Resistance value determination method and system of machine learning-based nbc resistance value sorting machine

Abstract

The application relates to the field of electronic component testing, and particularly discloses a resistance value determination method and system of an NBC resistance value sorting machine based on machine learning, which comprises the following steps: synchronously collecting probe contact, micro-environment, mechanism state and process code data, and constructing enhanced micro-time sequence profile data packets fusing multi-dimensional time sequence characteristics; training a physically guided double-path hybrid neural network model by using the data packets, so as to decouple the intrinsic resistance value characteristics of components from disturbed signals; outputting a structured diagnostic report containing intrinsic resistance value grades, quality scores, consistency indexes and health degree evaluations based on the model; finally triggering equipment compensation, maintenance warning or process optimization according to the report, and realizing system self-learning and updating through effect tracking; and the system comprises four functionally corresponding modules. The application realizes the intellectualization and explainability of resistance value determination, and upgrades the sorting equipment into an intelligent node of a manufacturing system with continuous optimization capability.

Description

Technical Field

[0001] This invention relates to the field of electronic component testing technology, specifically to a method and system for determining the resistance value of an NBC resistance value sorter based on machine learning. Background Technology

[0002] In the mass production of precision electronic components such as NBC (Negative Temperature Coefficient) thermistors, resistance sorting machines are key equipment to ensure product accuracy and consistency. Current technologies generally employ a measurement-threshold comparison method: after the test probe contacts the component, the measuring circuit obtains a stable resistance reading, typically the average of multiple samples. This reading is then compared to a preset fixed threshold range to determine the component's grade.

[0003] However, as the precision requirements for chips continue to increase, the inherent defects of this traditional method are becoming increasingly apparent, especially in the following deep and hidden technical aspects: Misjudgments caused by the black box approach to measurement: Existing methods only focus on the final resistance reading, completely discarding information such as the dynamic electrical signals during probe contact establishment and the micro-environmental fluctuations at the moment of testing. Microscopic bouncing of the contact probe, breakdown of the surface oxide layer, and instantaneous disturbances in ambient temperature and humidity all introduce unquantifiable noise and biases into the measurement signal. These biases, mixed with the inherent characteristics of the component itself, lead to a persistently high misjudgment rate, especially for "critical" components with resistance values ​​at the threshold.

[0004] Difficulty in attributing errors due to source confusion: When measurement results deviate, current technology cannot distinguish whether the deviation originates from a defect in the component itself or from a poor condition of the testing system, such as probe wear, temperature drift, or poor contact. This makes quality analysis and production line maintenance lack accurate data, often requiring only periodic, preventative overall calibration or replacement, which is inefficient and cannot eliminate the risk of batch misjudgment.

[0005] Information silos cause optimization stagnation: the testing phase is disconnected from upstream production processes and equipment maintenance. The sorting machine only outputs pass / fail conclusions; this data cannot be used to guide the optimization of upstream process parameters, nor can it provide early warnings of performance degradation in the testing equipment itself. This creates a one-way, closed information flow, hindering the continuous improvement of the overall precision of the manufacturing system.

[0006] Therefore, the industry urgently needs a new generation of resistance determination method that can see through the measurement process, decouple mixed errors, and form intelligent linkage with the production system in order to break through the current technical bottleneck of high-precision sorting. Summary of the Invention

[0007] The purpose of this invention is to provide a resistance value determination method and system for an NBC resistance value sorting machine based on machine learning, so as to solve the problems of high misjudgment rate and inability to trace error caused by the traditional threshold determination method in the prior art due to ignoring dynamic information of the test process.

[0008] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution: The resistance value determination method for the NBC resistance value sorting machine based on machine learning includes the following steps: S1. Synchronously collect probe contact, microenvironment and mechanism health time series data, and fuse them with component process feature encoding to construct micro-time series profile data package; S2. Using the micro-time profile data packet as input, train a dual-path hybrid neural network model and learn to decouple the intrinsic resistance of the components from the data through a physical guidance mechanism; S3. The trained dual-path hybrid neural network model is used to process real-time data and output a structured diagnostic report containing intrinsic resistance level, batch consistency deviation index and institutional health quantification assessment. S4. Based on the batch consistency deviation index and institutional health quantification assessment in the structured diagnostic report, trigger the optimization adjustment of production process parameters and the dynamic adaptive compensation of sorting machine test parameters.

[0009] As a preferred embodiment of the present invention, S1 specifically includes: S11. At the instant the test probe contacts the component pin, the synchronous clock source triggers the parallel activation of five data acquisition channels, which acquire the original waveforms of voltage and current of the contact circuit, the original signals of temperature and humidity at the test site, the original voltage differential signal of the resistance measurement module, and the original outputs of the micro-force sensor and displacement sensor built into the probe mechanism within a millisecond time window at a sampling rate of not less than 100kHz. S12. Perform real-time signal processing and feature extraction on the collected data to generate instantaneous power spectral density sequence of contact resistance, instantaneous gradient data of microenvironment temperature and humidity, high-speed sampling sequence of original resistance measurement, and time series data of mechanism health status, respectively. S13. The instantaneous power spectral density sequence of contact resistance, instantaneous gradient data of microenvironment temperature and humidity, high-speed sampling sequence of original resistance measurement, and timing data of mechanism health status are strictly aligned according to the unified millisecond-level timestamp provided by the synchronous clock source, and associated with the process feature encoding vector uniquely corresponding to the component obtained from the production management system, and finally encapsulated to generate a micro-timing profile data package.

[0010] As a preferred embodiment of the present invention, S12 specifically includes: S121. Perform a fast Fourier transform on the original voltage and current waveforms within a time window, calculate the impedance modulus and phase of each frequency component, and generate the instantaneous power spectral density sequence of the contact resistance. S122. For the raw temperature and humidity signals, calculate their instantaneous rate of change relative to the start of the measurement period to generate the instantaneous temperature and humidity gradient data of the microenvironment; S123. Perform synchronous analog-to-digital conversion and buffering on the original voltage differential signal to form the original resistance measurement high-speed sampling sequence; S124. After aligning and filtering the original outputs of the micro-force sensor and displacement sensor, calculate the eigenvalues ​​of their covariance matrix as a function of time to generate time-series data of the health status of the mechanism that characterizes contact stability.

[0011] As a preferred embodiment of the present invention, S2 specifically includes: S21. Construct the dual-path hybrid neural network model, which includes: The first path consists of a cascaded temporal convolutional network and a long short-term memory network, used to receive and process the temporal data in the micro-temporal profile data packet and output a dynamic feature vector. The second path consists of an embedding layer and a fully connected network, used to receive and process the process feature encoding vector and output a batch common feature vector. The fusion and output layer is used to concatenate the dynamic feature vector with the batch common feature vector, and map it to the intrinsic resistance level and other diagnostic labels through a fully connected layer; S22. In the training samples, each group of the micro-time profile data packets is labeled with its corresponding true intrinsic resistance level obtained by the high-precision static measurement system; at the same time, based on the time series data of the mechanism's health status, a theoretical contact deviation reference value is calculated and generated through a known probe wear mechanical model. S23. Use the training samples to train the model using a multi-objective loss function; S24. Use an independent validation set to evaluate the model's accuracy in determining resistance levels and its sensitivity in identifying scenarios such as poor contact and batch anomalies. Once the performance meets the standards, solidify the trained model parameters into the embedded intelligent judgment module of the sorting machine.

[0012] As a preferred embodiment of the present invention, S23 specifically includes: S231. Input a batch of the training samples into the dual-path hybrid neural network model, perform forward propagation calculation, and obtain the probability distribution of intrinsic resistance level predicted by the model and the contact state latent feature vector extracted from the first path respectively. S232. Based on the forward propagation results, the main loss term and the physical regularization term are weighted and summed according to preset weight coefficients to obtain the total loss value of this round of training; wherein, the main loss term is the cross-entropy loss between the probability distribution of the intrinsic resistance level predicted by the model and the labeled true intrinsic resistance level; the physical regularization term is the mean square error loss between the contact state latent feature vector extracted by the first path and the theoretical contact deviation reference value. S233. Based on the total loss value, calculate the gradient of each layer parameter of the dual-path hybrid neural network model using the backpropagation algorithm, and update all model parameters using an adaptive optimization algorithm; S234. Repeat the operation to optimize and update all model parameters until the total loss value no longer decreases significantly on the validation set and the physical regularization term is lower than the preset tolerance. Determine that the model training has converged and save the model parameters at this time.

[0013] As a preferred embodiment of the present invention, S3 specifically includes: S31. Input the micro-time profile data packet generated in real time during the test cycle of the current NBC component under test into the trained dual-path hybrid neural network model; the first path and the second path of the model are processed in parallel, and the dynamic feature vector and the batch common feature vector are output respectively, and the fusion layer calculates and generates the original output vector including the preliminary prediction value of the intrinsic resistance level. S32. Based on the internal layer output and original output vector of the dual-path hybrid neural network model, multiple diagnostic indicators are solved in parallel, including intrinsic resistance level, judgment confidence, dynamic contact quality score, environmental adaptation factor, batch consistency deviation index and institutional health quantification assessment. S33. The intrinsic resistance level, judgment confidence level, dynamic contact quality score, environmental adaptability factor, batch consistency deviation index and institutional health quantification assessment are packaged into a structured diagnostic report according to a predetermined format and simultaneously output to the sorting execution mechanism of the sorting machine and the central production management system.

[0014] As a preferred embodiment of the present invention, S32 specifically includes: S321. Extract the probability distribution belonging to the resistance category from the original output vector, determine the category corresponding to the maximum probability as the intrinsic resistance level; calculate the Shannon entropy of the probability distribution, and output the corresponding determination confidence based on the preset entropy-confidence mapping table; S322. Input the hidden layer feature vector related to the contact state in the first path into a pre-trained quality assessment fully connected network to generate the dynamic contact quality score; concatenate the hidden layer feature vector related to the environment in the first path with the instantaneous gradient data of microenvironment temperature and humidity, and calculate the environment adaptation factor through the environment compensation coefficient calculation unit. S323. Calculate the Mahalanobis distance between the current batch common feature vector output by the second path and the corresponding vector set of all process feature encoding vector elements in the historical database, and generate the batch consistency deviation index after normalizing the distance value; for a specific recurrent layer in the first path that processes the time series data of the health status of the institution, extract the activation spectrum of its neurons in time series, calculate the peak attenuation coefficient and frequency band energy ratio of the activation spectrum, and fuse them to generate the quantitative assessment of the health status of the institution.

[0015] As a preferred embodiment of the present invention, S4 specifically includes: S41. Receive and parse the structured diagnostic report in real time, extract the batch consistency deviation index and institutional health quantification assessment; match the indicators with a preset multi-level rule base, which defines the optimization action types and parameters corresponding to different indicator threshold ranges; S42. Based on the matching results, generate specific closed-loop optimization instructions: When the quantitative assessment of the mechanism's health falls below a first threshold, a dynamic compensation instruction is generated. The dynamic compensation instruction includes a new contact force setting or signal gain adjustment coefficient calculated for the sorting machine to compensate for the mechanism's performance degradation online. When five or more consecutive components with the same process feature code have a batch consistency deviation index that is consistently higher than the second threshold, a pre-maintenance alarm instruction is generated. The pre-maintenance alarm instruction identifies the abnormal batch and its associated process equipment. Aggregate the batch consistency deviation index statistical distribution of all batches within a production cycle. When its kurtosis and skewness exceed the third threshold, generate process parameter optimization suggestions. The process parameter optimization suggestions include the range of tuning parameters for the preceding sintering or forming process. S43. Execute specific closed-loop optimization instructions and track the execution results.

[0016] As a preferred embodiment of the present invention, S43 specifically includes: S431. Distribute the optimized instructions to different execution terminals according to their type: The new contact force setting value in the dynamic compensation command is sent to the servo driver of the sorter, and the signal gain adjustment coefficient is sent to the digital potentiometer of the measurement unit to complete the real-time hardware parameter reconfiguration. The pre-maintenance alarm command is pushed to the maintenance work order module of the production management system to generate a pending work order containing an abnormal batch identifier and suggested inspection items; The proposed process parameter optimization suggestions are pushed to the process management module of the manufacturing execution system as a recommended range for adjusting production parameters in the next batch. S432. After the optimization instruction is executed, a tracking cycle is initiated for the affected entity, such as a specific device, batch, or process: For a sorting machine that has implemented dynamic compensation, record the mean and variance changes of the mechanical health quantification assessment of more than fifty consecutive components. For process batches that trigger pre-maintenance alarms, record the mean and distribution changes of the batch consistency deviation index of the next batch of components after maintenance is completed. Record the overall statistical characteristics of the batch consistency deviation index of the components produced by the preceding process that adopted the process parameter optimization suggestions in subsequent tests; S433. Compare the quantitative performance data obtained from the tracking with the expected optimization target: If the effect meets or exceeds expectations, the weight of the rule that generates the instruction is strengthened. If the effect does not meet expectations or causes side effects, rule backtracking analysis is triggered to adjust the corresponding indicator thresholds in the rule base or optimize the action parameters, and a new rule to be verified is generated. The updated rule base will be synchronized to all relevant decision nodes to complete this self-learning optimization loop.

[0017] A resistance value determination system for an NBC resistance value sorting machine based on machine learning is provided to implement a resistance value determination method for an NBC resistance value sorting machine based on machine learning, including: The data construction module is used to simultaneously collect probe contact circuit, microenvironment, mechanism health time series data and component process feature encoding within milliseconds of the test probe contacting the NBC component, and fuse them to generate a micro-time profile data package. The model processing module, connected to the data construction module, is used to receive data packets and perform training and inference using a built-in dual-path hybrid neural network model; a multi-objective loss function with fused physical regularization terms is used for training to decouple the intrinsic resistance of components from the data; The intelligent diagnostic module, connected to the model processing module, is used to receive the feature vector and original results output by the model inference, and to solve the intrinsic resistance level, judgment confidence, dynamic contact quality score, environmental adaptation factor, batch consistency deviation index and institutional health quantification assessment in parallel, and encapsulate them into a structured diagnostic report. The closed-loop optimization module, connected to the intelligent diagnostic module, is used to parse the report, generate dynamic compensation instructions, pre-maintenance alarm instructions or process parameter optimization suggestions based on the diagnostic indicators, and drive the actuator to complete parameter adjustment, alarm triggering or suggestion push. At the same time, it tracks the optimization effect and updates the internal decision rule base.

[0018] Compared with the prior art, the present invention has the following advantages: 1. By constructing an enhanced micro-time profile data package that synchronously integrates the transient spectrum of contact resistance, micro-environment gradient, mechanism health time series and process coding, uncontrollable disturbances such as contact bounce and probe micro-wear are transformed into structured features that can be modeled and analyzed for the first time, laying the foundation for accurate error tracing and realizing a leap from single measurement value to full-process digital mapping.

[0019] 2. By employing a dual-path hybrid neural network incorporating physical regularization terms, the trained model decouples the intrinsic resistance characteristics of the chip from the disturbed signal, making the judgment process physically interpretable. This significantly improves the sorting accuracy and reliability of critical components, realizing a paradigm shift from threshold comparison to physical mechanism decoupling.

[0020] 3. The system can automatically convert diagnostic reports into equipment dynamic compensation, pre-maintenance alarms and process optimization suggestions. Through effect tracking and rule self-updating, it forms a self-learning cycle that continuously improves manufacturing yield and efficiency, upgrading the sorting machine from a detection tool to a manufacturing process optimization engine, realizing a system upgrade from passive detection to proactive optimization closed loop. Attached Figure Description

[0021] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the method described in Embodiment 1 of the present invention.

[0023] Figure 2 This is a framework diagram of the system described in Embodiment 2 of the present invention. Detailed Implementation

[0024] 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.

[0025] The concepts involved in this application will first be described with reference to the accompanying drawings. It should be noted that the following descriptions of various concepts are only for the purpose of making the content of this application easier to understand and do not constitute a limitation on the scope of protection of this application; furthermore, the embodiments and features in the embodiments of this application can be combined with each other unless otherwise specified. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] Example 1 like Figure 1 As shown, this invention provides a resistance value determination method for an NBC resistance value sorting machine based on machine learning, including the following steps: S1. Synchronously acquire probe contact, microenvironment, and structural health time-series data, and fuse them with component process feature encoding to construct a micro-time-series profile data package; specifically including: S11. At the transient moment when the test probe makes physical contact with the NBC component pin, a hardware-level trigger signal is generated by a high-precision synchronous clock source. This signal simultaneously and in parallel drives five independent data acquisition channels into a high-speed acquisition state through a low-jitter clock distribution network, specifically: The first channel is deployed across the precision sampling resistor in the contact circuit to continuously capture the transient voltage and current waveforms of the contact circuit within a millisecond-level time window at a sampling rate of not less than 100kHz. The second channel uses miniature thermocouples and capacitive humidity sensing elements deployed adjacent to the test site to simultaneously capture the original temperature and humidity signals in the microenvironment space at the same sampling rate. The third channel is directly coupled to the analog front-end output of the resistance measurement module, and the original voltage differential signal is captured in real time by a high-speed analog-to-digital converter. The fourth channel receives analog outputs from the micro-force sensor and displacement sensor integrated inside the probe mechanism, and records the mechanical response and mechanical displacement trajectory during the contact process between the probe tip and the component pin. The fifth channel retrieves the process feature encoding vector bound to the unique identifier of the component under test from the production management system database via a real-time industrial Ethernet link. This vector contains at least the normalized encoding information of the material batch number, sintering furnace number, and body forming mold number of the component.

[0027] The five channels achieve nanosecond-level synchronization under a unified clock reference, ensuring strict alignment of the raw data in the time domain and providing a complete multiphysics raw data foundation for the subsequent construction of micro-time-series profile data packets.

[0028] S12. Perform real-time signal processing and feature extraction on the acquired data, specifically as follows: S121. For the original waveforms of contact circuit voltage and current acquired by the first channel, firstly, according to the sampling theorem, a Hamming window function is applied within a millisecond time window to perform windowing truncation in order to suppress the spectral leakage effect. Then, a fast Fourier transform is performed to map the time-domain transient signal to the frequency domain, and the impedance modulus and phase angle at each discrete frequency point are calculated in the frequency domain. Then, by using the improved periodogram method or the Welch piecewise average power spectral density estimation algorithm, the energy distribution density of the contact resistance in the frequency domain during the transient contact process is calculated. Finally, a sequence of instantaneous power spectral density of contact resistance is generated. This sequence serves as time-series data characterizing the dynamic spectral properties of the contact interface impedance and is used for subsequent decoding of contact quality issues.

[0029] S122. For the raw temperature and humidity signals of the microenvironment acquired by the second channel, take the start time of the measurement cycle as the time reference zero point, perform a first-order difference operation on the raw temperature and humidity signals acquired continuously after that time point, and calculate the instantaneous rate of change of temperature and the instantaneous rate of change of humidity relative to the reference zero point at the current sampling time. By extracting the thermodynamic state fluctuation gradient of the microenvironment on a millisecond timescale, instantaneous temperature and humidity gradient data of the microenvironment are generated. This data quantitatively reflects the dynamic evolution trend of temperature and humidity of the microenvironment at the test site during the transient process of probe contact, and is used for the calculation of environmental adaptation factors in subsequent steps.

[0030] S123. For the raw voltage differential signal acquired by the third channel, the signal is latched by the front-end sample and hold circuit to ensure strict synchronization of the sampling time; Subsequently, a high-speed successive approximation analog-to-digital converter is used for real-time digital conversion. During the conversion process, a quantization precision of no less than 16 bits is used to preserve the details of weak signals. The converted digital quantity is processed by a digital anti-aliasing filter to eliminate high-frequency noise aliasing. It is then marked with a millisecond-level timestamp provided by a synchronous clock source and written into a first-in-first-out high-speed circular buffer to form the original high-speed sampling sequence for resistance measurement. This sequence completely preserves the original voltage response timing information of the resistance measurement module during the entire contact transient process, serving as the basis data for subsequent intrinsic resistance calculation.

[0031] S124. For the original outputs of the micro-force sensor and displacement sensor acquired by the fourth channel, strict spatiotemporal coordinate alignment is first performed based on the high-precision time reference provided by the synchronous clock source. The delay difference of the signal transmission path of each sensor is eliminated by the delay compensation algorithm to ensure strict synchronization of the force signal and displacement signal in the time domain. Subsequently, the extended Kalman filter algorithm is applied to perform state estimation and noise suppression on the aligned signal. Based on the filtered signal, a covariance matrix of microforce and displacement within the sliding time window is constructed. Eigenvalue decomposition is performed on the covariance matrix, and the ratio of the maximum eigenvalue to the minimum eigenvalue is extracted as a quantitative index of contact stability. Time series data of the mechanism's health status is generated. This data not only characterizes the mechanical coupling stability of the probe mechanism during the contact process, but also implicitly encodes the wear degree of the probe tip and the aging state of the mechanism through the time series evolution of the force-displacement coupling relationship.

[0032] S13. Data alignment and fusion encapsulation, specifically: S131. The instantaneous power spectral density sequence of contact resistance generated in step S121, the instantaneous gradient data of microenvironment temperature and humidity generated in step S122, the high-speed sampling sequence of original resistance measurement generated in step S123, and the timing data of mechanism health status generated in step S124 are uniformly loaded into the absolute time reference system established based on the synchronous clock source. Strict timing alignment is performed according to the unified millisecond-level timestamps marked by the synchronous clock source at the acquisition time of each data sequence, eliminating the timing deviations introduced by the delay of signal conditioning circuit, the setup time of analog-to-digital converter and the difference of transmission path in each physical field data acquisition channel, and ensuring accurate synchronization on the same time axis.

[0033] S132. At the same time, the time-aligned four-dimensional dynamic time series data is associated with and fused with the process feature encoding vector obtained from the production management system in step S11 and bound to the unique identifier of the NBC component under test. The process feature encoding vector contains normalized encoding information of material batch number, sintering furnace number and body forming mold number, which is embedded as a static attribute label into the header of the dynamic time series data or in the parallel data structure.

[0034] S133. Finally, through a structured data encapsulation protocol, the time-strictly aligned instantaneous power spectral density sequence of contact resistance, instantaneous gradient data of microenvironment temperature and humidity, high-speed sampling sequence of original resistance measurement, time series data of institutional health status, and their associated process feature encoding vectors are integrated into an enhanced micro-time series profile data package with a unified data format and indexing mechanism. This data package forms a complete data entity containing the entire process of multi-physics transient evolution and production process traceability information, which serves as the input sample for the subsequent physics-guided hybrid neural network model.

[0035] S2. A dual-path hybrid neural network model is trained using micro-time profile data packets as input, and through a physical guidance mechanism, it learns to decouple the intrinsic resistance values ​​of components from the data; specifically including: S21. Construct a dual-path hybrid neural network model. This model adopts a heterogeneous data parallel processing architecture and consists of three parts: a first path, a second path, and a fusion and output layer. Specifically: S211. The first path is designed for high-dimensional time-series data of millisecond-level transient processes. It adopts a deep architecture of cascading temporal convolutional networks and long short-term memory networks. It is specifically used to receive and deeply process four-dimensional dynamic data in micro-time-series profile data packets, including instantaneous power spectral density sequence of contact resistance, instantaneous gradient data of micro-environment temperature and humidity, high-speed sampling sequence of original resistance measurement, and time-series data of mechanism health status. In this path, the temporal convolutional network first extracts local features from the input multi-channel temporal signal through multiple layers of one-dimensional convolutional kernels with different receptive fields, capturing short-time pulse patterns, frequency and energy changes, and mechanical vibration features during the contact transient process; The extracted high-dimensional local feature sequences are then input into a long short-term memory network to model their temporal long-term dependencies. The evolutionary trend of the contact process is memorized and redundant noise is forgotten through a gating mechanism. Finally, global pooling is performed in the time dimension to output a compressed dynamic feature vector that represents the entire contact dynamics process.

[0036] S212. The second path is designed for static process attributes and consists of an embedded layer and a deep fully connected network. It is specifically used to process the process feature encoding vector associated in the enhanced micro-time profile data package. This vector contains high-dimensional sparse encoding of material batch number, sintering furnace number and body forming mold number. The embedding layer first maps the sparse code to a low-dimensional dense real vector space to learn the implicit similarities and common patterns between batches; Subsequently, the embedded vector is subjected to a depth transformation through a batch normalization layer and a multi-layer nonlinear fully connected network to extract batch common features related to material formulation, sintering process and molding die, and finally output batch common feature vector.

[0037] S213. The fusion and output layer is responsible for the deep fusion and end-to-end mapping of heterogeneous features. This layer first splices and fuses the dynamic feature vectors adjusted to a unified dimension through linear projection with the batch common feature vectors to form a joint representation that combines transient physical process characteristics and static process attributes. Subsequently, a multi-layer fully connected residual network containing batch normalization and activation functions is used to perform nonlinear transformation and dimensionality reduction on the joint representation; Finally, the data is mapped to the intrinsic resistance level probability distribution through a softmax classifier, and simultaneously outputs a mechanism health quantification assessment, contact quality score, and judgment confidence auxiliary diagnostic label through parallel branches, realizing an end-to-end mapping from multi-physics raw data to resistance judgment decision.

[0038] S22. For each sample used for model training, i.e., the micro-temporal profile data packet encapsulated in step S13, dual supervision signals need to be constructed in parallel to support the subsequent optimization of the multi-objective loss function, specifically as follows: S221. For the construction of the main supervised target, the NBC element under test corresponding to the micro-time profile data packet is subjected to offline static measurement in a laboratory environment away from the production line using a high-precision four-wire standard resistance measurement system. This measurement process is carried out in a strictly constant temperature and humidity shielded environment to completely eliminate the influence of instantaneous fluctuations in micro-environment temperature and humidity and contact transient processes. The measured accurate resistance value is discretized and mapped to the true intrinsic resistance level, which serves as the gold standard label for this sample in supervised learning.

[0039] S222. For the construction of physical guidance constraints, the timing data of the mechanism health status contained in the micro-time profile data package is extracted. This timing data contains mechanical response and mechanical stability information during the contact process between the probe tip and the component pin. It is input into the pre-established probe wear mechanical model. This model is based on the probe tip geometric morphology parameters, the electrical contact theory of the coating material and the historical wear accumulation effect. It calculates the theoretical contact deviation reference value introduced by the probe mechanical state under the current mechanism health state. This reference value characterizes the expected systematic measurement error caused by probe wear.

[0040] S223. Thus, each set of micro-time profile data packets used for training is associated to form a pair of supervision signals: namely, the true intrinsic resistance level provided by the high-precision static measurement system, and the theoretical contact deviation reference value calculated by the probe wear mechanics model based on the time series data of the mechanism's health status. Together, they constitute the true benchmark required for the calculation of the multi-objective loss function in step S23.

[0041] S23. Train the model using the training samples and a multi-objective loss function; specifically: S231. Perform forward propagation and feature extraction: The batch enhanced micro-temporal profile data packets are input into the dual-path hybrid neural network model. The first path receives the instantaneous power spectral density sequence of contact resistance, the instantaneous gradient data of microenvironment temperature and humidity, the high-speed sampling sequence of original resistance measurement, and the temporal data of the mechanism's health status from the data packets. After local feature extraction by the temporal convolutional network and long-term dependency modeling by the long short-term memory network, it not only outputs dynamic feature vectors, but also extracts contact state hidden feature vectors to characterize the physical state of the contact from the hidden layer states of the long short-term memory network. The second path receives the process feature encoding vector associated in the data packet, and outputs the batch common feature vector after being mapped by the embedding layer and transformed by the fully connected network. The fusion and output layer concatenates and fuses dynamic feature vectors with batch common feature vectors, and generates the probability distribution of intrinsic resistance level predicted by the model through a multi-layer fully connected network.

[0042] S232. Calculate multi-objective weighted loss: Construct a composite loss function based on the forward propagation results. Specifically, the main supervision loss term and the physical regularization loss term are weighted and summed according to preset weight coefficients to generate the total loss value of this round of training. The main supervision loss term uses the cross-entropy loss function to measure the difference between the probability distribution of the intrinsic resistance level output by the model and the probability distribution of the true intrinsic resistance level marked in step S22. The physical regularization loss term uses the mean square error loss function to calculate the Euclidean distance deviation between the contact state latent feature vector extracted within the first path and the theoretical contact deviation reference value calculated by the probe wear mechanics model based on the mechanism health state time series data in step S22. This regularization constraint forces the model to explicitly learn the physical mapping law between the mechanism health state time series data and the measurement deviation.

[0043] S233. Perform backpropagation and parameter optimization: Based on the calculated total loss, the gradients of all trainable parameters of the dual-path hybrid neural network model are calculated layer by layer from the output layer to the input layer using the backpropagation algorithm. This includes the kernel weights of the temporal convolutional network in the first path, the input gate weights and forget gate biases of the long short-term memory network, the lookup table parameters of the embedding layer and the weights of the fully connected layer in the second path, and the concatenation weights of the fusion and output layers and the classifier parameters. Subsequently, the adaptive moment estimation optimization algorithm is used to iteratively update all model parameters based on the calculated gradients and the adaptive learning rate.

[0044] S234. Iterative Training and Convergence Determination: Repeat the above-described forward propagation, loss calculation, and backpropagation optimization process. After each training epoch, calculate the total loss value of the model on the independent validation set and monitor the joint change trend of the main supervision loss term and the physical regularization term. When the total loss value on the validation set no longer decreases significantly and tends to stabilize over multiple consecutive training cycles, and the value of the physical regularization term decreases to below the preset tolerance threshold, it indicates that the model has fully embedded physical constraint laws while possessing the ability to accurately classify resistance levels. At this point, the model training is determined to be converged, and all network weight parameters and bias parameters of the current iteration cycle are saved as the firmware of the trained model.

[0045] S24. Model validation and deployment, specifically: S241. A comprehensive performance evaluation of the dual-path hybrid neural network model that has converged in training is performed using a validation sample set that is completely independent of the training process. The validation set consists of micro-time profile data packets collected on site, and its samples do not overlap with the training samples in terms of spatiotemporal distribution, covering multiple scenarios such as normal working conditions, poor contact, and batch anomalies.

[0046] S242. Input the micro-time profile data packets from the validation set into the trained model one by one, obtain the intrinsic resistance level prediction results, judgment confidence and batch consistency deviation index output by the model, and calculate the overall accuracy of resistance level judgment by comparing the prediction results with the true intrinsic resistance level marked by the high-precision standard measurement system. Simultaneously, based on the abnormal spectral characteristics of the instantaneous power spectral density sequence of contact resistance, poor contact scenarios are identified; based on the batch consistency deviation index exceeding a preset threshold, batch abnormal scenarios are determined; and then the model's sensitivity and recall rate for identifying these two types of abnormal scenarios are calculated.

[0047] S243. When the accuracy of resistance level determination reaches the preset engineering accuracy threshold and the sensitivity of identification of poor contact and batch abnormal scenarios meets the production reliability requirements, the trained model parameters are quantized, compressed, and optimized by operators to generate a firmware image adapted to the embedded processor architecture. The firmware image is burned and solidified into the embedded intelligent determination module built into the selector, so that the dual-path hybrid neural network model can receive micro-time profile data packets collected on site in real-time inference mode and output intrinsic resistance level determination results and multi-dimensional diagnostic information, realizing the end-side intelligent deployment of resistance selection function.

[0048] S3. Utilize the trained dual-path hybrid neural network model to process real-time data and output a structured diagnostic report containing intrinsic resistance level, batch consistency deviation index, and institutional health quantification assessment; specifically including: S31. Real-time data input and model inference, specifically: S311. After the NBC component under test is mechanically positioned at the sorting machine test station, during the entire test cycle in which the test probe contacts and stabilizes with the component pin, according to the synchronous acquisition and packaging process described in steps S11 to S13, a micro-time profile data packet containing complete multi-physics information of the component is generated in real time. The data packet contains a strictly time-aligned instantaneous power spectral density sequence of contact resistance, instantaneous gradient data of micro-environment temperature and humidity, high-speed sampling sequence of original resistance measurement, time-series data of mechanism health status, and its associated process feature encoding vector.

[0049] S312. Use this micro-temporal profile data package as model input and load it into the dual-path hybrid neural network model that has been solidified and deployed in step S24 to trigger the forward inference process of the model. At this point, the first and second paths of the model enter a parallel processing state. The first path receives four-dimensional time-series data from the data packet, extracts local spectral features and mechanical vibration modes in the contact transient process through multi-layer one-dimensional convolution operations of the temporal convolutional network, and then models the temporal dependency through the gating mechanism of the long short-term memory network to capture the long-term trend of the contact resistance power spectral density evolution over time and the dynamic changes in the health status of the mechanism. Finally, it outputs a compressed dynamic feature vector that characterizes the entire contact dynamics process. Meanwhile, the second path receives the process feature encoding vector in the data packet, and maps the high-dimensional sparse encoding to a low-dimensional dense real number vector space through the embedding layer to extract the implicit features of the material batch and process parameters. Then, through the nonlinear transformation of the fully connected network, it outputs the batch common feature vector that represents the common law of the batch to which the component belongs. The fusion layer receives the dynamic feature vectors and batch common feature vectors output from the two paths mentioned above, respectively, and performs vector concatenation to form a joint representation that combines transient physical characteristics and static process attributes. Then, through the nonlinear mapping and dimensional transformation of the multi-layer fully connected network, the original output vector is calculated and generated. This vector contains the probability distribution prediction value of the intrinsic resistance level and other intermediate hidden layer feature information required for diagnosis, which serves as the basic data for the parallel calculation of the six structured diagnostic indicators in the subsequent step S32.

[0050] S32. Based on the original output vector generated in step S31 and the hidden layer features of the inner layer outputs of the dual-path hybrid neural network model, generate complete structured diagnostic indicators; specifically: S321. Extract the intrinsic resistance level and decision confidence level from the original output vector: The probability distribution segment corresponding to the intrinsic resistance level prediction in the original output vector is normalized by softmax, and the category index with the largest probability value is extracted as the intrinsic resistance level of the NBC component under test. Simultaneously, Shannon information entropy is calculated for the probability distribution to measure the degree of uncertainty of the prediction result. Based on the preset entropy value and confidence level mapping table, the calculated entropy value is converted into a judgment confidence level, where low entropy value corresponds to high confidence level and high entropy value corresponds to low confidence level. This quantifies the certainty level of the model in determining the resistance value.

[0051] S322. Calculate dynamic contact quality score and environmental adaptation factor based on hidden layer features within the first path: The hidden layer feature vectors representing the physical state of the contact are generated by the temporal convolutional network and the long short-term memory network in the first path during the processing of the instantaneous power spectral density sequence of contact resistance and the temporal data of the mechanism's health status. These vectors are then input into a pre-trained fully connected quality evaluation network independent of the main network. This network establishes a nonlinear mapping between the contact state features and the contact quality score through supervised learning, and outputs a numerical dynamic contact quality score. This score reflects the physical coupling quality between the probe and the component pin during the current contact transient process.

[0052] Simultaneously, the environmental-related hidden layer feature vectors generated by the first path when processing the instantaneous gradient data of micro-environment temperature and humidity are extracted, and then spliced ​​and fused with the instantaneous gradient data of micro-environment temperature and humidity in the micro-time-series profile data package at the channel dimension. This data is then input into the environmental compensation coefficient calculation unit. Based on the physical mechanism model of the influence of temperature and humidity on the resistance measurement value, this unit calculates the compensation correction coefficient under the current micro-environment state and generates an environmental adaptation factor. This factor quantitatively represents the degree of deviation between the current micro-environment conditions and the standard test conditions and its influence weight on the resistance measurement.

[0053] S323. Quantitative assessment of batch consistency deviation index and institutional health based on second path output and first path loop layer state: The batch common feature vector of the current test component output by the second path is calculated, and the Mahalanobis distance between it and the set of batch common feature vectors of all historical components with the same process feature encoding vector stored in the historical database is calculated. This distance measures the degree of deviation of the process characteristics of the current component from the statistical distribution of the historical samples of the same batch in the high-dimensional feature space. The calculated Mahalanobis distance is mapped to the zero-one interval through a normalization function to generate a fractional batch consistency deviation index. The higher the index, the more the process characteristics of the component deviate from the consistency pattern of the batch.

[0054] Meanwhile, for a specific recurrent layer of the long short-term memory network in the first path that specifically processes temporal data on the health status of the institution, the activation response sequence of all hidden neurons in the time dimension is extracted to form an activation spectrum. The temporal characteristics of the activation spectrum are analyzed, and the attenuation coefficient of its peak value with time and the energy distribution ratio of the high-frequency band to the low-frequency band are calculated. The peak attenuation coefficient and the frequency band energy ratio are weighted and fused to generate a quantitative assessment of the institution's health. This assessment value reflects the current aging degree, mechanical wear state and contact stability level of the probe institution.

[0055] S33. Structured report encapsulation and output, specifically: S331. The six core diagnostic indicators calculated in step S32—intrinsic resistance level, judgment confidence level, dynamic contact quality score, environmental adaptability factor, batch consistency deviation index, and institutional health quantification assessment—along with the unique identifier of the NBC component under test, the corresponding process feature encoding vector, the test timestamp, and the index identifier of the micro-time profile data packet, are encapsulated and encoded according to a predetermined structured data protocol to generate a structured diagnostic report with a standard format.

[0056] S332. The report is transmitted in real time to the sorting execution mechanism via the high-speed data bus inside the sorting machine. Based on the intrinsic resistance level determination result in the report, the sorting execution mechanism drives the corresponding physical sorting device to perform classification actions, and guides the components into the corresponding resistance level receiving trough or abnormal product rejection channel, thereby realizing automated sorting operation based on intelligent judgment.

[0057] S333. Simultaneously, the structured diagnostic report is uploaded to the central production management system via industrial Ethernet or IoT communication links, where it is stored and aggregated in the database of the central production management system. This provides a data foundation for the adaptive optimization closed loop of production parameters in the subsequent step S4. Specifically, this includes using the batch consistency deviation index to identify the quality fluctuation trend of specific process batches, and using the quantitative assessment of institutional health to monitor the performance degradation status of sorting equipment, thereby triggering corresponding pre-maintenance alarms, dynamic adjustment of test parameters, or upstream process optimization instructions, thus realizing the end-to-end data connection and intelligent decision-making closed loop from field testing to production management.

[0058] S4. Based on the batch consistency deviation index and institutional health quantification assessment in the structured diagnostic report, trigger optimized adjustments to production process parameters and dynamic adaptive compensation for sorting machine test parameters; specifically including: S41. Diagnostic report parsing and rule matching, specifically: S411. Real-time reception of the structured diagnostic report uploaded via the industrial Ethernet or IoT communication link in step S33. The data parsing engine of the central production management system performs protocol parsing and field extraction on the report, accurately extracting the two core feedback indicators: batch consistency deviation index and institutional health quantification assessment. At the same time, the intrinsic resistance level, judgment confidence level, dynamic contact quality score, environmental adaptability factor and corresponding process feature encoding vector in the report are retained as auxiliary reference information.

[0059] S412. Input the extracted batch consistency deviation index and the institutional health metric assessment into a multi-level rule matching engine. The engine has a built-in preset multi-level rule library. Each rule in the rule library defines a specific indicator threshold range and the corresponding optimization action type and execution parameters. Specifically, it includes a first threshold range for the institutional health metric assessment, a second threshold range for the persistence of the mean of the batch consistency deviation index, and a third threshold range for the kurtosis and skewness of the statistical distribution of the batch consistency deviation index.

[0060] S413. The matching engine compares the input indicators with the threshold conditions in the rule base one by one. Based on the threshold range in which the indicator value falls, it determines the matching priority and activation status. When a single indicator meets multiple rule conditions at the same time, it resolves conflicts according to the preset priority of the rules and finally determines the unique or combined optimal rule matching result. This provides a decision basis for generating specific dynamic compensation instructions, pre-maintenance alarm instructions or process parameter optimization suggestions in the subsequent step S42.

[0061] S42. Based on the matching results output by the multi-level rule matching engine in step S41, specific closed-loop optimization instructions are generated for different abnormal indicator situations, specifically as follows: S421. When the quantitative assessment of the mechanism health is lower than the preset first threshold, it is determined that the probe mechanism of the current sorting machine has a trend of mechanical performance degradation or decreased contact stability. At this time, a dynamic compensation instruction is generated. This instruction contains a new set value of test contact force for online compensation of mechanism performance degradation, which is quantitatively calculated based on the degree of degradation, or a digital potentiometer gain adjustment coefficient for correcting the drift of the measurement signal. The measurement accuracy is maintained by adjusting the probe pressure or signal amplification factor in real time.

[0062] S422. When five or more consecutive components with the same process characteristic code are detected, and the average batch consistency deviation index is continuously higher than the preset second threshold, it is determined that there is a systematic quality abnormality in the specific process batch. At this time, a pre-maintenance alarm instruction is generated. The instruction clearly identifies the abnormal batch number, its associated upstream process equipment information including the sintering furnace number or the raw body forming mold number, and attaches a list of recommended inspection and maintenance items to trigger targeted equipment maintenance and process troubleshooting.

[0063] S423. In addition, by aggregating the batch consistency deviation index statistics of all production batches within a complete production cycle, the kurtosis and skewness characteristic values ​​of its distribution are calculated. When these statistical moment characteristic values ​​exceed the preset third threshold, it indicates that there is a systematic setting deviation in the process parameters of the preceding sintering or forming process. At this time, process parameter optimization suggestions are generated. These suggestions include specific adjustment parameter ranges for the sintering furnace temperature control curve, forming pressure parameters, or holding time, which are used to guide the adjustment of process parameters in the next production cycle.

[0064] S43. Based on the closed-loop optimization instruction type generated in step S42, accurately distribute it to the corresponding execution terminal and complete deployment and effect tracking; specifically including: S431. Distribute the optimized instructions to different execution terminals according to their type: For dynamic compensation commands, the new contact force setting value contained in the command is sent to the servo driver of the sorting machine through the real-time control bus. The driver adjusts the downward pressure control parameters of the probe mechanism according to the setting value, and at the same time sends the signal gain adjustment coefficient to the digital potentiometer of the resistance measurement unit to complete the online correction of the amplification factor of the measurement signal and realize the real-time hardware compensation for the degradation of mechanism performance. In response to pre-maintenance alarm commands, the commands are pushed to the maintenance work order module of the production management system through the work order interface. An electronic work order is automatically generated that includes the abnormal batch number, the associated sintering furnace number or the base body forming mold number, and the suggested inspection and maintenance items, triggering the preventive intervention process of maintenance personnel. Based on the optimization suggestions for process parameters, the suggestions are pushed to the process management module through the manufacturing execution system interface, and the adjustment range of the previous sintering temperature control parameters or molding pressure parameters is used as the recommended setting value for the next batch of production plan.

[0065] S432. After the optimization instruction is executed, a tracking cycle is initiated for the affected entity, such as a specific device, batch, or process: For a specific sorting machine that has executed dynamic compensation instructions, during the subsequent testing of more than fifty components, the average and variance trends of its mechanism health quantification assessment are continuously collected and recorded to verify the contribution of the compensation effect to the recovery of mechanism stability. For a specific process batch that triggered the pre-maintenance alarm command, after the maintenance work is completed, the batch consistency deviation index of the next batch of production components is statistically analyzed, and the changes in its mean and distribution characteristics are recorded to evaluate the effect of maintenance measures on batch quality consistency. For the preceding sintering or molding processes that have adopted process parameter optimization suggestions, continuously track the overall statistical characteristics of the batch consistency deviation index of the subsequent produced components in step S3 testing, including mean offset and dispersion.

[0066] S433. Quantitatively compare the quantification effect data obtained during the tracking period with the expected optimization target set when generating the instruction in step S42: If the actual measured improvement in the institutional health quantification assessment, the decrease in the batch consistency deviation index, or the improvement in the statistical distribution reaches or exceeds the expected threshold, then the weight coefficient of the rule on which the instruction is based will be increased in the rule base through reinforcement learning mechanism, thereby enhancing the priority of the rule in future decisions. If the actual effect does not meet the expected goal or produces negative side effects, the rule backtracking analysis mechanism will be automatically triggered to re-evaluate the corresponding indicator threshold settings or optimization action parameters in the rule base, adjust the numerical boundaries of the first threshold, second threshold or third threshold, correct the compensation coefficient calculation logic in the dynamic compensation instruction or the adjustment range in the process parameter optimization suggestion, and generate a new rule to be verified based on the adjusted logic and store it in the rule base. Finally, the verified and updated rule base is distributed to all relevant central production management system decision nodes and edge computing decision units through a distributed synchronization mechanism, completing this closed loop of adaptive optimization of production parameters based on diagnostic feedback.

[0067] Example 2 like Figure 2 As shown, a resistance value determination system for an NBC resistance value sorting machine based on machine learning is used to implement the resistance value determination method for the NBC resistance value sorting machine based on machine learning, including: A. Data construction module, used to simultaneously acquire probe contact circuit, microenvironment, mechanism health time-series data, and component process characteristic codes within milliseconds of the test probe contacting the NBC component, and fuse them to generate a micro-time-series profile data package; specifically including: The synchronous triggering unit is used to send a synchronous clock signal to each sensor and acquisition channel the instant the test probe contacts the NBC component; The signal processing unit, connected to the synchronization triggering unit, is used to receive the raw signals from each channel and process them in real time to generate the instantaneous power spectral density sequence of contact resistance, the instantaneous gradient data of microenvironment temperature and humidity, the high-speed sampling sequence of raw resistance measurement, and the time series data of the mechanism's health status. The fusion encapsulation unit, connected to the signal processing unit, is used to align and associate the processed timing data with the process feature encoding vector obtained from the production management system, and encapsulate it to generate an enhanced micro-timing profile data packet.

[0068] B. The model processing module, connected to the data construction module, receives data packets and performs training and inference using a built-in dual-path hybrid neural network model. It employs a multi-objective loss function incorporating physical regularization terms for training to decouple the intrinsic resistance values ​​of components from the data. Specifically, it includes: The architecture building unit is used to build a hybrid neural network model with two paths, where the first path is used to process time series data and the second path is used to process process feature encoding. The training execution unit, connected to the architecture building unit, is used to train the model using labeled data and a physically guided multi-objective loss function until the model converges. The deployment unit, connected to the training execution unit, is used to solidify the trained model parameters into an embedded intelligent decision module.

[0069] C. The intelligent diagnostic module, connected to the model processing module, receives the feature vectors and raw results output by the model inference, and performs parallel calculations of intrinsic resistance level, decision confidence level, dynamic contact quality score, environmental adaptability factor, batch consistency deviation index, and institutional health quantification assessment, and encapsulates them into a structured diagnostic report; specifically including: The inference unit is used to input real-time enhanced micro-time profile data packets into the intelligent decision module to obtain the original output vector of the model and the intermediate feature vectors of each path. The calculation unit, connected to the inference unit, is used to calculate in parallel intrinsic resistance level, decision confidence, dynamic contact quality score, environmental adaptation factor, batch consistency deviation index and institutional health quantification based on the original output vector and intermediate feature vector. The report generation unit, connected to the solution unit, is used to encapsulate various diagnostic indicators into a structured diagnostic report according to a predetermined format.

[0070] D. Closed-loop optimization module, connected to the intelligent diagnostic module, is used to parse reports, generate dynamic compensation instructions, pre-maintenance alarm instructions, or process parameter optimization suggestions based on the diagnostic indicators, and drive actuators to complete parameter adjustments, alarm triggering, or suggestion pushes. Simultaneously, it tracks optimization effects and updates the internal decision rule base; specifically including: The instruction generation unit is used to parse the structured diagnostic report, match and generate dynamic compensation instructions, pre-maintenance alarm instructions or process parameter optimization suggestions according to the preset rule base; The execution tracking unit, connected to the instruction generation unit, is used to distribute various instructions to the corresponding execution terminals and track and record the quantitative data of the effect after the instruction is executed. The self-learning unit, connected to the execution tracking unit, is used to dynamically update the thresholds and parameters in the rule base based on the comparison results of the effect quantification data and the expected target, thereby completing the iterative optimization of the system's decision-making capabilities.

[0071] As can be seen from the above description, the embodiments of the present invention achieve the following technical effects: At the data perception and representation level, by constructing a micro-time-series profile data package, a multi-dimensional, high-fidelity digital mapping of the testing process is achieved, fundamentally solving the industry problems of measurement results being black-boxed and error sources being untraceable. Traditional methods only collect the final resistance reading, discarding valuable dynamic process information. This invention, within a millisecond-level time window, simultaneously collects and fuses the instantaneous power spectral density sequence of contact resistance, instantaneous gradient data of microenvironmental temperature and humidity, time-series data of the mechanism's health status, and process feature encoding vectors. This is equivalent to establishing a complete health check report for each test, not only recording what value was measured, but also precisely recording the equipment state, environmental instant, and contact process at which the measurement was taken. Its direct effect is that, for the first time, interference factors such as contact bounce, probe micro-wear, and instantaneous environmental fluctuations, which are traditionally unquantifiable and considered random noise, are transformed into structured feature signals that can be analyzed and processed by subsequent models, laying an irreplaceable data foundation for accurate error tracing and compensation.

[0072] At the core algorithm and decision logic level, a physically guided dual-path hybrid neural network model achieves a paradigm shift from threshold-based comparison to mechanism-based decoupling, solving the core pain points of high false positive rates and insufficient model reliability in high-precision sorting. Traditional threshold methods or ordinary machine learning models essentially mix various interferences with real signals for fitting or comparison, which is prone to misjudgment under boundary conditions. The model of this invention introduces a physical regularization term, forcing the neural network to learn signal decomposition methods that conform to physical laws such as probe wear during training. Thus, it can reliably decouple the intrinsic resistance characteristics of the chip from contaminated measurement signals, just like a numerical filter. The direct effect is that the decision process has physical interpretability, and the accuracy of judging critical components at the edge of specifications is significantly improved. At the same time, the dual-path design allows the model to understand both the dynamic testing process and the static process background. Its output batch consistency deviation index can be directly correlated with upstream process variations, achieving accurate attribution of quality problems.

[0073] At the level of system value and adaptive capability, by forming a diagnostic-optimization self-learning closed loop, the resistance sorting machine is upgraded from a single quality screening tool to a manufacturing process optimization engine, achieving a qualitative leap from passive detection to proactive intervention and continuous evolution. Traditional sorting machines are information endpoints; this invention, through a closed-loop optimization module, transforms structured diagnostic reports into three types of executable instructions in real time: dynamic compensation, pre-maintenance alarms, and process parameter optimization suggestions, and establishes an execution effect tracking and rule base self-updating mechanism. The resulting systemic effects are: equipment possesses self-healing capabilities, automatically adjusting parameters based on the machine's health to compensate for performance degradation; quality control possesses early warning capabilities, detecting process drift in advance through batch consistency anomalies; and the entire manufacturing system possesses collaborative optimization capabilities, with data from the testing phase automatically guiding upstream process improvements and equipment maintenance plans. This creates a self-reinforcing cycle that continuously reduces variation and improves overall yield and efficiency, achieving true intelligent manufacturing.

[0074] The embodiments and / or implementation methods described above are merely preferred embodiments and / or implementation methods for implementing the technology of the present invention, and are not intended to limit the implementation methods of the technology of the present invention in any way. Any person skilled in the art can make some modifications or alterations to other equivalent embodiments without departing from the scope of the technical means disclosed in the present invention, but these should still be regarded as the technology or embodiments that are substantially the same as the present invention.

[0075] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. The above descriptions are only preferred embodiments of this application. It should be noted that due to the limitations of written expression, while there are objectively infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of this application, and can also combine the above technical features in an appropriate manner. These improvements, modifications, changes, or combinations, or the direct application of the inventive concept and technical solution to other situations without modification, should all be considered within the scope of protection of this application.

Claims

1. A resistance value determination method for an NBC resistance value sorting machine based on machine learning, characterized in that, include: Simultaneously collect probe contact, microenvironment and mechanism health time series data, and fuse them with component process feature encoding to construct micro-time series profile data package; The dual-path hybrid neural network model is trained using the micro-time profile data packet as input, and learns to decouple the intrinsic resistance of the components from the data through a physical guidance mechanism. The trained dual-path hybrid neural network model is used to process real-time data and output a structured diagnostic report that includes intrinsic resistance level, batch consistency deviation index and institutional health quantification assessment. Based on the batch consistency deviation index and institutional health quantification assessment in the structured diagnostic report, the optimization adjustment of production process parameters and the dynamic adaptive compensation of sorting machine test parameters are triggered.

2. The resistance value determination method for the NBC resistance value sorting machine based on machine learning according to claim 1, characterized in that, The synchronous acquisition of probe contact, microenvironment, and mechanism health time-series data, which is then fused with component process feature encoding to construct a micro-time-series profile data package, specifically includes: At the instant the test probe contacts the component pin, the original waveforms of voltage and current in the contact circuit within a millisecond time window, the original signals of temperature and humidity at the test site, the original voltage differential signal of the resistance measurement module, and the original outputs of the micro-force sensor and displacement sensor built into the probe mechanism are collected respectively. Real-time signal processing and feature extraction are performed on the collected data to generate instantaneous power spectral density sequence of contact resistance, instantaneous gradient data of microenvironment temperature and humidity, high-speed sampling sequence of original resistance measurement, and time series data of mechanism health status, respectively. The instantaneous power spectral density sequence of contact resistance, the instantaneous gradient data of microenvironment temperature and humidity, the high-speed sampling sequence of original resistance measurement, and the timing data of the mechanism health status are strictly aligned according to the synchronous clock source and associated with the process feature encoding vector obtained from the production management system. Finally, they are encapsulated to generate a micro-time profile data package.

3. The resistance value determination method for the NBC resistance value sorting machine based on machine learning according to claim 2, characterized in that, The real-time signal processing and feature extraction of the collected data specifically includes: For the original voltage and current waveforms, a fast Fourier transform is performed within a time window to calculate the impedance modulus and phase of each frequency component, thereby generating the instantaneous power spectral density sequence of the contact resistance. For the raw temperature and humidity signals, calculate their instantaneous rate of change relative to the start of the measurement period to generate the instantaneous temperature and humidity gradient data of the microenvironment; The original voltage differential signal is synchronously converted to digital and buffered to form the original resistance measurement high-speed sampling sequence; After aligning and filtering the raw outputs of the micro-force sensor and displacement sensor, the eigenvalues ​​of their covariance matrix over time are calculated to generate time-series data of the mechanism's health status characterizing contact stability.

4. The resistance value determination method for the NBC resistance value sorting machine based on machine learning according to claim 2, characterized in that, Using the aforementioned micro-time profile data packet as input, a dual-path hybrid neural network model is trained. Through a physical guidance mechanism, it learns to decouple the intrinsic resistance values ​​of components from the data, specifically including: Construct the dual-path hybrid neural network model, which includes a first path, a second path, a fusion and output layer; In the training samples, each set of micro-time profile data packets is labeled with its corresponding true intrinsic resistance level obtained by a high-precision static measurement system; at the same time, based on the time series data of the mechanism's health status, a theoretical contact deviation reference value is calculated and generated through a known probe wear mechanical model. The model is trained using the training samples to perform a multi-objective loss function. The accuracy of the model in determining resistance levels and its sensitivity in identifying scenarios such as poor contact and batch anomalies are evaluated using an independent validation set. Once the performance meets the standards, the trained model parameters are solidified into the embedded intelligent judgment module of the sorting machine.

5. The resistance value determination method and system for the NBC resistance value sorting machine based on machine learning according to claim 4, characterized in that, The model is trained using the training samples to perform a multi-objective loss function training, specifically including: A batch of training samples is input into a dual-path hybrid neural network model, and forward propagation calculation is performed to obtain the probability distribution of intrinsic resistance level predicted by the model and the contact state latent feature vector extracted from the first path. Based on the forward propagation results, the main loss term and the physical regularization term are weighted and summed according to preset weight coefficients to obtain the total loss value of this round of training. Based on the total loss value, the gradients of the parameters of each layer of the dual-path hybrid neural network model are calculated using the backpropagation algorithm, and all model parameters are updated using an adaptive optimization algorithm. Repeat the operation to optimize and update all model parameters until the total loss value no longer decreases significantly on the validation set and the physical regularization term is lower than the preset tolerance. At this point, the model training is considered to have converged, and the model parameters at this time are saved.

6. The resistance value determination method for the NBC resistance value sorting machine based on machine learning according to claim 1, characterized in that, The trained dual-path hybrid neural network model is used to process real-time data and output a structured diagnostic report containing intrinsic resistance level, batch consistency deviation index, and institutional health quantification assessment, specifically including: The micro-time profile data packet generated in real time during the selector test cycle of the current NBC component under test is input into the trained dual-path hybrid neural network model to generate an original output vector containing the preliminary prediction value of the intrinsic resistance level. Based on the internal layer output and original output vector of the dual-path hybrid neural network model, multiple diagnostic indicators are solved in parallel, including intrinsic resistance level, judgment confidence, dynamic contact quality score, environmental adaptation factor, batch consistency deviation index and institutional health quantification assessment. The intrinsic resistance level, judgment confidence level, dynamic contact quality score, environmental adaptability factor, batch consistency deviation index, and institutional health quantification assessment are packaged into a structured diagnostic report according to a predetermined format and simultaneously output to the sorting execution mechanism of the sorting machine and the central production management system.

7. The resistance value determination method for the NBC resistance value sorting machine based on machine learning according to claim 6, characterized in that, Based on the inner layer outputs and original output vectors of the aforementioned dual-path hybrid neural network model, multiple diagnostic indicators are calculated in parallel, specifically including: Extract the probability distribution belonging to the resistance category from the original output vector, determine the category corresponding to the maximum probability as the intrinsic resistance level; calculate the Shannon entropy of the probability distribution, and output the corresponding determination confidence based on the preset entropy-confidence mapping table; The hidden layer feature vectors related to the contact state in the first path are input into a pre-trained quality assessment fully connected network to generate the dynamic contact quality score; the hidden layer feature vectors related to the environment in the first path are concatenated with the instantaneous gradient data of microenvironment temperature and humidity, and the environmental adaptation factor is calculated by the environmental compensation coefficient calculation unit. Calculate the Mahalanobis distance between the current batch common feature vector output by the second path and the corresponding vector set of all process feature encoding vector elements in the historical database. Normalize the distance value to generate the batch consistency deviation index. For a specific recurrent layer in the first path that processes the time series data of the institution's health status, extract the activation spectrum of its neurons in time series, calculate the peak attenuation coefficient and frequency band energy ratio of the activation spectrum, and fuse them to generate the quantitative assessment of the institution's health.

8. The resistance value determination method for the NBC resistance value sorting machine based on machine learning according to claim 6, characterized in that, Based on the batch consistency deviation index and institutional health quantification assessment in the structured diagnostic report, optimization adjustments to production process parameters and dynamic adaptive compensation for sorting machine test parameters are triggered, specifically including: The system receives and parses the structured diagnostic report in real time, extracts the batch consistency deviation index and institutional health quantification assessment, and matches the indicators with a preset multi-level rule base. Based on the matching results, specific closed-loop optimization instructions are generated: when the quantitative assessment of the health of the mechanism is lower than the first threshold, a dynamic compensation instruction is generated; when the average batch consistency deviation index of five or more consecutive components with the same process feature code is continuously higher than the second threshold, a pre-maintenance alarm instruction is generated; the statistical distribution of the batch consistency deviation index of all batches in a production cycle is aggregated, and when its kurtosis and skewness exceed the third threshold, process parameter optimization suggestions are generated. Execute specific closed-loop optimization instructions and track the execution results.

9. The resistance value determination method for the NBC resistance value sorting machine based on machine learning according to claim 8, characterized in that, The execution of specific closed-loop optimization instructions and the tracking of execution results specifically include: Based on the type of the optimization instruction, it is distributed to different execution terminals: the new contact force setting value in the dynamic compensation instruction is sent to the servo driver of the sorting machine, and the signal gain adjustment coefficient is sent to the digital potentiometer of the measurement unit; the pre-maintenance alarm instruction is pushed to the maintenance work order module of the production management system to generate a pending work order containing abnormal batch identifiers and suggested inspection items; the process parameter optimization suggestion is pushed to the process management module of the manufacturing execution system. After the optimization instruction is executed, a tracking cycle is initiated for the affected entities: for the sorting machine that has performed dynamic compensation, the mean and variance changes of the mechanical health quantification assessment of more than fifty consecutive components are recorded; for the process batch that triggers the pre-maintenance alarm, the mean and distribution changes of the batch consistency deviation index of the next batch of components are recorded after maintenance is completed; the overall statistical characteristics of the batch consistency deviation index of the components produced by the preceding process that adopted the process parameter optimization suggestion are recorded in subsequent tests. The quantitative effect data obtained from tracking is compared with the expected optimization target: if the effect reaches or exceeds the expectation, the rule weight of the generated instruction is strengthened; if the effect does not meet the expectation or produces side effects, rule backtracking analysis is triggered, the corresponding indicator threshold or optimization action parameter in the rule base is adjusted, and a new rule to be verified is generated; the updated rule base is synchronized to all relevant decision nodes to complete this self-learning optimization loop.

10. A resistance value determination system for an NBC resistance value sorting machine based on machine learning, characterized in that, The resistance determination method for implementing the machine learning-based NBC resistance value sorting machine as described in any one of claims 1-9 includes: The data construction module is used to simultaneously collect probe contact circuit, microenvironment, mechanism health time series data and component process feature encoding within milliseconds of the test probe contacting the NBC component, and fuse them to generate a micro-time profile data package. The model processing module, connected to the data construction module, is used to receive data packets and perform training and inference using a built-in dual-path hybrid neural network model; a multi-objective loss function with fused physical regularization terms is used for training to decouple the intrinsic resistance of components from the data; The intelligent diagnostic module, connected to the model processing module, is used to receive the feature vector and original results output by the model inference, and to solve the intrinsic resistance level, judgment confidence, dynamic contact quality score, environmental adaptation factor, batch consistency deviation index and institutional health quantification assessment in parallel, and encapsulate them into a structured diagnostic report. The closed-loop optimization module, connected to the intelligent diagnostic module, is used to parse the report, generate dynamic compensation instructions, pre-maintenance alarm instructions or process parameter optimization suggestions based on the diagnostic indicators, and drive the actuator to complete parameter adjustment, alarm triggering or suggestion push. At the same time, it tracks the optimization effect and updates the internal decision rule base.

Images