Ultrasonic welding quality prediction method based on multi-sensor data decision fusion

Abstract

The present application relates to an ultrasonic welding quality prediction method based on multi-sensor data decision fusion, comprising the following steps: step one: signal acquisition, data acquisition; Step two: data preprocessing; Step three: feature extraction and resistance value prediction; Step four: sensor data fusion. According to the requirement of wire harness connection to ultrasonic welding quality, the welding quality is evaluated by predicting the resistance value of the welding sample; A SE-VMD method based on minimum sample entropy is proposed, which can decompose the complex amplitude data in the welding process, effectively filter out the amplitude component conforming to the ultrasonic working frequency, and reconstruct the data based on the component; The decision fusion model based on amplitude, power and pressure three kinds of sensor data and MobileNetV2 is determined, and the electrical characteristics of the welding sample are accurately predicted.

Description

Technical Field

[0001] This invention relates to a method for predicting ultrasonic welding quality based on multi-sensor data decision fusion, belonging to the field of ultrasonic welding technology. Background Technology

[0002] Ultrasonic welding is a special method that uses the mechanical vibration energy of ultrasonic frequencies to join similar or dissimilar materials. During ultrasonic welding, neither current nor high-temperature heat sources are applied to the workpiece; instead, under static pressure, the vibration energy of the wire frame is converted into frictional work, deformation energy, and a limited temperature rise between the workpieces. Ultrasonic welding is widely used in various industrial fields such as power battery manufacturing, wire harness connection, and automotive component welding.

[0003] Traditional ultrasonic welding quality inspection methods for lithium batteries are mostly offline destructive testing methods, which measure the joint shear strength of the sample to determine weld quality. Individual samples are discarded after inspection, making it impossible to inspect the entire batch. Furthermore, in cases of over-welding, the peeling peak deviation range is large, requiring secondary confirmation using electron microscopy. Existing offline destructive testing methods are no longer suitable for the needs of large-scale industrial production of electric vehicles due to problems such as sample damage, low sampling rate, low accuracy, poor real-time performance, and low efficiency. Moreover, current quality inspection methods only consider the mechanical properties of welded samples based on the joint strength, thus neglecting the crucial role of electrical conductivity.

[0004] Ultrasonic welding quality is affected by multiple factors, including welding pressure, ultrasonic amplitude, welding energy, and random external interference. The amount of welding status data is enormous, and abnormal data is far less than normal data, making online monitoring of welding quality difficult using traditional methods. In recent years, with breakthroughs in machine learning technology, especially deep learning, analyzing, mining, and effectively utilizing the evolution patterns of faults based on large amounts of status monitoring data has become a practical approach. Summary of the Invention

[0005] To overcome the shortcomings of existing research, this invention provides an end-to-end online quality monitoring method for ultrasonic metal welding based on multi-source sensor fusion and deep learning. Based on the requirements of wire harness connections for ultrasonic welding quality, it evaluates welding quality for the first time by predicting the resistance value of the welded sample. A variational mode decomposition (SE-VMD) method based on minimum sample entropy is proposed to decompose the complex amplitude data during welding, effectively filtering out amplitude components that conform to the ultrasonic operating frequency, and reconstructing the data based on these components. A decision fusion model based on amplitude, power, and pressure sensor data and MobileNet V2 is determined to accurately predict the electrical characteristics of the welded sample.

[0006] Regarding the sources of multi-source sensor data, the amplitude data of the ultrasonic welding head is measured by a Doppler laser vibrometer, and the changes in key parameters during the welding process, such as power, pressure, and depth, are obtained from the welding machine. The aim is to find the optimal and most efficient sensor combination and determine the fusion strategy with the highest accuracy among the three sensor data fusion methods: data fusion, feature fusion, and decision fusion.

[0007] The specific steps of the ultrasonic welding quality prediction method based on multi-sensor data decision fusion are as follows:

[0008] Step 1: Data Acquisition: Use a laser vibrometer to collect the amplitude data of the ultrasonic welding head in real time, denoted as H; export the changes of key parameters during the welding process through the welding machine, where W represents power, P represents pressure, and D represents depth; measure the resistance value of the welded sample through a loop current generator.

[0009] Step 2: Data preprocessing: To facilitate the training of the deep learning model later, the four types of data, H, W, P, and D, are filled with 0 values ​​to equal length and transformed into two-dimensional image data. Among them, H needs to be downsampled and SE-VMD processed first.

[0010] Step 3: Feature Extraction and Resistance Prediction: Extract the data obtained in Step 2 and put them into the Mobile netV2 model to extract features. Then, use the four extracted features to predict the resistance value through a fully connected layer network and compare the prediction results with those of different sensor data.

[0011] Step 4: Sensor Data Fusion: Verify the different combinations of H, W, P, and D data types, and evaluate the effectiveness of the three fusion strategies: data fusion, feature fusion, and decision fusion. Select the optimal data combination method and corresponding fusion strategy.

[0012] Step two specifically includes the following steps:

[0013] Step 2.1: This invention uses four key data points from the welding process: the ultrasonic welding head amplitude (H) measured by a laser sensor, and the pressure (P), depth (D), and power (W) variation curves derived from inside the welding machine. These three types of data are then normalized to distribute the data within the range [0,1].

[0014]

[0015] In the formula, any one of the three types of data X = {x1, x2, x3, ..., x...} n}, x i Let X represent the i-th sample data in a certain class of data, n represent the length of the class of data, max(X) represent the maximum value in the class of data, and min(X) represent the minimum value in the class of data;

[0016] Step 2.2: The laser sensor has a sampling frequency of 312.5 kHz, which will collect two million points in a welding process, but the effective data length is only about 300,000. Therefore, it is necessary to manually truncate this effective data length, and retain one point every 100 points for downsampling to reduce the dimensionality of the data. Then, fill the amplitude data and the three types of data in Step 2.1 with 0 values ​​to make them the same length.

[0017] Step 2.3: Decompose and reconstruct amplitude data based on SE-VMD. Due to the high complexity of amplitude data, involving multimodal aliasing and significant noise interference, directly feeding it into the model for training would negatively impact prediction performance. This invention uses Variational Mode Decomposition (VMD) to eliminate the impact of multimodal aliasing on prediction results. It is worth noting that the number of decompositions, K, significantly affects the decomposition results when using the VMD method. If K is too small, the decomposed subsequences will lose information or lead to mode aliasing. Conversely, when K is large, over-decomposition and wasted computational resources can occur, potentially affecting subsequent predictions. Therefore, this invention uses the minimum sample entropy (SE) criterion to determine the optimal K value. To evaluate the complexity of the sequence after VMD decomposition and obtain the optimal number of decomposition levels, sample entropy (SE) is introduced. When the amplitude subsequences are more complex, the calculated SE value will be larger, and vice versa. Accordingly, after applying VMD to decompose the amplitude, the SE value of each subsequence is calculated, and the sequence with the smallest SE represents the trend term of the amplitude data. When the number of decompositions K is small, it may lead to over-decomposition of the data, resulting in the loss of valid data; other distracting terms may also be mixed in with the trend term, causing the SE value to increase. When an appropriate value of K is chosen, the SE of the trend term will decrease. Subsequently, as the number of decompositions K increases, the SE gradually stabilizes. Therefore, the K value at the turning point where the SE of the trend term tends to stabilize is considered the optimal number of decompositions for VMD.

[0018] After determining K, the amplitude data is processed by VMD to obtain K modal components. Among the K modal components, the modal components within the ultrasonic operating frequency range (19KHz-20KHz) are selected, and the amplitude data is reconstructed by accumulation to remove noise interference.

[0019] Step 2.4: Since convolutional networks have excellent image feature extraction and representation capabilities, the amplitude data processed in Step 2.3 is converted into a 68×68×1 three-dimensional image. Because the power, pressure, and depth data have relatively low complexity, they are directly converted into 68×68×1 three-dimensional images.

[0020] Step three specifically includes the following steps:

[0021] Step 3.1: The four types of image data obtained in Step 2.4 are fed into Mobilenet V2 for further feature extraction. Mobilenet V2 has a 17-layer Bottleneck, each layer containing depthwise separable convolutions. The network is optimized using a linear bottleneck and inverse residual structure to achieve a deeper number of layers, a smaller model size, and faster execution speed. Each Bottleneck layer has a different number of convolutional kernels and stride to obtain feature vectors of different scales. Specifically, H is transformed into a 1000×1 feature vector after feature extraction, while W, P, and D are all transformed into 512×1 feature vectors.

[0022] Step 3.2: The four feature vectors obtained in Step 3.1 are fed into a three-layer fully connected network for resistance value prediction. The predicted resistance value is output by the last layer of the fully connected network. The Adam optimizer is used, and the loss is the root mean square error (RMSE).

[0023] Step four specifically includes the following steps:

[0024] To verify whether multi-source data fusion can yield more accurate results than single-data prediction, this invention combines the four sets of image data obtained in step 2.4, resulting in 11 possible combinations. The prediction performance of data fusion, feature fusion, and decision fusion strategies on these 11 data combinations is then verified.

[0025] Step 4.1: In the data fusion stage, this invention first stacks the data, resulting in a data dimension of 68×68×N (N being the number of sensor data points to be fused). The fused data is then fed into Mobile Net V2 to extract features, and finally passed through a fully connected layer network to output the predicted resistance value.

[0026] Step 4.2: In the feature fusion stage, each type of sensor data is fed into MobileNet V2. After feature extraction in the first 14 layers of the network, the dimension of each data type becomes 3×3×160, and then stacked into image features with a dimension of 3×3×(160×N) (N is the number of fused sensor data). These features are then fed into the last three layers of the MobileNet V2 network to obtain three sets of features. Finally, the features are merged and fed into the fully connected layer to output the predicted resistance value.

[0027] Step 4.3: In the decision fusion stage, the data from each type of sensor is fed into MobileNet V2. After feature extraction through a 17-layer network, the amplitude data becomes a 1000×1 one-dimensional vector, and the power, pressure, and depth data become 512×1 one-dimensional vectors. These are then concatenated to form a one-dimensional feature. Unlike feature fusion, decision fusion occurs after the 17-layer feature extraction network of MobileNetV2. The concatenated features are then fed into a fully connected layer to output the predicted resistance value.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0029] To avoid the inconvenience and drawbacks of evaluating welded samples through shear strength, this invention passes a 100A current through both ends of the welded sample and measures its resistance, using the resistance value as an evaluation index for weld quality. Since the Doppler laser vibrometer's acquisition frequency is around 312.5kHz, the acquired amplitude data is large and complex. Therefore, this invention proposes a variational mode decomposition (SE-VMD) method based on minimum sample entropy to process the amplitude data. To effectively extract the effective features from each sensor's data, the lightweight network MobileNet V2 is used as the feature extraction model, and the data, feature, and decision fusion strategies are validated based on this model. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a diagram illustrating the overall structure of the ultrasonic welding quality prediction method based on multi-sensor data decision fusion according to the present invention.

[0032] Figure 2 The curve showing the change of sample entropy of the trend term as a function of the K value after VMD decomposition is shown.

[0033] Figure 3 For VMD decomposition and reconstruction diagram;

[0034] Figure 4 This is a diagram of the internal structure of Mobile Net V2. Detailed Implementation

[0035] 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.

[0036] Figure 1 As shown, the ultrasonic welding quality prediction method based on multi-sensor data decision fusion includes four steps: 1) signal acquisition, 2) signal processing, 3) feature extraction, and 4) decision fusion. In the signal processing step, the raw signal is preprocessed and then fed into the MobileNetV2 model for feature extraction. During the decision-making process, features extracted from multiple sensors are combined to predict resistance.

[0037] Step 1: Signal Acquisition and Data Acquisition: The amplitude data of the ultrasonic welding head is acquired in real time using a laser vibrometer and denoted as H; the changes in key parameters during the welding process are exported through the welding machine, where W represents power, P represents pressure, and D represents depth; the resistance value of the welded sample is measured through a loop current generator.

[0038] Step 2: Signal and data preprocessing: To facilitate the training of the deep learning model later, the four types of data, H, W, P and D, are filled with 0 values ​​to equal length and converted into two-dimensional image data. Among them, H needs to be downsampled and SE-VMD processed first.

[0039] Step two specifically includes the following steps:

[0040] Step 2.1: Use four key data points from the welding process: the ultrasonic welding head amplitude (H) measured by the laser sensor, and the pressure (P), depth (D), and power (W) variation curves derived from inside the welding machine. Normalize these three types of data so that the data distribution is between [0,1].

[0041]

[0042] In the formula, any one of the three types of data X = {x1, x2, x3, ..., x...} n}, x i Let X represent the i-th sample data in a certain class of data, n represent the length of the class of data, max(X) represent the maximum value in the class of data, and min(X) represent the minimum value in the class of data;

[0043] Step 2.2: The laser sensor has a sampling frequency of 312.5 kHz, which will collect two million points in a welding process, but the effective data length is only about 300,000. Therefore, it is necessary to manually truncate this effective data length, and retain one point every 100 points for downsampling to reduce the dimensionality of the data. Then, fill the amplitude data and the three types of data in Step 2.1 with 0 values ​​to make them the same length.

[0044] Step 2.3: Decompose and reconstruct amplitude data based on SE-VMD. Due to the high complexity of amplitude data, involving multimodal aliasing and significant noise interference, directly feeding it into the model for training would negatively impact prediction performance. This invention uses Variational Mode Decomposition (VMD) to eliminate the impact of multimodal aliasing on prediction results. It is worth noting that the number of decompositions, K, significantly affects the decomposition results when using the VMD method. If K is too small, the decomposed subsequences will lose information or lead to mode aliasing. Conversely, when K is large, over-decomposition and wasted computational resources can occur, potentially affecting subsequent predictions. Therefore, this invention uses the minimum sample entropy (SE) criterion to determine the optimal K value. To evaluate the complexity of the sequence after VMD decomposition and obtain the optimal number of decomposition levels, sample entropy (SE) is introduced. When the amplitude subsequences are more complex, the calculated SE value will be larger, and vice versa. Accordingly, after applying VMD to decompose the amplitude, the SE value of each subsequence is calculated, and the sequence with the smallest SE represents the trend term of the amplitude data. When the number of decompositions K is small, it may lead to over-decomposition of the data, resulting in the loss of valid data; other distracting terms may also be mixed in with the trend term, causing the SE value to increase. When an appropriate value of K is chosen, the SE of the trend term will decrease. Subsequently, as the number of decompositions K increases, the SE gradually stabilizes. Therefore, the K value at the turning point where the SE of the trend term tends to stabilize is considered the optimal number of decompositions for VMD.

[0045] After determining K, the amplitude data is processed by VMD to obtain K modal components. Among the K modal components, the modal components within the ultrasonic operating frequency range (19KHz-20KHz) are selected, and the amplitude data is reconstructed by accumulation to remove noise interference.

[0046] Step 2.4: Since convolutional networks have excellent image feature extraction and representation capabilities, the amplitude data processed in Step 2.3 is converted into a 68×68×1 three-dimensional image. Because the power, pressure, and depth data have relatively low complexity, they are directly converted into 68×68×1 three-dimensional images.

[0047] The VMD algorithm is a completely non-recursive decomposition that can simultaneously decompose and extract modes. The model seeks multiple sets of modes and their respective center frequencies so that these modes can collectively reproduce the input signal. Essentially, VMD extends the classical Wiener filter to multiple adaptive bands and then obtains K narrowband mode components (IMFs), mainly consisting of two steps: constructing the variational problem and solving the variational problem.

[0048] First, perform a Hilbert transform on the amplitude signal to obtain v. k (t), calculate its one-sided spectrum, and compare it with the operator. Multiply, v k The center frequency of (t) is modulated to the corresponding baseband. Calculate the square norm L of the demodulation gradient. 2 The bandwidth of each modal component is estimated and calculated as follows:

[0049]

[0050]

[0051] In the above formula, {v k}={v1,···,v k} represents the decomposed IMF sub-modes, {ω k}={ω1,···,ω k} represents the center frequency of a submode, s is the original signal, δ(t) is the unit impulse function, and * represents the convolution operation.

[0052] The above equation is transformed from a constrained variational problem into an unconstrained variational problem by using the penalty factor α and the Lagrange multiplication operator λ. The corresponding equation is:

[0053]

[0054] In the formula, <> represents the dot product of two vectors. The unconstrained variational problem is then solved using the Alternating Direction Multipliers (ADMM) iterative algorithm to obtain λ and v. k ω k The iterative formula in the corresponding frequency domain.

[0055] Then input the initial λ and v k ω k The Fourier transform of the expression is iterated to obtain the following formula.

[0056]

[0057]

[0058]

[0059] Repeat the above formula until the following condition is met:

[0060]

[0061] In the formula, the superscripts (n) and (n+1) represent the number of iterations, ε represents the allowable error, and τ is the noise tolerance.

[0062] Given an amplitude sequence x(n) of length N, 1≤n≤N, based on the sliding window concept, the sequence is divided into a vector sequence X of length m. m (1),···,X m (N-m+1), for a given X m (i), Statistics X m (i) and X m The distance d[X] between (j) m (i),X m The number of j less than or equal to r, denoted as B. i For 1≤i≤Nm, we have:

[0063] d[X m (i),X m (j)]=

[0064] max k＝0,...m-1 (|x(i+k)-x(j+k)|)

[0065]

[0066] Therefore B (m) (r) is defined as:

[0067]

[0068] When the dimension increases to m+1, A is calculated using the same method described above. (m) (r):

[0069]

[0070] Therefore, the formula for sample entropy is as follows:

[0071]

[0072] The K value at the turning point where the trend term SE tends to stabilize is considered the optimal decomposition order of VMD. According to... Figure 2 It can be seen that when K=8, the sample entropy of the trend term tends to stabilize, indicating that the decomposition is thorough and there is no mode aliasing. To avoid losing effective data, modal components with frequencies in the range of (19KHz-20KHz) are selected, and the amplitude data is reconstructed by accumulation to remove noise interference. The results are as follows. Figure 3As shown, (a) is a timing diagram and (b) is a spectrum diagram.

[0073] Step 3: Feature Extraction and Resistance Prediction: Extract the data obtained in Step 2 and put them into the Mobile netV2 model to extract features. Then, use the four extracted features to predict the resistance value through a fully connected layer network and compare the prediction results with those of different sensor data.

[0074] Step three specifically includes the following steps:

[0075] Step 3.1: The four types of image data obtained in Step 2.4 are fed into Mobilenet V2 for further feature extraction. Mobilenet V2 has a 17-layer Bottleneck, each layer containing depthwise separable convolutions. The network is optimized using a linear bottleneck and inverse residual structure to achieve a deeper number of layers, a smaller model size, and faster execution speed. Each Bottleneck layer has a different number of convolutional kernels and stride to obtain feature vectors of different scales. Specifically, H is transformed into a 1000×1 feature vector after feature extraction, while W, P, and D are all transformed into 512×1 feature vectors.

[0076] Step 3.2: The four feature vectors obtained in Step 3.1 are fed into a three-layer fully connected network for resistance value prediction. The predicted resistance value is output by the last layer of the fully connected network. The Adam optimizer is used, and the loss is the root mean square error (RMSE).

[0077] Mobilenet V2 introduces depthwise separable convolutions, and the network has 17 Bottleneck layers. For example... Figure 4 MobileNet V2 optimizes the network using linear bottlenecks and inverse residual structures, resulting in deeper network layers, smaller model size, and faster execution speed. Depthwise separable convolution is a convolutional structure composed of a depthwise convolution layer and a pointwise convolution layer. Each convolution layer is followed by batch normalization and a ReLU6 activation function. Compared to standard convolution, it significantly reduces parameters and computational cost by approximately three times while maintaining roughly the same accuracy. The final pointwise convolution layer introduces a linear activation function, further reducing the number of training parameters.

[0078] like Figure 4 Compared to ordinary residual connections, the inverse residual structure's training process is "expansion-convolution-compression," which expands the number of data channels by six times, extracting more information. At the same time, the inverse residual structure retains features from shallow networks, facilitating the stacking of deeper models and improving feature extraction capabilities.

[0079] Key network parameters selected: width factor α = 1.0, resolution factor 68×68. The formula for the ReLU6 activation function is as follows:

[0080]

[0081] The subsequent fully connected network used for prediction consisted of three layers with 128, 64, and 1 neuron counts, respectively. The optimizer was Adam, and the loss was RMSE.

[0082] Step 4: Sensor Data Fusion: Verify the different combinations of H, W, P, and D data types, and evaluate the effectiveness of the three fusion strategies: data fusion, feature fusion, and decision fusion. Select the optimal data combination method and corresponding fusion strategy.

[0083] Step four specifically includes the following steps:

[0084] To verify whether multi-source data fusion can yield more accurate results than single-data prediction, this invention combines the four sets of image data obtained in step 2.4, resulting in 11 possible combinations. The prediction performance of data fusion, feature fusion, and decision fusion strategies on these 11 data combinations is then verified.

[0085] Step 4.1: In the data fusion stage, this invention first stacks the data, resulting in a data dimension of 68×68×N (N being the number of sensor data points to be fused). The fused data is then fed into Mobile Net V2 to extract features, and finally passed through a fully connected layer network to output the predicted resistance value.

[0086] Step 4.2: In the feature fusion stage, each type of sensor data is fed into MobileNet V2. After feature extraction in the first 14 layers of the network, the dimension of each data type becomes 3×3×160, and then stacked into image features with a dimension of 3×3×(160×N) (N is the number of fused sensor data). These features are then fed into the last three layers of the MobileNet V2 network to obtain three sets of features. Finally, the features are merged and fed into the fully connected layer to output the predicted resistance value.

[0087] Step 4.3: In the decision fusion stage, the data from each type of sensor is fed into MobileNet V2. After feature extraction through a 17-layer network, the amplitude data becomes a 1000×1 one-dimensional vector, and the power, pressure, and depth data become 512×1 one-dimensional vectors. These are then concatenated to form a one-dimensional feature. Unlike feature fusion, decision fusion occurs after the 17-layer feature extraction network of MobileNetV2. The concatenated features are then fed into a fully connected layer to output the predicted resistance value.

[0088] In the data fusion stage, this invention first stacks the data, resulting in a data dimension of 68×68×N (where N is the number of sensor data points to be fused). The fused data is then fed into Mobile Net V2 to output the predicted resistance value.

[0089] In the feature fusion stage, each type of sensor data is fed into MobileNet V2. After feature extraction in the first 14 layers of the network, the dimension of each data type becomes 3×3×160, and then stacked into image features with a dimension of 3×3×(160×N) (N is the number of fused sensor data), which are then fed into the last three layers of the MobileNet V2 network. Finally, the predicted resistance value is output through the fully connected layer.

[0090] In the decision fusion stage, data from each sensor type is fed into MobileNet V2. After feature extraction through a 17-layer network, amplitude data becomes a 1000×1 one-dimensional vector, while power, pressure, and depth data become 512×1 one-dimensional vectors. These are then concatenated to form a one-dimensional feature. Unlike feature fusion, decision fusion occurs after the 17-layer feature extraction network of MobileNet V2. The concatenated features are then fed into a fully connected layer to output the predicted resistance value.

[0091] As shown in Table 1, MobileNet V2 introduces inverse residual connections and depthwise separable convolutions, which increases the model depth, acquires more data features, and reduces the number of training parameters, preventing the network from overfitting. Therefore, MobileNet V2 can achieve better prediction results when using amplitude data for prediction.

[0092] Table 1 Comparison of data results from a single sensor

[0093]

[0094] Table 2 shows the prediction results of multi-source data fusion. In the data fusion stage, the fusion of amplitude, power, and pressure data yielded the best prediction results: the root mean square error (RMSE) was 14.10, the correlation coefficient was 0.643, and the fitting coefficient was 0.406. However, compared to the resistance value prediction results based on single sensor amplitude data, the RMSE increased by 0.55, the correlation coefficient decreased by 0.027, and the R² value... 2 The accuracy decreased by 0.044. This demonstrates that data fusion before inputting into the network model reduces prediction accuracy. This is because while data fusion retains all the features of the original data, it also retains interference and redundant terms. This simple pixel-level data fusion is not conducive to processing complex sensor signals in ultrasonic metal welding; the superposition of a large amount of data can actually reduce the prediction effect of resistance values.

[0095] Table 2 Comparison of Multi-Source Sensor Data Fusion Results

[0096]

[0097]

[0098] In the feature fusion stage, the fusion of amplitude and power data yielded the best prediction results: a root mean square error (RMSE) of 13.7, a correlation coefficient of 0.669, and a fitting coefficient of 0.439. This is similar to the best prediction results achieved with single-sensor data. This is because the feature fusion stage occurs at layer 14 of the MobileNet V2 network. In other words, each type of sensor data is processed through only 14 layers of the Bottleneck network before being stacked and then fed into the last three layers of the Bottleneck network to obtain the final fused features. While this fusion strategy, occurring in the middle layers of the network, reduces interference and redundancy in the original data, the feature extraction is insufficient because the data has not undergone complete network model processing, thus the prediction accuracy of the resistance value is not improved.

[0099] In contrast, decision fusion processes each type of sensor data through a complete 17-layer Mobile NetV2 network, preserving the richest features while minimizing interference and redundancy. Therefore, the decision fusion stage achieves better prediction results than single-sensor data fusion, multi-sensor data fusion, and feature fusion. The optimal results are: Root Mean Square Error (RMSE) 12.46, correlation coefficient 0.732, and fit coefficient 0.55. This applies to decision fusion of amplitude, power, and pressure data. Compared to resistance prediction based on single-sensor amplitude data, the RMSE is reduced by 1.09, the correlation coefficient is increased by 0.062, and the R² value is improved. 2 It increased by 0.1.

[0100] In summary, this explains why the combination of H, W, and P data types achieves optimal results. It's worth noting that during the fusion process, whether depth data is used as a supplement or to replace another data type, the prediction results will deteriorate. Conversely, the prediction performance is better with amplitude data present than without it.

[0101] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A method for predicting ultrasonic welding quality based on multi-sensor data decision fusion, characterized in that: Includes the following steps: Step 1: Signal Acquisition and Data Acquisition: The amplitude data of the ultrasonic welding head is acquired in real time using a laser vibrometer and denoted as H; the changes in key parameters during the welding process are exported through the welding machine, where W represents power, P represents pressure, and D represents depth; the resistance value of the welded sample is measured through a loop current generator. Step 2: Data preprocessing: Fill the four types of data (H, W, P, and D) with 0 values ​​to make them equal in length, and convert them into two-dimensional image data. H data needs to be downsampled and processed by SE-VMD first. Step 3: Feature extraction and resistance prediction: Extract the data obtained in Step 2 and put them into the Mobilenet V2 model to extract features. Then, use the four sets of extracted features to predict the resistance value through a fully connected layer network and compare the prediction results with those of different sensor data. Step 4: Sensor Data Fusion: Verify the different combinations of H, W, P, and D data types, assess the effectiveness of three fusion strategies—data fusion, feature fusion, and decision fusion—and select the optimal data combination method and corresponding fusion strategy. Step two specifically includes: Step 2.1: Use four key data points from the welding process: the ultrasonic welding head amplitude H measured by the laser sensor, the pressure P, depth D, and power W variation curves derived from inside the welding machine, and normalize these three types of data so that the data distribution is between [0,1]. In the formula, any one of the three types of data X = {x1, x2, x3, ..., x...} n }, x i Let X represent the i-th sample data in a certain class of data, n represent the length of the class of data, max(X) represent the maximum value in the class of data, and min(X) represent the minimum value in the class of data; Step 2.2: Truncate the effective data length, and retain 1 point every 100 points for downsampling to reduce the dimensionality of the data. Fill the amplitude data and the three types of data in Step 2.1 with 0 values ​​to make them the same length. Step 2.3: Decompose and reconstruct the amplitude data based on SE-VMD. Use Variational Mode Decomposition (VMD) to eliminate the impact of multimodal aliasing on the prediction effect. Use the minimum sample entropy (SE) criterion to determine the optimal K value, evaluate the complexity of the sequence after VMD decomposition, obtain the optimal number of decomposition layers K, introduce sample entropy SE, apply VMD to decompose the amplitude, calculate the SE value of each subsequence, the sequence with the smallest SE is the trend term of the amplitude data, and the K value of the turning point where the trend term SE tends to stabilize is regarded as the optimal number of decompositions of VMD. After determining K, the amplitude data is processed by VMD to obtain K modal components. Among the K modal components, the modal components within the ultrasonic working frequency range are selected, and the amplitude data is reconstructed by accumulation to remove noise interference. Step 2.4: Convert the amplitude data processed in Step 2.3 into a 68×68×1 three-dimensional image.

2. The ultrasonic welding quality prediction method based on multi-sensor data decision fusion according to claim 1, characterized in that: Step three specifically includes: Step 3.1: The four types of image data obtained in Step 2.4 are put into Mobilenet V2 to further extract data features. Mobilenet V2 has 17 layers of Bottleneck, each layer of Bottleneck contains depthwise separable convolutions, and the network is optimized through linear bottleneck and inverse residual structure. Each layer of Bottleneck has different numbers of convolution kernels and stride to obtain feature vectors of different scales. Step 3.2: The four sets of feature vectors obtained in Step 3.1 are put into a three-layer fully connected network to predict the resistance value. The last layer of the fully connected network outputs the predicted resistance value.

3. The ultrasonic welding quality prediction method based on multi-sensor data decision fusion according to claim 2, characterized in that: Step four specifically includes: The four sets of image data obtained in step 2.4 are combined to verify the prediction effects of the three fusion strategies of data fusion, feature fusion, and decision fusion on the above data combination. Step 4.1: In the data fusion stage, the data is first stacked, and the data dimension becomes 68×68×N, where N is the number of sensor data to be fused; the fused data is put into MobileNet V2 to extract features, and then outputs the predicted resistance value through a fully connected layer network. Step 4.2: In the feature fusion stage, each type of sensor data is put into MobileNet V2. After feature extraction in the first 14 layers of the network, the dimension of each type of data becomes 3×3×160. Then, they are stacked into image features with a dimension of 3×3×(160×N). They are put into the last three layers of the MobileNet V2 network to obtain three sets of features. Finally, the features are merged and put into the fully connected layer to output the predicted resistance value. Step 4.3: In the decision fusion stage, the data of each type of sensor is put into MobileNet V2. After feature extraction by 17 layers of network, the amplitude data dimension becomes a 1000×1 one-dimensional vector, and the power, pressure and depth data become 512×1 one-dimensional vectors. Then, they are spliced ​​together to form a one-dimensional feature. The spliced ​​feature is put into the fully connected layer to output the predicted resistance value.

Images