Multi-unit integrated test system for comprehensive performance testing

Abstract

This invention belongs to the field of testing technology, specifically disclosing a multi-unit integrated performance testing system. The system includes a main control module, a multi-channel stimulus generation module, a multi-channel data acquisition module, an online feature extraction and modeling module, and an adaptive testing strategy engine. By introducing the online feature extraction and modeling module, this invention achieves real-time perception and digital modeling of the dynamic response characteristics of the unit under test. This system can break free from dependence on fixed test scripts, continuously learning and updating the characteristic model of each unit during the testing process, providing an accurate data foundation for subsequent adaptive decision-making, and solving the problem of mismatch between test parameters and the characteristics of the unit under test.

Description

Technical Field

[0001] This invention belongs to the field of testing technology, specifically relating to an integrated testing system for the comprehensive performance of multi-unit transmitters. Background Technology

[0002] In high-end equipment fields such as aerospace and precision manufacturing, accurate and efficient testing of the performance of key functional units is a core aspect of ensuring system reliability and safety. This type of testing typically involves applying excitation signals to multiple independent units and simultaneously acquiring their response data to evaluate their dynamic characteristics, accuracy, and consistency, falling under the category of automated testing and performance evaluation technologies.

[0003] Multi-unit parallel testing technology aims to improve testing efficiency and reduce testing costs by simultaneously stimulating and measuring multiple units under test through an integrated system. The basic principle of this technology is to use a unified control platform to coordinate hardware resources such as signal generators and data acquisition units, and to perform cyclic or concurrent operations on multiple unit channels according to a preset test sequence.

[0004] Existing multi-unit test systems often rely on pre-set scripts or human experience for their test procedures and parameter configurations. When faced with various types of test units (UTs) with different dynamic response characteristics, accuracy requirements, or interface protocols, the system lacks the intelligent capability to dynamically analyze and adjust parameters based on real-time test data. This results in a rigid testing process, making it impossible to adaptively optimize the amplitude, frequency, or waveform of the excitation signal to suit the optimal test range for a specific unit, and also making it difficult to identify abnormal units and adjust subsequent test strategies in real time during testing.

[0005] Poor matching between test parameters and unit characteristics can lead to a series of problems, such as insufficient test data validity, prolonged test cycles, or the risk of overstress to sensitive units. Therefore, how to construct an integrated test system that can intelligently sense unit characteristics and dynamically optimize the test process has become a key technical challenge to improve the efficiency, accuracy, and adaptability of multi-unit testing. Summary of the Invention

[0006] The purpose of this invention is to provide an integrated testing system for the comprehensive performance of multi-unit test units, in order to solve the problems of low testing efficiency, insufficient data validity, and potential overstress risk in existing multi-unit parallel testing systems due to rigid testing processes and parameter configurations, which prevent dynamic analysis and adaptive adjustment based on the real-time response characteristics of the tested units.

[0007] This invention provides an integrated testing system for the comprehensive performance of multi-unit transmitters, comprising:

[0008] The system main control module, as the central scheduling and decision-making core of the entire testing system, is used to parse test tasks, coordinate the timing and data interaction of various functional modules, and execute the final test decisions and issue instructions.

[0009] A multi-channel excitation generation module, connected to the system main control module, is used to receive excitation parameter instructions from the system main control module, generate and synchronously output excitation signals to the corresponding unit under test;

[0010] A multi-channel data acquisition module, connected to the system main control module, is used to synchronously acquire the response signals of all tested units under the action of excitation signals, and transmit the acquired raw response data to the system main control module;

[0011] The feature online extraction and modeling module is connected to the system main control module. It is used to receive raw response data from the multi-channel data acquisition module in real time, process and analyze it online, extract feature parameters that characterize the dynamic performance of the unit under test, and build or update the online response model of the unit under test based on these feature parameters.

[0012] An adaptive testing strategy engine, connected to the system's main control module, is used to perform real-time decision calculations based on the online response model output by the feature online extraction and modeling module, combined with preset test objectives and constraints. This generates optimized incentive parameters and test process adjustment instructions for the next test step, which are then fed back to the system's main control module.

[0013] Preferably, the working process of the online feature extraction and modeling module is as follows:

[0014] The raw response data is preprocessed, including noise reduction filtering and baseline correction;

[0015] The rise time, peak value, steady-state value, and overshoot of the response signal are extracted using time-domain analysis.

[0016] The frequency domain analysis method is used to calculate the spectrum of the response signal through fast Fourier transform, and its main frequency, bandwidth and the amplitude of key harmonic components are extracted.

[0017] The extracted time-domain and frequency-domain feature parameter sets are input into a pre-defined recurrent neural network model.

[0018] The recurrent neural network model takes the model state of the previous time step and the current set of feature parameters as input, and outputs an online response model that represents the dynamic characteristics of the tested unit at the current time step through the state update of its internal hidden layer. The online response model is represented in the form of a set of time-varying state vectors and parameter matrices.

[0019] Preferably, the decision calculation process of the adaptive testing strategy engine is implemented based on a reinforcement learning framework;

[0020] This reinforcement learning framework defines the system state as the feature vector of the current online response model and defines actions as adjustments to the output parameters of the multi-channel excitation generation module, including the amplitude, frequency, duty cycle, or waveform type of the excitation signal.

[0021] The reward function is defined as a weighted sum of test efficiency reward, test accuracy reward, and safety penalty;

[0022] The test efficiency reward is positively correlated with the amount of effective feature information acquired per unit time, the test accuracy reward is positively correlated with the signal-to-noise ratio of the feature parameters extracted under the current stimulus, and the safety penalty is triggered when the amplitude of the unit response predicted by the online response model is greater than its preset safety threshold.

[0023] The adaptive testing strategy engine internally maintains a deep Q-network, which continuously receives system states, executes actions, observes the next state, and calculates rewards to update its network weights. In this way, it learns the optimal testing strategy that can maximize long-term cumulative rewards and generates optimized incentive parameters accordingly.

[0024] Preferably, the system's main control module integrates a dynamic test process orchestrator;

[0025] The test process dynamic orchestrator receives test process adjustment instructions from the adaptive test strategy engine. These instructions include adding or deleting subsequent test steps, rearranging their order, or modifying the number of iterations. Based on these instructions, the dynamic orchestrator reconstructs the logical structure of the test sequence in real time and ensures that the reconstructed test sequence meets the overall task completeness constraints and the temporal dependencies between each test step.

[0026] Preferably, the multi-channel excitation generation module includes a waveform synthesis unit and a power amplification unit; the waveform synthesis unit generates a high-precision baseband digital waveform according to digital instructions;

[0027] The power amplifier unit amplifies and impedance-matches the baseband waveform signal according to the interface requirements and driving capability requirements of the unit under test, and finally outputs a physical excitation signal suitable for driving the unit under test.

[0028] Multiple channels are synchronized using a common clock source to ensure strict timing consistency of the excitation signal output.

[0029] Preferably, the multi-channel data acquisition module includes a high-precision analog-to-digital converter array and a synchronous sampling controller;

[0030] The synchronous sampling controller receives a unified trigger signal from the system's main control module to ensure that the analog-to-digital converters of all channels start sampling at the same time.

[0031] After initial formatting, the collected data is transmitted in real time to the system's main control module via a high-speed data bus.

[0032] Preferably, the system further includes an anomaly diagnosis and isolation module;

[0033] The anomaly diagnosis and isolation module monitors the online response model output by the feature online extraction and modeling module in real time. When a sudden change is detected in the response model parameters of a certain unit under test, and the magnitude of the change is greater than the preset anomaly threshold, an anomaly alarm signal is immediately sent to the system main control module.

[0034] In response to the abnormal alarm signal, the system main control module instructs the adaptive test strategy engine to suspend subsequent stimulation of the abnormal unit, and instructs the multi-channel data acquisition module to continue monitoring and acquiring the status of the abnormal unit. At the same time, the test resources are centrally allocated to the remaining normal units to ensure the continued execution of the overall test task.

[0035] Preferably, the system operates in a hierarchical iterative test cycle; each test cycle includes six stages: stimulus application, data acquisition, feature extraction, model update, strategy decision-making, and process orchestration; the system main control module controls the sequential execution of these six stages and applies the optimization instructions output from the strategy decision-making stage to the stimulus application stage of the next test cycle, thereby achieving closed-loop online optimization of the test process.

[0036] Preferably, the network weights of the deep Q-network in the adaptive testing strategy engine are updated through a temporal difference learning algorithm and an experience replay mechanism.

[0037] Preferably, the anomaly diagnosis and isolation module performs anomaly detection using a sliding window statistical process control method. This method continuously records the historical sequence of key parameters in the online response model of each tested unit and calculates the deviation between the current value and the average of the past N periods. If the deviation is greater than 3 times the standard deviation and lasts for M consecutive periods, it is determined to be a parameter mutation.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] 1. This invention introduces an online feature extraction and modeling module, enabling real-time perception and digital modeling of the dynamic response characteristics of the tested unit. This system eliminates reliance on fixed test scripts, continuously learning and updating the characteristic model of each unit during testing, providing a precise data foundation for subsequent adaptive decision-making, and solving the problem of mismatch between test parameters and the characteristics of the tested unit.

[0040] 2. This invention integrates an adaptive testing strategy engine based on a reinforcement learning framework, endowing the system with the ability to autonomously optimize the testing process. This adaptive testing strategy engine can dynamically calculate and adjust the optimal stimulus parameters and testing procedures based on the real-time updated unit model, aiming to maximize testing efficiency and accuracy while minimizing security risks. This enables the testing system to intelligently adapt to units under test with different characteristics and states.

[0041] 3. This invention achieves flexible management and fault tolerance of the test process through a dynamic test process orchestrator in the system's main control module and an independent anomaly diagnosis and isolation module. The system can reorganize test steps in real time according to the instructions of the strategy engine, and quickly isolate fault points and adjust resource allocation when unit anomalies are detected, ensuring the robustness and continuity of complex test task execution and avoiding interruption of the entire test task or generation of invalid data due to problems with individual units. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0043] Figure 2 This is a schematic diagram of the core principle framework of the adaptive testing strategy engine based on reinforcement learning in this invention;

[0044] Figure 3 This is a logical flowchart of the feature online extraction and modeling module in this invention;

[0045] Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the system main control module and various functional modules in this invention;

[0046] Figure 5 This is a schematic diagram of the hierarchical iterative test loop of the system of this invention. Detailed Implementation

[0047] Example 1: Reference Figures 1 to 5 This invention proposes a multi-unit integrated performance testing system. The system's main control module establishes bidirectional communication connections with a multi-channel excitation generation module, a multi-channel data acquisition module, a feature online extraction and modeling module, an adaptive test strategy engine, and an anomaly diagnosis and isolation module via a high-speed data bus, forming a closed-loop feedback and dynamically optimized test execution system. During system operation, all functional modules work collaboratively under the unified scheduling of the main control module, following preset logic or real-time generated instructions, to achieve parallel, high-precision, and adaptive performance testing of multiple tested units.

[0048] In this embodiment, the system main control module serves as the central scheduling and decision-making core of the entire testing system, undertaking multiple responsibilities including task parsing, timing control, resource allocation, instruction issuance, and status monitoring. Before the test task starts, the system main control module first loads the user-defined initial test task description file, which contains metadata such as the number of units under test, interface type, initial stimulus parameter range, safety threshold boundary, and test target priority weight. The system main control module initializes its internal subsystems and sends initial stimulus configuration instructions to the multi-channel stimulus generation module, while configuring parameters such as sampling rate, range, and trigger mode to the multi-channel data acquisition module. During the test execution phase, the system main control module strictly follows a layered iterative test loop mechanism, sequentially driving the execution of six stages: stimulus application, data acquisition, feature extraction, model update, strategy decision-making, and process orchestration, ensuring seamless data flow and state synchronization between each stage.

[0049] In a preferred implementation, the system's main control module integrates a dynamic test flow orchestrator. This orchestrator receives flow adjustment instructions from the adaptive test strategy engine and reconstructs the logical structure of subsequent test steps in real time. For example, when the strategy engine determines that a certain type of test step no longer provides information gain to the currently tested unit, it can instruct the deletion of the step; if a nonlinear abrupt change in the response of a unit is detected, additional fine-scan steps can be inserted. When performing such reconstruction operations, the dynamic orchestrator automatically verifies whether the new flow meets the task completeness constraints and the temporal dependencies between steps, thereby ensuring the engineering rationality and technical integrity of the test logic.

[0050] In a preferred implementation, the multi-channel excitation generation module is directly connected to the system's main control module, responsible for generating and synchronously outputting multiple high-precision, programmable physical excitation signals to the corresponding units under test. This multi-channel excitation generation module internally consists of a waveform synthesis unit and a power amplification unit. The waveform synthesis unit generates baseband digital waveforms based on digital commands issued by the system's main control module, using direct digital frequency synthesis technology. These baseband digital waveforms can be sine waves, square waves, triangle waves, pulse sequences, or any user-defined waveform, with parameters including amplitude, frequency, phase, duty cycle, and modulation scheme, all of which can be switched within microseconds. The generated digital waveform is converted into an analog signal by a digital-to-analog converter and then sent to the power amplification unit.

[0051] As a preferred implementation, the power amplification unit performs power amplification and impedance matching according to the electrical interface specifications of the unit under test, ensuring that the output signal has sufficient driving capability and is not distorted due to reflection. To ensure strict timing consistency of the multi-channel excitation signals, all channels share the same high-stability clock source, which is uniformly provided by the system main control module and distributed to the waveform synthesis unit of each channel through a low-jitter clock distribution network. In actual testing, when the adaptive test strategy engine outputs a new excitation parameter command, the system main control module parses the excitation parameter command and distributes it to the waveform synthesis unit of the corresponding channel. The latter completes the waveform parameter update within the next clock cycle, thereby achieving millisecond-level dynamic adjustment of the excitation signal.

[0052] In a preferred implementation, the multi-channel data acquisition module also establishes a high-speed data link with the system's main control module to synchronously acquire the response signals of all units under test (DUTs) under excitation. This multi-channel data acquisition module consists of a high-precision analog-to-digital converter (ADC) array and a synchronous sampling controller. The output port of each DUT is connected to an independent ADC channel with a resolution of at least 16 bits and a sampling rate exceeding 1 megasamples per second to meet the requirements for accurate capture of rapid dynamic responses. The synchronous sampling controller receives a unified trigger signal from the system's main control module, which is emitted immediately after the excitation application phase of each test cycle, ensuring that all ADCs start sampling within the same nanosecond-level time window, thereby eliminating time offset errors between channels.

[0053] In this embodiment, the acquired raw response data is first pre-formatted locally, including adding metadata such as timestamps, channel identifiers, and excitation parameter snapshots. It is then transmitted in real-time to the system's main control module's memory buffer via PCIe 4.0 or a similar high-speed serial bus. To handle high-throughput data streams, the multi-channel data acquisition module incorporates a dual-buffering mechanism: while one set of data is being uploaded, the next set is already prepared in the local buffer, preventing sampling loss due to bus congestion. Furthermore, this multi-channel data acquisition module supports dynamic range adjustment, automatically switching gain levels based on the signal amplitude acquired in the previous cycle to maximize the signal-to-noise ratio and prevent signal saturation.

[0054] In this embodiment, the online feature extraction and modeling module is a key component for realizing the system's intelligent perception capability, and its logical flow is as follows: Figure 3 As shown. This online feature extraction and modeling module receives raw response data streams from the multi-channel data acquisition module in real time and processes and analyzes them online cycle by cycle. The processing flow is divided into three levels:

[0055] The first level is preprocessing, which includes filtering the original signal using a wavelet threshold denoising algorithm to suppress high-frequency electromagnetic interference and quantization noise; at the same time, baseline correction is performed, and DC drift or slow trend terms are eliminated by sliding window least squares fitting.

[0056] The second level is feature extraction, performed in both the time and frequency domains. In the time domain, the rise time, peak value, steady-state value, and overshoot of the response signal are calculated. In the frequency domain, a Fast Fourier Transform is performed on the preprocessed signal to obtain its spectral distribution, from which the dominant frequency, the -3dB bandwidth (the frequency range where the energy drops to half on both sides of the dominant frequency), and the amplitudes of the first five key harmonic components are extracted.

[0057] The third layer is model construction, which involves inputting the extracted time-domain and frequency-domain feature parameter sets (a total of 12-dimensional vectors) into a pre-defined recurrent neural network model. This recurrent neural network uses a long short-term memory (LSTM) unit structure, and its internal state is continuously updated in each test loop. Let the input feature vector of the k-th test loop be... Let represent the set of real numbers, and let be the hidden state at the previous time step. The current hidden state Updated by the following formula:

[0058] ;

[0059] This represents standard Long Short-Term Memory (LSTM) operations. Ultimately, the online response model output by the online feature extraction and modeling module represents the current hidden state. Together with their corresponding output layer weight matrices, they constitute a time-varying, low-dimensional representation of the dynamic characteristics of the tested unit. This online response model not only includes the static features of the current response but also implicitly contains the evolutionary trend of historical responses, providing a high-information-density state representation for subsequent adaptive decision-making.

[0060] As a preferred implementation method, the adaptive testing strategy engine achieves intelligent decision-making based on a reinforcement learning framework, the principle of which is as follows: Figure 2As shown, this adaptive testing strategy engine models the testing process as a Markov decision process. The system state is defined as the feature vector corresponding to the online response model output by the online feature extraction and modeling module, i.e., the hidden state of the recurrent neural network. The action space is defined as the adjustment amount of the output parameters of the multi-channel excitation generation module, including the amplitude increment, frequency offset, duty cycle change, or waveform type switching command of the excitation signal. The reward function comprehensively considers three dimensions: testing efficiency, testing accuracy, and operational safety. Specifically, the testing efficiency reward is positively correlated with the amount of effective feature information acquired per unit time, which is quantified by calculating the KL divergence between the current feature vector and the historical feature vector set. The testing accuracy reward is positively correlated with the signal-to-noise ratio of the feature parameters extracted under the current excitation; the higher the signal-to-noise ratio, the greater the reward. The safety penalty is triggered when the amplitude of the unit response predicted by the online response model is greater than its preset safety threshold. At this time, the reward function value is significantly reduced to suppress dangerous operations.

[0061] As a preferred implementation, the adaptive testing strategy engine internally maintains a deep Q-network. This deep Q-network takes the system state as input and outputs the expected cumulative reward value for each possible action. During the policy decision phase of each test loop, the engine first selects the action that maximizes the Q-value based on the current state, then executes the action and observes the next state and immediate reward. Finally, it updates the weights of the deep Q-network using an experience replay mechanism and a temporal difference learning algorithm. After several iterations, the deep Q-network converges to the optimal policy that maximizes the long-term cumulative reward, thereby achieving autonomous optimization of the testing process. Notably, the adaptive testing strategy engine supports multi-unit parallel decision-making, meaning that each tested unit independently calculates its optimal incentive parameters, thus achieving refined testing with a "one unit, one policy" approach.

[0062] In this embodiment, the anomaly diagnosis and isolation module acts as a security guardian of the system, continuously monitoring the online response model output by the feature online extraction and modeling module. This module internally maintains an anomaly detector, implemented using sliding window statistical process control. Specifically, for each tested unit, the module continuously records the historical sequence of key parameters (such as main frequency, steady-state value, and overshoot) in its online response model and calculates the deviation between the current value and the average of the past N periods (N is typically 20). If this deviation is greater than 3 times the standard deviation and persists for M consecutive periods (M is typically 3), it is determined to be a parameter mutation. Furthermore, if the mutation magnitude exceeds a preset anomaly threshold, the module immediately sends an anomaly alarm signal to the system's main control module. Upon receiving the abnormal alarm signal, the system's main control module first instructs the adaptive test strategy engine to suspend subsequent updates of the excitation parameters for the abnormal unit, maintaining its excitation signal in a safe standby state. Simultaneously, it instructs the multi-channel data acquisition module to continue low-frequency monitoring of the abnormal unit to track its state evolution. Furthermore, the main control module marks the abnormal unit as "isolated" and excludes it from subsequent resource scheduling, reallocating the computing and communication bandwidth originally allocated to the abnormal unit to other normal units, thus ensuring that the overall test throughput and schedule are not affected. If, after several monitoring cycles, the abnormal unit's response returns to normal, the operator can manually remove the isolation status and reinstate it into the test process.

[0063] As a preferred implementation, the entire system runs in a hierarchical iterative test loop, such as... Figure 5 As shown, each test cycle strictly comprises six stages: the first stage is stimulus application, where the system main control module sends the stimulus parameters for the current cycle to the multi-channel stimulus generation module, which synchronously outputs physical signals; the second stage is data acquisition, where the multi-channel data acquisition module synchronously captures the responses of all tested units under unified triggering; the third stage is feature extraction, where the online feature extraction and modeling module preprocesses and calculates features from the raw data; the fourth stage is model updating, where new features are input into the recurrent neural network to generate an updated online response model; the fifth stage is strategy decision-making, where the adaptive test strategy engine calculates the optimal stimulus parameters for the next cycle based on the new model; and the sixth stage is process orchestration, where the system main control module's dynamic orchestrator updates the logical structure of subsequent test steps according to the strategy engine's process adjustment instructions.

[0064] These six stages are tightly linked in time, forming a closed-loop feedback loop. The system's main control module ensures that each stage is completed within a predetermined time window through a precise timing controller, and performs data integrity verification and state consistency checks during stage switching. Typically, the cycle time of a single test loop can be controlled within 100 milliseconds, thereby supporting high-frequency tracking and optimization of fast dynamic characteristics.

[0065] In practical deployments, this system can be applied to batch performance verification scenarios for complex electromechanical products such as aerospace electronic equipment, electric drive systems for new energy vehicles, and joint modules for industrial robots. For example, when conducting comprehensive performance testing on a batch of servo motor drivers, the system first applies a standard step voltage excitation and collects their current and speed responses. The feature online extraction and modeling module then extracts indicators such as rise time, overshoot, and steady-state error, and constructs an initial dynamic model. After evaluating the initial dynamic model, the adaptive test strategy engine automatically increases the high-frequency sweep excitation to accurately locate the cutoff frequency if it finds that the bandwidth of a certain unit is significantly low. If another unit shows signs of abnormal heating, its excitation power is immediately reduced and isolation monitoring is initiated. The entire process requires no manual intervention; the system autonomously completes the entire process from coarse screening to fine diagnosis, improving testing efficiency and data value.

[0066] In summary, this embodiment constructs a highly integrated, closed-loop optimized, and resilient multi-unit test platform through centralized scheduling of the system main control module, high-precision synchronization of multi-channel excitation and acquisition modules, real-time perception of the online feature extraction and modeling module, intelligent decision-making of the adaptive test strategy engine, and fault-tolerant assurance of the anomaly diagnosis and isolation module. This multi-unit test platform not only solves the inherent problems of rigid parameters, low efficiency, and uncontrollable risks in traditional test systems.

[0067] Example 2: Based on the previous examples, this example enhances the online feature extraction and modeling module and the adaptive testing strategy engine for special application scenarios where the unit under test has strong nonlinear or time-varying characteristics, so as to further improve the system's modeling accuracy and adaptability to complex dynamic behaviors.

[0068] In the online feature extraction and modeling module, a new nonlinear feature extraction subunit is added in addition to the existing time-domain and frequency-domain features. This nonlinear feature extraction subunit is specifically used to identify and quantify nonlinear phenomena in the response of the tested unit, such as hysteresis, dead zone, saturation, and limit cycle oscillations. The nonlinear feature extraction subunit first performs phase plane analysis on the preprocessed response signal and plots its phase trajectory diagram. Subsequently, it characterizes the nonlinear strength and stability of the system by calculating the enclosing area of ​​the phase trajectory, the trajectory asymmetry, and the Lyapunov exponent. Furthermore, this nonlinear feature extraction subunit introduces mutual information calculation to measure the nonlinear correlation between the excitation signal and the response signal, thereby determining whether there are dynamic couplings not captured by the linear model. These nonlinear feature parameters (a total of 8 dimensions) are added to the original 12-dimensional feature vector, forming a 20-dimensional enhanced feature set as input to the recurrent neural network. Correspondingly, the hidden layer dimension of the recurrent neural network is also expanded from 64 dimensions to 128 dimensions to accommodate higher-dimensional input information and improve its representational ability. During the model update phase, the network not only updates the hidden states, but also dynamically adjusts the activation function type of its output layer to better fit different types of nonlinear responses.

[0069] The adaptive testing strategy engine has also been upgraded, expanding its action space to include nonlinear excitation modes. For example, a new "hysteresis scan" action has been added, which superimposes a slowly changing triangular wave onto the excitation signal to excite and measure the hysteresis loop of the unit under test; or a "limit cycle detection" action, which applies a sinusoidal excitation of specific frequency and amplitude to observe whether the system enters a self-sustaining oscillation state. A nonlinear information gain term has also been added to the reward function. This nonlinear information gain term is positively correlated with the variance of the newly extracted nonlinear features, encouraging the strategy engine to actively explore excitation regions that reveal nonlinear characteristics. The structure of the deep Q-network has also been adjusted accordingly. Its input layer has been expanded to receive a 20-dimensional state vector, the hidden layers use residual connections to alleviate the gradient vanishing problem, and the output layer supports a discrete-continuous hybrid action space, allowing for both selection of excitation mode categories and continuous adjustment of its parameters.

[0070] As a preferred implementation, a nonlinear characteristic verification sub-process template is added to the test process dynamic orchestrator of the system's main control module. When the adaptive test strategy engine detects significant nonlinear characteristics in a unit under test, it can instruct the dynamic orchestrator to automatically insert this sub-process. The nonlinear characteristic verification sub-process contains a series of predefined nonlinear excitation sequences, such as multi-frequency composite excitation, random noise excitation, and chaotic signal excitation, designed to comprehensively stimulate the nonlinear dynamics of the unit under test. During this period, the multi-channel data acquisition module increases the sampling rate to 2 megasamples per second and enables 18-bit high-resolution mode to ensure accurate capture of subtle nonlinear phenomena.

[0071] The anomaly diagnosis and isolation module also enhances its ability to identify nonlinear anomalies. For example, when the Lyapunov exponent of a certain unit changes from negative to positive and remains above the threshold, even if its response amplitude does not exceed the limit, it is judged as a potential instability risk, triggering an early warning isolation. The system's main control module then reduces the unit's excitation intensity, starts low-frequency monitoring, and notifies the operator to conduct a manual review.

[0072] Through the aforementioned enhanced design, this embodiment improves the system's testing capability for strongly nonlinear, time-varying units under test. In testing a certain type of piezoelectric ceramic actuator, traditional methods can only obtain its gain and bandwidth in the linear operating region. However, this embodiment successfully identifies its hysteresis characteristics and saturation inflection point under high-voltage excitation and automatically generates targeted compensation parameters, providing a basis for the subsequent design of the control system. This testing capability has significant application value in cutting-edge fields such as precision instruments, microelectromechanical systems (MEMS), and smart materials.

[0073] Example 3: This example focuses on large-scale parallel testing scenarios, specifically the extreme case where the number of tested units reaches hundreds or even thousands of channels. In this scenario, the system faces challenges such as computational resource limitations, communication bandwidth pressure, and decreased test synchronization accuracy. Therefore, this example implements a distributed refactoring of the system architecture to support high-concurrency, low-latency testing requirements.

[0074] The system's main control module is divided into a central master control node and multiple regional coordination nodes. The central master control node is responsible for global task scheduling, policy learning, and process orchestration; each regional coordination node manages a subset (e.g., 64) of the tested units, responsible for local stimulus generation, data acquisition, feature extraction, and preliminary anomaly detection. The regional coordination nodes are interconnected with the central master control node via Gigabit Ethernet or InfiniBand, exchanging status summaries and control commands using a publish-subscribe communication model. The multi-channel stimulus generation module and multi-channel data acquisition module are also divided by region, with each region equipped with independent waveform synthesis and power amplification hardware, as well as an independent analog-to-digital converter array, thereby achieving channel isolation and resource decoupling at the physical level.

[0075] As a preferred implementation, the online feature extraction and modeling module adopts an edge-cloud collaborative architecture. Primary feature extraction is performed on the embedded processor of the regional coordination node to reduce data transmission volume; advanced modeling (such as recurrent neural network inference) is centrally executed by the GPU cluster of the central control node. To reduce communication latency, regional nodes only upload feature vectors, not the original waveforms, and employ differential coding compression technology to transmit only the incremental changes from the previous cycle.

[0076] The adaptive testing strategy engine implements a hierarchical decision-making mechanism. The central master node maintains a global deep Q-network, responsible for learning common policies across units; each regional node maintains a lightweight local policy network to quickly respond to individual differences within its local unit. The local policy networks periodically synchronize weights from the global network and fine-tune them locally. During decision-making, regional nodes first generate preliminary actions based on their local networks, then report to the central node for conflict detection and resource arbitration, and finally, the central node issues a confirmation command.

[0077] As a preferred implementation, the anomaly diagnosis and isolation module adopts a distributed deployment. Regional nodes monitor their local units in real time. Once an anomaly is detected, isolation operations are immediately performed locally, and an anomaly event summary is reported to the central node. The central node aggregates anomaly information from all regions, performs a global health status assessment, and dynamically adjusts the overall testing strategy, such as reducing the testing intensity of batches with high failure rates.

[0078] Through this distributed architecture, this embodiment successfully scales the system to a parallel testing scale of 1024 channels, while keeping the single cycle time within 150 milliseconds, meeting the cycle time requirements of large-scale production lines. In an application on a semiconductor packaging and testing production line, this system enables parallel electrical performance and thermal characteristic testing of all chips on an entire batch of wafers. The testing efficiency is improved compared to traditional serial solutions, and adaptive optimization reduces invalid testing steps, thereby reducing energy consumption and equipment wear.

[0079] The above three embodiments together constitute the complete technical system of the present invention, comprehensively covering various application needs of integrated testing of multi-unit comprehensive performance, from core algorithms and special scenario adaptation to large-scale deployment.

[0080] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-unit integrated performance testing system for conductors, characterized in that, include: The system's main control module is used to parse test tasks, coordinate the timing and data interaction of various functional modules, and execute the final test decisions and issue instructions. A multi-channel excitation generation module, connected to the system main control module, is used to receive excitation parameter instructions from the system main control module, generate and synchronously output excitation signals to the corresponding unit under test; A multi-channel data acquisition module, connected to the system main control module, is used to synchronously acquire the response signals of all tested units under the action of excitation signals, and transmit the acquired raw response data to the system main control module; The feature online extraction and modeling module is connected to the system main control module. It is used to receive raw response data from the multi-channel data acquisition module in real time, process and analyze it online, extract feature parameters that characterize the dynamic performance of the unit under test, and build or update the online response model of the unit under test based on these feature parameters. An adaptive test strategy engine, connected to the system main control module, is used to perform real-time decision calculations based on the online response model output by the feature online extraction and modeling module, combined with preset test objectives and constraints, to generate optimized incentive parameters and test process adjustment instructions for the next test step, and then feed them back to the system main control module. The working process of the online feature extraction and modeling module is as follows: The raw response data is preprocessed, including noise reduction filtering and baseline correction; The rise time, peak value, steady-state value, and overshoot of the response signal are extracted using time-domain analysis. The frequency domain analysis method is used to calculate the spectrum of the response signal through fast Fourier transform, and its main frequency, bandwidth and the amplitude of key harmonic components are extracted. The extracted time-domain and frequency-domain feature parameter sets are input into a pre-defined recurrent neural network model. The recurrent neural network model takes the model state of the previous time step and the current feature parameter set as input, and outputs an online response model that represents the dynamic characteristics of the tested unit at the current time step through the state update of its internal hidden layer. The online response model is represented in the form of a set of time-varying state vectors and parameter matrices. The decision-making process of the adaptive testing strategy engine is implemented based on a reinforcement learning framework; The reinforcement learning framework defines the system state as the feature vector of the current online response model and defines the action as the adjustment amount of the output parameters of the multi-channel excitation generation module, including the amplitude, frequency, duty cycle or waveform type of the excitation signal; The reward function is defined as a weighted sum of test efficiency reward, test accuracy reward, and safety penalty; The test efficiency reward is positively correlated with the amount of effective feature information acquired per unit time, the test accuracy reward is positively correlated with the signal-to-noise ratio of the feature parameters extracted under the current stimulus, and the safety penalty is triggered when the amplitude of the unit response predicted by the online response model is greater than its preset safety threshold. The adaptive test strategy engine maintains a deep Q-network internally. By continuously receiving system status, executing actions, observing the next state, and calculating rewards, it updates its network weights, thereby learning the optimal test strategy that can maximize long-term cumulative rewards and generating optimized incentive parameters accordingly. The system operates in a hierarchical iterative test loop; each test loop includes six stages: stimulus application, data acquisition, feature extraction, model update, strategy decision-making, and process orchestration; the system's main control module controls the sequential execution of these six stages and applies the optimization instructions output from the strategy decision-making stage to the stimulus application stage of the next test loop, thereby achieving closed-loop online optimization of the test process.

2. The integrated testing system for multi-unit conductive comprehensive performance according to claim 1, characterized in that, The system's main control module integrates a dynamic test process orchestrator. The test process dynamic orchestrator receives test process adjustment instructions from the adaptive test strategy engine. These instructions include adding or deleting subsequent test steps, rearranging their order, or modifying the number of iterations. Based on the test process adjustment instructions, the dynamic orchestrator reconstructs the logical structure of the test sequence in real time and ensures that the reconstructed test sequence meets the overall completeness constraints of the task and the temporal dependencies between each test step.

3. The integrated testing system for multi-unit conductive comprehensive performance according to claim 1, characterized in that, The multi-channel excitation generation module includes a waveform synthesis unit and a power amplification unit; the waveform synthesis unit generates a high-precision baseband digital waveform according to digital instructions; The power amplifier unit amplifies and impedance-matches the baseband waveform signal according to the interface requirements and driving capability requirements of the unit under test, and finally outputs a physical excitation signal suitable for driving the unit under test. Multiple channels are synchronized using a common clock source to ensure strict timing consistency of the excitation signal output.

4. The integrated testing system for multi-unit conductive comprehensive performance according to claim 1, characterized in that, The multi-channel data acquisition module includes a high-precision analog-to-digital converter array and a synchronous sampling controller; The synchronous sampling controller receives a unified trigger signal from the system's main control module to ensure that the analog-to-digital converters of all channels start sampling at the same time. After initial formatting, the collected data is transmitted in real time to the system's main control module via a high-speed data bus.

5. The integrated testing system for multi-unit conductive comprehensive performance according to claim 1, characterized in that, The system also includes an anomaly diagnosis and isolation module; The anomaly diagnosis and isolation module monitors the online response model output by the feature online extraction and modeling module in real time. When a sudden change is detected in the response model parameters of a certain unit under test, and the magnitude of the change is greater than the preset anomaly threshold, an anomaly alarm signal is immediately sent to the system main control module. In response to the abnormal alarm signal, the system main control module instructs the adaptive test strategy engine to suspend subsequent stimulation of the abnormal unit, and instructs the multi-channel data acquisition module to continue monitoring and acquiring the status of the unit. At the same time, the test resources are centrally allocated to the remaining normal units to ensure the continued execution of the overall test task.

6. The integrated testing system for multi-unit conductive comprehensive performance according to claim 1, characterized in that, The deep Q-network in the adaptive testing strategy engine updates its network weights through a temporal difference learning algorithm and an experience replay mechanism.

7. The integrated testing system for multi-unit conductive comprehensive performance according to claim 5, characterized in that, The anomaly diagnosis and isolation module performs anomaly detection using a sliding window statistical process control method. This method continuously records the historical sequence of key parameters in the online response model of each tested unit and calculates the deviation between the current value and the average of the past N periods. If the deviation is greater than 3 times the standard deviation and lasts for M consecutive periods, it is determined to be a parameter mutation.

Images