Piezoresistive sensor thick film detection method and detection device

Abstract

The application discloses a piezoresistive sensor thick film detection method and a detection device, and relates to the technical field of sensor manufacturing detection, which comprises the following steps: collecting response data at multiple measurement points on a thick film surface, and constructing a thick film surface state probability distribution map according to the response data; correcting isolated abnormal classification labels on the distribution map to generate an optimized state distribution map; calculating the area, the longest axis length and the shape factor of abnormal regions according to the connected regions of different state labels in the optimized map; comparing the geometric properties with preset criterion thresholds to determine the defect grade of each abnormal region; and finally, generating a structured thick film detection report. The application realizes systematic detection of thick film quality from probability distribution to morphological quantification and grading evaluation.

Description

Technical Field

[0001] This invention belongs to the field of sensor manufacturing and testing technology, specifically a method and equipment for thick film testing of piezoresistive sensors. Background Technology

[0002] In the manufacturing process of piezoresistive sensors, the quality of the thick-film resistive layer is crucial in determining the sensor's performance, reliability, and consistency. Microscopic defects on the thick-film surface, such as cracks, pores, material inhomogeneity, or poor adhesion, directly affect the sensor's resistive characteristics, long-term stability, and pressure response accuracy. Therefore, accurate detection and quantitative evaluation of the thick-film surface condition during the production process are essential.

[0003] Currently, the inspection of thick film quality mainly relies on two types of technologies. One type is visual inspection methods based on optical or electron microscopy, which identify obvious physical defects by acquiring surface morphology images. These methods are highly dependent on operator experience, have low inspection efficiency, and are difficult to detect subsurface or microscopic electrical performance anomalies, let alone quantify the potential impact of defects on sensor functionality. The other type is screening methods based on electrical performance testing, such as measuring the resistance or temperature coefficient of the thick film. While this method can reflect electrical performance, it only provides the overall performance parameters of the end product and cannot pinpoint the specific location, distribution range, and morphological characteristics of defects on the thick film surface. When there are localized minor defects in the thick film, the overall electrical parameters may still be within the acceptable range, leading to missed defects and posing a potential threat to the long-term reliability of the product. In addition, random noise interference may be introduced during the measurement process, generating discrete outliers in single-point measurement data. Simply classifying these as defects can lead to misjudgments and affect product yield.

[0004] Current technologies lack an analytical method that can combine discrete point measurement data with the spatial morphology of thick film surfaces. The problem lies in how to construct a picture reflecting the probability distribution of the overall surface state of the thick film from limited point measurement data, rather than simply judging whether it is qualified or not. Simultaneously, in data processing, effectively distinguishing between anomalies caused by real physical defects and isolated anomalies caused by measurement noise, and avoiding the latter interfering with the final defect identification and assessment, is another key to improving the accuracy and reliability of detection. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a method and equipment for detecting thick films in piezoresistive sensors, including: Multiple measurement points are preset on the thick film surface of the piezoresistive sensor under test. Based on the response data of the measurement points, a probability distribution map of the state of the thick film surface is constructed. On the probability distribution map of the state on the thick film surface, a spatial continuity correction algorithm is applied to correct the isolated anomaly classification labels so that the corrected state labels satisfy the continuity constraint in spatial distribution, thereby generating an optimized state distribution map. Based on the connected regions of different state labels in the optimized state distribution map, calculate the geometric properties of each abnormal region, including the area, longest axis length, and shape factor of the abnormal region; The geometric properties of the abnormal regions are compared with preset criterion thresholds to determine the defect level of each abnormal region; By combining the final status labels of all measurement points, the defect level and location information of all abnormal areas, a structured thick film inspection report is generated.

[0006] Furthermore, the method of constructing a probability distribution map of the thick film surface state based on the response data of multiple preset measurement points on the thick film surface of the piezoresistive sensor under test includes: Multiple measurement points are preset on the thick film surface of the piezoresistive sensor under test. A standard excitation electrical signal is applied through an electrical signal acquisition device, and the initial voltage response of each measurement point is recorded synchronously to form a set of initial response signals. The initial response signal set is preprocessed to eliminate environmental noise and baseline drift, generating a denoised response signal set. Multiple signal features of each measurement point are extracted from the set of denoised response signals to form a multidimensional signal feature vector, which includes amplitude features, phase features and frequency domain energy distribution features. The multidimensional signal feature vector of each measurement point is input into the fully trained thick film state classification model to obtain the preliminary state classification label of each measurement point. The preliminary state classification label includes normal, crack, void and uneven thickness. Based on the preliminary state classification labels of all measurement points and their spatial coordinates on the thick film surface, a probability distribution map of the state of the thick film surface is constructed. The initial response signal set is preprocessed to remove environmental noise and baseline drift, generating a denoised response signal set, including: For each initial voltage response signal in the set of initial response signals, a zero-phase digital filter is applied for bandpass filtering to retain the characteristic frequency band signals related to the piezoresistive material response; Adaptive baseline correction is performed on the bandpass filtered signal to eliminate signal baseline drift components caused by inherent sensor characteristics or slow environmental changes. The baseline-corrected signal is processed using wavelet threshold denoising technology to suppress high-frequency random noise in the signal, thereby obtaining a single denoised response signal in the set of denoised response signals.

[0007] Furthermore, multiple signal features for each measurement point are extracted from the denoised response signal set to form a multi-dimensional signal feature vector, including: Time-domain analysis is performed on a single denoised response signal in the set of denoised response signals. The average amplitude of the signal within the standard excitation period is calculated as the amplitude feature, and the time difference between the zero crossing point of the signal and the excitation signal is calculated as the phase feature. A fast Fourier transform is performed on the single denoised response signal to obtain the spectral distribution of the signal, and the spectral energy integral within a preset characteristic frequency band is calculated as the frequency domain energy distribution feature. The amplitude feature, the phase feature, and the frequency domain energy distribution feature are combined in a fixed order to form the multidimensional signal feature vector corresponding to the response characteristics of the measurement point; The step of inputting the multidimensional signal feature vector of each measurement point into the fully trained thick film state classification model includes: The fully trained thick-film state classification model is based on a feedforward neural network as its core architecture, and includes an input layer, multiple hidden layers, and an output layer. The number of nodes in the input layer is the same as the dimension of the multidimensional signal feature vector. The multidimensional signal feature vector is input from the input layer, undergoes layer-by-layer nonlinear transformation and feature abstraction in the multiple hidden layers, and finally generates a probability distribution corresponding to different thick film states in the output layer. The state category with the highest probability value in the output layer is selected as the preliminary state classification label for the measurement point.

[0008] Furthermore, based on the preliminary state classification labels of all measurement points and their spatial coordinates on the thick film surface, a probability distribution map of the thick film surface state is constructed, including: Using the thick film surface as a two-dimensional plane, the spatial coordinates of each measurement point are mapped to the grid point positions in the probability distribution map of the thick film surface state. The preliminary state classification label of each measurement point is converted into a state probability vector, in which the probability of the corresponding state category is a fixed high value, and the probabilities of the other categories are fixed low values; Each grid point is assigned the state probability vector corresponding to the measurement point. For grid points without measurement points, their state probability vectors are generated by bilinear interpolation of the state probability vectors of neighboring grid points, thus completing the construction of the state probability distribution map of the thick film surface.

[0009] Furthermore, on the probability distribution map of the thick film surface state, a spatial continuity correction algorithm is applied to correct isolated anomaly classification labels, including: On the probability distribution map of the surface state of the thick film, scan all grid points marked as abnormal states; For each abnormal state grid point, count the number of similar abnormal state grid points within its preset neighborhood range; If the number of grid points of the same type of abnormal state is lower than the connectivity threshold, the abnormal state grid point is determined to be an isolated abnormal point, and the state probability vector of the abnormal state grid point is corrected to the probability vector corresponding to the dominant state category in its neighborhood. After traversing all abnormal state grid points and completing the correction, the state probability vectors of all grid points in the map are smoothed and filtered again to generate the optimized state distribution map.

[0010] Furthermore, based on the connected regions of different state labels in the optimized state distribution map, the geometric properties of each abnormal region are calculated, including: The optimized state distribution map is binarized, and all grid points belonging to the same abnormal state are marked as foreground, while the remaining grid points are marked as background. The connected component labeling algorithm is used to identify all interconnected pixel blocks in the foreground, with each pixel block corresponding to an independent anomalous region. Calculate the total number of grid points contained in each independent anomaly region, and combine this with the actual physical area of ​​a single grid point to calculate the area of ​​the anomaly region; Extract the boundary of each independent anomaly region, and calculate the Euclidean distance between the two farthest points on the boundary as the longest axis length of the anomaly region; Based on the area and perimeter of the abnormal region, its shape factor is calculated, which is used to quantify the degree to which the shape of the region deviates from a circle.

[0011] Furthermore, the geometric properties of the abnormal regions are compared with preset criterion thresholds to determine the defect level of each abnormal region, including: Set area-level thresholds for abnormal regions, length-level thresholds for the longest axis length, and shape-level thresholds for the shape factor; The area of ​​the abnormal region is compared with the area level threshold to determine the level score in the area dimension; The longest axis length of the abnormal region is compared with the length level threshold to determine the level score in the length dimension; The shape factor of the abnormal region is compared with the shape level threshold to determine the level score in the shape dimension; The comprehensive defect score of the abnormal region is calculated by combining the grade scores in the area dimension, the length dimension, and the shape dimension using a weighted summation method. The final defect level is determined based on the interval to which the comprehensive defect score belongs.

[0012] Furthermore, by integrating the final status labels of all measurement points, the defect levels and location information of all abnormal areas, a structured thick film inspection report is generated, including: Summarize the spatial coordinates of all measurement points and their corresponding final state labels on the optimized state distribution map to form a measurement point state list; Summarize the geometric properties of all abnormal regions, the calculated defect level, and their positional contour coordinates on the thick film surface to form a detailed list of abnormal regions; The list of measurement point statuses and the list of abnormal area details, along with the identification information of the piezoresistive sensor under test and the detection time information, are formatted and organized according to a predefined report template to generate the thick film detection report containing plain text and structured data.

[0013] Furthermore, the training process of the fully trained thick-film state classification model includes: Acquire labeled thick film sample data accumulated in historical detection, wherein the labeled thick film sample data includes multidimensional signal feature vectors collected from thick films in known states and their true state labels; The labeled thick film sample data is divided into a training sample subset and a validation sample subset; A feedforward neural network is initialized as a classification model to be trained. The feedforward neural network includes an input layer, multiple hidden layers and an output layer, and the number of nodes in the input layer is the same as the dimension of the multidimensional signal feature vector. Using the aforementioned training sample subset, the cross-entropy between the output of the feedforward neural network and the true state label is used as the loss function, and the connection weights and bias parameters of the feedforward neural network are iteratively updated through the backpropagation algorithm. After each iteration, the classification accuracy of the current classification model is evaluated using the validation sample subset. The iteration stops when the classification accuracy no longer improves for several consecutive iterations. The feedforward neural network with fixed connection weights and bias parameters obtained at this time is used as the fully trained thick film state classification model.

[0014] Furthermore, the present invention also includes a thick film detection device for a piezoresistive sensor, the device including a detection body (1) and a tray arrangement machine (2), the detection body (1) having a workpiece (3) to be detected directly below it, the detection body (1) being used to detect the workpiece (3) and distinguish defective products, the tray arrangement machine (2) being used to arrange the workpiece (3) into sections, characterized in that a conveying mechanism (11) is installed on the detection body (1), the conveying mechanism (11) being used to convey the workpiece (3); A spacing mechanism (12) is provided at the top of the middle section of the first conveying mechanism (11), which is used to divide the workpiece (3) by spacing; a flipping mechanism (13) is provided at the end of the conveying stroke of the first conveying mechanism (11), which is used to flip the workpiece (3). The first conveying mechanism (11) is provided with a second conveying mechanism (14) at one end near the flipping mechanism (13). The second conveying mechanism (14) and the first conveying mechanism (11) are arranged perpendicularly to each other. The second conveying mechanism (14) is used to transfer the workpiece (3) flipped by the flipping mechanism (13). A visual inspection mechanism (15) is provided at the top of the middle section of the second conveying mechanism (14). The visual inspection mechanism (15) is used to visually inspect whether the workpiece (3) is qualified or not. A storage frame (17) is provided on the first conveying mechanism (11). The storage frame (17) is used to collect and store unqualified workpieces (3). The second conveying mechanism (14) is connected to the tray-laying machine (2) from the side away from the flipping mechanism (13). The second conveying mechanism (14) is provided with a diversion mechanism (16). The diversion mechanism (16) is used to divert the qualified workpiece (3) into a channel, so that the workpiece (3) is diverted into the tray-laying machine (2) by the second conveying mechanism (14) into two channels. The tray arrangement machine (2) is equipped with a dual-camera positioning mechanism (21), which is located at the end of the conveying stroke of the second conveying mechanism (14). The dual-camera positioning mechanism (21) assists the tray arrangement machine (2) in performing partitioned tray arrangement through visual detection.

[0015] 11. The piezoresistive sensor thick film detection device according to claim 1, characterized in that the flipping mechanism (13) flips the workpiece (3) from the first conveying mechanism (11) to the second conveying mechanism (14), so that the surface of the thick film to be detected of the workpiece (3) faces the visual inspection mechanism (15); the visual inspection mechanism (15) includes a processor, the processor being used to perform the following steps: Multiple measurement points are preset on the thick film surface of the piezoresistive sensor under test. Based on the response data of the measurement points, a probability distribution map of the state of the thick film surface is constructed. On the probability distribution map of the state on the thick film surface, a spatial continuity correction algorithm is applied to correct the isolated anomaly classification labels so that the corrected state labels satisfy the continuity constraint in spatial distribution, thereby generating an optimized state distribution map. Based on the connected regions of different state labels in the optimized state distribution map, calculate the geometric properties of each abnormal region, including the area, longest axis length, and shape factor of the abnormal region; The geometric properties of the abnormal regions are compared with preset criterion thresholds to determine the defect level of each abnormal region; By combining the final status labels of all measurement points, the defect levels of all abnormal areas, and their location information, a structured thick film inspection report is generated.

[0016] 12. The piezoresistive sensor thick film inspection device according to claim 2, wherein the storage frame (17) is disposed on the conveying mechanism (11) and located downstream of the vision inspection mechanism (15), and is used to collect and store the workpieces (3) that are judged to be unqualified according to the thick film inspection report generated by the vision inspection mechanism (15). According to the thick film inspection report, the diversion mechanism (16) diverts the qualified workpieces (3) into dual channels and transports them to the tray arrangement machine (2). The dual camera positioning mechanism (21) positions the qualified workpieces (3) entering through the dual channels and assists the tray arrangement machine (2) in arranging the qualified workpieces (3) into trays according to the inspection level.

[0017] Compared with the prior art, the beneficial effects of the present invention are: Based on response data from multiple discrete measurement points, a probability distribution map of the thick film surface state is constructed, expanding the detection perspective from traditional single-point acceptance judgment to statistical state assessment of the entire surface. This technology transforms discrete measurement data into a continuous probability field, intuitively revealing the macroscopic distribution patterns and spatial variation trends of the thick film surface state. Characterizing the state through a probability distribution accommodates the uncertainty of the measurement data, enabling more precise assessment of areas in critical states. Compared to relying solely on single-point threshold judgments, this method can grasp the overall quality status of the thick film, identifying potential, spatially correlated weak anomalies or gradual defects, resulting in more comprehensive and forward-looking detection results.

[0018] A spatial continuity correction algorithm is applied to correct isolated anomaly classification labels in the initially generated state probability distribution map. This technique leverages prior knowledge that real physical defects typically possess a certain degree of spatial continuity. The algorithm identifies and analyzes the spatial clustering of anomaly labels, correcting isolated "anomaly" labels—those lacking association with surrounding normal areas and more likely caused by random measurement noise or transient interference—to a state consistent with the surrounding areas. This process effectively filters data noise, resulting in a final optimized state distribution map that more accurately reflects the physical condition of the thick film surface. This avoids misjudgments caused by sporadic noise, improves the accuracy of identifying real defect areas, makes detection conclusions more stable and reliable, and reduces false product rejections or missed defects due to detection fluctuations. Attached Figure Description

[0019] Figure 1This is a state diagram of the piezoresistive sensor thick film detection device and detection method described in this invention; Figure 2 A fishbone diagram of multiple measurement points preset on the thick film surface of the piezoresistive sensor under test; Figure 3 This is a schematic diagram showing the positions of the testing machine and the tray arrangement machine of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] See Figure 1 This invention provides a thick-film inspection device and method for piezoresistive sensors. The specific implementation is as follows: Multiple measurement points are pre-set on the thick-film surface of the piezoresistive sensor under test. A probability distribution map of the thick-film surface state is constructed based on the response data of these measurement points. On the probability distribution map of the thick-film surface state, a spatial continuity correction algorithm is applied to correct isolated anomaly classification labels in the map, ensuring that the corrected state labels satisfy continuity constraints in spatial distribution, thereby generating an optimized state distribution map. Based on the connected regions of different state labels in the optimized state distribution map, the geometric attributes of each anomaly region are calculated. These geometric attributes include the area, longest axis length, and shape factor of the anomaly region. The geometric attributes of each anomaly region are compared with a preset criterion threshold to determine the defect level of each anomaly region. A structured thick-film inspection report is generated by combining the final state labels of all measurement points, the defect levels of all anomaly regions, and their location information on the thick-film surface.

[0022] In one embodiment of the present invention, see [reference] Figure 2Multiple measurement points are preset on the thick film surface of the piezoresistive sensor under test. A standard excitation electrical signal is applied to these measurement points through an electrical signal acquisition device, and the initial voltage response of each measurement point is recorded simultaneously to form an initial response signal set. The initial response signal set is preprocessed to eliminate environmental noise and baseline drift, generating a denoised response signal set. Specifically, each initial voltage response signal in the initial response signal set is bandpass filtered using a zero-phase digital filter to retain characteristic frequency band signals related to the piezoresistive material response. Adaptive baseline correction is performed on the bandpass-filtered signal to eliminate signal baseline drift components caused by inherent sensor characteristics or slow environmental changes. Wavelet threshold denoising technology is used to process the baseline-corrected signal to suppress high-frequency random noise in the signal, obtaining a single denoised response signal in the denoised response signal set. Multiple signal features of each measurement point are extracted from the denoised response signal set to form a multidimensional signal feature vector, which includes amplitude features, phase features, and frequency domain energy distribution features. The multidimensional signal feature vector of each measurement point is input into a fully trained thick film state classification model to obtain a preliminary state classification label for each measurement point. The preliminary state classification labels include normal, cracked, voided, and uneven thickness. Based on the preliminary state classification labels of all measurement points and their spatial coordinates on the thick film surface, a probability distribution map of the thick film surface state is constructed.

[0023] In practice, multiple measurement points are pre-planned and marked on the thick-film surface of the piezoresistive sensor under test. These measurement points uniformly cover the effective area of ​​the thick film in space. A probe of a multi-channel electrical signal acquisition device contacts these measurement points sequentially or simultaneously, applying a calibrated standard excitation electrical signal of frequency and amplitude to each measurement point, and simultaneously recording the initial voltage response waveform of that measurement point within the excitation period. The initial voltage response waveforms recorded at all measurement points are aggregated to form an initial response signal set. In practice, each initial voltage response signal in the initial response signal set needs to be pre-processed to improve the signal-to-noise ratio. A zero-phase digital filter is applied to each initial voltage response signal for bandpass filtering. The passband range of this filter is set to match the characteristic response frequency of the piezoresistive material, aiming to retain the characteristic frequency band signal related to the piezoresistive material response while suppressing interference outside the passband. Adaptive baseline correction is performed on the bandpass-filtered signal. In some embodiments, this is achieved by fitting a trend line of the signal during inactive time periods and subtracting it from the original signal, thereby eliminating signal baseline drift components caused by inherent sensor characteristics or slow environmental changes. By using wavelet thresholding denoising technology to process the baseline-corrected signal, it can be understood that this step, by selecting an appropriate wavelet basis and threshold function, can effectively suppress high-frequency random noise in the signal, and finally obtain a single denoised response signal in the set of denoised response signals.

[0024] The signal features of each measurement point are extracted from the set of denoised response signals. In specific implementation, time-domain analysis is performed on a single denoised response signal to calculate the average voltage value of the signal over a complete standard excitation cycle. This average value is used as the amplitude feature of the response signal at that measurement point. The time difference between the zero-crossing moments of the denoised response signal and the standard excitation signal is calculated and used as the phase feature. A Fast Fourier Transform is performed on the same denoised response signal to obtain the signal's spectral distribution. The spectral amplitude is integrated within a preset characteristic frequency band, and the integration result is used as the frequency domain energy distribution feature. It can be understood that combining the calculated amplitude feature, phase feature, and frequency domain energy distribution feature in a fixed order, such as [amplitude, phase, frequency domain energy], constitutes a three-dimensional multidimensional signal feature vector. This vector characterizes the response characteristics of the corresponding measurement point.

[0025] The multidimensional signal feature vector of each measurement point is input into a fully trained thick film state classification model. In some embodiments, the core of the fully trained thick film state classification model is a feedforward neural network with three neurons in its input layer, corresponding to the three dimensions of the multidimensional signal feature vector. The multidimensional signal feature vector is input from the input layer and processed by nonlinear activation functions and weight calculations in multiple hidden layers to achieve layer-by-layer abstraction and transformation of the features. Finally, four output values ​​are generated in the output layer, which correspond to four thick film states: "normal," "crack," "void," and "non-uniform thickness." The output values ​​of the output layer are normalized by the Softmax function and represent the probability of belonging to each state category. The state category with the highest probability value is selected as the preliminary state classification label for this measurement point. For example, if the output probability value corresponding to "crack" is the highest, then the measurement point is labeled "crack." Based on the preliminary state classification labels of all measurement points and their pre-recorded two-dimensional spatial coordinates on the thick film surface, a probability distribution map of the thick film surface state is constructed. In specific implementations, the thick film surface is regarded as a two-dimensional planar coordinate system, and the coordinates of each measurement point correspond to a location point in this coordinate system. Each location point is assigned a state probability vector, generated based on its initial state classification label. For example, a measurement point classified as "crack" might have a state probability vector of [0.1, 0.8, 0.05, 0.05], where the second element (corresponding to "crack") has a significantly higher value than the others. For locations in the coordinate system without measurement points, their state probability vectors are generated by bilinear interpolation of the state probability vectors of their four nearest neighbors. This results in a thick film surface state probability distribution map covering the entire surface, with each grid point having its own state probability vector. Optionally, the mathematical process of adaptive baseline correction can be expressed by the following formula: S b (t)=Sf (t)-(m•t+c)

[0026] Wherein: S b (t) represents the signal amplitude at time point t after adaptive baseline correction, S f (t) represents the signal amplitude after bandpass filtering at time point t, and m and c are the slope and intercept parameters obtained by linear regression of the signal data in a preset inactive time period.

[0027] In one embodiment of the present invention, the fully trained thick-film state classification model uses a feedforward neural network as its core architecture. This model includes an input layer, multiple hidden layers, and an output layer. The number of nodes in the input layer is the same as the dimension of the multidimensional signal feature vector. The multidimensional signal feature vector is input from the input layer and undergoes layer-by-layer nonlinear transformation and feature abstraction through multiple hidden layers, ultimately generating a probability distribution corresponding to different thick-film states in the output layer. The state category with the highest probability value in the output layer is selected as the initial state classification label for that measurement point. The training process of this fully trained thick-film state classification model is implemented through the following steps: acquiring labeled thick-film sample data accumulated from historical detections. The labeled thick-film sample data includes multidimensional signal feature vectors collected from known-state thick films and their true state labels. Dividing the labeled thick-film sample data into a training sample subset and a validation sample subset. Initializing a feedforward neural network as the classification model to be trained. This feedforward neural network includes an input layer, multiple hidden layers, and an output layer, and the number of nodes in the input layer is the same as the dimension of the multidimensional signal feature vector. Using a subset of training samples, the cross-entropy between the output of the feedforward neural network and the true state labels is used as the loss function. The connection weights and bias parameters of the feedforward neural network are iteratively updated using the backpropagation algorithm. After each iteration, the classification accuracy of the current model is evaluated using a subset of validation samples. Iteration stops when the classification accuracy no longer improves after several consecutive iterations. The feedforward neural network with fixed connection weights and bias parameters obtained at this point is used as the fully trained thick-film state classification model.

[0028] In practice, the fully trained thick-film state classification model uses a feedforward neural network as its core architecture. The feedforward neural network comprises an input layer, three fully connected hidden layers, and an output layer. The number of nodes in the input layer strictly matches the dimension of the multidimensional signal feature vector extracted from the measurement point signal. In some embodiments, the multidimensional signal feature vector includes amplitude features, phase features, and frequency domain energy distribution features; therefore, the input layer is designed with three neurons. When the fully trained thick-film state classification model is working, the multidimensional signal feature vector of a measurement point is input from the input layer. After weighted summation and nonlinear activation function transformation in the first hidden layer, the first layer of feature abstraction is generated. The output of the first layer of feature abstraction serves as the input to the second hidden layer. It can be understood that through this layer-by-layer nonlinear transformation and feature abstraction across multiple hidden layers, the original multidimensional signal feature vector is transformed into a higher-level feature representation. High-level feature representations are processed in the output layer, which has four neurons corresponding to four thick film state categories: "normal," "cracked," "void," and "uneven thickness." The output layer generates a four-dimensional probability distribution vector through a Softmax activation function, where each element represents the probability that the input multidimensional signal feature vector belongs to the corresponding state category. The state category with the highest probability value in the probability distribution vector generated by the output layer is selected as the initial state classification label for the measurement point corresponding to this input multidimensional signal feature vector.

[0029] The training process for a fully trained thick-film state classification model needs to be completed beforehand. In practice, the starting point of the training process is to acquire labeled thick-film sample data accumulated from historical detections. The labeled thick-film sample data consists of two parts: one part is the multi-dimensional signal feature vector collected and extracted from a large number of thick-film samples with known real states; the other part is the real state labels of these thick-film samples confirmed by authoritative methods such as metallographic analysis. In practice, all collected labeled thick-film sample data is randomly shuffled and then divided into training sample subsets and validation sample subsets according to a preset ratio, such as a 7:3 ratio. A feedforward neural network classification model with the same structure as described above is initialized. The connection weights and bias parameters of the feedforward neural network are usually initialized to small random numbers conforming to a certain statistical distribution. The feedforward neural network classification model is trained using the training sample subset. In some embodiments, the training process uses the cross-entropy between the predicted probability distribution output by the feedforward neural network classification model and the real state labels of the samples as the loss function, and the gradient of the loss function with respect to the connection weights and bias parameters of each layer of the network is calculated using the error backpropagation algorithm. Using optimization algorithms, such as stochastic gradient descent, the connection weights and bias parameters in the feedforward neural network classification model are iteratively updated based on the calculated gradients. Each iteration updating the entire subset of training samples is considered a completed training cycle. After each training cycle, the feedforward neural network classification model under the current parameter state is evaluated using a validation subset, and its classification accuracy is calculated. The training process continues. Optionally, the cross-entropy loss function H can be calculated as follows:

[0030] Where: H represents the average cross-entropy loss calculated for a batch of samples, N represents the total number of samples in the current batch, C represents the total number of state categories (4 here), and y k,c It is an indicator function whose value is 1 when the true class of sample k is c, and 0 otherwise. k,c This refers to the probability that a sample k belongs to class c, predicted by the feedforward neural network classification model. When the classification accuracy of the feedforward neural network classification model on the validation sample subset no longer shows an increasing trend over several consecutive training epochs, the training process stops iterating. It can be understood that at this point, the connection weights and bias parameters of the feedforward neural network classification model are fixed. This feedforward neural network classification model with fixed parameters is considered a fully trained thick-film state classification model, used for subsequent automatic classification of new measurement point data.

[0031] In one embodiment of the present invention, the thick film surface is used as a two-dimensional plane, and the spatial coordinates of each measurement point are mapped to grid point positions in the thick film surface state probability distribution map. The preliminary state classification label of each measurement point is converted into a state probability vector, where the probability of the corresponding state category is a fixed high value, and the probabilities of other categories are fixed low values. Each grid point position is assigned to the state probability vector corresponding to the measurement point. For grid positions without measurement points, their state probability vectors are generated by bilinear interpolation of the state probability vectors of neighboring grid points, thereby completing the construction of the thick film surface state probability distribution map.

[0032] In practical implementation, the thick film surface is modeled as a two-dimensional plane. A Cartesian coordinate system is established with a corner point of the thick film as the origin and two mutually perpendicular edges as the X and Y axes. The actual position of each measurement point on the thick film surface is obtained by measurement, and its two-dimensional spatial coordinates are represented as (x, y). When constructing the probability distribution map of the thick film surface state, this two-dimensional plane is discretized into a regular grid. The number of rows and columns of the grid is determined according to the physical size of the thick film and the required resolution. In practical implementation, the grid spacing is usually set to be less than the minimum distance between measurement points. The spatial coordinates of each measurement point are mapped to the nearest grid point position in the probability distribution map of the thick film surface state. The mapping rule can be rounding or rounding down. For example, a measurement point with coordinates (1.2mm, 3.7mm) is mapped to the grid point position with grid index (2, 7) in the probability distribution map of the thick film surface state state with a grid spacing of 0.5mm. It can be understood that this mapping establishes a direct correspondence between the measurement point and the grid cell in the probability distribution map of the thick film surface state state.

[0033] The initial state classification label of each measurement point is converted into a state probability vector. In specific implementations, the initial state classification label is a discrete category identifier, such as "normal," "crack," "void," or "uneven thickness." The state probability vector is a multi-dimensional vector with the same number of categories, where each element corresponds to the probability value of a state category. The conversion rule is that the category position corresponding to the initial state classification label is assigned a fixed high value in the state probability vector, while the positions of other categories are assigned a fixed low value. In some embodiments, the fixed high value is set to 0.85, and the fixed low value is set to 0.05. For a measurement point classified as "crack," its state probability vector is converted to [0.05, 0.85, 0.05, 0.05], where the second element corresponds to the "crack" category. Each grid point position is assigned to the state probability vector of the corresponding measurement point. In specific implementations, if a grid point position exactly maps to a measurement point, then that grid point is directly assigned the converted state probability vector of that measurement point.

[0034] For grid locations in the thick film surface state probability distribution map that are not directly mapped by measurement points, their state probability vectors need to be generated through interpolation. In some embodiments, a bilinear interpolation algorithm is used to calculate the state probability vector of the blank grid location using the state probability vectors of neighboring grid points. Optionally, the bilinear interpolation process is performed independently for each dimension of the state probability vector (i.e., each state category). The state probability vector of the blank grid location Q represents the probability value P of the i-th category. i (Q), its interpolation calculation can be expressed as:

[0035] Where: P i (A), P i (B) P i (C) and P i (D) represents the probability value of the i-th class in the state probability vector at the four nearest neighbor grid points A, B, C, and D surrounding the blank grid position Q. Parameters u and v are the normalized horizontal and vertical displacements of the blank grid position Q relative to grid point A, with values ​​ranging from [0,1). In this way, each grid point in the thick film surface state probability distribution map, regardless of whether it is directly mapped by a measurement point, is assigned a complete state probability vector. After assigning values ​​to all grid points, a complete thick film surface state probability distribution map covering the entire thick film surface is constructed. This thick film surface state probability distribution map can be understood as expressing the probability estimate of each point on the thick film surface belonging to different states in the form of grid data.

[0036] Referring to Table 1, the spatial coordinates of the four measurement points, their initial state classification labels, and their transformed state probability vectors, as well as their corresponding grid positions mapped onto the thick film surface state probability distribution map, are described: Table 1: Mapping Table of Measurement Point Information and State Probability Vector

[0037] In practical implementation, for the position of grid index (3,3) in the table above, since there is no direct mapping from the measurement point, its state probability vector needs to be calculated using bilinear interpolation, using the state probability vectors of grid points (2,2), (2,4), (4,2), and (4,4). Optionally, the specific values ​​of the fixed high and fixed low values ​​can be adjusted according to the output confidence of the classification model, but the constraint that the sum of all elements within the same state probability vector is 1 must be satisfied.

[0038] In one embodiment of the present invention, all grid points marked as anomalous states are scanned on the state probability distribution map of the thick film surface. For each anomalous state grid point, the number of similar anomalous state grid points within its preset neighborhood is counted. If the number of similar anomalous state grid points is lower than the connectivity threshold, the anomalous state grid point is determined to be an isolated anomalous point, and the state probability vector of the anomalous state grid point is corrected to the probability vector corresponding to the dominant state category in its neighborhood. After traversing all anomalous state grid points and completing the correction, the state probability vector of all grid points in the entire map is smoothed and filtered again to generate an optimized state distribution map. The optimized state distribution map is binarized, and all grid points belonging to the same anomalous state are marked as foreground, and the remaining grid points are marked as background. A connected component labeling algorithm is applied to identify all interconnected pixel blocks in the foreground, and each pixel block corresponds to an independent anomalous region. The total number of grid points contained in each independent anomalous region is calculated, and the area of ​​the anomalous region is calculated by combining the actual physical area of ​​a single grid point. The boundary of each independent anomalous region is extracted, and the Euclidean distance between the two farthest points on the boundary is calculated as the longest axis length of the anomalous region. Based on the area and perimeter of the anomalous region, its shape factor is calculated. This shape factor is used to quantify the degree to which the shape of the region deviates from a circle.

[0039] In practical implementation, a spatial continuity correction algorithm is applied to the thick film surface state probability distribution map. This algorithm visits each grid point in the map line by line. When a grid point is scanned where the state category corresponding to the maximum probability value in the state probability vector belongs to a preset set of anomalous states, this grid point is marked as an anomalous state grid point to be inspected. For each marked anomalous state grid point, a preset neighborhood is defined centered on itself. In some embodiments, this neighborhood can be a 3x3 grid window. The number of grid points within this preset neighborhood whose state category corresponding to the maximum probability value in the state probability vector is the same as that of the anomalous state grid point is counted. This number is the number of anomalous state grid points of the same type. In practical implementation, the counted number of anomalous state grid points of the same type is compared with a preset connectivity threshold. The connectivity threshold can be set according to the actual application scenario, for example, 2. If the number of anomalous state grid points of the same type is lower than the connectivity threshold, the anomalous state grid point is determined to be an isolated anomalous point. For grid points identified as isolated anomalies, their state probability vectors need to be corrected. This correction involves replacing their state probability vectors with the standard probability vector corresponding to the statistically dominant category of the state probability vectors of all grid points (including themselves) within a preset neighborhood. In some embodiments, the statistically dominant category is determined by calculating the mean of the probabilities of each category in the state probability vectors of all grid points within the neighborhood and selecting the category with the largest mean. After traversing all marked anomalous state grid points in the thick film surface state probability distribution map and performing the above correction operation, the state probability vectors of all grid points in the entire map are then subjected to smoothing filtering. Optionally, the smoothing filtering can use a mean filter to average the probability vectors of each grid point and its surrounding neighborhood to generate the final optimized state distribution map.

[0040] The optimized state distribution map is binarized. In specific implementations, an independent binarization process is performed for each anomalous state requiring separate analysis, such as the "crack" state. The probability value corresponding to the "crack" category in the state probability vector of each grid point in the optimized state distribution map is extracted. If this probability value is greater than a set binarization threshold (e.g., 0.5), the grid point is marked as foreground in this processing; otherwise, it is marked as background. A connected component labeling algorithm is applied to identify all interconnected foreground pixel blocks in the binarized image. This algorithm iterates through the image, assigning a unique label to each connected foreground pixel block. In some embodiments, an 8-connectivity-based region growing algorithm is used, where each labeled connected foreground pixel block corresponds to an independent "crack" anomalous region on the thick film surface. For each identified independent anomalous region, the total number of grid points it contains is calculated. In specific implementations, each grid point corresponds to a fixed physical area on the actual thick film surface, such as a square region with a side length of 10 micrometers. The formula for calculating the area of ​​the anomalous region is: Anomalous region area = Total number of grid points × Actual physical area of ​​a single grid point. Extract the boundary pixel coordinates of each independent anomaly region. In practice, this is achieved by scanning the outer contour of the connected region to obtain the boundary point set. Calculate the Euclidean distance between any two points in the boundary point set and find the largest distance value. This largest Euclidean distance is defined as the longest axis length of the anomaly region. Calculate its shape factor based on the area and perimeter of the anomaly region. Optionally, the formula for calculating the shape factor R can be expressed as:

[0041] Where R represents the shape factor, A represents the area of ​​the anomalous region, and L represents the perimeter of the anomalous region. The value of the shape factor R is used to quantify the degree of deviation of the region's shape from a circle. For a perfectly circular region, R equals 1; for irregular or elongated regions, the value of R is less than 1. Refer to Table 2, which illustrates the local process of spatial continuity correction for a hypothetical thick film surface state probability distribution map, where the connectivity threshold is set to 2. Table 2: Local Process Table for Spatial Continuity Correction

[0042] In the specific implementation, Table 2 shows the local region centered on grid (5,5). Grid (5,5) was classified as a "crack" before correction, but within its 3x3 neighborhood, it has only one other grid point of the same type ("crack"), which is below the connectivity threshold of 2. Therefore, it is judged as an isolated outlier and corrected. During correction, the dominant state category in its neighborhood is "normal," so the state probability vector of grid (5,5) is corrected to the standard vector representing "normal." The correction logic for grid (6,6) is similar. Grids (5,7) and (6,7) retain their original "crack" classification because the number of similar outliers in their neighborhood meets or reaches the threshold. It can be understood that through this process, the isolated outlier classification labels are corrected, making the spatial distribution of state labels more consistent with the continuity constraint.

[0043] In one embodiment of the present invention, area-level thresholds, length-level thresholds, and shape-level thresholds are set for the area of ​​the abnormal region, the longest axis length, and the shape factor. The area of ​​the abnormal region is compared with the area-level thresholds to determine the level score in the area dimension. The longest axis length of the abnormal region is compared with the length-level thresholds to determine the level score in the length dimension. The shape factor of the abnormal region is compared with the shape-level thresholds to determine the level score in the shape dimension. The comprehensive defect score of the abnormal region is calculated by combining the area-level, length-level, and shape-level scores using a weighted summation method. The final defect level is determined based on the interval to which the comprehensive defect score belongs. The spatial coordinates of all measurement points and their corresponding final state labels on the optimized state distribution map are summarized to form a measurement point state list. The geometric attributes of all abnormal regions, the calculated defect levels, and their positional contour coordinates on the thick film surface are summarized to form an abnormal region detail list. The measurement point state list and the abnormal region detail list, along with the identification information of the measured piezoresistive sensor and the detection time information, are formatted and organized according to a predefined report template to generate a thick film detection report containing plain text and structured data.

[0044] In practical implementation, thresholds are set for the area of ​​abnormal regions, the length of the longest axis, and the shape factor. The area thresholds include multiple values; for example, an area less than 0.01 square millimeters is classified as Level 1, an area between 0.01 and 0.05 square millimeters as Level 2, and an area greater than 0.05 square millimeters as Level 3. The length thresholds also include grading values; for example, a longest axis length less than 0.1 millimeters is Level 1, between 0.1 and 0.3 millimeters is Level 2, and greater than 0.3 millimeters is Level 3. The shape thresholds are set based on the shape factor. The closer the shape factor is to 1, the closer the shape is to a circle. A shape factor greater than 0.7 is Level 1 (relatively regular), between 0.4 and 0.7 is Level 2 (relatively irregular), and less than 0.4 is Level 3 (extremely irregular). The area of ​​an abnormal region is compared with the area thresholds to determine its grade in the area dimension. For example, if the area is 0.03 square millimeters, its grade in the area dimension is 2. The longest axis length of the same anomaly region is compared with the length level threshold to determine the level score in the length dimension. The shape factor of the same anomaly region is compared with the shape level threshold to determine the level score in the shape dimension.

[0045] The comprehensive defect score of the abnormal region is calculated by combining the rating scores in the area, length, and shape dimensions using a weighted summation method. In some embodiments, the weights of the area, length, and shape dimensions can be set to 0.5, 0.3, and 0.2, respectively. The comprehensive defect score S final The calculation formula can be expressed as: S final =W1•S area +W2•S length +W3•S shape

[0046] Wherein: S final S represents the overall defect score for the abnormal area. area S represents the rating based on the area dimension. length S represents the rating scale along the length dimension. shape The rating represents the level of shape dimension, where W1, W2, and W3 represent the corresponding weight coefficients, and W1 + W2 + W3 = 1. The calculated comprehensive defect score S... final The predefined range determines the final defect level. For example, a comprehensive defect score between 1.0 and 1.5 can be defined as a minor defect level, between 1.5 and 2.5 as a moderate defect level, and greater than 2.5 as a severe defect level.

[0047] The spatial coordinates of all measurement points and their corresponding final state labels on the optimized state distribution map are summarized to form a measurement point state list. This list is organized in tabular form, with each row recording the measurement point's number, two-dimensional spatial coordinates, and final state classification label. The geometric attributes of all abnormal regions, the calculated defect levels, and the positional contour coordinates of the abnormal regions on the thick film surface are summarized to form an abnormal region detail list. In some embodiments, each row of the abnormal region detail list records a unique identifier, abnormality type, area, longest axis length, shape factor, comprehensive defect score, final defect level, and a series of contour point coordinates describing its boundary for each abnormal region. The measurement point state list and the abnormal region detail list, along with the unique identification information of the tested piezoresistive sensor and the start and end timestamps of the detection task, are formatted and organized according to a predefined report template. Optionally, the predefined report template uses an Extensible Markup Language (EXPLAIN) structure, including an identification information area, a detection summary area, a measurement point detail area, and an abnormal region detail area. The final generated file contains plain text descriptions and structured list data, constituting a complete structured thick film detection report.

[0048] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0049] In practical implementation, the workpiece to be inspected is placed directly below the inspection body of the piezoresistive sensor thick film inspection equipment. The inspection body is used to inspect the workpiece and distinguish defective products, while the tray-arranging machine is used to divide and arrange the workpieces into sections. A conveyor mechanism one is installed on the inspection body, used to transport the workpieces. A spacing mechanism is installed at the top of the middle section of conveyor mechanism one, used to distribute the workpieces at preset intervals on conveyor mechanism one. A flipping mechanism is installed at the end of the conveying stroke of conveyor mechanism one, used to flip the workpieces. A conveyor mechanism two is installed near the flipping mechanism of conveyor mechanism one, perpendicular to conveyor mechanism one, used to transfer the workpieces flipped by the flipping mechanism. A vision inspection mechanism is installed at the top of the middle section of conveyor mechanism two, used to visually inspect whether the workpieces are qualified or not. A storage frame is installed on conveyor mechanism one, used to collect and store defective workpieces. The second conveyor mechanism enters the tray-laying machine from the side furthest from the flipping mechanism. A diversion mechanism is installed on the second conveyor mechanism to separate qualified workpieces into two channels, allowing them to enter the tray-laying machine. The tray-laying machine is equipped with a dual-camera positioning mechanism located at the end of the conveying stroke of the second conveyor mechanism. This mechanism assists the tray-laying machine in performing zoned tray arrangement through visual inspection.

[0050] In some embodiments, the flipping mechanism flips the workpiece from conveying mechanism one to conveying mechanism two, so that the thick film surface of the workpiece to be inspected faces the vision inspection mechanism. The vision inspection mechanism includes a processor for performing the following steps: constructing a probability distribution map of the thick film surface state based on the response data of multiple preset measurement points on the thick film surface of the piezoresistive sensor under test; applying a spatial continuity correction algorithm to correct isolated anomaly classification labels on the probability distribution map of the thick film surface state, so that the corrected state labels satisfy the continuity constraint in spatial distribution, generating an optimized state distribution map; calculating the geometric attributes of each anomaly region according to the connected regions of different state labels in the optimized state distribution map, the geometric attributes including the area, longest axis length, and shape factor of the anomaly region; comparing the geometric attributes of the anomaly region with a preset criterion threshold to determine the defect level of each anomaly region; and generating a structured thick film inspection report by integrating the final state labels of all measurement points, the defect levels of all anomaly regions, and their location information.

[0051] Optionally, when the processor performs calculations on the geometric properties of each anomalous region, the shape factor is calculated using the following formula:

[0052] Where A represents the area of ​​the abnormal region, P represents the perimeter of the abnormal region, and F represents the shape factor. The shape factor is used to quantify the degree of deviation of the region shape from a circle. When the shape factor is close to 1, the region shape is close to a circle, and when the shape factor is less than 1, the region shape deviates from a circle.

[0053] In practice, a storage frame is installed on the conveying mechanism and located downstream of the vision inspection mechanism. The storage frame is used to collect and store workpieces deemed unqualified based on the thick-film inspection report generated by the vision inspection mechanism. The diversion mechanism, based on the thick-film inspection report, diverts qualified workpieces into dual channels for transport to the tray-laying machine. A dual-camera positioning mechanism positions the qualified workpieces entering through both channels, assisting the tray-laying machine in arranging the qualified workpieces into trays according to their inspection level.

[0054] It is understandable that the spacing mechanism, when dividing workpieces by spacing, adjusts the interval between adjacent workpieces to ensure they enter the flipping mechanism at a uniform distance, preventing workpiece accumulation. The flipping mechanism uses a rotating clamping method, picking up the workpiece from conveyor one, flipping it 180 degrees, and placing it on conveyor two, ensuring the surface of the thick film to be inspected faces upwards. The vision inspection mechanism triggers an image capture as the workpiece passes by, and the processor executes the aforementioned inspection steps based on the acquired image data. The dual-camera positioning mechanism includes two industrial cameras, each aimed at the end of a dual-channel system, capturing the workpiece's position and orientation in real time. This position information is sent to the robotic arm of the tray-laying machine, which then places the workpiece into the corresponding tray based on the position information and inspection level.

[0055] In some embodiments, the vision inspection mechanism further includes a light source module, which provides uniform illumination to the workpiece to ensure image acquisition quality. When constructing the probability distribution map of the thick film surface state, the processor first applies a standard excitation electrical signal through an electrical signal acquisition device and records the initial voltage response at each measurement point. After signal preprocessing, a multi-dimensional signal feature vector is extracted and then input into a fully trained thick film state classification model to obtain preliminary state classification labels. Finally, the probability distribution map is generated by combining spatial coordinates. The spatial continuity correction algorithm scans abnormal grid points and their neighborhoods, correcting the state labels of isolated abnormal points to the dominant states of their neighborhoods, and then generates an optimized state distribution map after smoothing filtering. Connectivity region identification uses a four-connectivity or eight-connectivity criterion to ensure the integrity of abnormal region boundaries. The defect level is determined based on a weighted comprehensive score of area, longest axis length, and shape factor, with different score intervals corresponding to different defect levels.

[0056] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for detecting thick films using piezoresistive sensors, characterized in that, The method includes: Multiple measurement points are preset on the thick film surface of the piezoresistive sensor under test. Based on the response data of the measurement points, a probability distribution map of the state of the thick film surface is constructed. On the probability distribution map of the state on the thick film surface, a spatial continuity correction algorithm is applied to correct the isolated anomaly classification labels so that the corrected state labels satisfy the continuity constraint in spatial distribution, thereby generating an optimized state distribution map. Based on the connected regions of different state labels in the optimized state distribution map, calculate the geometric properties of each abnormal region, including the area, longest axis length, and shape factor of the abnormal region; The geometric properties of the abnormal regions are compared with preset criterion thresholds to determine the defect level of each abnormal region; By combining the final status labels of all measurement points, the defect level and location information of all abnormal areas, a structured thick film inspection report is generated.

2. The piezoresistive sensor thick film detection method according to claim 1, characterized in that, The method involves pre-setting multiple measurement points on the thick film surface of the piezoresistive sensor under test, and constructing a probability distribution map of the thick film surface state based on the response data of the measurement points, including: Multiple measurement points are preset on the thick film surface of the piezoresistive sensor under test. A standard excitation electrical signal is applied through an electrical signal acquisition device, and the initial voltage response of each measurement point is recorded synchronously to form a set of initial response signals. The initial response signal set is preprocessed to eliminate environmental noise and baseline drift, generating a denoised response signal set. Multiple signal features of each measurement point are extracted from the set of denoised response signals to form a multidimensional signal feature vector, which includes amplitude features, phase features and frequency domain energy distribution features. The multidimensional signal feature vector of each measurement point is input into the fully trained thick film state classification model to obtain the preliminary state classification label of each measurement point. The preliminary state classification label includes normal, crack, void and uneven thickness. Based on the preliminary state classification labels of all measurement points and their spatial coordinates on the thick film surface, a probability distribution map of the state of the thick film surface is constructed. The initial response signal set is preprocessed to remove environmental noise and baseline drift, generating a denoised response signal set, including: For each initial voltage response signal in the set of initial response signals, a zero-phase digital filter is applied for bandpass filtering to retain the characteristic frequency band signals related to the piezoresistive material response; Adaptive baseline correction is performed on the bandpass filtered signal to eliminate signal baseline drift components caused by inherent sensor characteristics or slow environmental changes. The baseline-corrected signal is processed using wavelet threshold denoising technology to suppress high-frequency random noise in the signal, thereby obtaining a single denoised response signal in the set of denoised response signals.

3. The piezoresistive sensor thick film detection method according to claim 2, characterized in that, Multiple signal features for each measurement point are extracted from the denoised response signal set to form a multidimensional signal feature vector, including: Time-domain analysis is performed on a single denoised response signal in the set of denoised response signals. The average amplitude of the signal within the standard excitation period is calculated as the amplitude feature, and the time difference between the zero crossing point of the signal and the excitation signal is calculated as the phase feature. A fast Fourier transform is performed on the single denoised response signal to obtain the spectral distribution of the signal, and the spectral energy integral within a preset characteristic frequency band is calculated as the frequency domain energy distribution feature. The amplitude feature, the phase feature, and the frequency domain energy distribution feature are combined in a fixed order to form the multidimensional signal feature vector corresponding to the response characteristics of the measurement point; The step of inputting the multidimensional signal feature vector of each measurement point into the fully trained thick film state classification model includes: The fully trained thick-film state classification model is based on a feedforward neural network as its core architecture, and includes an input layer, multiple hidden layers, and an output layer. The number of nodes in the input layer is the same as the dimension of the multidimensional signal feature vector. The multidimensional signal feature vector is input from the input layer, undergoes layer-by-layer nonlinear transformation and feature abstraction in the multiple hidden layers, and finally generates a probability distribution corresponding to different thick film states in the output layer. The state category with the highest probability value in the output layer is selected as the preliminary state classification label for the measurement point.

4. The piezoresistive sensor thick film detection method according to claim 2, characterized in that, Based on the preliminary state classification labels of all measurement points and their spatial coordinates on the thick film surface, a probability distribution map of the thick film surface state is constructed, including: Using the thick film surface as a two-dimensional plane, the spatial coordinates of each measurement point are mapped to the grid point positions in the probability distribution map of the thick film surface state. The preliminary state classification label of each measurement point is converted into a state probability vector, in which the probability of the corresponding state category is a fixed high value, and the probabilities of the other categories are fixed low values; Each grid point is assigned the state probability vector corresponding to the measurement point. For grid points without measurement points, their state probability vectors are generated by bilinear interpolation of the state probability vectors of neighboring grid points, thus completing the construction of the state probability distribution map of the thick film surface.

5. The piezoresistive sensor thick film detection method according to claim 1, characterized in that, On the probability distribution map of the thick film surface state, a spatial continuity correction algorithm is applied to correct isolated anomaly classification labels, including: On the probability distribution map of the surface state of the thick film, scan all grid points marked as abnormal states; For each abnormal state grid point, count the number of similar abnormal state grid points within its preset neighborhood range; If the number of grid points of the same type of abnormal state is lower than the connectivity threshold, the abnormal state grid point is determined to be an isolated abnormal point, and the state probability vector of the abnormal state grid point is corrected to the probability vector corresponding to the dominant state category in its neighborhood. After traversing all abnormal state grid points and completing the correction, the state probability vectors of all grid points in the map are smoothed and filtered again to generate the optimized state distribution map.

6. The piezoresistive sensor thick film detection method according to claim 1, characterized in that, Based on the connected regions of different state labels in the optimized state distribution map, calculate the geometric properties of each abnormal region, including: The optimized state distribution map is binarized, and all grid points belonging to the same abnormal state are marked as foreground, while the remaining grid points are marked as background. The connected component labeling algorithm is used to identify all interconnected pixel blocks in the foreground, with each pixel block corresponding to an independent anomalous region. Calculate the total number of grid points contained in each independent anomaly region, and combine this with the actual physical area of ​​a single grid point to calculate the area of ​​the anomaly region; Extract the boundary of each independent anomaly region, and calculate the Euclidean distance between the two farthest points on the boundary as the longest axis length of the anomaly region; Based on the area and perimeter of the abnormal region, its shape factor is calculated, which is used to quantify the degree to which the shape of the region deviates from a circle.

7. The piezoresistive sensor thick film detection method according to claim 1, characterized in that, The geometric properties of the abnormal regions are compared with preset criterion thresholds to determine the defect level of each abnormal region, including: Set area-level thresholds for abnormal regions, length-level thresholds for the longest axis length, and shape-level thresholds for the shape factor; The area of ​​the abnormal region is compared with the area level threshold to determine the level score in the area dimension; The longest axis length of the abnormal region is compared with the length level threshold to determine the level score in the length dimension; The shape factor of the abnormal region is compared with the shape level threshold to determine the level score in the shape dimension; The comprehensive defect score of the abnormal region is calculated by combining the grade scores in the area dimension, the length dimension, and the shape dimension using a weighted summation method. The final defect level is determined based on the interval to which the comprehensive defect score belongs.

8. The piezoresistive sensor thick film detection method according to claim 1, characterized in that, The final status labels of all measurement points, the defect levels and location information of all abnormal areas are integrated to generate a structured thick film inspection report, including: Summarize the spatial coordinates of all measurement points and their corresponding final state labels on the optimized state distribution map to form a measurement point state list; Summarize the geometric properties of all abnormal regions, the calculated defect level, and their positional contour coordinates on the thick film surface to form a detailed list of abnormal regions; The list of measurement point statuses and the list of abnormal area details, along with the identification information of the piezoresistive sensor under test and the detection time information, are formatted and organized according to a predefined report template to generate the thick film detection report containing plain text and structured data.

9. The piezoresistive sensor thick film detection method according to claim 3, characterized in that, The training process of the fully trained thick-film state classification model includes: Acquire labeled thick film sample data accumulated in historical detection, wherein the labeled thick film sample data includes multidimensional signal feature vectors collected from thick films in known states and their true state labels; The labeled thick film sample data is divided into a training sample subset and a validation sample subset; A feedforward neural network is initialized as a classification model to be trained. The feedforward neural network includes an input layer, multiple hidden layers and an output layer, and the number of nodes in the input layer is the same as the dimension of the multidimensional signal feature vector. Using the aforementioned training sample subset, the cross-entropy between the output of the feedforward neural network and the true state label is used as the loss function, and the connection weights and bias parameters of the feedforward neural network are iteratively updated through the backpropagation algorithm. After each iteration, the classification accuracy of the current classification model is evaluated using the validation sample subset. The iteration stops when the classification accuracy no longer improves for several consecutive iterations. The feedforward neural network with fixed connection weights and bias parameters obtained at this time is used as the fully trained thick film state classification model.

10. A thick-film piezoresistive sensor inspection device, comprising an inspection body (1) and a tray arrangement machine (2), wherein a workpiece (3) to be inspected is disposed directly below the inspection body (1), the inspection body (1) is used to inspect the workpiece (3) and distinguish defective products, and the tray arrangement machine (2) is used to arrange the workpiece (3) into sections on a tray, characterized in that, The detection body (1) is equipped with a conveying mechanism (11), which is used to convey the workpiece (3). A spacing mechanism (12) is provided at the top of the middle section of the first conveying mechanism (11), which is used to divide the workpiece (3) by spacing; a flipping mechanism (13) is provided at the end of the conveying stroke of the first conveying mechanism (11), which is used to flip the workpiece (3). The first conveying mechanism (11) is provided with a second conveying mechanism (14) at one end near the flipping mechanism (13). The second conveying mechanism (14) and the first conveying mechanism (11) are arranged perpendicularly to each other. The second conveying mechanism (14) is used to transfer the workpiece (3) flipped by the flipping mechanism (13). A visual inspection mechanism (15) is provided at the top of the middle section of the second conveying mechanism (14). The visual inspection mechanism (15) is used to visually inspect whether the workpiece (3) is qualified or not. A storage frame (17) is provided on the first conveying mechanism (11). The storage frame (17) is used to collect and store unqualified workpieces (3). The second conveying mechanism (14) is connected to the tray-laying machine (2) from the side away from the flipping mechanism (13). The second conveying mechanism (14) is provided with a diversion mechanism (16). The diversion mechanism (16) is used to divert the qualified workpiece (3) into a channel, so that the workpiece (3) is diverted into the tray-laying machine (2) by the second conveying mechanism (14) into two channels. The tray arrangement machine (2) is equipped with a dual-camera positioning mechanism (21), which is located at the end of the conveying stroke of the second conveying mechanism (14). The dual-camera positioning mechanism (21) assists the tray arrangement machine (2) in performing partitioned tray arrangement through visual detection.

11. The piezoresistive sensor thick film detection device according to claim 1, characterized in that, The flipping mechanism (13) flips the workpiece (3) from the first conveying mechanism (11) onto the second conveying mechanism (14), so that the surface of the thick film to be inspected on the workpiece (3) faces the visual inspection mechanism (15); the visual inspection mechanism (15) includes a processor, which is used to perform the following steps: Multiple measurement points are preset on the thick film surface of the piezoresistive sensor under test. Based on the response data of the measurement points, a probability distribution map of the state of the thick film surface is constructed. On the probability distribution map of the state on the thick film surface, a spatial continuity correction algorithm is applied to correct the isolated anomaly classification labels so that the corrected state labels satisfy the continuity constraint in spatial distribution, thereby generating an optimized state distribution map. Based on the connected regions of different state labels in the optimized state distribution map, calculate the geometric properties of each abnormal region, including the area, longest axis length, and shape factor of the abnormal region; The geometric properties of the abnormal regions are compared with preset criterion thresholds to determine the defect level of each abnormal region; By combining the final status labels of all measurement points, the defect level and location information of all abnormal areas, a structured thick film inspection report is generated.

12. The piezoresistive sensor thick film detection device according to claim 2, characterized in that, The storage frame (17) is disposed on the conveying mechanism (11) and located downstream of the vision inspection mechanism (15), and is used to collect and store the workpieces (3) that are judged to be unqualified according to the thick film inspection report generated by the vision inspection mechanism (15). According to the thick film inspection report, the diversion mechanism (16) diverts the qualified workpieces (3) into dual channels and transports them to the tray arrangement machine (2). The dual camera positioning mechanism (21) positions the qualified workpieces (3) entering through the dual channels and assists the tray arrangement machine (2) in arranging the qualified workpieces (3) into trays according to the inspection level.

Images