A method and device for health assessment of an ultrasonic welding horn and a storage medium

Abstract

This invention discloses a method, device, and storage medium for health assessment of ultrasonic welding heads in the field of ultrasonic welding technology. The method includes: acquiring multi-source data signals from the ultrasonic welding equipment within the assessment period; preprocessing the acquired multi-source data signals to obtain a feature vector to be assessed; inputting the feature vector to be assessed into a trained deep forest model and outputting the assessment result, which includes a health status category, a predicted remaining lifespan, and a feature importance ranking obtained based on model analysis; and dynamically adjusting an alarm threshold using an adaptive threshold algorithm based on historical assessment results within a sliding window, triggering a maintenance prompt when the alarm threshold is reached. This invention achieves adaptive feature fusion of multi-source sensor data, capturing the nonlinear laws of welding head health degradation; constructs an assessment model with both high precision and strong generalization ability, adapting to different welding head models and welding scenarios without requiring a large amount of labeled data; and realizes quantitative assessment of the welding head health status.

Description

Technical Field

[0001] This invention relates to a health assessment method, device, and storage medium for ultrasonic welding heads, belonging to the field of ultrasonic welding technology. Background Technology

[0002] Ultrasonic welding, as a highly efficient and clean joining technology, is widely used in precision manufacturing fields such as power batteries, microelectronics, and aerospace. The welding head, as the core component of ultrasonic welding equipment, directly determines the stability of welding quality. Wear, cracks, or resonant frequency shifts in the welding head can lead to defects such as insufficient welding strength and incomplete welds, even causing production interruptions. Existing welding head health assessment methods mainly rely on offline detection, such as dimensional measurements after periodic disassembly and flaw detection; or threshold judgment based on single sensor data, such as fixed threshold alarms based on current and amplitude signals. However, the ultrasonic welding process involves multi-physical field coupling, including mechanical vibration, heat conduction, and material deformation, resulting in nonlinear and multimodal characteristics in welding head health degradation. For example, welding heads at different wear stages exhibit significant differences in the spectral distribution of their vibration acceleration signals under the same welding parameters; and factors such as electromagnetic interference and workpiece differences in the production environment can cause dynamic drift in health characteristics. Traditional methods struggle to capture this complex degradation pattern, either due to the lag in offline detection leading to quality risks, or due to the limitations of single features resulting in numerous false alarms, such as misinterpreting normal signal fluctuations as abnormalities. Therefore, there is an urgent need for an assessment method that can integrate multi-source data in real time and adaptively learn health features to achieve accurate identification and early warning of the health status of welding heads. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method, device, and storage medium for health assessment of ultrasonic welding heads. This addresses the problems of low assessment accuracy, weak generalization ability, and insufficient real-time performance in complex welding scenarios. The invention introduces the deep forest algorithm into the field of welding head health assessment, automatically learning high-order features from multi-source data through a multi-level tree ensemble structure. While maintaining the interpretability and efficiency of traditional machine learning models, it possesses feature representation capabilities approaching those of deep learning, making it particularly suitable for handling complex nonlinear data in the welding process. It achieves adaptive feature fusion of multi-source sensor data, capturing the nonlinear patterns of welding head health degradation. It constructs an assessment model with both high accuracy and strong generalization ability, adapting to different welding head models and welding scenarios without requiring a large amount of labeled data. It achieves quantitative assessment of the welding head's health status, rather than simple abnormal alarms.

[0004] To achieve the above objectives, the present invention is implemented using the following technical solution: In a first aspect, the present invention provides a health assessment method for ultrasonic welding heads, comprising: Collect multi-source data signals from the ultrasonic welding equipment during the evaluation period; The collected multi-source data signals are preprocessed to obtain the feature vector to be evaluated; The feature vector to be evaluated is input into the trained deep forest model, and the evaluation results are output. The evaluation results include health status category, remaining life prediction value, and feature importance ranking obtained based on model analysis. Based on historical evaluation results within the sliding window, an adaptive threshold algorithm is used to dynamically adjust the alarm threshold, and a maintenance prompt is triggered when the alarm threshold is reached.

[0005] The aforementioned technical solution achieves real-time, online, and automated assessment of the welding head's health status through a complete process from multi-source data acquisition to deep forest model evaluation, completely changing the traditional lagging mode that relies on offline detection. This method comprehensively utilizes a multi-modal feature adaptive fusion mechanism and a hierarchical health status assessment framework, enabling not only high-precision identification of the current state but also quantifiable prediction of remaining lifespan, providing a direct basis for predictive maintenance. Furthermore, through an adaptive threshold algorithm based on a sliding window, discrete assessment results are transformed into robust and reliable decision logic. Adjusting alarm thresholds using the adaptive threshold algorithm to trigger early warnings effectively filters out random fluctuations in single assessments, significantly reducing the false alarm rate. This provides timely and reliable decision-making support for predictive maintenance of the production line, achieving a leap from condition monitoring to intelligent early warning.

[0006] Furthermore, the multi-source data signals include at least two of the following: longitudinal, transverse, and axial vibration signals of the welding head; real-time temperature signals of the contact area between the welding head and the workpiece; output current signals of the ultrasonic generator; and clamping force signals of the welding process.

[0007] The aforementioned technical solution, by explicitly specifying the fusion of at least two or more of the vibration, temperature, current, and clamping force signals, lays the data foundation for realizing a multi-modal feature adaptive fusion mechanism. This combination of multi-source signals can comprehensively capture the degradation information of the welding head from multiple physical dimensions such as mechanical vibration, thermal effects, electrical parameters, and mechanical state, overcoming the one-sidedness of single-signal analysis. Utilizing the complementarity between different signals, the model can more comprehensively and robustly learn the health degradation laws, thereby improving the reliability of the evaluation results.

[0008] Furthermore, the preprocessing includes noise removal, preliminary feature extraction, and data standardization.

[0009] In the above technical solution, a standardized preprocessing procedure provides high-quality, standardized input features for subsequent model analysis. This step effectively filters out inherent electromagnetic and mechanical noise interference in the industrial environment and initially extracts time-domain, frequency-domain, and time-frequency information from the signal. Standardized data creates conditions for the adaptive fusion of multimodal features and efficient training of the deep forest model, which is an important prerequisite for ensuring the accuracy of the final evaluation and the stability of the model.

[0010] Furthermore, the noise removal includes performing wavelet threshold denoising on the multi-source data signal to filter out power frequency interference and mechanical noise; The preliminary feature extraction includes extracting time-domain features, frequency-domain features, and time-frequency features from the noise-removed signal; The data standardization includes using Z-score standardization to map the extracted features to a zero-mean, unit-variance distribution.

[0011] The aforementioned technical solution further refines the specific preprocessing techniques. Wavelet threshold denoising effectively filters out specific noises such as power frequency interference while preserving the essential signal characteristics. Comprehensive extraction of time-domain, frequency-domain, and time-frequency features provides a complete description of the dynamic behavior and state changes of the welding head. Z-score standardization eliminates the influence of features with different dimensions on the model. These specific techniques work together to provide the deep forest model with a purer and more comparable feature set that better reflects the health of the welding head, directly improving the model's feature learning ability and evaluation accuracy.

[0012] Furthermore, the deep forest model includes an input layer, at least one cascaded forest layer, a feature splicing layer, and an output layer; The input layer will be used to input the feature vector to be evaluated after dividing it according to the signal mode. Each level of the cascaded forest layer contains multiple random forests, which are used to output the probability distribution of samples in the health status category; The feature concatenation layer is used to concatenate the output probability of the previous forest with the original features and use it as the input of the next forest, thereby realizing the hierarchical abstraction of features. The output layer includes a classifier and a regressor, which are used to output the health status category and the remaining life expectancy prediction value, respectively. After training, the deep forest model can perform global statistics based on the feature splitting gains of all decision trees in the model to generate the feature importance ranking.

[0013] In the aforementioned technical solution, the cascaded structure of the deep forest model naturally realizes a hierarchical health status assessment framework and a multimodal feature adaptive fusion mechanism. The input layer partitions data according to modality, and the cascaded forest layer automatically learns and fuses high-order abstract features across modalities through layer-by-layer voting and feature concatenation. The classifier and regressor in the output layer work together, balancing the real-time performance of state recognition with the refinement of lifespan prediction. Furthermore, based on this model structure, feature importance ranking can be easily generated by calculating the decision tree feature split gain, providing interpretability for the evaluation results.

[0014] Furthermore, the training process of the deep forest model includes: The historical data is divided into a training set and a validation set, and the historical data includes health status labels and remaining lifespan values ​​marked by experts based on offline detection results. The model is trained using the training set through a loss function, which is a weighted sum of classification loss and regression loss; the formula is: L = αL_class + (1-α) L_reg; Where L is the loss function, α is the weight coefficient, L_class is the classification loss, and L_reg is the regression loss; Training stops when the validation set loss does not decrease for a preset number of consecutive rounds.

[0015] In the above technical solution, the health status label can be divided into four levels: healthy, mildly degraded, severely degraded, and faulty. Training is performed using expert-annotated real data combined with a joint loss function, ensuring that the model simultaneously learns state classification and lifetime regression tasks, perfectly supporting the hierarchical health status assessment framework. The early stopping mechanism based on validation set loss effectively prevents model overfitting and ensures its generalization ability on real-world unknown data. This training method results in a model that is not only highly accurate but also stable and reliable, adaptable to complex industrial scenarios.

[0016] Furthermore, the evaluation method also includes triggering incremental learning when the difference between the feature distribution of the newly collected data and the feature distribution of the training set used for model training exceeds a preset threshold. The deep forest model is updated by retaining some tree models from the original forest and adding newly trained tree models.

[0017] The above technical solution introduces an incremental learning mechanism to achieve dynamic model updates. When new data distribution drift is detected, the mechanism can update the model quickly and at low cost by partially retaining historical knowledge and incorporating new knowledge. This avoids the drawbacks of traditional methods that require relabeling a large amount of data and retraining the entire system. It significantly enhances the long-term adaptability and generalization ability of the evaluation system under different batches and operating conditions, and reduces operation and maintenance costs.

[0018] Furthermore, the deep forest model is a lightweight model, which removes redundant decision trees or tree nodes by pruning the trained model and quantizes the feature splitting threshold in the model.

[0019] In the above technical solution, pruning can effectively remove redundant structures and parameters in the model, while quantization significantly reduces the resource requirements for model computation and storage. These two technologies work together to greatly improve the inference efficiency of the model while maintaining the basic evaluation accuracy, thereby ensuring that the evaluation method can meet the stringent real-time requirements of the production line and providing a solid technical foundation for achieving uninterrupted online health assessment and early warning.

[0020] Secondly, the present invention provides a health assessment device for ultrasonic welding heads, the device comprising: Data acquisition module: used to collect multi-source data signals from the ultrasonic welding equipment during the evaluation period; Data processing module: used to preprocess the acquired multi-source data signals to obtain the feature vector to be evaluated; Health assessment module: This module takes the feature vector to be assessed as input into a trained deep forest model and outputs the assessment results, which include the health status category, the predicted remaining life expectancy, and the ranking of feature importance obtained based on model analysis. The early warning module is used to dynamically adjust the alarm threshold based on the historical evaluation results within the sliding window using an adaptive threshold algorithm. When the alarm threshold is reached, a maintenance prompt is triggered.

[0021] The aforementioned technical solution modularizes and materializes data acquisition, processing, and evaluation steps, forming a dedicated ultrasonic welding head health assessment device. This device integrates a series of innovative functions, including multimodal feature adaptive fusion, hierarchical health status assessment, incremental learning and updating, and feature importance interpretation. It facilitates deployment and implementation on production lines, providing users with a ready-to-use solution that effectively ensures welding quality stability and reduces maintenance costs and product defect rates.

[0022] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the health assessment method for the ultrasonic welding head described in the first aspect.

[0023] In the above technical solution, the innovative ultrasonic welding head health assessment method is stored in software form, which can be easily copied, distributed and deployed on various computing devices or embedded systems. By running the program, general industrial computers or edge computing devices can have professional intelligent assessment capabilities for welding head health, which greatly reduces the threshold for technology promotion and system integration. This is conducive to the widespread application of this high-precision, interpretable and adaptive advanced assessment technology in the industry.

[0024] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: The ultrasonic welding head health assessment algorithm provided by this invention introduces the deep forest algorithm into the field of welding head health assessment. Through multimodal feature fusion and hierarchical learning of deep forest, the accuracy of identifying the health status of the welding head reaches over 98%, and the remaining life prediction error is ≤5%. Compared with the traditional machine learning method of single random forest, the accuracy is significantly improved, effectively avoiding production losses caused by misjudgment. By providing early warning of welding head degradation trends, it avoids batch welding defects caused by welding head failure, greatly reduces the defect rate, and significantly improves product reliability. The evaluation method described in this invention enables real-time online evaluation. Its lightweight design keeps the model inference latency within 50ms, meeting the real-time monitoring requirements of the production line. Health evaluation can be completed without interrupting production, solving the lag problem of offline detection. This invention employs an incremental learning mechanism that enables the model to quickly adapt to different welding head models and workpiece materials without the need to re-label large amounts of data, significantly reducing deployment costs. This invention provides interpretable decision support, which can visualize the ranking of feature importance, help engineers understand the physical basis of evaluation results, guide the optimization of welding head maintenance strategies, and significantly reduce maintenance costs. Attached Figure Description

[0025] Figure 1 This is a flowchart of a method for health assessment algorithm of ultrasonic welding head provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.

[0027] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0028] Traditional welding head health assessment methods rely on manually designed features, such as peak amplitude and energy entropy, which struggle to capture the deep nonlinear characteristics of the welding head degradation process, resulting in low assessment accuracy. Furthermore, while vibration, temperature, and sound pressure levels exhibit strong correlations during welding, existing methods often employ single-signal analysis, neglecting the complementarity of cross-modal information. The distribution of health features varies significantly across different welding head models and workpiece materials, requiring existing models to be retrained with extensive labeled data, making them ill-suited for complex production environments. While deep learning methods can extract deep features, they suffer from model complexity, slow inference speeds, and their black-box nature makes it impossible to trace the physical meaning of assessment results, hindering their ability to meet the decision-making needs of industrial scenarios.

[0029] To address the aforementioned issues, this invention provides a deep forest-based method for assessing the health of ultrasonic welding heads. This method solves the problems of low assessment accuracy, weak generalization ability, and insufficient real-time performance in complex welding scenarios of existing technologies. It introduces the deep forest algorithm into the field of welding head health assessment: through a multi-level tree ensemble structure, it automatically learns high-order features from multi-source data. While maintaining the interpretability and efficiency of traditional machine learning models, it possesses feature representation capabilities close to deep learning, making it particularly suitable for processing complex nonlinear data in the welding process.

[0030] Example 1 Figure 1 This is a flowchart illustrating a health assessment method for an ultrasonic welding head provided in this embodiment. This flowchart merely shows the logical sequence of the method described in this embodiment; however, in other possible embodiments of the invention, different methods may be used, provided there are no conflicts. Figure 1 Complete the steps shown or described in the order indicated.

[0031] The health assessment method for ultrasonic welding heads provided in this embodiment can be applied to a terminal and can be executed by a health assessment device for ultrasonic welding heads. This device can be implemented in software and / or hardware and can be integrated into the terminal, such as any smartphone, tablet, or computer device with communication capabilities. See also... Figure 1 The health assessment method for ultrasonic welding heads described in this embodiment includes the following steps: Collect multi-source data signals from the ultrasonic welding equipment during the evaluation period; The collected multi-source data signals are preprocessed to obtain the feature vector to be evaluated; The feature vector to be evaluated is input into the trained deep forest model, and the evaluation results are output. The evaluation results include health status category, remaining life prediction value, and feature importance ranking obtained based on model analysis. Based on historical evaluation results within the sliding window, an adaptive threshold algorithm is used to dynamically adjust the alarm threshold. When the alarm threshold is reached, a maintenance prompt is triggered.

[0032] In this embodiment, a complete process from multi-source data acquisition to deep forest model evaluation is implemented to achieve real-time, online, and automated assessment of the welding head's health status, completely changing the traditional lagging mode that relies on offline detection. This method comprehensively utilizes a multimodal feature adaptive fusion mechanism and a hierarchical health status assessment framework, which can not only accurately identify the current state but also quantitatively predict the remaining lifespan, providing a direct basis for predictive maintenance. Furthermore, through an adaptive threshold algorithm based on a sliding window, discrete assessment results are transformed into robust and reliable decision logic. Early warnings are triggered by monitoring the trend of severe degradation states in multiple consecutive iterations, effectively filtering out random fluctuations in single assessments and significantly reducing the false alarm rate. This provides timely and reliable decision-making basis for predictive maintenance of the production line, achieving a leap from condition monitoring to intelligent early warning.

[0033] Example 2 This embodiment provides a health assessment method for ultrasonic welding heads. Based on Embodiment 1, the technical solution is further refined to improve the technical effect. The health assessment method for ultrasonic welding heads specifically includes the following steps: 1. Data Acquisition and Preprocessing 1.1 Multi-source data acquisition: Deploying a sensor array on the ultrasonic welding equipment, including: A triaxial accelerometer with a sampling rate of 10kHz is used to acquire longitudinal, transverse, and axial vibration signals of the welding head. An infrared temperature sensor with a sampling rate of 1kHz is used to measure the real-time temperature of the contact area between the welding head and the workpiece. A current sensor with a sampling rate of 1kHz is used to record the output current of the ultrasonic generator. The pressure sensor, with a sampling rate of 500Hz, is used to monitor the clamping force during the welding process. Here, clamping force refers to the pressure applied between the welding head and the workpiece by the ultrasonic welding equipment through the welding head (or welding torch) during the welding process.

[0034] The collected data is synchronized with timestamps to form the original dataset D = {X1, X2, ..., X}. n}, where X i The data sample for the i-th welding cycle consists of multi-source data including vibration time-domain waveform, temperature sequence, current value, and clamping force.

[0035] 1.2 Data Preprocessing: Noise Removal: Wavelet threshold denoising is used to process multi-source data signals, filtering out 50Hz power frequency interference and mechanical noise; Preliminary feature extraction: Time-domain features, frequency-domain features, and time-frequency features are extracted from the preprocessed signal to form an initial feature set F_raw = {f1, f2, ..., f m}, where m is the feature dimension; among which, time-domain features include peak value, root mean square value, and kurtosis; frequency-domain features include peak frequency and frequency band energy percentage; and time-frequency features include wavelet packet decomposition coefficients. Data standardization: Z-score standardization maps features to a zero-mean, unit-variance distribution, resulting in a standardized feature set F_std, as shown in the formula: f'_jk = (f_jk-μ_k) / σ_k; Where f_jk is the k-th feature of the j-th sample, and μ_k and σ_k are the mean and standard deviation of the feature, respectively.

[0036] 2. Deep Forest Model Construction and Training 2.1 Model Structure: This embodiment adopts a cascaded deep forest architecture, which includes the following modules: A. Input layer: The standardized feature set F_std is divided into vibration feature subset Fv, temperature feature subset Ft, and current feature subset Fi, which are then input into different random forest sub-modules. The random forest sub-module in the input layer is a pre-processing step of the cascaded forest layer. After extracting features by mode, the output probabilities are concatenated and used as the input of the cascaded forest layer. The cascaded layer then deepens the features layer by layer.

[0037] B. Cascaded Forest Layer: Each layer contains 4 random forests, including 2 random forests and 2 completely random tree forests. The probability distribution of each forest's output sample in the health status category is as follows: In this embodiment, the health status category is divided into four categories: healthy, mildly degraded, severely degraded, and faulty. In the model structure, the tree model of the deep forest cascaded forest layer naturally has the ability to calculate feature split gain, providing underlying support for global feature importance.

[0038] C. Feature Concatenation Layer: The output probability of each forest level is concatenated with the original features and used as the input for the next level, realizing hierarchical abstraction of features; D. Output layer: Contains a classifier (softmax regression) and a regressor (gradient boosting tree), which output the health status category and the remaining lifespan prediction value, respectively.

[0039] 2.2 Model Training: A. Dataset partitioning: The historical data is divided into a training set and a validation set in a 7:3 ratio. The training set contains health status labels and remaining life expectancy values ​​labeled by experts based on offline detection results. B. Loss Function: The loss function is a weighted sum of classification loss (cross-entropy) and regression loss (mean squared error), expressed as follows: L = αL_class + (1-α) L_reg; Where L is the loss function, α is the weight coefficient, L_class is the classification loss, and L_reg is the regression loss; C. Training strategy: An early stopping mechanism is adopted, that is, training is stopped when the validation set loss does not decrease for 5 consecutive rounds; after the training of each level of forest is completed, the parameters are fixed before training the next level to avoid overfitting. After training, the deep forest model performs global statistics based on the feature splitting gains of all decision trees in the model to generate a ranking of feature importance.

[0040] 3. Health Assessment Process 3.1 Real-time feature extraction: For newly acquired welding cycle data, features are extracted and standardized according to the method in step 1.2 to obtain the feature vector f_new to be evaluated; 3.2 Model Inference: Input the feature vector f_new to be evaluated into the trained deep forest model, and output: A. Health status categories are divided into four levels: healthy, mildly deteriorated, severely deteriorated, and malfunctioning; The health level is: the welding head has stable performance, no degradation, and meets the requirements of the rated welding parameters; The mild degradation level is as follows: the welding head shows slight wear, the resonant frequency shifts slightly, the welding parameters are normal, and the welding quality is not affected at present. The severe degradation level is: accelerated wear of the welding head / the appearance of microcracks, excessive deviation of the resonant frequency, and potential welding quality problems. The fault level is: excessive wear of the welding head / crack propagation / severe frequency deviation, which directly leads to defects such as poor welding and insufficient welding strength, making it impossible to use normally.

[0041] B. Remaining life forecast, including two quantitative indicators: (1) Remaining effective welding times, such as: 500 welding times remaining; (2) Remaining service time, such as remaining continuous service time: 72 hours; The two are linked and output based on welding conditions (single welding time, daily welding frequency) to adapt to different production scheduling needs.

[0042] C. Feature importance ranking, covering all dimensions of multi-source sensing features, clearly defining core contributing features and their proportions; specifically including: (1) Vibration-related, such as 35% energy contribution of vibration in the 20kHz frequency band, 22% root mean square contribution of axial vibration, and 18% kurtosis contribution of transverse vibration; (2) Temperature-related factors, such as the contribution of peak temperature in the welding head contact area (10%) and the contribution of temperature fluctuation rate during welding (6%). (3) Current and pressure factors, such as the contribution of ultrasonic generator output current stability to 5%, the contribution of average welding clamping force to 3%, and the contribution of peak current to 1%; Sort by contribution from highest to lowest, and simultaneously output the key impact features corresponding to the single-sample SHAP value heatmap.

[0043] In model inference, the importance ranking of output features is clearly defined, which is the core presentation of the visualization module. Combined with SHAP (SHapley Additive exPlanations) values, an explanatory report of the single-sample evaluation results is generated, corresponding to the additional output of this inference stage. In the visualization stage, bar charts (feature-contribution), heatmaps (sample-feature weights), and text lists (key features + percentages) can be used to intuitively present the core features affecting the health of the welding head and their respective percentages of influence.

[0044] 3.3 Dynamic adjustment of decision threshold: Based on the historical evaluation results within the sliding window, an adaptive threshold algorithm is used to adjust the alarm threshold. In this embodiment, a maintenance prompt is triggered when three consecutive evaluations show severe degradation.

[0045] 4. Model Adaptive Update 4.1 Incremental learning trigger condition: Incremental update is initiated when the KL divergence between the distribution of newly collected data and the distribution of the training set exceeds the threshold (0.1); 4.2 Update Strategy: A. Retain the 60% of trees that perform best in the cascaded forest model; B. Use the new data to train a new tree model and add it to the forest, keeping the number of trees at each level of the forest constant; C. Recalculate feature importance and optimize feature selection strategy.

[0046] Example 3 This embodiment provides a health assessment device for ultrasonic welding heads, used to implement the health assessment method described in Embodiment 1 or Embodiment 2. The device includes: Data acquisition module: used to collect multi-source data signals from the ultrasonic welding equipment during the evaluation period; Data processing module: used to preprocess the acquired multi-source data signals to obtain the feature vector to be evaluated; Health assessment module: This module takes the feature vector to be assessed as input into a trained deep forest model and outputs the assessment results, which include the health status category, the predicted remaining life expectancy, and the ranking of feature importance obtained based on model analysis. The early warning module is used to dynamically adjust the alarm threshold based on the historical evaluation results within the sliding window using an adaptive threshold algorithm. When the alarm threshold is reached, a maintenance prompt is triggered.

[0047] The health assessment device for ultrasonic welding heads provided in this embodiment can execute the health assessment method for ultrasonic welding heads provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the method.

[0048] Example 4 This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the health assessment method for the ultrasonic welding head described in Embodiment 1 or Embodiment 2.

[0049] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0050] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0051] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0052] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0053] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for health assessment of ultrasonic welding heads, characterized in that, include: Collect multi-source data signals from the ultrasonic welding equipment during the evaluation period; The collected multi-source data signals are preprocessed to obtain the feature vector to be evaluated; The feature vector to be evaluated is input into the trained deep forest model, and the evaluation results are output. The evaluation results include health status category, remaining life prediction value, and feature importance ranking obtained based on model analysis. Based on historical evaluation results within the sliding window, an adaptive threshold algorithm is used to dynamically adjust the alarm threshold, and a maintenance prompt is triggered when the alarm threshold is reached.

2. The health assessment method for ultrasonic welding heads according to claim 1, characterized in that, The multi-source data signals include at least two of the following: longitudinal, transverse, and axial vibration signals of the welding head; real-time temperature signals of the contact area between the welding head and the workpiece; output current signals of the ultrasonic generator; and clamping force signals of the welding process.

3. The health assessment method for ultrasonic welding heads according to claim 1, characterized in that, The preprocessing includes noise removal, preliminary feature extraction, and data standardization.

4. The health assessment method for ultrasonic welding heads according to claim 3, characterized in that, The noise removal includes performing wavelet threshold denoising on multi-source data signals to filter out power frequency interference and mechanical noise. The preliminary feature extraction includes extracting time-domain features, frequency-domain features, and time-frequency features from the noise-removed signal; The data standardization includes using Z-score standardization to map the extracted features to a zero-mean, unit-variance distribution.

5. The health assessment method for ultrasonic welding heads according to claim 1, characterized in that, The deep forest model includes an input layer, at least one cascaded forest layer, a feature splicing layer, and an output layer. The input layer will be used to input the feature vector to be evaluated after dividing it according to the signal mode. Each level of the cascaded forest layer contains multiple random forests, which are used to output the probability distribution of samples in the health status category; The feature concatenation layer is used to concatenate the output probability of the previous forest with the original features and use it as the input of the next forest, thereby realizing the hierarchical abstraction of features. The output layer includes a classifier and a regressor, which are used to output the health status category and the remaining life expectancy prediction value, respectively. After training, the deep forest model can perform global statistics based on the feature splitting gains of all decision trees in the model to generate the feature importance ranking.

6. The health assessment method for ultrasonic welding heads according to claim 5, characterized in that, The training process of the deep forest model includes: The historical data is divided into a training set and a validation set, and the historical data includes health status labels and remaining lifespan values ​​marked by experts based on offline detection results. The model is trained using the training set through a loss function, which is a weighted sum of classification loss and regression loss; the formula is: L = αL_class + (1-α) L_reg; Where L is the loss function, α is the weight coefficient, L_class is the classification loss, and L_reg is the regression loss; Training stops when the validation set loss does not decrease for a preset number of consecutive rounds.

7. The health assessment method for ultrasonic welding heads according to claim 5, characterized in that, The evaluation method further includes triggering incremental learning when the difference between the feature distribution of newly collected data and the feature distribution of the training set used for model training exceeds a preset threshold. The deep forest model is updated by retaining some tree models from the original forest and adding newly trained tree models.

8. The health assessment method for ultrasonic welding heads according to claim 1, characterized in that, The deep forest model is a lightweight model. It removes redundant decision trees or tree nodes by pruning the trained model and quantizes the feature splitting threshold in the model.

9. A health assessment device for an ultrasonic welding head, characterized in that, The device includes: Data acquisition module: used to collect multi-source data signals from the ultrasonic welding equipment during the evaluation period; Data processing module: used to preprocess the acquired multi-source data signals to obtain the feature vector to be evaluated; Health assessment module: This module takes the feature vector to be assessed as input into a trained deep forest model and outputs the assessment results, which include the health status category, the predicted remaining life expectancy, and the ranking of feature importance obtained based on model analysis. The early warning module is used to dynamically adjust the alarm threshold based on the historical evaluation results within the sliding window using an adaptive threshold algorithm. When the alarm threshold is reached, a maintenance prompt is triggered.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the health assessment method for ultrasonic welding heads according to any one of claims 1 to 8.

Images