changes in environmental states, ensuring the welding process remains stable under different environmental conditions, reducing welding defects caused by environmental interference, improving the consistency of welded product quality, and reducing reliance on manual parameter adjustments. This promotes the development of welding processes towards intelligence and automation, and is applicable to welding production scenarios of different scales and types, possessing broad application prospects. Attached Figure Description

[0054] Figure 1 This is a timing diagram of the multi-sensor data fusion processing system for the welding process described in this invention;

[0055] Figure 2 The flowchart for the feature preprocessing module;

[0056] Figure 3 A flowchart illustrating the process of generating a feature profile of the welding process;

[0057] Figure 4 A flowchart for processing by the cloud retrieval module;

[0058] Figure 5 A flowchart for the process parameter mapper. Detailed Implementation

[0059] 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.

[0060] Please see Figure 1 This invention provides a multi-sensor data fusion processing system for welding processes, the system comprising:

[0061] The multimodal welding data acquisition module is responsible for acquiring various types of sensor data in real time during the welding process, including welding current, welding voltage, molten pool image, weld morphology, ambient temperature, ambient humidity, and equipment vibration frequency. The welding process feature profile generation module generates welding quality feature vectors and process parameter feature vectors based on the acquired multimodal welding data, which together constitute the welding process feature profile. The feature preprocessing module performs time-domain segmentation and frequency-domain transformation on the welding current and welding voltage to form a standard welding process parameter dataset. The environmental encoding module performs spectral feature encoding on the ambient temperature and equipment vibration frequency to generate an environmental state encoding sequence. The process parameter mapper matches and maps the environmental state encoding sequence with the standard welding process parameter dataset, outputting process parameter mapping feature values ​​for real-time adjustment of welding process parameters.

[0062] Example 1: See Figure 2The time-domain segmentation unit of the feature preprocessing module first processes the welding current and welding voltage signals. This unit presets a 200-millisecond welding time window, which is wide enough to capture a complete welding droplet transition cycle. The system uses this window as the basic unit to continuously and non-overlappingly segment the high-speed sampled raw current and voltage signals. The time-domain segmentation and frequency-domain transformation of multimodal welding data mentioned in the feature preprocessing module specifically refers to segmenting the welding current and welding voltage electrical signal data in the multimodal welding data into a set of fragments according to the welding time window through the time-domain segmentation unit, and then performing a fast Fourier transform through the frequency-domain transformation unit to extract the frequency-domain energy distribution features. For the molten pool image data in the multimodal welding data, the feature preprocessing module does not perform time-domain segmentation and frequency-domain transformation, but instead extracts gray-level co-occurrence matrix features (such as contrast, energy, homogeneity, and correlation features) through the process parameter integration unit. The process parameter integration unit integrates the frequency-domain energy distribution features of the electrical signals with the gray-level co-occurrence matrix features of the molten pool image to form a standard welding process parameter dataset. The above processing method covers the core process-related data in multimodal welding data and conforms to the overall logic of multimodal data preprocessing. Current signal segments within each time window are aggregated into a current time segment set, and voltage signal segments form a voltage time segment set. These time segments retain the original waveform details of the signals, providing a foundation for subsequent frequency domain analysis. The frequency domain transformation unit then processes these time segment sets. This unit independently performs a Fast Fourier Transform (FFT) on each current and voltage time segment. The transform converts the current and voltage fluctuation signals in the time domain into an energy distribution representation in the frequency domain. Through this transformation, the energy intensity of different frequency components in each segment can be extracted, forming frequency domain energy distribution characteristics. These characteristics reveal deeper information such as arc stability and energy input uniformity during the welding process; for example, the concentration of energy in a specific high-frequency band may indicate a specific state of the arc.

[0063] The process parameter integration unit processes molten pool image data in parallel. This unit performs texture feature analysis on molten pool images captured by a high-speed camera, using a gray-level co-occurrence matrix algorithm for calculation. This algorithm extracts multiple feature indicators to describe the surface state of the molten pool by analyzing the spatial relationship of gray-level values ​​between image pixels, including contrast, energy, homogeneity, and correlation. The contrast feature of the molten pool image reflects the degree of unevenness of the molten pool surface, the energy feature characterizes the uniformity of the image texture, homogeneity describes the local consistency of the texture, and correlation measures the spatial dependence of the texture. These features extracted from visual information differ significantly in dimensionality and physical meaning from the frequency domain energy distribution features obtained from electrical signals.

[0064] The core task of the process parameter integration unit is to fuse these two heterogeneous features into a unified, standardized dataset. This unit employs a feature concatenation and normalization strategy, concatenating the current frequency domain feature vector (5-dimensional), voltage frequency domain feature vector (5-dimensional), and molten pool image feature vector (4-dimensional) corresponding to each time window in the order of current frequency domain feature - voltage frequency domain feature - molten pool image feature, forming a longer 14-dimensional comprehensive feature vector. During concatenation, the internal dimensional order of each vector remains unchanged, and all feature values ​​are subjected to max-min normalization to the [0,1] interval. Subsequently, max-min normalization is performed on all vectors to eliminate the influence of differences in feature dimensions and numerical ranges, ultimately generating a standard welding process parameter dataset. Each data point in this dataset uniquely corresponds to a time window and comprehensively reflects the electrical signal and visual state information within that time period.

[0065] The environmental coding module processes data concurrently with the feature preprocessing module, but focuses on the potential impact of environmental factors on the welding process. The vibration spectrum analysis unit within this module processes the equipment vibration frequency signals. During welding, mechanical vibrations from the wire feeding mechanism, robotic arm, and other equipment are captured by accelerometers, forming a vibration signal sequence. The vibration spectrum analysis unit performs a three-level wavelet packet decomposition on this signal, distributing the energy of the vibration signal into a complete set consisting of multiple frequency bands. By calculating the energy proportion of the signal within each frequency band, the vibration energy spectrum features are extracted. This feature can finely characterize the frequency distribution of equipment vibration; vibrations from different sources often exhibit significant energy differences across different frequency bands. The temperature frequency domain conversion unit processes ambient temperature data. This unit does not simply record instantaneous or average temperature values, but rather focuses on the dynamic characteristics of temperature changes over time. It treats the temperature data sequence collected over a period of time as a signal and performs a Fourier transform on it, converting it to the frequency domain to obtain the temperature change frequency spectrum. This spectrum reveals the periodicity and main frequency components of ambient temperature fluctuations; for example, whether there are periodic temperature fluctuations caused by the workshop ventilation system.

[0066] The encoding integration unit is the final stage of the environment coding module. Its task is to fuse the feature representations of vibration and temperature into a unified environment state code. This unit weights and concatenates the vibration energy spectrum feature vector and the temperature change frequency spectrum vector to generate a multi-dimensional environment state coding sequence. The weighting coefficients are set based on prior knowledge of the influence of vibration and temperature on specific welding processes. For example, in vibration-sensitive precision welding, the weight of vibration features may be set higher. The generated environment state coding sequence, as a whole, quantifies the instantaneous state of external environmental conditions, providing crucial environmental context information for subsequent matching and mapping of process parameters.

[0067] Example 2: See Figure 3 The historical feature profile unit serves to establish a high-quality benchmark reference for the welding process. This unit sets a fixed welding quality assessment cycle, for example, one hundred welded joints completed per cycle. The welding process feature profile generation process is as follows: receiving welding current, voltage, molten pool image, weld morphology, ambient temperature and humidity, and equipment vibration frequency data output from the multimodal welding data acquisition module; calculating the mean, variance, and peak statistical features of current and voltage; extracting area, perimeter, and roundness features from the molten pool image through edge detection; extracting reinforcement height, weld width, and weld depth features from the weld morphology; calculating the deviation rate between real-time values ​​and standard values ​​for ambient temperature and humidity and equipment vibration frequency; and sorting the extracted features according to preset dimensions to form a feature vector. The vector content includes a welding quality feature vector containing molten pool roundness, weld reinforcement height deviation rate, weld width consistency, and porosity percentage (4-dimensional, numerical range [0,1]); and a process parameter feature vector containing welding current mean, voltage variance, current peak deviation, and voltage fluctuation frequency (4-dimensional, numerical range [0,1]). At the end of each evaluation cycle, the system integrates data from all welding processes within that cycle to generate a historical welding process feature profile representing the average process level of that cycle. This profile consists of two core vectors: a historical welding quality feature vector and a historical process parameter feature vector. The historical welding quality feature vector is not generated through direct measurement, but rather by inferring from the results of non-destructive testing and performance testing of all welded joints completed within that cycle. These results may include the number of internal porosity and slag inclusions determined by sampled X-ray inspections, the statistical values ​​of weld penetration and width measurements, and strength samples obtained from tensile tests. These quality indicators, after normalization, are combined into a multi-dimensional vector to quantify the overall welding quality level of that production cycle. The historical process parameter feature vector is directly derived from the statistical characteristics of the standard welding process parameter dataset output by the feature preprocessing module within that cycle. The system calculates the mean, variance, extreme values, and other statistics of the standard feature vector corresponding to each time window within that cycle, sorts and concatenates these statistics to form a fixed-dimensional vector describing the typical process state and fluctuations of that cycle. All these historical welding process feature profiles are assigned timestamps and production batch identifiers and stored in an orderly manner in the feature database nodes, forming a continuously growing high-quality process history library.

[0068] The real-time feature profiling unit operates in parallel with historical data generation, but its focus is entirely on the current welding process. This unit continuously acquires the latest output from the multimodal welding data acquisition module at intervals much shorter than the historical evaluation cycle (e.g., per second). It uses the exact same feature extraction and construction algorithms as the historical unit to generate real-time welding quality feature vectors and real-time process parameter feature vectors for the current moment. However, a crucial difference is that the generation of the real-time welding quality feature vector cannot rely on post-event detection results. Therefore, it employs an indirect quality inference method based on process signals. This unit integrates molten pool image features (such as morphological stability and area change rate) and electrical signal frequency domain features (such as specific frequency band energy of arc acoustic emission signals and short-time stability indices of current and voltage waveforms) within the current and recent time windows. A pre-trained model maps these features to estimate the key quality attributes that the current weld may form, thus constructing the real-time welding quality feature vector. The construction of the real-time process parameter feature vector is more direct. It involves performing a moving average and feature extraction on a standard welding process parameter dataset from the most recent several time windows. Its dimensions and structure are completely consistent with the historical process parameter feature vectors, ensuring comparability.

[0069] The profile deviation analysis unit is the output of the entire module, and its function is to quantify the deviation between the current actual operation and the historical excellent benchmark. This unit receives one or more recent historical benchmark profiles from the historical feature profile unit, and the current real-time profile from the real-time feature profile unit. Its calculation process is divided into two parallel paths. The first path calculates the first deviation vector: this unit selects the most representative historical welding quality feature vector as the benchmark (e.g., the vector corresponding to the batch rated "excellent" in the most recent evaluation cycle), and performs element-wise subtraction between the real-time welding quality feature vector and this benchmark vector. Since all vectors have been normalized during construction, this difference directly reflects the absolute gap between the currently estimated weld quality and the historical excellent level in each dimension. The second path calculates the second deviation vector: this unit uses the same benchmark selection strategy to obtain the corresponding historical process parameter feature vector, and performs element-wise subtraction between it and the real-time process parameter feature vector. This calculation result quantifies the difference between the overall state of the current process parameters (such as the average level and stability of energy input) and the historical excellent operating conditions.

[0070] The profile deviation analysis unit merges the calculated first deviation vector (4-dimensional, corresponding to the deviations in each dimension of the welding quality feature vector) and the second deviation vector (4-dimensional, corresponding to the deviations in each dimension of the process parameter feature vector) in dimensional order of quality deviation - process parameter deviation, generating a comprehensive 8-dimensional welding process feature profile deviation vector. After merging, each dimension's deviation value is retained to three decimal places; a positive value indicates that the real-time value is higher than the historical benchmark, and a negative value indicates that the real-time value is lower than the historical benchmark. This deviation vector is a high-dimensional signal; the magnitude and sign of each dimension's value precisely indicate the direction and magnitude of the current welding process's deviation from the historical high-quality benchmark in each quality attribute and process parameter feature. This deviation vector is output to other modules of the system in real time. For example, it can be used by the cloud retrieval module to find historical cases with similar deviation patterns and their solutions in the historical database; more directly, it can provide a high-level optimization target for the process parameter mapper, indicating how it should adjust parameters to reduce the current overall deviation and bring the welding process back to a known high-quality state. Through this continuous comparison and feedback mechanism, the entire module transforms historical experience into precise quantitative guidance for the current process.

[0071] Example 3: See Figure 4 After the welding process feature profile generation module generates a welding process feature profile deviation vector, this vector is transmitted to the cloud retrieval module. The vector encryption unit in the cloud retrieval module first encrypts this vector. The encryption process uses an order-preserving encryption algorithm, which maintains the order relationship between values, allowing the encrypted data to still be compared and manipulated. The encryption function is expressed as:

[0072]

[0073] in: This represents the deviation vector of the original welding process feature profile, and each component of it... This represents the deviation value on a specific feature dimension; Represents the encrypted deviation vector; This represents the order-preserving encryption algorithm function. During encryption, a pre-generated key pair is used to ensure that only authorized feature database nodes can perform decryption operations.

[0074] The node retrieval unit distributes the encrypted feature vectors to all distributed feature database nodes. Each node, upon receiving the encrypted vector, performs a ciphertext domain similarity calculation locally. This calculation is based on a distance metric between encrypted vectors; the node uses a ciphertext processing algorithm to calculate the approximate distance between the query vector and each historical encrypted vector stored locally. This process does not require data decryption, protecting data privacy. Based on the calculation results, each node returns the index identifiers of the k most similar historical data records, forming an initial feature data index set. The outlier processing unit cleanses the collected initial feature data index set. This unit uses statistical analysis methods to calculate the distribution characteristics of the distance values ​​corresponding to all returned indices. By calculating the quartiles and interquartile ranges of the distance values, it identifies and removes outlier index values ​​that significantly deviate from the main data distribution. The processing employs an adaptive threshold mechanism, with the threshold dynamically calculated based on the distribution characteristics of the current dataset. After outlier processing, a refined target feature data index set is obtained, improving the accuracy and consistency of the retrieval results.

[0075] The feature extraction module initiates a data extraction request to the corresponding feature database node based on the final obtained feature data index. Upon receiving the request, the node decrypts the previously received encrypted vector using its decryption unit. The decryption process uses the key matching the encryption key to restore the encrypted vector to the original welding process feature profile deviation vector. The feature retrieval unit uses the decrypted vector to perform a precise matching search in the local database to find the corresponding welding process feature data. The index feedback unit packages the retrieved feature data and returns it to the feature extraction module, completing the entire retrieval process.

[0076] The system monitors the operational status of each feature database node through a load balancing module. This module collects real-time metrics such as node processing load, network latency, and resource usage, dynamically adjusting the allocation strategy for query requests. A priority allocation unit assigns appropriate priorities to different query requests based on the real-time load of the nodes, ensuring that high-priority requests are processed quickly. A distributed storage unit manages the physical storage distribution of data, optimizing data storage locations based on data access patterns and node storage capacity to improve data access efficiency. Data security mechanisms are implemented throughout the entire process; all sensitive data transmitted over the network is encrypted, and communication between nodes is protected by secure protocols. When processing data, feature database nodes adhere to the principle of least privilege, exposing only necessary interfaces and functions. The system also maintains a complete security audit log, recording all data access and operation activities for easy post-event auditing and troubleshooting.

[0077] Example 4: See Figure 5The process parameter mapper begins its operation with a feature grouping unit, which preprocesses and classifies the feature values ​​of the process parameters to be matched and mapped based on the welding material type. The system's preset material types include low-carbon steel, stainless steel, and aluminum alloys. Each material corresponds to a unique set of grouping and coding rules, based on its physical properties (such as thermal conductivity and melting point) and chemical properties (such as alloy composition). The process parameter mapper's matching and mapping method involves: establishing a mapping database to store standard process parameters for different materials (low-carbon steel, stainless steel, aluminum alloys) under different environmental conditions (3D vibration energy spectrum + 2D temperature change frequency spectrum, for a total of 5 dimensions); and calculating similarity using the Euclidean distance formula. The Euclidean distance formula is as follows:

[0078]

[0079] in, Represents Euclidean distance. For real-time encoding dimension, For standard coding number The system selects the standard parameter with the highest similarity and corrects it according to the welding stage (current correction coefficient 1.1 for the arc initiation stage, 1.0 for the stabilization stage, and 0.9 for the arc termination stage). The feature value includes four dimensions: matching similarity ([0,1]), current correction value ([180,250]A), voltage correction value ([24,29]V), and matching confidence ([0.6,1.0]). For example, for aluminum alloys with high thermal conductivity, the encoding may focus more on the parameter channels related to heat dissipation; while for stainless steel, the encoding may focus more on maintaining parameter stability within the temperature range. Before entering the core matching process, all input process parameter mapping feature values ​​are assigned a clear grouping code label, which will determine the parameter weights and matching rule base used in subsequent matching calculations.

[0080] The dynamic matching unit is crucial for the mapper to achieve precise timing control. This unit receives process parameter mapping feature values ​​with grouped coding tags and matches them to the corresponding welding time intervals based on the welding stage sequence. The welding process is divided into three stages: arc initiation, stabilization, and arc termination. Each stage has its unique process objectives and control logic. The arc initiation stage requires rapid establishment of a stable arc, and its matching rules tend to select feature values ​​that allow for rapid current rise and precise voltage matching. The stabilization stage prioritizes process consistency and quality stability, and its matching rules focus on feature value combinations that minimize parameter fluctuations. The arc termination stage needs to prevent defects such as crater cracks, and its matching rules seek feature values ​​with smooth current decay curves. The dynamic matching unit maintains a time-rule mapping table, ensuring that at every millisecond, the most suitable matching algorithm for the current stage is invoked to calculate the final output. The parameter optimization unit performs the final process parameter adjustments based on the calculation results of the dynamic matching unit. This unit receives the matched feature values ​​and converts them into specific control commands for actuators such as the welding power source and wire feed mechanism. The adjustment process employs closed-loop control logic, using the calculated ideal characteristic value as the target setpoint and real-time acquired current and voltage as process measurements. The output is dynamically adjusted through a control algorithm to continuously bring the actual process parameters closer to the target setpoint. This unit can fine-tune or significantly correct the set thresholds for welding current and welding voltage to cope with interference from environmental fluctuations.

[0081] The mapper training module is responsible for optimizing the performance of the process parameter mapper. This module collects multiple sets of welding process sample data for training, forming a training set covering various working conditions. Each set of sample data is a paired data element, containing a sample environment state encoding sequence and its corresponding sample standard welding process parameters. The sample environment state encoding sequence comes from the historical records of the environment encoding module, while the sample standard welding process parameters come from historical data output by the feature preprocessing module, which has been verified as high-quality by actual weld quality inspection. The training process is implemented through an iterative learning algorithm, continuously adjusting the matching rules and parameter weights within the mapper so that when an environment state encoding is input, the mapper's output gets closer and closer to the optimal process parameters under that environment. The energy efficiency evaluation module quantifies the performance of the trained process parameter mapper. This module calculates the mapper's mapping energy efficiency characterization value, which is a comprehensive performance index reflecting the degree of agreement between the mapper's output and the sample standard welding process parameters. The mapping energy efficiency characterization value is defined as follows: current deviation rate:

[0082]

[0083] in, For mapped current, This is the standard current. The voltage deviation rate is:

[0084]

[0085] in, For mapped voltage, Standard voltage. Quality compliance rate:

[0086]

[0087] The mapped energy efficiency characterization value is: (Value range [0,1], preset threshold 0.2, (Qualified).

[0088] The evaluation process is conducted on a separate test dataset containing historical samples that were not used in the training process to avoid overfitting. The mapped energy efficiency characterization value is a dimensionless scalar; the lower the value, the closer the mapper's output is to the ideal process parameters, and the better its performance.

[0089] The parameter binding module, as the core execution unit of the system configuration, undertakes the task of accurately matching welding process parameter mappers with material types. This module filters process parameter mappers through preset mapping energy efficiency characterization thresholds, and only binds mapper instances that meet the threshold conditions with specified welding material types. System administrators can set dynamically adjustable performance thresholds based on welding quality control standards, and quantitatively evaluate the steady-state error rate, dynamic response speed, and process robustness indicators of the mappers through an integrated evaluation engine. Only mapper instances that pass the evaluation can enter the production line deployment process. Each validated mapper instance will be assigned a two-dimensional label system: the material type label adopts the ASTM metal material classification code, and the version label follows the semantic version control specification. It is stored in the mapper library through a distributed hash table algorithm to achieve O(1) complexity production calls. This mechanism constructs a full-link quality control system of "threshold screening - multi-dimensional evaluation - label management - hash storage", which fundamentally ensures the process reliability and performance stability of the mappers during material welding, and meets the traceability and verifiability requirements of intelligent manufacturing systems.

[0090] Table 1: Training sample dataset for process parameter mapper.

[0091]

[0092] Example 5: The feature database node, as the basic unit of the distributed storage network, contains multiple functional components. The node decryption unit is responsible for processing encrypted feature vectors from the cloud retrieval module. This unit is equipped with a decryption algorithm and key management mechanism that matches the encryption system. When an encrypted feature vector arrives at the node, the decryption unit first verifies the legitimacy of the request source, and then uses the corresponding private key to decrypt the encrypted data. The decryption process restores the encrypted feature vector to the original welding process feature profile deviation vector, which contains the deviation information between the current welding process and the historical benchmark in the multi-dimensional feature space. The decryption operation is performed within the node's secure isolation zone, ensuring that the key and plaintext data are not leaked.

[0093] Upon obtaining the decrypted deviation vector, the feature retrieval unit immediately initiates the retrieval process. This unit accesses the node's local feature database, which stores a large amount of historical welding process feature data. Each data entry contains a complete set of feature vectors and corresponding process parameter information. The retrieval process employs a vector similarity-based matching algorithm to calculate the Euclidean distance between the query vector and each historical vector in the database. To improve retrieval efficiency, the database pre-establishes an efficient spatial index structure, enabling the system to quickly locate the few historical data records closest to the query vector. The retrieval process considers additional conditions such as welding material type and process type, ensuring that the returned results are not only highly similar but also context-relevant.

[0094] The index feedback unit plays a crucial role, primarily responsible for formatting and returning search results. Specifically, this unit performs a series of transformations on the matching results obtained by the feature retrieval unit, ultimately generating a standardized index format. These standardized indexes not only uniquely identify data records in the database but also ensure data consistency and traceability. The index information contains several key elements, such as node identifiers, data block addresses, and metadata information like data versions. To improve transmission efficiency, this index information is encoded in a compact binary format, ensuring data integrity while minimizing transmission time and bandwidth.

[0095] Before sending the index information back to the requester, the index feedback unit performs a series of preprocessing steps on the search results. First, it conducts an initial screening to exclude data records that do not meet the requirements or are irrelevant. Then, based on preset sorting rules, typically based on similarity, the filtered results are sorted to ensure that the most relevant data records are presented first. Furthermore, to avoid returning too much index information and overburdening subsequent processing, the index feedback unit also manages the number of returned indexes reasonably, thus ensuring the manageability and efficiency of subsequent processing. The filtered and sorted index information is then sent back to the cloud retrieval module through a secure channel, ensuring the security and reliability of the data during transmission. This series of operations not only improves the overall performance of the retrieval system but also provides users with more accurate and efficient search services.

[0096] The load balancing module continuously monitors the operational status of the entire distributed storage system. This module assesses system load by periodically collecting performance metrics from each characteristic database node. These metrics include node CPU utilization, memory usage, disk I / O throughput, network bandwidth usage, and the number of concurrent queries. Monitoring data is stored in time-series format for analyzing load patterns and trends of each node. The module employs a sliding window mechanism to calculate recent load metrics, avoiding the impact of instantaneous fluctuations on the evaluation results. The priority allocation unit dynamically calculates the storage priority of each node based on real-time data provided by the load balancing module. The priority calculation algorithm comprehensively considers the node's current load status, hardware performance metrics, and historical load patterns. Nodes with lighter loads and better performance are assigned higher storage priorities, while nodes nearing saturation are assigned lower priorities. Priority evaluation is a continuous and dynamic process, constantly adjusted as the system's operational status changes. The priority value of each node is updated periodically and broadcast throughout the system.

[0097] The distributed storage unit is responsible for executing the actual data distribution and storage strategy according to the instructions of the priority allocation unit. This unit receives a subset of the target welding process feature data that needs to be stored and distributes the data blocks to different feature database nodes according to the current priority value of each node. The data distribution algorithm considers factors such as the access frequency of data blocks, data correlation, and the geographical distribution of nodes. For frequently accessed "hot" data, it will be preferentially stored on lightly loaded, high-performance nodes; while for infrequently accessed "cold" data, it can be stored on nodes with higher loads or lower storage costs. This unit is also responsible for data redundancy backup to ensure data reliability and availability.

[0098] The various components communicate and exchange data through clearly defined interface protocols. An encrypted communication protocol ensures secure data transmission between the node decryption unit and the cloud retrieval module. The feature retrieval unit interacts efficiently with the local database through an optimized data access interface. The load balancing module maintains state synchronization with each feature database node through a lightweight heartbeat mechanism. The priority allocation unit and the distributed storage unit exchange data rapidly through shared memory or message queues.

[0099] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0100] 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. A multi-sensor data fusion processing system for welding processes, characterized in that, include: The multimodal welding data acquisition module is used to acquire multimodal welding data during the welding process. The multimodal welding data includes welding current, welding voltage, molten pool image, weld morphology, ambient temperature, ambient humidity, and equipment vibration frequency. A welding process feature profile generation module is used to generate a welding process feature profile based on the multimodal welding data. The welding process feature profile includes a welding quality feature vector and a process parameter feature vector. The feature preprocessing module is used to perform time-domain segmentation and frequency-domain transformation on the multimodal welding data to generate a standard welding process parameter dataset. The environment coding module is used to perform spectral feature coding on the ambient temperature and the vibration frequency of the equipment according to the environmental state attributes, and generate an environmental state coding sequence. A process parameter mapper is used to match and map the environmental state encoding sequence with the standard welding process parameter dataset, and output process parameter mapping feature values; The feature preprocessing module includes: The time-domain segmentation unit is used to segment the welding current and the welding voltage according to the welding time window, and generate a current time segment set and a voltage time segment set; The frequency domain transformation unit is used to perform a fast Fourier transform on the current time segment set and the voltage time segment set to extract frequency domain energy distribution features; The process parameter integration unit is used to integrate the frequency domain energy distribution features with the gray-level co-occurrence matrix features of the molten pool image to generate the standard welding process parameter dataset.

2. The welding process multi-sensor data fusion processing system as described in claim 1, characterized in that, The environment coding module includes: The vibration spectrum analysis unit is used to perform wavelet packet decomposition on the vibration frequency of the equipment and extract vibration energy spectrum features. A temperature frequency domain conversion unit is used to convert the ambient temperature into a temperature change frequency spectrum; The encoding and integration unit is used to fuse the vibration energy spectrum features and the temperature change frequency spectrum to generate the environmental state encoding sequence.

3. The welding process multi-sensor data fusion processing system as described in claim 1, characterized in that, The welding process feature profile generation module includes: The historical feature profiling unit is used to extract historical welding process feature profiles based on a preset welding quality assessment cycle. The historical welding process feature profiles include historical welding quality feature vectors and historical process parameter feature vectors. The real-time feature profiling unit is used to acquire the current welding process feature profiling, which includes a real-time welding quality feature vector and a real-time process parameter feature vector. The image deviation analysis unit is used to calculate the first deviation vector between the historical welding quality feature vector and the real-time welding quality feature vector, and the second deviation vector between the historical process parameter feature vector and the real-time process parameter feature vector, and merge them to generate a welding process feature image deviation vector.

4. The welding process multi-sensor data fusion processing system as described in claim 3, characterized in that, Also includes: The feature database node is used to store the welding process feature dataset associated with the welding process feature profile deviation vector; The cloud retrieval module is used to send an encrypted retrieval request to the cloud server to obtain the target feature data index that matches the deviation vector of the welding process feature profile; The feature extraction module is used to extract a subset of target welding process feature data from the feature database node based on the target feature data index.

5. The welding process multi-sensor data fusion processing system as described in claim 4, characterized in that, The cloud-based retrieval module includes: A vector encryption unit is used to perform homomorphic encryption on the deviation vector of the welding process feature profile to generate an encrypted feature vector. A node retrieval unit is used to distribute the encrypted feature vector to distributed feature database nodes to obtain an initial feature data index set; An outlier processing unit is used to delete outlier index values ​​from the initial feature data index set and generate the target feature data index.

6. The welding process multi-sensor data fusion processing system as described in claim 1, characterized in that, The process parameter mapper includes: A feature grouping unit is used to group and encode the feature values ​​of the process parameters according to the type of welding material. The dynamic matching unit is used to match the grouped and coded process parameters to the corresponding welding time intervals based on the welding stage sequence. The parameter optimization unit is used to dynamically adjust the set thresholds of the welding current and the welding voltage based on the matching results.

7. The multi-sensor data fusion processing system for welding processes as described in claim 6, characterized in that, Also includes: The mapper training module is used to collect multiple sets of welding process sample data to train the process parameter mapper. The multiple sets of welding process sample data include sample environmental state encoding sequences and sample standard welding process parameters. The energy efficiency assessment module is used to calculate the mapped energy efficiency characterization value of the process parameter mapper; The parameter binding module is used to bind process parameter mappers that meet the preset mapping energy efficiency characterization value thresholds to preset welding material types.

8. The welding process multi-sensor data fusion processing system as described in claim 4, characterized in that, The feature database nodes include: The node decryption unit is used to decrypt the received encrypted feature vector and restore the welding process feature profile deviation vector. The feature retrieval unit is used to retrieve the local feature database based on the deviation vector of the reconstructed welding process feature profile; The index feedback unit is used to return the feature data index corresponding to the search results to the cloud search module.

9. The welding process multi-sensor data fusion processing system as described in claim 8, characterized in that, Also includes: The load balancing module is used to count the real-time query load of each characteristic database node; The priority allocation unit is used to dynamically allocate the storage priority of welding process feature data based on the real-time query load. Distributed storage units are used to distribute and store subsets of target welding process feature data to corresponding feature database nodes according to storage priority.

Images