Standardized test method and test platform for power module assembly

Abstract

This invention discloses a standardized testing method and platform for modular power supply components. The method includes constructing a probabilistic time automaton model to uniformly represent test specifications, platform resources, response probabilities, and observation labels; performing formal verification on the model to check its consistency and resource matching; adaptive pruning based on belief states and information gain, terminating the test early when a confidence threshold is met or the information gain is below the threshold; when a failure is detected, backtracking the actual path and constructing a counterfactual path, calculating and outputting the root cause score; real-time acquisition of transient waveforms, extraction of time-frequency feature fingerprints, matching with a fault database, and updating the belief state; incrementally updating the fingerprint database and model parameters and re-verifying when there is no match or a specification change. This invention achieves dynamic optimization of the test process, precise root cause localization, real-time identification of transient faults, and platform self-evolution, significantly improving test efficiency and diagnostic accuracy.

Description

Technical Field

[0001] This invention belongs to the field of power modular component testing technology, and particularly relates to standardized testing methods and testing platforms for power modular components. Background Technology

[0002] Standardized testing of modular power supply components (such as DC-DC converters, communication power modules, and redundant server power modules) is crucial for ensuring their performance, reliability, and safety. Currently, the industry widely employs automated testing systems that perform item-by-item testing on power modules using predefined test sequences. Test items include voltage accuracy, load regulation, ripple, overvoltage / overcurrent protection, and timing. Some advanced systems integrate data acquisition and statistical process control functions, capable of recording test data and generating reports. Furthermore, the International Electrotechnical Commission (IEC) and national standardization organizations have published numerous testing standards (such as GB / T and AQG324), providing a unified basis for testing procedures. Some manufacturers have also developed modular testing platforms that support multi-station parallel testing and achieve automated control through instruments such as programmable electronic loads, oscilloscopes, and signal generators. However, the aforementioned existing technologies still have the following shortcomings in practical applications:

[0003] First, existing systems lack the ability to self-consistently verify the test specifications themselves. Inconsistencies, redundancies, or incompatibility with platform resources in the specifications often rely on manual review, which is prone to oversight and leads to batch misjudgments. Second, the fixed-sequence test process cannot be dynamically adjusted according to the actual quality of the modules. A large number of redundant items cause waste of time and energy, and there is a lack of a risk-controlled early termination mechanism. Third, when the test fails, the system only reports the parameters that exceed the standard, and cannot distinguish the root cause (such as module defects, poor fixture contact, environmental drift, or instrument deviation), requiring maintenance personnel to spend time troubleshooting one by one. Fourth, for millisecond-level transient fault waveforms, existing single features in the frequency or time domain are difficult to classify effectively, and fault identification relies on expert experience. Fifth, the system cannot automatically evolve in the face of new faults or standard version updates, requiring manual modification of the test program. Sixth, each functional module (verification, optimization, diagnosis, identification) operates independently, lacking unified model support, resulting in prominent information silos. Therefore, there is an urgent need for a standardized testing method and platform for power modular components that can achieve self-consistent verification of specifications, adaptive test pruning, rapid root cause location, real-time identification of transient faults, and incremental learning. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, this invention provides a standardized testing method and testing platform for power modular components, which solves the problems of lack of standard self-consistency, fixed testing procedures, difficulty in root cause location, difficulty in transient fault identification, inability to incrementally evolve, and isolated functional modules in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Standardized testing methods and testing platforms for modular power supply components, including:

[0007] S1: Construct a probabilistic time-time automaton model; the probabilistic time-time automaton model is used to uniformly represent test specifications, platform resource constraints, module response probability distribution, and observation labels;

[0008] S2: Perform formal verification on the probabilistic time automata model, check the self-consistency, redundancy and matching with platform resources of the test specification, and generate counterexamples and correction suggestions when violations are detected;

[0009] S3: During test execution, adaptive pruning is performed based on the probabilistic time automata model: a belief state is defined, which represents the probability distribution of the module's pass / fail status; test items are dynamically selected based on the current belief state and the information gain of each test item, and the test is terminated in advance when the pass / fail probability or fail / fail probability in the belief state reaches a first threshold, or when the maximum information gain of the remaining test items is lower than a second threshold.

[0010] S4: When the test is deemed unqualified, perform counterfactual root cause localization: backtrack the actual execution path in the probabilistic time automaton model, construct counterfactual paths for the variables on the path, calculate the root cause score for each variable, and output the root cause list.

[0011] S5: During the test execution process, transient waveforms are acquired in real time, the time-frequency feature fingerprints of the waveforms are extracted, and they are matched with the pre-stored fault fingerprint database. The belief state is updated according to the matching result.

[0012] S6: When a mismatched waveform occurs or the test specification changes, perform incremental learning: update the parameters of the fingerprint database and the probabilistic time automata model, and re-perform formal verification.

[0013] Preferably, the formal verification includes: verifying the reachability, contradiction, redundancy, and resource constraint satisfaction of the model; wherein, contradiction verification is achieved by checking whether there is a test execution path that simultaneously satisfies a certain specification requirement and its logical negation; redundancy verification is achieved by comparing the observable equivalence of the model after deleting test actions; when a violation is detected, a minimum conflict set algorithm is used to output a correction suggestion.

[0014] Preferably, the counterfactual root cause localization includes: backtracking the actual execution path in the probabilistic time automata model; constructing counterfactual paths for variables that affect the test results on the path, setting the variable as an ideal reference value, and keeping other variables unchanged; calculating the attribution causal effect of each variable, and obtaining the root cause score by combining the likelihood ratio of the actual path and the counterfactual path; and outputting the root cause list and correction suggestions in descending order of score.

[0015] Preferably, the adaptive pruning includes: calculating the information gain of each executable test item based on the current belief state and probability function; and selecting the test item with the largest information gain to execute, provided that the hard constraints marked by the formal verification are satisfied; the hard constraints are test items that are required to be executed in the specification.

[0016] Preferably, the incremental learning includes: when waveform matching fails and is labeled, adding the new fingerprint to the index library and updating the observation label function and probability function of the probabilistic time automata model; re-performing formal verification to check whether the new fault mode is covered by the existing specification, and generating supplementary specification suggestions if not; and relearning the probability function after completing a preset number of module tests, and re-verifying and adjusting the model structure when the test specification changes.

[0017] Preferably, a standardized test platform for power modular components includes:

[0018] A unified model building module is used to build and store probabilistic time automata models. The models uniformly represent test specifications, platform resource constraints, module response probability distributions, and observation labels.

[0019] A formal verification engine, connected to the unified model building module, is used to perform temporal logic verification on the probabilistic time automaton model, detect self-consistency, redundancy and reachability, and output counterexamples and correction suggestions.

[0020] An adaptive pruning controller, connected to the unified model building module and the formal verification engine, is used to select test items based on belief state and information gain during test execution, and terminate the test when the pass probability or fail probability in the belief state reaches a first threshold or the maximum information gain of the remaining test items is lower than a second threshold.

[0021] The counterfactual reasoning locator, connected to the adaptive pruning controller and the unified model building module, is used to backtrack the actual path, construct the counterfactual path, calculate the root cause score, and output the root cause list when the test fails.

[0022] The time-frequency feature fingerprint recognition and online matching module is connected to the adaptive pruning controller and the unified model construction module. It is used to collect transient waveforms in real time, extract time-frequency feature fingerprints, match them with the fault fingerprint database, and feed back the matching results to the adaptive pruning controller to update the belief state.

[0023] The incremental learning and model adaptation module is connected to the time-frequency feature fingerprint recognition and online matching module, the formal verification engine, and the unified model construction module. It is used to update the parameters and structure of the fingerprint database and the probabilistic time automaton model when unmatched waveforms or test specifications change.

[0024] Preferably, the formal verification engine and the adaptive pruning controller have a bidirectional data channel: the formal verification engine pre-calculates whether each test action is marked as a hard constraint that must be executed, and sends the mark to the adaptive pruning controller; the adaptive pruning controller retains the test items marked with hard constraints when making pruning decisions.

[0025] Preferably, the time-frequency fingerprint recognition and online matching module includes: a signal processing unit, an approximate nearest neighbor index library, and a matching decision unit; the signal processing unit is used to perform time-frequency decomposition and energy spectrum calculation on the sampled waveform; the matching decision unit outputs the fault type and matching confidence, and sends the matching result to the adaptive pruning controller; the incremental learning and model adaptation module is configured to: after a matching failure and annotation, add a new fingerprint vector to the index library, and trigger the formal verification engine to re-examine the test specification coverage.

[0026] Preferably, the probabilistic time automaton model includes a state, an initial state, a clock variable, a data variable, an action, a transition edge, a probability function, an observation label function, and a termination state; the construction of the model includes: parsing the standardized test document to generate the state and transition edge, setting the clock guard condition of the transition edge according to the time requirements in the test specification, determining the probability function based on historical test data statistics, and determining the observation label function based on transient waveform samples.

[0027] Preferably, the incremental learning includes: when waveform matching fails and is labeled, adding the new fingerprint to the index library and updating the observation label function and probability function of the probabilistic time automata model; re-performing formal verification to check whether the new fault mode is covered by the existing specification, and generating supplementary specification suggestions if not; and relearning the probability function after completing a preset number of module tests, and re-verifying and adjusting the model structure when the test specification changes.

[0028] The technical effects and advantages of the standardized testing method and testing platform for the power modular component of this invention are as follows:

[0029] 1. This invention, by constructing a probabilistic time automaton model and performing formal verification, can automatically detect contradictions, redundancies, and unreachable states with platform resources in test specifications, generating counterexample paths and minimum conflict set correction suggestions. This avoids batch misjudgments caused by errors in the specifications themselves, ensuring the testing process is logically consistent and physically executable, significantly reducing the workload of manual specification review.

[0030] 2. This invention utilizes belief states and information gain-driven adaptive pruning to dynamically select test items with the highest information gain for execution, and terminates the test early when the probability of passing or failing reaches a first threshold or the maximum information gain of the remaining items falls below a second threshold. Combined with hard constraint marking to force the retention of mandatory items, this ensures test compliance while saving an average of 30% to 60% of testing time within the user-defined maximum false positive rate, achieving a balance between efficiency and risk.

[0031] 3. This invention is based on a counterfactual reasoning-based root cause localization method. It backtracks the actual execution path within a probabilistic time automaton, constructs counterfactual paths for variables affecting test results, calculates the attribution causal effect and likelihood ratio to obtain root cause scores, and outputs a root cause list with confidence levels and corrective suggestions in sequence. This mechanism can distinguish between correlation and causation, reducing fault location time from tens of minutes to minutes, and avoiding misleading repairs.

[0032] 4. This invention extracts waveform fingerprints through multi-resolution time-frequency decomposition and matches them with a fault fingerprint database using an approximate nearest neighbor index. This enables millisecond-level identification of transient fault types such as output capacitor short circuits, MOSFET breakdowns, and inductor saturation. The matching results directly update the belief state and drive subsequent test pruning, achieving online fault diagnosis and dynamic adaptation of the testing process.

[0033] 5. This invention employs an incremental learning and model adaptation mechanism. When a mismatched waveform appears, the fingerprint database and observation label function are updated after annotation; the probability function is periodically re-estimated; and when the test specification changes, automatic re-verification and supplementary specification suggestions are generated. This mechanism enables the platform to adapt to new fault modes and standard version updates without manual reprogramming, significantly improving the detection rate of new faults and ensuring the accuracy and stability of long-term operation.

[0034] 6. All functions of this invention share the same probabilistic time automata model, unifying the representation of test specifications, platform resources, probability distributions, and observation labels. Formal verification results directly serve the pruning hard constraints, the belief states during the pruning process provide priors for counterfactual reasoning, and fingerprint matching results update probability parameters, forming a complete closed-loop collaboration that avoids information gaps and interface mismatches during multi-system integration. Attached Figure Description

[0035] Figure 1 This is a system architecture diagram of the standardized testing method and testing platform for power modular components proposed in this invention;

[0036] Figure 2 This is a schematic diagram of the probabilistic-time automaton model of the standardized testing method and testing platform for power modular components proposed in this invention. Detailed Implementation

[0037] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0038] 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 entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "includes..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0039] refer to Figure 1-2This invention provides a standardized testing method and platform for modular power supply components, aiming to solve problems in existing technologies such as lack of self-consistency in testing specifications, fixed testing procedures, difficulty in locating root causes of faults, lag in transient fault identification, and the inability of the system to adaptively evolve. The method includes: constructing a probabilistic time automaton model to uniformly represent testing specifications, platform resource constraints, module response probability distribution, and observation labels; performing formal verification on the model to detect self-consistency, redundancy, and resource matching, generating counterexamples and correction suggestions; defining belief states during test execution, dynamically selecting test items based on information gain, and terminating the test early when the probability of passing or failing reaches a first threshold or the maximum information gain of the remaining items is lower than a second threshold; when a failure is determined, backtracking the actual path in the model and constructing counterfactual paths for key variables, calculating root cause scores, and outputting a root cause list; real-time acquisition of transient waveforms, extraction of time-frequency feature fingerprints, and matching with a fault fingerprint database, updating the belief state based on the matching results; updating the fingerprint database and model parameters and re-verifying when unmatched waveforms appear or testing specifications change. This invention's platform includes a unified model building module, a formal verification engine, an adaptive pruning controller, a counterfactual reasoning locator, a time-frequency feature fingerprint recognition and online matching module, and an incremental learning and model adaptation module. This invention enables adaptive pruning of the testing process, precise root cause localization, real-time identification of transient faults, and long-term adaptive evolution of the system, significantly improving testing efficiency and diagnostic accuracy.

[0040] The modular power supply components described in this invention include, but are not limited to, DC-DC converters, AC-DC power modules, communication power modules, server redundant power modules, and vehicle power modules. Standardized testing items include voltage accuracy testing, load regulation testing, ripple testing, overvoltage protection testing, undervoltage protection testing, overcurrent protection testing, short-circuit protection testing, and timing testing. The invention will be described in detail below with reference to specific embodiments and comparative examples, but the implementation of this invention is not limited thereto.

[0041] Example 1

[0042] Construction and formal verification of probabilistic time automata models.

[0043] Objective: To construct a unified probabilistic time automaton model and formally verify the test specifications to ensure their self-consistency, non-redundancy, and compatibility with platform resources, and to prevent batch misjudgments due to specification errors.

[0044] Implementation Platform: A unified model building module and a formal verification engine are employed. The unified model building module is used to build and store probabilistic time automata models; the formal verification engine is connected to the unified model building module and is used to perform temporal logic verification on the probabilistic time automata models, detect self-consistency, redundancy, and reachability, and output counterexamples and correction suggestions.

[0045] Implementation steps:

[0046] S1: Construct a probabilistic time automaton model. Taking the standardized test procedure of the power supply module under test as an example, the test specifications include: waiting 20 milliseconds for the output voltage to stabilize after power-on; performing ripple measurement, requiring the peak-to-peak value of the ripple to be less than or equal to one percent of the output voltage; performing load regulation test, with the load jumping from zero to 100%, requiring the voltage drop to be less than or equal to three percent of the output voltage and recovering within 5 milliseconds; performing overcurrent protection test, with protection triggered when the load current reaches 1.2 times the rated current. The above specifications are constructed as a probabilistic-time automaton model, specifically including: defining a set of nine states: idling, power-on, waiting for stabilization, ripple measurement, load step, recovery monitoring, overcurrent test, pass, and fail. The initial state is idling. Two clock variables are defined. Data variables include output voltage, output current, ripple peak-to-peak value, drop percentage, recovery time, ambient temperature, and contact resistance. The action set includes six actions: power-on, ripple measurement, applying load step, monitoring recovery, injecting overcurrent, and decision. Transition edges are defined (e.g., the transition condition from waiting for stabilization to ripple measurement is that the clock variable is greater than or equal to 20 milliseconds). A probability function is defined (derived from historical test data, e.g., the overcurrent protection trigger probability is 99.5%). An observation label function is defined (associating the ripple measurement state with a normal waveform fingerprint). The termination states are defined as pass and fail. After the model is constructed, it is stored as an Extensible Markup Language (Extreme Markup Language) file.

[0047] The observation label function is constructed as follows: For each state or transition, at least one hundred typical transient waveform samples (including normal mode and various fault modes) are collected. An eight-dimensional energy spectrum fingerprint is extracted using the time-frequency decomposition method described in Example 4. The mean value of each mode fingerprint is taken as the centroid, and the centroid vector is stored in the observation label function. For example, for the normal ripple state, the observation label is an eight-dimensional vector with values ​​of 0.70, 0.18, 0.05, 0.02, 0.01, 0.01, 0.01, and 0.01, respectively. For the inductor saturation fault, the observation label is another eight-dimensional vector with values ​​of 0.60, 0.20, 0.08, 0.05, 0.03, 0.02, 0.01, and 0.01, respectively. During matching, the real-time fingerprint is compared with the observation label of each state or transition; the closest one below the threshold is the matched state.

[0048] S2: Perform formal verification of the probabilistic-time automata model. Use a model checker to verify the following:

[0049] Reachability verification: Check if there are any terminal states that are unreachable from the initial state. The test found that all paths from idle to qualified are reachable, and there are no unreachable states.

[0050] Conflict verification: The specification requirements and their logical negations are encoded into probabilistic time computation tree logic formulas. Taking "voltage drop less than or equal to three percent of output voltage" as an example, it is encoded as follows: for all execution paths, the final percentage drop when a load step occurs is less than or equal to 0.03%. Its logical negation is that there exists an execution path where the percentage drop when a load step occurs is greater than 0.03. The model checker constructs a product automaton to search for whether there exists a test execution path that simultaneously satisfies the specification requirements and their logical negations. The verification result is a conflict-free path.

[0051] Redundancy verification: After deleting the "ripple measurement" action, a new model is generated. The observable equivalence is compared with the original model. If inequivalence is found, it is determined to be non-redundant.

[0052] Resource constraint satisfaction: Set the platform's maximum sampling rate to two million points per second and the maximum power to five hundred watts, and verify that the resource variables do not exceed the limits under all reachable states.

[0053] When a violation is detected (e.g., a specification requires an input voltage of 50 volts while the platform's maximum is 30 volts), the model checker outputs a counterexample path and recommends, based on a minimum conflict set algorithm, "adjust the input voltage threshold to 30 volts." Operators can then modify the test specifications according to this recommendation.

[0054] A bidirectional data channel with hard constraints is established between the formal verification engine and the adaptive pruning controller. During the offline phase, the formal verification engine analyzes the test specification and generates a Boolean hard constraint flag for each test action: if the specification uses mandatory terms such as "must," "shall," or "cannot be omitted" to require the action, it is flagged as true; otherwise, it is flagged as false. This flag is sent to the adaptive pruning controller in the form of a key-value pair list and stored. During online testing, the adaptive pruning controller calculates the information gain and checks the hard constraint flags: if an action is flagged as true, it is retained in the candidate set regardless of its information gain; if it is flagged as false and the information gain is below the second threshold, it can be skipped. In this embodiment, the overcurrent protection test is flagged as true.

[0055] Implementation Results: Verification results show that all termination states are achievable, with no contradictions, unnecessary redundancy, and resource constraints are met. Operators can adjust the test specifications based on the suggested corrections to avoid subsequent misjudgments. Hard constraint markers are correctly transmitted and stored.

[0056] Example 2

[0057] Adaptive pruning based on belief states and information gain.

[0058] Objective: To dynamically skip test items with low information gain while ensuring the confidence level of the judgment, thereby significantly shortening the test time and improving the efficiency of batch testing.

[0059] Implementation platform: An adaptive pruning controller is used, which is connected to a unified model building module and a formal verification engine. It is used to select test items based on belief state and information gain during test execution, and terminate the test when the pass probability or fail probability in the belief state reaches the first threshold or the maximum information gain of the remaining test items is lower than the second threshold.

[0060] Implementation steps:

[0061] S1: Model construction has been completed (same as Example 1).

[0062] S2: Formal verification has been performed, and the overcurrent protection test has been marked as a hard constraint (true).

[0063] S3: Test a batch of one thousand DC-to-DC converter modules of the same model.

[0064] Belief State Definition and Update: The belief state is a two-dimensional vector. The first component represents the probability of passing, and the second component represents the probability of failing. The sum of the two components is one. The initial value is set to a pass probability of 98% and a failure probability of 2% based on historical pass rates. The information entropy is calculated as follows: the negative pass probability multiplied by the logarithm of the pass probability (base 2), minus the failure probability multiplied by the logarithm of the failure probability (base 2). The initial information entropy is approximately 0.14. The belief state is updated using Bayes' theorem: the posterior pass probability equals the conditional probability multiplied by the prior pass probability divided by the normalization factor, where the normalization factor is the sum of the conditional probability multiplied by the prior pass probability and the conditional probability multiplied by the prior failure probability; the posterior failure probability is updated similarly. The conditional probability represents the probability of obtaining the test result given the true pass status and current environmental parameters (temperature, estimated contact resistance). This value is obtained from the probability function. Based on historical data, when the module is in a qualified state, the probability that the ripple measurement value is lower than 1% of the output voltage is 99%, and the probability that it is higher than 1% is 1%. When the module is in a non-qualified state, the probability that the ripple measurement value is lower than 1% of the output voltage is 10%, and the probability that it is higher than 1% is 90%.

[0065] Information gain calculation: The information gain of a test item is defined as the expected value of the information entropy of the current belief state minus the information entropy of the belief state after the test. This expected value is obtained by weighting and summing all possible test results according to the probability predicted by the current belief. Specifically, a random forest classifier is trained using a historical test dataset (complete test records of the same module model). The current belief state and context variables are input, and the estimated information gain for each test item is output. In this embodiment, the estimated results are: ripple measurement information gain is 0.08, load regulation measurement information gain is 0.05, overcurrent protection test information gain is 0.03, and timing test information gain is 0.01.

[0066] Pruning execution: The controller selects the ripple measurement with the maximum information gain to perform. The measured ripple value is 0.8% of the output voltage (lower than the 1% threshold). Apply the Bayesian formula: the posterior qualified probability is equal to 99% multiplied by 98% divided by (99% multiplied by 98% plus 10% multiplied by 2%), and the calculation result is 99%. The belief state is updated to a qualified probability of 99% and an unqualified probability of 1%. Among the remaining items, the information gain of load regulation measurement is 0.03, the information gain of overcurrent protection test is 0.02, the information gain of timing test is 0.005, and the maximum information gain is 0.03. The user sets the first threshold (qualified or unsatisfactory probability threshold) to 95%, the second threshold (lower limit of information gain) to 0.01, and the maximum allowable misjudgment rate to 0.1%. The current qualified probability of 99% is greater than or equal to 95%, the termination condition is satisfied, the test is terminated in advance, and the product is determined to be qualified.

[0067] Threshold calibration method: Offline cross-validation is used to determine the first threshold and the second threshold. The historical test dataset (containing the complete test results of 10,000 modules) is divided into a training set and a validation set at a ratio of 8:2. The pruning process is simulated on the training set, the first threshold is traversed from 90% to 99% (with a step size of 1%), the second threshold is traversed from 0.01 to 0.10 (with a step size of 0.01), and the misjudgment rate of each parameter combination on the validation set is calculated, that is, the proportion of products that are determined to be qualified after pruning but are actually unqualified. The parameter combination that meets the condition that the misjudgment rate is less than or equal to the maximum allowable misjudgment rate and has the shortest test time is selected. The calibration results are the first threshold of 95% and the second threshold of 0.01.

[0068] S4, S5, S6: Since the test determination result is qualified, root cause localization, fingerprint matching and incremental learning are not triggered.

[0069] Implementation effect: The average test time is reduced from 120 seconds per module of the traditional method to 48 seconds per module, saving 60%. The estimated misjudgment rate on the validation set is 0.05%, which is lower than the user-set 0.1%.

[0070] Embodiment 3

[0071] Test failure root cause localization based on counterfactual reasoning.

[0072] Implementation purpose: When the test is unqualified, quickly locate the root cause leading to the failure (whether it is a problem with the module itself, poor fixture contact, ambient temperature offset, or uncalibrated instrument), and output a root cause list with confidence levels and correction suggestions.

[0073] Implementation platform: It adopts a counterfactual reasoning locator, which is connected to an adaptive pruning controller and a unified model building module. It is used to backtrack the actual path, construct counterfactual paths, calculate root cause scores, and output a root cause list and correction suggestions when the test fails.

[0074] Implementation steps:

[0075] S1 to S3: During the load regulation test, the voltage drop of a certain module is 4.5% of the output voltage (the standard is less than or equal to 3%), which is unacceptable. After the belief state is updated, the probability of failure reaches 95%, triggering step S4.

[0076] S4: Perform counterfactual root cause analysis. Backtrack the actual execution path in the probabilistic-time automaton model: from idling to power-on, then to stabilization, then to load step, and finally to failure. Variables affecting the test results along the path include ambient temperature, contact resistance, and the equivalent series resistance of the output capacitor. Construct a counterfactual path for each variable:

[0077] For ambient temperature: set this variable to 25 degrees Celsius (ideal value, taken from standard environmental conditions), keep other variables unchanged, simulate the counterfactual path on the model, and obtain a voltage drop of 4.2 percent of the output voltage.

[0078] For contact resistance: set this variable to five milliohms (standard upper limit), keep other variables unchanged, and the simulation shows that the voltage drop is four point zero percent of the output voltage.

[0079] For the equivalent series resistance of the output capacitor: set this variable to 0.1 ohms (nominal value), the simulation shows that the voltage drop is 2.8 percent of the output voltage (qualified).

[0080] Calculate the attribution causality effect for each variable: The attribution causality effect equals the absolute value of the deviation of the counterfactual path minus the absolute value of the deviation of the actual path, then divided by the absolute value of the deviation of the actual path. The deviation of the actual path is (4.5 percent - 3 percent) divided by 3 percent equals 0.5 percent. The attribution causality effect of ambient temperature is the absolute value of (0.4 percent - 0.5 percent) divided by 0.5 percent equals 0.2 percent; the attribution causality effect of contact resistance is the absolute value of (0.33 percent - 0.5 percent) divided by 0.5 percent equals 0.34 percent; the attribution causality effect of the equivalent series resistance of the output capacitor is the absolute value of (0 percent - 0.5 percent) divided by 0.5 percent equals 1.0 percent.

[0081] Calculate the likelihood ratio: The likelihood ratio equals the probability of the counterfactual path divided by the probability of the actual path. The probabilities of each path are obtained by multiplying the probability functions in the probabilistic-time automaton model. In this embodiment, the likelihood ratio for ambient temperature is 0.8, the likelihood ratio for contact resistance is 0.6, and the likelihood ratio for the equivalent series resistance of the output capacitor is 0.9.

[0082] The root cause score equals the attribution causal effect multiplied by the likelihood ratio: Ambient temperature score is 0.16, contact resistance score is 0.20, and output capacitor equivalent series resistance score is 0.90. The root cause list is output in descending order of score: First, increased output capacitor equivalent series resistance (score 0.90), it is recommended to replace the capacitor batch or extend the lifespan test; Second, high contact resistance (score 0.20), it is recommended to clean the fixture probes; Third, high ambient temperature (score 0.16), it is recommended to turn on the air conditioner or check the heat dissipation.

[0083] S5, S6: This embodiment does not trigger fingerprint matching and incremental learning.

[0084] Implementation Results: After maintenance personnel replaced the capacitors based on the root cause list, the module passed the test. The fault location time was reduced from 45 minutes in traditional manual troubleshooting to 5 minutes.

[0085] Example 4

[0086] Transient fault fingerprint recognition based on multi-resolution time-frequency decomposition.

[0087] Objective: To identify the fault type corresponding to transient waveforms in real time (such as inductor saturation, output capacitor short circuit, and metal-oxide-semiconductor field-effect transistor breakdown) without manual waveform analysis, and to achieve millisecond-level fault classification.

[0088] Implementation Platform: A time-frequency feature fingerprint recognition and online matching module is employed, connected to an adaptive pruning controller and a unified model building module. This module is used to acquire transient waveforms in real time, extract time-frequency feature fingerprints, match them with a fault fingerprint database, and feed the matching results back to the adaptive pruning controller to update the belief state. This module includes a signal processing unit, an approximate nearest neighbor index library, and a matching decision unit. The signal processing unit performs time-frequency decomposition and energy spectrum calculation on the sampled waveforms, which can be achieved using various time-frequency analysis methods such as wavelet packet decomposition (in this embodiment), wavelet transform, short-time Fourier transform, or empirical mode decomposition. The approximate nearest neighbor index library is constructed using local sensitive hashing or product quantization. The matching decision unit outputs the fault type and matching confidence level.

[0089] Implementation steps:

[0090] S1 to S4: Perform overcurrent protection tests according to the procedure and collect transient waveforms.

[0091] S5: Real-time acquisition of transient waveforms at a sampling rate of two million points per second, with a waveform length of 2048 points. Extraction of time-frequency feature fingerprints: The waveform undergoes three-level wavelet packet decomposition (using the db4 wavelet basis) to obtain eight frequency band nodes. The energy of each node is calculated as the sum of the squares of all wavelet packet coefficients for that node. The total energy is the sum of the energies of all eight nodes. The normalized fingerprint vector is the energy of each of the eight nodes divided by the total energy. In this embodiment, the normalized fingerprint vector is: 0.727, 0.168, 0.049, 0.019, 0.010, 0.005, 0.002, 0.001. Similarity matching is performed between the locality-sensitive hash index and a pre-stored fault fingerprint database. The fault database contains centroids for three types of faults: output capacitor short-circuit centroids (0.45, 0.25, 0.12, 0.08, 0.05, 0.03, 0.01, 0.01); MOSFET breakdown centroids (0.03, 0.02, 0.15, 0.12, 0.10, 0.07, 0.04, 0.02); and inductor saturation centroids (0.60, 0.20, 0.08, 0.05, 0.03, 0.02, 0.01, 0.01). Euclidean distances are calculated as follows: distance to the output capacitor short-circuit centroid is 0.31, distance to the MOSFET breakdown centroid is 0.52, and distance to the inductor saturation centroid is 0.18. The preset matching threshold is 0.20. If the distance to the centroid of inductor saturation is less than 0.20, the match is successful, and the fault type "Inductor Saturation" is output. The matching confidence level is defined as an exponential function of the negative square of the distance to the base of the natural constant divided by the variance; in this embodiment, it is set to 0.92. The matching result is fed back to the adaptive pruning controller to update the belief state, marking the cause of failure as inductor saturation, and skipping subsequent test items unrelated to inductor saturation.

[0092] S6: Match successful, no incremental learning required.

[0093] Implementation results: The time from waveform acquisition to fault output is less than 0.5 milliseconds, the matching accuracy is over 95%, and the system automatically skips subsequent irrelevant tests, further shortening the test time.

[0094] Example 5

[0095] Incremental learning and model adaptation.

[0096] Objective: To enable the testing platform to adapt to new failure modes and changes in testing specifications, achieve long-term self-evolution, and reduce manual maintenance.

[0097] Implementation platform: It adopts an incremental learning and model adaptation module, which is connected to the time-frequency feature fingerprint recognition and online matching module, the formal verification engine and the unified model building module. It is used to update the parameters and structure of the fingerprint database and the probabilistic time automaton model when there are unmatched waveforms or changes in test specifications.

[0098] Implementation steps:

[0099] S1 to S5: The system ran for one month, processed 5,000 modules, and accumulated 30 waveforms that failed to match. The minimum distance between the fingerprint vectors of these waveforms and the centroids of all known faults was greater than the threshold of 0.20.

[0100] S6: Perform incremental learning, which includes the following four parts:

[0101] (i) Labeling and Fingerprint Database Update: Engineers view the unmatched waveforms through the graphical user interface, labeling twenty of them as the new fault mode "Driver IC Failure" and ten as known modes but with noise interference (not added to the new class). After labeling, the system adds the twenty new fingerprint vectors to the Local Sensitive Hash Index, re-clusters and updates the centroids, and simultaneously updates the observation label function of the probabilistic time automata model to associate new fingerprints with relevant states.

[0102] (II) Probability Function Update: After completing every 100 module tests, the system automatically relearns the probability function. The actual trigger success rate of the overcurrent protection test within this period was 99.2%, while the historical rate was 99.5%. A time-decay weighted average was adopted: new data weighted at 0.3, old data weighted at 0.7. After the update, the probability parameter equals 0.3 multiplied by 0.992 plus 0.7 multiplied by 0.995, resulting in 0.9941. The model parameters are updated accordingly.

[0103] (III) Re-verification of Specification Coverage: When the new fault mode "Driver IC Failure" is added, the system triggers the formal verification engine to re-examine whether the existing test specifications cover this mode. Verification revealed that existing test items (ripple, load regulation, overcurrent protection, timing, etc.) cannot effectively induce the failure characteristics of the driver IC. The engine automatically generated a supplementary specification suggestion: "It is recommended to add a duty cycle test for the light-load drive signal under high-voltage input, with the threshold set to a duty cycle change rate exceeding 15%." After operator review, the new test item is added to the standardized test suite, and the probabilistic-time automaton model structure is adjusted accordingly: a new state "Driver Signal Test" and corresponding transition edges are added.

[0104] (iv) Response to Test Standard Changes: When the test standard version is upgraded from version 1 to version 2 (the ripple limit is tightened from 1% to 0.8% of the output voltage), the operator updates the test documentation, and the system re-parses and generates the probabilistic-time automaton model and performs formal verification. Verification revealed that the new standard requires ripple to be less than or equal to 0.8% of the output voltage, but the platform's current sampling accuracy is 0.5% of the output voltage, with a quantization error of approximately ±0.25%, making it impossible to accurately distinguish between 0.8% and 0.79% ripple values. The minimum conflict set algorithm outputs a correction suggestion: "Adjust the ripple limit to 0.9% or upgrade the oscilloscope to 12-bit resolution." After the user selects to upgrade the hardware, the system automatically adjusts the guard conditions in the model, updating the ripple comparison threshold from 0.8% to 0.9%.

[0105] Implementation Results: The system operates without manual reprogramming for extended periods, automatically adapting to new faults and specification changes, maintaining testing accuracy and efficiency. Verification shows that incremental learning increased the detection rate of new faults from 0% to 98%.

[0106] Comparative Example 1

[0107] Traditional fixed-sequence testing method.

[0108] Purpose of implementation: To compare with the adaptive pruning method of this invention, highlighting the efficiency improvement and root cause localization advantages of this invention.

[0109] Implementation platform: Traditional automated testing systems use hard-coded fixed test sequences, lacking probabilistic time automata, adaptive pruning, counterfactual reasoning, fingerprint matching, and incremental learning.

[0110] Implementation steps (fixed order, no step numbers):

[0111] Power-on stability in 20 milliseconds;

[0112] Ripple measurement;

[0113] Load regulation test;

[0114] Overcurrent protection test;

[0115] Timing tests;

[0116] Short circuit protection test;

[0117] Efficiency testing;

[0118] Determine whether it is qualified or unqualified.

[0119] Test subjects: one thousand DC-DC converters of the same model from the same batch as in Example 2, with a pass rate of 98%.

[0120] Implementation Results: Each module had to perform all eight tests, with an average test time of 120 seconds per module and a total test time of 33.3 hours. The false positive rate was zero. However, when tests failed, only direct results such as "ripple exceeds the limit" and "voltage drop exceeds the limit" were reported, without pinpointing the root cause. Operators had to manually change fixtures, adjust the environment, and replace components one by one to troubleshoot, with an average fault location time of 45 minutes. The system was unable to adapt to new faults or specification changes, requiring manual modification of the test program.

[0121] Compared with Examples 1-5 and Comparative Example 1, the technical advantages of this invention can be summarized as follows:

[0122] The comparative example uses a fixed-order, hard-coded test sequence, requiring each module to complete all eight tests, resulting in an average test time of 120 seconds per module and a total time consumption of 33.3 hours. Even when module quality is good, many redundant tests (such as the high-current protection test for low-power modules, which is guaranteed to pass) are still forced to run, wasting time and energy. Example 2 uses an adaptive pruning controller to dynamically select test items based on belief state and information gain. Tests are terminated early when the pass probability reaches the first threshold (95%) or the information gain of remaining items falls below the second threshold (0.01), reducing the average test time to 48 seconds per module, saving 60% of the time. Simultaneously, offline cross-validation controls the false positive rate to 0.05% (lower than the user-set 0.1%). Example 5 further introduces real-time matching of time-frequency feature fingerprints, automatically skipping irrelevant test items after identifying the fault type, further shortening the test cycle. In contrast, the comparative example cannot dynamically adjust the process according to the actual module quality, resulting in low efficiency and no risk control mechanism.

[0123] The comparative example can only report surface results such as "ripple exceeds the standard," and cannot pinpoint the root cause. When the test fails, the operator needs to manually check possible causes such as fixtures, environment, and components, which takes an average of 45 minutes. When the test specification changes or a new type of fault occurs, the comparative example needs to manually modify the test program, which has a long response cycle and is prone to errors. Example 1 uses a formal verification engine to automatically detect the self-consistency, redundancy, and matching with platform resources of the specification in the offline stage, and outputs counterexamples and correction suggestions to avoid specification errors from the source. Example 3 uses a counterfactual root cause locator to backtrack the actual execution path in the PTA model, construct counterfactual paths, calculate attribution causal effects and likelihood ratios, and output a root cause list with confidence (such as "output capacitor ESR increased, score 0.90"), shortening the fault location time to less than 5 minutes. Example 5 utilizes an incremental learning module to automatically update the fingerprint database, re-estimate the probability function, trigger specification coverage re-verification, and generate supplementary suggestions. This enables the platform to adapt to new types of faults (such as driver IC failure) and specification version upgrades (such as tightening ripple limits). Long-term operation requires no manual reprogramming, and the detection rate of new faults increases from 0% to 98%. The above comparisons fully demonstrate that this invention significantly outperforms traditional methods in terms of testing efficiency, diagnostic accuracy, and system adaptability.

[0124] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0125] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A standardized testing method for modular power supply components, characterized in that, include: S1: Construct a probabilistic time-time automaton model; the probabilistic time-time automaton model is used to uniformly represent test specifications, platform resource constraints, module response probability distribution, and observation labels; S2: Perform formal verification on the probabilistic time automata model, check the consistency, redundancy and matching of the test specification with platform resources, and generate counterexamples and correction suggestions when violations are detected; S3: During test execution, adaptive pruning is performed based on the probabilistic time automata model: a belief state is defined, which represents the probability distribution of the module's pass / fail status; test items are dynamically selected based on the current belief state and the information gain of each test item, and the test is terminated in advance when the pass / fail probability or fail / fail probability in the belief state reaches a first threshold, or when the maximum information gain of the remaining test items is lower than a second threshold. S4: When the test is deemed unqualified, perform counterfactual root cause localization: backtrack the actual execution path in the probabilistic time automaton model, construct counterfactual paths for the variables on the path, calculate the root cause score for each variable, and output the root cause list. S5: During the test execution process, transient waveforms are acquired in real time, the time-frequency feature fingerprints of the waveforms are extracted, and they are matched with the pre-stored fault fingerprint database. The belief state is updated according to the matching result. S6: When a mismatched waveform occurs or the test specification changes, perform incremental learning: update the parameters of the fingerprint database and the probabilistic time automata model, and re-perform formal verification.

2. The standardized testing method for power modular components as described in claim 1, characterized in that, The formal verification includes verifying the reachability, contradiction, redundancy, and resource constraint satisfaction of the model; among which, contradiction verification is achieved by checking whether there is a test execution path that simultaneously satisfies a certain specification requirement and its logical negation; redundancy verification is achieved by comparing the observable equivalence of the model after deleting test actions; when a violation is detected, a minimum conflict set algorithm is used to output a correction suggestion.

3. The standardized testing method for power modular components as described in claim 1, characterized in that, The counterfactual root cause localization includes: backtracking the actual execution path in the probabilistic time automata model; constructing counterfactual paths for variables that affect the test results on the path, setting the variable as the ideal reference value, and keeping other variables unchanged; calculating the attribution causal effect of each variable, and obtaining the root cause score by combining the likelihood ratio of the actual path and the counterfactual path; and outputting the root cause list and correction suggestions in descending order of score.

4. The standardized testing method for power modular components as described in claim 1, characterized in that, The adaptive pruning includes: calculating the information gain of each executable test item based on the current belief state and probability function; and selecting the test item with the largest information gain to execute, provided that the hard constraints marked by the formal verification are satisfied; the hard constraints are the test items that are required to be executed in the specification.

5. The standardized testing method for power modular components as described in claim 1, characterized in that, The incremental learning includes: when waveform matching fails and is labeled, adding the new fingerprint to the index library and updating the observation label function and probability function of the probabilistic time automata model; re-performing formal verification to check whether the new fault mode is covered by the existing specification, and generating supplementary specification suggestions if not; and relearning the probability function after completing a preset number of module tests, and re-verifying and adjusting the model structure when the test specification changes.

6. A standardized test platform for the power modular component according to any one of claims 1-5, characterized in that, include: A unified model building module is used to build and store probabilistic time automata models. The models uniformly represent test specifications, platform resource constraints, module response probability distributions, and observation labels. A formal verification engine, connected to the unified model building module, is used to perform temporal logic verification on the probabilistic time automaton model, detect self-consistency, redundancy and reachability, and output counterexamples and correction suggestions. An adaptive pruning controller, connected to the unified model building module and the formal verification engine, is used to select test items based on belief state and information gain during test execution, and terminate the test when the pass probability or fail probability in the belief state reaches a first threshold or the maximum information gain of the remaining test items is lower than a second threshold. The counterfactual reasoning locator, connected to the adaptive pruning controller and the unified model building module, is used to backtrack the actual path, construct the counterfactual path, calculate the root cause score, and output the root cause list when the test fails. The time-frequency feature fingerprint recognition and online matching module is connected to the adaptive pruning controller and the unified model construction module. It is used to collect transient waveforms in real time, extract time-frequency feature fingerprints, match them with the fault fingerprint database, and feed back the matching results to the adaptive pruning controller to update the belief state. The incremental learning and model adaptation module is connected to the time-frequency feature fingerprint recognition and online matching module, the formal verification engine, and the unified model construction module. It is used to update the parameters and structure of the fingerprint database and the probabilistic time automaton model when unmatched waveforms or test specifications change.

7. The standardized testing platform for the power modular component as described in claim 6, characterized in that, The formal verification engine and the adaptive pruning controller have a bidirectional data channel: the formal verification engine pre-calculates whether each test action is marked as a hard constraint that must be executed, and sends the mark to the adaptive pruning controller; the adaptive pruning controller retains test items marked with hard constraints when making pruning decisions.

8. The standardized testing platform for the power modular component as described in claim 6, characterized in that, The time-frequency fingerprint recognition and online matching module includes a signal processing unit, an approximate nearest neighbor index library, and a matching decision unit. The signal processing unit is used to perform time-frequency decomposition and energy spectrum calculation on the sampled waveform. The matching decision unit outputs the fault type and matching confidence level, and sends the matching result to the adaptive pruning controller. The incremental learning and model adaptation module is configured to add a new fingerprint vector to the index library after a matching failure and annotation, and trigger the formal verification engine to re-examine the test specification coverage.

9. The standardized testing platform for the power modular component as described in claim 6, characterized in that, The probabilistic time automaton model includes a state, an initial state, a clock variable, a data variable, an action, a transition edge, a probability function, an observation label function, and a termination state. The construction of the model includes: parsing the standardized test document to generate the state and transition edge, setting the clock guard condition of the transition edge according to the time requirements in the test specification, determining the probability function based on historical test data statistics, and determining the observation label function based on transient waveform samples.

10. The standardized testing platform for the power modular component as described in claim 6, characterized in that, The incremental learning includes: when waveform matching fails and is labeled, adding the new fingerprint to the index library and updating the observation label function and probability function of the probabilistic time automata model; re-performing formal verification to check whether the new fault mode is covered by the existing specification, and generating supplementary specification suggestions if not; and relearning the probability function after completing a preset number of module tests, and re-verifying and adjusting the model structure when the test specification changes.

Images