High-precision control module detection and data analysis system and method

Abstract

The invention relates to the technical field of electronic control system testing and diagnosis, and discloses a high-precision control module detection and data analysis system and method, and the system comprises a data collection and excitation execution module which is used for carrying out data interaction with a tested control module and applying an excitation signal; the dynamic divergence calculation module is used for calculating the dynamic divergence representing the internal dynamic relation deviation degree of the tested control module according to the real-time state data of the tested control module and by comparing with a preset reference state space tensor; and the predicted deviation degree calculation module is used for calculating the predicted deviation degree between the actual state and the predicted state of the tested control module according to a preset state observer model. According to the method, a closed-loop self-adaptive test framework is adopted, active exploration is carried out by taking maximization of dynamic divergence as a target, deep faults which are difficult to expose in a conventional test and are caused by dynamic coupling failure can be efficiently found, and the depth and coverage rate of fault detection are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of electronic control system testing and diagnostic technology, specifically to a high-precision control module testing and data analysis system and method. Background Technology

[0002] With the rapid development of modern industry, especially in automotive electronics, aerospace, and high-end manufacturing, high-precision control modules, such as electronic control units (ECUs) and flight controllers, are becoming increasingly complex in their internal structure and function. These modules integrate a large number of hardware circuits and millions of lines of software code, and bear increasingly stringent functional safety responsibilities. Therefore, how to conduct comprehensive, efficient, and in-depth testing and verification to ensure their reliability under various extreme and complex operating conditions has become a key link in product development and quality control processes, and also a huge technical challenge.

[0003] Currently, the mainstream testing method for such high-precision control modules in the industry mainly relies on open-loop functional testing based on design requirements. This method typically involves test engineers pre-writing a series of fixed test cases or test sequences according to the product specifications. During testing, the system applies preset stimulus signals to the control module under test according to the script and verifies whether its output response meets the expected specifications. This method is effective for verifying known functional points that are explicitly defined in the design specifications.

[0004] However, with the exponential increase in the complexity of control modules, the inherent limitations of this traditional testing method have become increasingly apparent. The breadth and depth of its test coverage heavily rely on the prior knowledge and engineering experience of the test engineers. Faced with the vast, even infinite, state space of the module under test, pre-set test cases often only cover the tip of the iceberg. Especially for deep-seated, interconnected defects caused by the dynamic coupling of multiple system variables at specific time sequences or by weak transient disturbances, this "follow-the-map" open-loop testing approach often proves ineffective, because these failure modes are often unknown, unexpected, and not explicitly described in the design documents.

[0005] In addition, the industry also employs some auxiliary analysis methods, such as setting simple static or dynamic thresholds for monitoring key output signals, or recording massive amounts of data during testing for subsequent offline manual or semi-automatic analysis. These methods also have significant shortcomings. While simple threshold monitoring can detect some superficial faults, it is essentially a "knowing what, but not why" approach, failing to reveal the root cause of the fault and resulting in severely insufficient diagnostic depth. Offline analysis of massive amounts of data often traps engineers in a "data ocean, information island" dilemma, leading to low analysis efficiency. More importantly, for those elusive, intermittent faults, even if accidentally captured in data records, the precise boundary conditions that triggered them are extremely difficult to reconstruct, making stable fault reproduction a daunting task and posing a significant obstacle to subsequent debugging and repair work.

[0006] Therefore, existing technologies generally lack a closed-loop intelligent testing and diagnosis method that can proactively explore unknown defects, conduct in-depth analysis from the perspective of the system's inherent laws, accurately correlate fault phenomena with physical roots, and ultimately achieve stable fault reproduction. This has become a key technical bottleneck restricting the efficiency and quality improvement of high-reliability control module R&D. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a high-precision control module detection and data analysis system and method, aiming to solve the technical problem that existing testing technologies, due to their reliance on preset test cases, are unable to proactively discover, accurately diagnose, and stably reproduce deep-seated, transient, and correlated defects in control modules caused by multi-variable dynamic coupling.

[0008] To achieve the above objectives, the present invention provides a high-precision control module detection and data analysis system and method, comprising:

[0009] The data acquisition and excitation execution module is used to interact with the control module under test and apply excitation signals;

[0010] The dynamic divergence calculation module is used to calculate the dynamic divergence, which characterizes the degree of deviation of the internal dynamic relationship of the control module under test, based on the real-time state data of the control module under test and by referring to a preset reference state space tensor.

[0011] The prediction deviation calculation module is used to calculate the prediction deviation between the actual state and the predicted state of the controlled module under test according to a preset state observer model.

[0012] An adversarial stimulus generation module, connected to the dynamic divergence calculation module and the prediction deviation calculation module, is used to generate and control the data acquisition and stimulus execution module to apply new stimulus signals based on the dynamic divergence and the prediction deviation.

[0013] Preferably, the adversarial incentive generation module is specifically used for:

[0014] When the dynamic divergence shows an increasing trend and the prediction deviation exceeds a preset deviation threshold, the effectiveness of the current excitation signal adjustment is confirmed, and a new excitation signal is generated based on the effective adjustment.

[0015] Preferably, the dynamic divergence calculation module includes:

[0016] The instantaneous state tensor reconstruction unit is used to calculate the instantaneous state space tensor of the control module under test in real time based on the real-time state data.

[0017] The difference tensor calculation unit is used to calculate the difference tensor between the instantaneous state space tensor and the reference state space tensor;

[0018] A dynamic divergence calibration unit is used to calculate the dynamic divergence based on the difference tensor.

[0019] Preferably, the prediction deviation calculation module is specifically used for:

[0020] By deploying the state observer model in an online environment, the predicted state is generated based on the real-time historical state and current excitation signal of the controlled module under test, and the prediction deviation is calculated.

[0021] Preferably, it also includes a reference benchmark offline modeling module for use before the test begins:

[0022] The reference state space tensor is generated based on healthy sample data;

[0023] And based on the healthy sample data, the state observer model is trained.

[0024] Preferably, the offline modeling module for the reference benchmark includes:

[0025] A reference state space tensor generation unit is used to calculate the dynamic coupling relationship between state variables in the health sample data under different time delays, so as to generate the reference state space tensor; and

[0026] A predictive state observer training unit is used to train the state observer model to learn the behavioral response patterns of the health sample data.

[0027] Preferably, it also includes a root cause correlation diagnosis module, used to perform root cause analysis after the adversarial stimulus generation module determines the final adversarial stimulus vector.

[0028] Preferably, the root cause correlation diagnostic module includes:

[0029] The fault reproduction path recording unit is used to record the final adversarial stimulus vector;

[0030] The root cause correlation analysis unit is used to locate the dynamic coupling relationship that causes functional failure by analyzing the final differential tensor generated under the final adversarial excitation vector.

[0031] Preferably, the data acquisition and stimulus execution module includes:

[0032] The signal excitation unit is used to apply excitation to the control module under test according to the new excitation signal generated by the adversarial excitation generation module;

[0033] The data acquisition unit is used to acquire the real-time status data of the control module under test in a highly synchronous manner.

[0034] This invention also provides a method for high-precision control module detection and data analysis, the method comprising the following steps:

[0035] The data acquisition and excitation execution module interacts with the control module under test and applies excitation signals.

[0036] The step of calculating the dynamic divergence includes: calculating the dynamic divergence, which characterizes the degree of deviation of the internal dynamic relationship of the control module under test, based on the real-time state data of the control module under test and by referring to a preset reference state space tensor.

[0037] The step of calculating the prediction deviation includes: calculating the prediction deviation between the actual state and the predicted state of the controlled module under test according to a preset state observer model;

[0038] The step of generating adversarial stimuli includes: responding to the calculation results of the dynamic divergence and the predicted deviation, and generating a new stimuli signal based on the dynamic divergence and the predicted deviation, and controlling the data acquisition and stimuli execution module to apply the new stimuli signal, wherein the effectiveness of generating the new stimuli signal depends on the dynamic divergence showing an increasing trend and the predicted deviation exceeding a preset deviation threshold.

[0039] This invention provides a high-precision control module detection and data analysis system and method. It has the following beneficial effects:

[0040] 1. This invention constructs a closed-loop adaptive testing framework consisting of modules for data acquisition and stimulus execution, dynamic divergence calculation, predicted deviation calculation, and adversarial stimulus generation. This framework proactively and intelligently explores the state space of the control module under test, rather than relying on pre-set, limited open-loop test cases. This adversarial exploration mechanism, by aiming to maximize dynamic divergence, guides the test system to systematically approach the most vulnerable operating points of the module under test. This allows for the discovery of deep-seated, transient, and correlated faults that are difficult to expose under normal operating conditions, resulting from the failure of multivariable dynamic coupling relationships. This significantly improves the depth and coverage of fault detection.

[0041] 2. This invention, by setting up a root cause correlation diagnosis module, after determining the final adversarial stimulus vector that can stably reproduce the fault, further analyzes the final difference tensor to achieve automatic correlation from the fault phenomenon to the root cause. This module can directly analyze and locate the element with the largest deviation in the high-dimensional tensor, pinpointing which two specific state variables, at what specific time delay, have experienced functional failure in their dynamic coupling relationship. This transforms the diagnostic results from vague phenomenological descriptions into concrete and actionable engineering instructions, providing precise guidance for subsequent repairs and design optimizations, and significantly improving the accuracy of the diagnosis.

[0042] 3. This invention introduces a dual collaborative constraint mechanism of dynamic divergence and prediction deviation, providing a highly robust decision-making basis for the generation and confirmation of adversarial stimuli. The effectiveness of an stimulus adjustment must simultaneously satisfy two independent conditions: "the intrinsic physical laws are violated" (dynamic divergence increases) and "the external behavioral pattern is unpredictable" (prediction deviation exceeds a threshold). This combined "spear and shield" strategy effectively filters out artifacts or non-critical anomalies caused by fluctuations in a single indicator, ensuring that the system seeks real and significant potential defects, thereby significantly improving the confidence of the detection results and reducing the false alarm rate.

[0043] 4. This invention establishes an offline modeling module for a reference benchmark during the offline phase before testing begins. From a data-driven perspective, this module creates a reference state space tensor representing "absolute health" and a predictive state observer. This modeling-then-comparison approach provides the system with an objective and quantitative basis for defining "normal" and "abnormal" conditions, overcoming the limitations of traditional rule-based or threshold-based methods. This not only provides a stable and reliable benchmark for online real-time diagnosis but also makes the method readily applicable to different types of control modules, exhibiting good portability and scalability.

[0044] 5. This invention, through its fault reproduction path recording unit, accurately records and archives the adversarial stimulus vectors that ultimately trigger faults. This provides an effective technical means to address the long-standing pain point in industry where intermittent and sporadic faults are difficult to reproduce. Engineers can use the recorded stimulus vectors to deterministically reproduce fault scenarios at any time, thereby enabling detailed analysis, debugging, and verification. This not only greatly facilitates subsequent R&D and repair work but also lays a solid foundation for establishing an enterprise-level fault case library and automated regression testing system, improving the efficiency and determinism of the overall R&D testing process. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the system framework of the present invention;

[0046] Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Please see the appendix Figure 1 and Figure 2 This invention provides a high-precision control module detection and data analysis system and method, including:

[0049] Data acquisition and incentive execution module;

[0050] In this embodiment, the data acquisition and excitation execution module serves as the interface layer for physical interaction between the system described in this invention and the control unit under test (DUT). It is the physical basis for realizing online closed-loop test diagnosis and provides real-time data sources and precise instruction execution capabilities for subsequent processing such as dynamic divergence calculation, prediction deviation calculation, and adversarial excitation generation.

[0051] Functionally, this module can be further divided into two main parts that cooperate with each other: the signal excitation unit and the data acquisition unit.

[0052] Signal excitation unit:

[0053] The core responsibility of the signal excitation unit in this embodiment is to accurately convert the digital and abstract excitation commands output by the upper-level adversarial excitation generation module into physical electrical signals or bus communication messages with time sequence applied to the control module under test.

[0054] The excitation command generated by the adversarial excitation generation module is, in form, a multidimensional controllable excitation vector u(t), which can be expressed as:

[0055] u(t) = [u1(t), u2(t), ..., u M (t)] u (t)=[u1(t),u2(t),...,u M (t)] T ;

[0056] Where t represents the current time, and M is the total number of excitation signals that can be actively controlled by this system. Each element u in the vector m (t)(m=1,...,M), each corresponds to a specific, quantifiable, and adjustable physical excitation parameter, such as the power supply voltage of a certain pin, the output signal frequency of a certain analog sensor, or the transmission cycle of a certain message on the CAN bus.

[0057] To achieve the aforementioned conversion from digital vectors to physical excitation, in a preferred embodiment, this signal excitation unit can consist of a set of programmable standard instruments or functional boards. These instruments or boards receive the excitation vector u(t) from the system's main control unit and output the corresponding physical signal. For example:

[0058] Programmable DC power supply: used to simulate the working conditions that the control module under test may encounter in the real operating environment of the vehicle, such as the gradual rise or fall of the power supply voltage or the superposition of specific noise. Its output voltage value is controlled by a certain element in the excitation vector u(t).

[0059] Arbitrary waveform generator (AWG): Used to simulate various sensor signals with high fidelity, such as crankshaft position sensors and wheel speed sensors. By changing the corresponding elements in the excitation vector u(t), the amplitude, frequency, and phase of the output waveform can be adjusted in real time, and even specific distortions or noise can be introduced to explore the robustness of the control module under test to non-ideal sensor signals.

[0060] Bus communication interface card (such as CAN / LIN card): Used to simulate the vehicle communication network environment in which the control module under test is located. This interface card can not only send and receive normal messages, but more importantly, it can actively generate bus errors (such as error frames, bit errors), adjust the bus load rate, or change the transmission timing of critical messages according to the instructions of the excitation vector u(t) in order to detect the behavior of the module under test under complex electromagnetic interference or network anomalies.

[0061] Digital / Analog I / O Card: Used to simulate switching signals or slowly changing analog signals provided by other ECUs or switch contacts.

[0062] Through the collaborative work of the above units, this system is able to transform the "attack" intent of the adversarial stimulus generation module into a physical stress test on the control module under test, down to the specific signals and timing.

[0063] Data acquisition unit:

[0064] Corresponding to the application of excitation, the core responsibility of the data acquisition unit in this module is to comprehensively, accurately, and synchronously capture the complete state response of the control module under test under excitation, providing high-quality raw data for subsequent analysis modules.

[0065] The data collected in this unit is organized into an N-dimensional real-time state vector x(t), which can be represented as:

[0066] x(t) = [x1(t), x2(t), ..., x N (t)] x (t)=[x1(t),x2(t),...,x N (t)] T ;

[0067] Where t represents the current time, N is the total number of variables required to characterize the complete state of the control module under test, and x i (t)(i=1,...,N) represents the observed value of the i-th state variable at time t. This state vector x(t) will be simultaneously and in real-time transmitted to the "Real-time State Tensor Reconstruction and Dynamics Divergence Calculation Module" and the "Predictive State Observer Module". To ensure the completeness and comprehensiveness of the state vector x(t), the collected variables typically cover the following categories:

[0068] Input variables: These are the actual sampled values ​​of the excitation signals applied by the signal excitation unit of this module, used to form a control closed loop.

[0069] Output variables: The direct output of the control module under test, such as the driving voltage / current of the actuator (motor, valve, etc.), the duty cycle and frequency of the PWM signal, etc.

[0070] Internal state variables: To achieve deeper diagnostics, it is preferable to read key internal variables of the firmware in real time through debugging interfaces (such as JTAG, SWD, DAP, etc.), such as intermediate calculated values ​​of the PID controller, the current state of the state machine, and the output of key filters.

[0071] Bus variables: Key message data sent by the control module under test, collected through the bus interface card.

[0072] In the technical solution of this invention, since the subsequent construction of the "reference state space tensor" depends on the analysis of dynamic coupling relationships between variables at the microsecond or even nanosecond level, the synchronization of data acquisition is crucial. In a preferred embodiment, all data acquisition channels can achieve strict synchronization by sharing the same high-precision clock reference. For example, the backplane trigger and synchronization clock of the PXI or LXI chassis can be used to ensure that the acquisition actions of all channels on different boards (such as analog acquisition cards, digital acquisition cards, and bus cards) are highly aligned in time, and the synchronization accuracy can reach the sub-nanosecond level.

[0073] Furthermore, to ensure the accuracy of data processing, this unit appends a high-resolution, unified timestamp to each acquired state vector x(t). This timestamp is used by subsequent modules for data alignment, constructing rolling time windows, and calculating the time delay τ. k The absolute time base.

[0074] In summary, this data acquisition and stimulus execution module, through its dual functions of stimulus and acquisition, forms a bridge connecting the algorithmic world and the physical world. It ensures that the intent of the adversarial stimulus generation module can be accurately executed, while guaranteeing that the complete and realistic response of the controlled module under test can be captured without distortion. This provides a solid data foundation and physical guarantee for the effective operation of the entire closed-loop adaptive testing and diagnostic method.

[0075] Reference Benchmark Offline Modeling Module

[0076] In this embodiment, the core function of the offline modeling module for the reference benchmark is to pre-build and solidify all necessary prior knowledge models for the online diagnostic phase of this invention. Specifically, this module runs in the offline phase before the formal start of the test, and its outputs—the "preset reference state space tensor" and the "preset state observer model"—constitute the benchmark and reference system for all subsequent online real-time analysis and decision-making. The setting of this module is the fundamental basis for this invention to define the "absolute health" state of the control unit under test (DUT) from a data-driven perspective.

[0077] Functionally, this module can be further divided into two complementary processing units: a reference state space tensor generation unit and a predictive state observer training unit.

[0078] Reference state space tensor generation unit:

[0079] The purpose of this unit is to extract and solidify, from known healthy sample data, a mathematical object that can comprehensively and precisely describe the inherent dynamic physical laws between the state variables within the tested control module, namely the reference state space tensor T. refThis tensor represents the "fingerprint" or "blueprint" of the internal signal flow and coupling relationships of the system under ideal and healthy conditions.

[0080] Specifically, the generation process is as follows: First, data preparation is required. Preferably, this unit utilizes historical time series data collected from multiple rigorously verified and flawless "golden samples" under various typical operating conditions. Alternatively, in some implementations, simulation data generated from a well-validated high-fidelity simulation model (e.g., a physical model built in the MATLAB / Simulink environment) can also be used. This data must have a sufficiently high sampling rate and synchronization accuracy to capture the details of the system's transient response. Second, the state variables are defined. From the prepared data, a set of N key state variables that can most comprehensively characterize the dynamic characteristics of the control module under test are selected to form the state vector x(t) = [x1(t), x2(t), ..., x N (t)] T The selection of these variables is crucial to the effectiveness of subsequent analysis and typically includes input signals, output signals, and key internal variables obtained through the debugging interface.

[0081] Next, this unit constructs a tensor by quantizing the dynamic coupling relationship between any two state variables. To this end, a dynamic relationship function g(x) is introduced. i ,x j ,τ k ), used to calculate variable x i (t) and variable x j (t) at a specific time delay τ k The strength of the cross-correlation. In a preferred embodiment, the function can take the form of normalized cross-correlation, and its mathematical expression is:

[0082]

[0083] Where: x i (t) and x j (t) represent the values ​​of state variables i and j at time t, respectively; τ k Denotes the k-th discrete time delay step; E[·] represents the mathematical expectation over the entire observation time series; μ i and μ j These are time series x i (t) and x j The mean of (t); σ i and σ j These are time series x i (t) and xj The standard deviation of (t). The output value of this function is in the range of [-1, 1], and it can effectively measure the degree of correlation between two signals in time.

[0084] Finally, by applying all state variable pairs (i,j) (where i,j∈{1,...,N}) and all preset time delays τ k (where k∈{1,...,P}, and P is the preset maximum delay step) is traversed and calculated, and the calculation result g(x) is used to calculate the result. i ,x j ,τ k Fill in the corresponding position of a third-order tensor to construct the final reference state space tensor T. ref The tensor has dimensions N×N×P, and each element T ref (i,j,k) all imply a clear physical meaning, namely, the variable x in the health system. i With variable x j Between them, with a time delay of τ k The standard coupling strength at that time. Once generated, this tensor is permanently stored and available for use by the online diagnostic module.

[0085] Predictive State Observer Training Unit:

[0086] In parallel with the above unit, the goal of this unit is to train a time-series prediction model that can learn and reproduce the "behavioral patterns" of the health system based on the same health sample data. During the online phase, the model will act as a "virtual health reference" to determine whether the actual behavior of the currently tested module deviates from its expected normal trajectory that conforms to historical experience.

[0087] The training process is as follows: First, model selection is performed. Based on the characteristics of the control module under test, a suitable time-series prediction model can be selected. Preferably, an artificial neural network model with long-term memory capabilities, such as a Long Short-Term Memory (LSTM) network or a Gated Recurrent Unit (GRU), can be used. In other cases where model interpretability is crucial, model-based filtering algorithms, such as the Extended Kalman Filter (EKF), can also be employed. Second, model training is performed. The training process is a typical supervised learning process. The model input typically consists of two parts: one part is the historical state vector sequence [x(t-L+1),...,x(t)] over the past L time steps, and the other part is the controllable excitation vector u(t) at the current time. The model output is its prediction of the state vector at the next time step t+1. The objective function of the entire training process is to minimize the error between the predicted value and the true healthy sample data, for example, to minimize the mean squared error (MSE).

[0088] After training, a state observer model is formed that can accurately capture the nonlinear mapping relationship between the current state and input to the next state of the health system. The model was then established. It was also stored in a fixed format for later use by the online deviation calculation module.

[0089] Real-time state tensor reconstruction and dynamic divergence calculation module;

[0090] In this embodiment, the real-time state tensor reconstruction and dynamic divergence calculation module is one of the core analysis engines that runs in parallel with the predictive state observer module during the online diagnostic phase of the system described in this invention. Its fundamental task is to quantify, in real-time, the deviation between the current operating state of the controlled under test (DUT) and a preset "absolute health" benchmark from the perspective of the dynamic physical laws within the DUT. The output of this module, namely the dynamic divergence, is one of the core bases for decision-making by the subsequent adversarial stimulus generation module.

[0091] In its specific implementation, this module can be further decomposed into an instantaneous state tensor reconstruction unit, a difference tensor calculation unit, and a dynamic divergence calibration unit. The functions of each unit will be described in detail below.

[0092] Instantaneous state tensor reconstruction unit:

[0093] The function of this unit is to capture the dynamic characteristics of the control module under test in real time under the current excitation and express them in a mathematical structure that is completely consistent with the reference benchmark.

[0094] Specifically, this unit continuously receives real-time state vector data stream x(t) with high-precision timestamps from the aforementioned "Data Acquisition and Stimulus Execution Module". For dynamic analysis, this unit processes this data stream within a rolling time window of a preset length. Crucially, the methods, parameters, and computational logic used to reconstruct the instantaneous state tensor are strictly consistent with those used in the aforementioned "Reference Benchmark Offline Modeling Module" for generating the reference state space tensor. This is to ensure the fairness and validity of subsequent comparisons.

[0095] In a preferred embodiment, this unit also uses a normalized cross-correlation function as the dynamic relationship function g(x). i ,x j ,τ k ), to calculate any two state variables x within the current scrolling time window. i (t) and x j (t) at different time delays τ k The strength of the coupling relationship. Its mathematical expression is:

[0096]

[0097] Where: x i (t) and x j (t) represent the time series of state variables i and j within the current time window, respectively; τ k The k-th discrete time delay step is represented, and its range is exactly the same as when the reference benchmark was generated; E[·] represents the mathematical expectation within the current rolling time window; μ i and μ j These are the time series x within the current time window. i (t) and x j The mean of (t); σ i and σ j These are the time series x within the current time window. i (t) and x j The standard deviation of (t).

[0098] By considering all variable pairs (i,j) and all time delays τ k By performing traversal calculations, this unit constructs an instantaneous state space tensor T in real time, representing the current internal dynamic relationships of the module under test. inst (t). This tensor is continuously updated as the time window rolls, thus dynamically reflecting the real-time operating conditions of the module under test.

[0099] Differential tensor computation unit:

[0100] The function of this unit is to directly compare the instantaneous system state with the ideal health benchmark, thereby explicitly revealing all dynamic coupling relationships that have deviated.

[0101] Specifically, this unit receives T generated by the aforementioned instantaneous state tensor reconstruction unit. inst (t), and at the same time, it retrieves the reference state space tensor T, which is pre-stored by the "Reference Benchmark Offline Modeling Module". ref Subsequently, this unit performs element-wise subtraction to calculate the difference tensor ΔT(t), whose mathematical expression is:

[0102] ΔT(t)=T inst (t)-T ref ;

[0103] The generated difference tensor ΔT(t) and T ref and T inst (t) have the same dimension. Each of its elements ΔT(t) i,j,k The values ​​of all precisely quantify "variable x". i With variable x j With a time delay of τk The dynamic coupling relationship is a specific physical law that is compared to the deviation in its absolutely healthy state. An element value close to zero indicates that the dynamic relationship is normal, while an element value with a large absolute value directly indicates that there may be a functional abnormality or failure at that point.

[0104] Dynamic divergence calibration unit:

[0105] The function of this unit is to transform the high-dimensional, detail-rich difference tensor ΔT(t) into a single scalar value that can macroscopically characterize the overall anomaly of the system, namely the "dynamic divergence" D. K (t). This mapping from high dimension to one dimension is necessary because it provides a concise and explicit optimization objective and trigger signal for the subsequent adversarial stimulus generation module. To achieve this, this unit obtains the dynamic divergence by calculating the norm of the difference tensor ΔT(t). In a preferred embodiment, the Frobenius Norm is used for calculation because it comprehensively considers the sum of squares of all elements and is more sensitive to overall bias. Its mathematical expression is:

[0106]

[0107] Where: N is the total number of state variables; P is the maximum time delay step under consideration; ΔT(t) i,j,k It is the element value of the difference tensor at index (i,j,k).

[0108] The final calculated dynamic divergence D K (t) is a non-negative real number. Ideally, when the module under test is perfectly healthy, this value is close to zero. When the module under test experiences any fault that disrupts its internal dynamic relationships, this value will increase accordingly. This unit outputs this dynamic divergence value to the "Adversarial Excitation Generation and Cooperative Constraint Module" in real time and continuously, as one of the most important input signals for it to judge system anomalies, calculate adversarial gradients, and execute cooperative decisions.

[0109] Predictive state observer module;

[0110] In this embodiment, the predictive state observer module is another core analysis engine running during the online diagnostic phase of the system described in this invention. It operates in parallel with the aforementioned "real-time state tensor reconstruction and dynamic divergence calculation module," but assesses the system's health status from a complementary dimension—namely, the "external behavior pattern" of the controlled under test (DUT). The fundamental task of this module is to use a pre-trained predictive model capable of characterizing the behavior pattern of a healthy system to determine in real time whether the actual state response of the current tested module deviates from its expected normal trajectory based on historical experience. Its final output, the predicted deviation, is a key input for realizing the core collaborative constraint strategy of this invention.

[0111] Specifically, the functionality of this module depends on the online deployment and operation of a pre-defined state observer model.

[0112] First, at the start of the online diagnostic process, this module will load the state observer model trained and stored by the aforementioned "Reference Benchmark Offline Modeling Module" during its offline phase. The model Essentially a mathematical function, it has accurately captured the nonlinear mapping relationship between the tested control module and its future state from its historical state and current input through learning from a large amount of health sample data.

[0113] During system operation, this module continuously and in real-time receives two data streams from the "Data Acquisition and Stimulus Execution Module": one is the historical state vector sequence over the past L time steps, and the other is the controllable stimulus vector u(t) at the current time t. Subsequently, this module feeds these input data into the deployed state observer model. In this context, the predicted value of the state vector for the next time step t+1 is calculated, denoted as... This prediction process can be represented by the following functional form:

[0114]

[0115] in: It is the model's prediction of the state vector at the next time step t+1; The model represents the trained state observer model; [x(t-L+1),…,x(t)] is a historical sequence of actual state vectors of length L from time t-L+1 to time t, provided by the “Data Acquisition and Excitation Execution Module”; u(t) is the excitation vector consisting of M controllable variables applied to the control module under test at the current time t.

[0116] This predicted value The physical meaning of this is that it represents the behavioral pattern established based on historical health data, the next state the system "should" enter after receiving the current input. However, simply generating a predicted value is not the ultimate goal of this module. Its deeper function lies in quantifying a specific anomaly, namely "behavioral unpredictability," by comparing the predicted value with the system's actual response. For this purpose, this module also needs to obtain the state vector x(t+1) actually measured at the next time step t+1 from the "Data Acquisition and Stimulus Execution Module." After obtaining the predicted value... After comparing the actual value x(t+1), this module calculates the difference between the two, thus obtaining a scalarized "prediction deviation", denoted as E. P (t+1). In a preferred embodiment, this deviation can be obtained by calculating the p-norm distance between the two vectors, for example, the Euclidean distance when p = 2:

[0117]

[0118] Among them: E P (t+1) is the prediction deviation calculated at time t+1; x(t+1) is the N-dimensional state vector actually collected at time t+1; It is the model's prediction of the N-dimensional state vector at time t+1; ||·|| p This represents the p-norm operator.

[0119] This prediction deviation E P (t+1) is a crucial indicator. A lower E P The value indicates that the current behavior of the tested module is predictable and conforms to its known healthy behavior patterns. Conversely, a significantly increased E value indicates... P The value strongly indicates that the tested module is in an abnormal state that has never appeared in its historical health data and cannot be explained by the normal behavior model.

[0120] Within the overall framework of this invention, the prediction deviation E P It acts as a "shield" or "constraint." It is continuously output in real-time to the "Adversarial Excitation Generation and Cooperative Constraint Module." In this module, it interacts with the "Dynamic Divergence" D from another analysis engine. K These factors work together to form a collaborative decision. This dual verification mechanism ensures that the system seeks not only any anomalies deviating from the baseline, but also those that simultaneously violate the inherent physical laws (D...). K (Increases) and exhibits unpredictable external behavior (E) P (Increase) High-value, deep-seated potential defects.

[0121] Module for generating adversarial incentives and coordinating constraints;

[0122] In this embodiment, the adversarial incentive generation and collaborative constraint module is the "decision center" for the closed-loop adaptive testing of the system described in this invention. Its fundamental purpose is to abandon the open-loop mode of preset fixed test cases in traditional testing methods, and instead use a proactive and intelligent "exploration-verification" mechanism to interact with the control unit under test (DUT) in real time, so as to automatically and convergently discover and approximate those deep-seated potential defects that are difficult to trigger by conventional means in the most efficient way.

[0123] The core working principle of this module can be understood as a constrained, adversarial optimization process. It continuously receives real-time outputs from two other parallel analysis modules: one is the dynamic divergence D, provided by the "Real-time State Tensor Reconstruction and Dynamic Divergence Calculation Module," which characterizes the degree of deviation from the system's intrinsic physical laws. K (t); secondly, the prediction deviation E provided by the "predictive state observer module" characterizes the unpredictability of the system's external behavioral patterns. P (t). This module integrates these two very different indicators in a "spear and shield" synergy to guide the adjustment direction of the excitation signal.

[0124] Specifically, the operation of this module can be described as follows:

[0125] First, the core objective of this module, its "antagonistic" or "spear-like" aspect, lies in actively and maximally amplifying the abnormal characteristics of the system by iteratively adjusting the controllable excitation vector u(t) applied to the controlled module under test, that is, maximizing the dynamic divergence D. K (t). This is a typical gradient ascent optimization process.

[0126] In order to find D K (t) indicates the optimal direction for increasing; this module needs to calculate D. K The gradient of (t) with respect to the current activation vector u(t). However, since the controlled module under test is usually a physical entity or a "black box" model, its internal analytical function is unknown, and this gradient cannot be directly obtained analytically. Therefore, in a preferred embodiment, this module uses numerical methods, such as the finite difference method, to approximate the gradient. Specifically, for the controllable excitation vector u(t) = [u1(t),...,u...],... M (t)] T Each component u in m (t), apply a small perturbation Δu m A new excitation vector is formed, and the dynamic divergence value after perturbation is obtained from the "Real-time State Tensor Reconstruction and Dynamic Divergence Calculation Module", and its change ΔD is calculated.K The partial derivative on this component can be approximated as:

[0127]

[0128] By perturbing all M controllable components, a complete gradient vector can be constructed. After obtaining the gradient vector, this module generates a new candidate activation vector u(t+1) based on the principle of gradient ascent. Its update rule is as follows:

[0129]

[0130] Where u(t+1) is the new activation vector proposed to be applied in the next time step; u(t) is the activation vector at the current time. It is the gradient of the calculated dynamic divergence with respect to the current excitation; α is a preset learning rate or step size factor used to control the magnitude of adjustment in each iteration.

[0131] However, merely pursuing D K Maximizing (t) might cause the system to get stuck in some meaningless or unrealistic outlier regions. To ensure the effectiveness and realism of the exploration, this module introduces its "cooperative constraint" or "shield." This constraint comes from the prediction deviation E calculated by the "predictive state observer module." P (t).

[0132] This module applies the new excitation vector u(t+1) generated by the above update rules to the controlled module under test through the "Data Acquisition and Excitation Execution Module". Subsequently, this module obtains the prediction deviation E generated under the new excitation from the "Predictive State Observer Module". P (t+1).

[0133] At this point, this module enters its core collaborative decision-making stage. The effectiveness of an incentive adjustment is not simply determined by D. K The determination is not based on whether (t) increases, but rather on the simultaneous satisfaction of a dual constraint condition: the dynamic divergence exhibits an increasing trend if and only if D... K (t+1)>D K (t), and the prediction deviation exceeds a preset deviation threshold δ. E E P (t+1)>δ E Only then is the incentive adjustment guided by gradient ascent recognized as a "valid attack".

[0134] The setting of this collaborative constraint is of critical technical significance. It ensures that what this system seeks is a state that simultaneously violates the "intrinsic physical laws" (manifested as disordered internal dynamic relationships, i.e., D).K Increased) and "external behavioral patterns" (manifested as behaviors that cannot be predicted by health models, i.e., E) P High-confidence genuine defects exceeding a threshold are identified. Any scenario that meets only a single condition will be filtered out, thus avoiding wasted computing power and incorrect diagnoses.

[0135] If the collaborative decision determines that the current step is valid, the system adopts a new incentive vector u(t+1) and begins the next iteration based on it. If the determination is invalid, the system may reject the update and may try new directions by reducing the step size factor α or adopting other exploration strategies.

[0136] This iterative process repeats until the dynamic divergence D is reached. K The value of (t) reaches a local or global maximum and cannot increase further (i.e., the gradient approaches zero), or reaches the preset upper limit of the number of iterations. When the iteration terminates, the final activation vector output by this module is the "final adversarial activation vector" u. final This vector represents a set of precise excitation conditions discovered in this invention, capable of stably reproducing a specific deep-level fault. Subsequently, this u... final Together with its corresponding final difference tensor, it will be transmitted to the "Root Cause Association Diagnosis Module" for subsequent root cause localization and analysis.

[0137] Root cause correlation diagnostic module;

[0138] In this embodiment, the root cause correlation diagnosis module serves as the final interpretation and analysis stage after the system of the present invention completes proactive, adversarial defect exploration. Its core purpose is not to discover new anomalies, but to deeply analyze the final results converged and confirmed by the aforementioned "adversarial incentive generation and collaborative constraint module," transforming a macroscopic, quantitative anomaly indicator (i.e., the maximized dynamic divergence) into a specific, understandable, and accurate diagnostic report on the root cause that can guide repair work. This module is a crucial bridge in the present invention's transition from "problem discovery" to "problem localization."

[0139] Functionally, this module can be further divided into a fault reproduction path recording unit and a root cause correlation analysis unit. These two units work together to ensure the accuracy and traceability of the diagnostic results.

[0140] Fault Reproduction Path Recording Unit:

[0141] The primary responsibility of this unit is to accurately and permanently store the "reproduction conditions" of defects.

[0142] Once the "Adversarial Incentive Generation and Cooperative Constraint Module" has determined a set of incentive conditions that can stably trigger a certain deep-level fault through its closed-loop iterative process, it will output a final adversarial incentive vector u. final The fault reproduction path recording unit receives this vector and archives it along with relevant test metadata (such as test time, serial number of the module under test, etc.).

[0143] This u final A vector is not a random or isolated data point; it contains extremely valuable information and is the precise "key" to unlocking a specific fault phenomenon. By storing this vector, this invention provides a means to reproduce a specific fault with 100% certainty at any time on any platform with the same testing environment. This is of paramount value for subsequent engineering activities, such as manual debugging and verification, regression testing after firmware repair, and the establishment of a standardized test case library for this fault.

[0144] Root cause correlation analysis unit:

[0145] This unit is the core of the final diagnosis, and its responsibility is to answer the fundamental question of "where exactly is the problem?" It directly correlates the fault phenomena with specific physical or logical root causes by analyzing the internal state of the control module under final adversarial excitation.

[0146] This unit receives final data from two upstream modules: one of which is the final excitation vector u stored in the aforementioned recording unit. final Secondly, and more importantly, in applying u final At that instant, the final instantaneous state space tensor T generated by the "Real-time State Tensor Reconstruction and Dynamics Divergence Calculation Module" inst_final This unit first calculates the final difference tensor ΔT. final The calculation method is as follows:

[0147] ΔT final =T inst_final -T ref ;

[0148] Wherein: T inst_final It is an instantaneous depiction of the dynamic relationships within the tested module under final adversarial excitation; T ref It is a baseline tensor representing an absolute healthy state, pre-generated by the "Reference Benchmark Offline Modeling Module".

[0149] This ΔT final Each element ΔT in the tensor final (i,j,k) quantitatively represent the degree of deviation of a specific dynamic coupling relationship from its health baseline under the final, most problematic operating condition.

[0150] Next, this unit will focus on the third-order difference tensor ΔT. final Perform a search operation to locate one or more elements with the largest absolute value. Suppose the search finds that the element with the largest absolute value is located at index (p, q, r), i.e., |ΔT|. final (p,q,r)| represents the global maximum value.

[0151] Finally, and most importantly, the core step in achieving "association diagnosis" in this unit lies in translating this abstract mathematical index (p, q, r) back to its corresponding concrete physical or logical meaning. This translation follows the rules pre-defined when constructing the tensor:

[0152] Index p corresponds to the p-th state variable x in the state vector x(t). p ;

[0153] Index q corresponds to the q-th state variable x in the state vector x(t). q ;

[0154] Index r corresponds to the r-th discrete time delay τ under consideration. r .

[0155] Through this correlation, the system can automatically generate a highly accurate diagnostic conclusion, such as: "Through analysis, the root cause of the detected functional failure has been located as: state variable 'x'." p 'with state variable' x q Between ', there exists 'τ' r 'The dynamic coupling relationship under time delay has undergone a severe functional deviation.'

[0156] In summary, the root cause correlation diagnostic module, by recording the reproduction path and analyzing the final difference tensor, successfully reduces the dimensionality of complex system-level fault phenomena and traces them back to one or more specific failures in the dynamic relationships between variables. This transforms the diagnostic results from vague phenomenological descriptions into verifiable, locatable, and repair-guided engineering information, thus completing the full closed loop of this invention from intelligent detection to precise diagnosis.

[0157] This invention also provides a method for high-precision control module detection and data analysis, the method comprising the following steps:

[0158] The data acquisition and excitation execution module interacts with the control module under test and applies excitation signals.

[0159] The step of calculating the dynamic divergence includes: calculating the dynamic divergence, which characterizes the degree of deviation of the internal dynamic relationship of the control module under test, based on the real-time state data of the control module under test and by referring to a preset reference state space tensor.

[0160] The step of calculating the prediction deviation includes: calculating the prediction deviation between the actual state and the predicted state of the controlled module under test according to a preset state observer model;

[0161] The step of generating adversarial stimuli includes: responding to the calculation results of the dynamic divergence and the predicted deviation, and generating a new stimuli signal based on the dynamic divergence and the predicted deviation, and controlling the data acquisition and stimuli execution module to apply the new stimuli signal, wherein the effectiveness of generating the new stimuli signal depends on the dynamic divergence showing an increasing trend and the predicted deviation exceeding a preset deviation threshold.

[0162] The method in this embodiment can be used to execute the above system embodiment, and its principle and technical effect are similar, so it will not be described again here.

[0163] 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 high precision control module detection and data analysis system, characterized in that, The method comprises the following steps: a data acquisition and excitation execution module for interacting with the measured control module and applying excitation signals; a dynamic divergence calculation module for calculating the dynamic divergence of the measured control module according to the real-time state data of the measured control module and the preset reference state space tensor; a predicted divergence calculation module for calculating the predicted divergence between the actual state and the predicted state of the measured control module according to the preset state observer model; an adversarial excitation generation module connected with the dynamic divergence calculation module and the predicted divergence calculation module for generating and controlling the data acquisition and excitation execution module to apply new excitation signals based on the dynamic divergence and the predicted divergence.

2. The high precision control module test and data analysis system of claim 1, wherein, The adversarial excitation generation module is specifically used for: when the dynamic divergence shows an increasing trend and the predicted divergence exceeds the preset divergence threshold, confirming the effectiveness of the current excitation signal adjustment and generating the new excitation signal based on the effective adjustment.

3. The high precision control module test and data analysis system of claim 1, wherein, The dynamic divergence calculation module comprises: a transient state tensor reconstruction unit for calculating the transient state space tensor of the measured control module in real time according to the real-time state data; a difference tensor calculation unit for calculating the difference tensor between the transient state space tensor and the reference state space tensor; a dynamic divergence calibration unit for calculating the dynamic divergence according to the difference tensor.

4. The high precision control module test and data analysis system of claim 1, wherein, The predicted divergence calculation module is specifically used for: by deploying the state observer model in an online environment, generating the predicted state according to the real-time historical state and the current excitation signal of the measured control module, and calculating the predicted divergence.

5. The high precision control module test and data analysis system of claim 1, wherein, It also includes a reference benchmark offline modeling module for generating the reference state space tensor based on healthy sample data before the test starts; and training the state observer model based on the healthy sample data. The reference benchmark offline modeling module comprises:

6. The high precision control module test and data analysis system of claim 5, wherein, a reference state space tensor generation unit for calculating the dynamic coupling relationship between each state variable in the healthy sample data at different time delays to generate the reference state space tensor; and a predictive state observer training unit for training the state observer model to learn the behavior response mode of the healthy sample data. It also includes a root cause correlation diagnosis module for performing root cause analysis after the adversarial excitation generation module determines the final adversarial excitation vector.

7. The high precision control module test and data analysis system of claim 1, wherein, The root cause correlation diagnosis module comprises:

8. The high precision control module test and data analysis system of claim 7, wherein, a fault reproduction path recording unit for recording the final adversarial excitation vector; a root cause correlation analysis unit for locating the dynamic coupling relationship that causes functional failure by analyzing the final difference tensor generated under the final adversarial excitation vector. The data acquisition and excitation execution module comprises:

9. The high precision control module test and data analysis system of claim 1, wherein, a signal excitation unit for applying excitation to the measured control module according to the new excitation signal generated by the adversarial excitation generation module; a data acquisition unit for synchronously acquiring real-time state data of the measured control module. ​ 10. The method of claim 1-9, wherein the method is applied to the system of claim 1-9. The method comprises the following steps: interacting with the measured control module and applying excitation signals through a data acquisition and excitation execution module; calculating a dynamic divergence, which comprises calculating the dynamic divergence representing the deviation degree of the internal dynamic relationship of the measured control module according to the real-time state data of the measured control module and against a preset reference state space tensor; calculating a prediction deviation degree, which comprises calculating the prediction deviation degree between the actual state and the predicted state of the measured control module according to a preset state observer model; generating an antagonistic excitation, which comprises generating a new excitation signal in response to the calculation results of the dynamic divergence and the prediction deviation degree and based on the dynamic divergence and the prediction deviation degree, and controlling the data acquisition and excitation execution module to apply the new excitation signal, wherein the effectiveness of generating the new excitation signal depends on the dynamic divergence showing an increasing trend and the prediction deviation degree exceeding a preset deviation degree threshold.

Images