An end-cloud cooperative programmable test method, system and portable device for an automobile instrument

Abstract

The application discloses a kind of end cloud cooperation's automobile instrument's program-controlled test method, system and portable device, it is related to end cloud cooperation field, including: cloud server receives the test task submitted by user, and from cloud test case library, call basic test case, input basic test case into test strategy optimization model, generate initial program-controlled instruction sequence;Cloud server issues initial program-controlled instruction sequence to end side test equipment, based on initial program-controlled instruction sequence, to be measured automobile instrument output test signal, and collect the response data of measured automobile instrument;Response data is uploaded to cloud server in real time by end side test equipment;Dynamically adjust initial program-controlled instruction sequence in subsequent parameter step and test point density, generate optimization program-controlled instruction sequence, issue and execute;Finally, response data is compared with standard response, and generate test report.The problems, such as low efficiency in prior art, poor abnormal detection capability and poor test scheme adaptability, are solved.

Description

Technical Field

[0001] This invention relates to the field of edge-cloud collaboration, specifically to a method, system, and portable device for edge-cloud collaborative testing of automotive instruments. Background Technology

[0002] Currently, automotive instrument panels typically require manual input from testers during the R&D phase. Although production workshops are equipped with factory testing equipment that has functions such as hard-wired high and low level output, frequency output, power output, CAN signal output, and voltage input detection, and supports independent software programming and archiving for different product models, such equipment is expensive, bulky, and requires external procurement, making it extremely impractical to widely use such equipment for product testing within the R&D department.

[0003] Because the testing strategy is rigid and lacks adaptability, it cannot dynamically adjust the parameter step size and test point density according to the real-time response during the testing process, resulting in low testing efficiency, poor anomaly detection capability, and poor adaptability of the testing solution. Summary of the Invention

[0004] This application provides a method, system, and portable device for edge-cloud collaborative programmable testing of automotive instruments, addressing the problems of low testing efficiency, poor anomaly detection capability, and poor adaptability of testing schemes in existing technologies.

[0005] In view of the above problems, this application provides a method, system and portable device for edge-cloud collaborative testing of automotive instruments.

[0006] Firstly, this application provides a method for testing the programmable control of an automotive instrument cluster using an edge-cloud collaborative approach, the method comprising: The cloud server receives the test task submitted by the user, calls the basic test cases from the cloud test case library according to the test task, and inputs the basic test cases into the pre-trained test strategy optimization model to generate an initial program control instruction sequence. The test strategy optimization model aims to maximize the anomaly detection rate and minimize the test time. The cloud server sends the initial programmable instruction sequence to the end-side testing device. The end-side testing device outputs test signals to the instrument panel of the vehicle under test based on the initial programmable instruction sequence and simultaneously collects the response data of the instrument panel of the vehicle under test. The end-side testing device uploads the response data to the cloud server in real time. The cloud server evaluates the current test progress and anomaly detection in real time based on the response data, dynamically adjusts the parameter step size and test point density of the initial programmable instruction sequence, generates an optimized programmable instruction sequence, and sends the optimized programmable instruction sequence to the end-side test device for continued execution. The cloud server compares the final collected response data with the standard responses in historical data to generate a test report.

[0007] Secondly, the present invention provides a cloud-edge collaborative automotive instrument cluster control testing system, the system comprising: The programmable instruction generation module is used to receive test tasks submitted by users on the cloud server, call basic test cases from the cloud test case library according to the test tasks, and input the basic test cases into a pre-trained test strategy optimization model to generate an initial programmable instruction sequence. The test strategy optimization model aims to maximize the anomaly detection rate and minimize the test time. The response data acquisition module is used for the cloud server to send the initial programmable instruction sequence to the end-side test device, and the end-side test device to output test signals to the instrument panel of the vehicle under test based on the initial programmable instruction sequence, and simultaneously collect the response data of the instrument panel of the vehicle under test. A response data upload module is used by the end-side test device to upload the response data to the cloud server in real time. The programmable instruction optimization module is used by the cloud server to evaluate the current test progress and anomaly detection in real time based on the response data, dynamically adjust the parameter step size and test point density of the initial programmable instruction sequence in the subsequent steps, generate an optimized programmable instruction sequence, and send the optimized programmable instruction sequence to the end-side test device for continued execution. The comparison test module is used by the cloud server to compare the final collected response data with the standard responses in historical data and generate a test report.

[0008] Thirdly, the present invention provides a portable device for programmable testing of automotive instruments with edge-cloud collaboration, comprising: a memory for storing computer software programs; and a processor for reading and executing the computer software programs, thereby realizing a programmable testing method for automotive instruments with edge-cloud collaboration.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, the cloud server receives the test task submitted by the user and calls basic test cases from the cloud test case library according to the test task. Based on the pre-trained test strategy, it optimizes the model and generates an initial programmable instruction sequence to achieve a pre-balance between test efficiency and anomaly detection rate. Second, the cloud server sends the initial programmable instruction sequence to the edge test equipment, outputs test signals and collects response data, realizing the physical execution and timing synchronization of test stimulus and response capture to ensure the authenticity and effectiveness of the test. Next, the edge test equipment uploads the response data to the cloud server in real time, establishing a data pipeline from the physical test site to the cloud intelligent analysis center. Then, based on the response data, it evaluates the current test progress and anomaly detection status in real time, generates an optimized programmable instruction sequence, and sends it to the edge test equipment for continued execution to improve the anomaly detection rate and test efficiency. Finally, the cloud server compares all the finally collected response data with the standard responses in historical data to generate a structured test report, generating structured data with clear judgment conclusions.

[0010] In summary, this method, through an edge-cloud collaborative architecture design, deploys computationally intensive strategy optimization in the cloud and deploys real-time signal output and acquisition on the edge, realizing intelligent, adaptive, and closed-loop optimization of test strategies. This significantly improves the automation level, anomaly detection capability, and test efficiency of automotive instrument testing, while reducing the complexity and cost of edge devices. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating a cloud-edge collaborative method for programmable testing of automotive instruments provided in this application. Figure 2 This is a schematic diagram of the structure of a cloud-edge collaborative automotive instrument control test system provided in this application; Figure 3 This is a schematic diagram of the structure of a portable device for testing the programmable instrument panel of an automobile with end-to-cloud collaboration, as provided in this application. Figure 4 , Figure 5 This is a schematic diagram of a prototype of a portable device for testing the programmable instrument panel of an automobile that is based on end-to-end cloud collaboration, as provided in this application.

[0012] In the attached diagram, the components represented by each label are as follows: The device includes a programmable instruction generation module 11, a response data acquisition module 12, a response data upload module 13, a programmable instruction optimization module 14, a comparison test module 15, a portable device 200 for programmable testing of automotive instruments with end-to-cloud collaboration, a memory 210, a processor 220, and a computer program 211. Detailed Implementation

[0013] This application provides a cloud-edge collaborative method, system, and portable device for programmable testing of automotive instruments, which specifically addresses the problems of low testing efficiency, poor anomaly detection capability, and poor adaptability of testing schemes in the prior art.

[0014] The present invention will now be described in detail with reference to the accompanying drawings.

[0015] Example 1, as Figure 1 As shown, this application provides a method for testing the programmable control of an automotive instrument cluster using edge-cloud collaboration, the method comprising: S10: The cloud server receives the test task submitted by the user, calls the basic test cases from the cloud test case library according to the test task, and inputs the basic test cases into the pre-trained test strategy optimization model to generate an initial program control instruction sequence. The test strategy optimization model aims to maximize the anomaly detection rate and minimize the test time. In this embodiment, the cloud server is one or more servers deployed in a cloud computing center; the test task is a work instruction submitted by the user to the system through a client, which is to test a specific automotive instrument; the cloud test case library is a structured database stored on the cloud server; the basic test cases are the original test plans that are not optimized and correspond to the standard test specifications; the initial programmable instruction sequence is a set of instructions output by the optimized model that can be directly executed by the end-side test equipment.

[0016] Specifically, the cloud server first submits and parses the test task, identifying its core elements. Then, based on these elements, it retrieves corresponding basic test cases from the cloud-based test case library. These basic test cases might define the input signal for the speedometer test as a square wave frequency signal, with a frequency range corresponding to vehicle speeds from 0 km / h to 220 km / h, and standard test points every 20 km / h. Subsequently, the cloud server uses these retrieved basic test cases as input to a pre-trained test strategy optimization model, which analyzes the basic test cases. The model then determines and outputs an executable initial sequence of programmable control instructions.

[0017] Step S10 in the method provided in this application embodiment includes: The cloud server parses the test task and extracts the instrument model and test items from the test task. The test items include at least indicator light detection, speedometer detection, tachometer detection, fuel gauge detection, and water temperature gauge detection. The cloud server matches corresponding basic test cases from the cloud test case library according to the model of the instrument under test and the test item. The basic test cases include at least signal type, parameter range, test sequence and judgment criteria. The signal type includes at least high and low level signals, PWM wave signals, frequency signals, resistance signals and voltage signals. The cloud server inputs the basic test cases into a pre-trained test strategy optimization model. The test strategy optimization model aims to maximize the anomaly detection rate and minimize the test time. It iteratively optimizes the parameter range and test timing in the basic test cases, outputs the optimized parameter step sequence and test point density, and generates an initial program control instruction sequence based on the optimized parameter step sequence and test point density.

[0018] In this embodiment, the cloud server first parses the test task and extracts the instrument under test model and test items from the test task. The instrument under test model is the code or name of the automotive instrument product type. The test items are the specific functions or performance verification categories that need to be performed on the instrument under test. Each test item focuses on a certain subsystem or specific function of the instrument. The test items include at least indicator light detection, speedometer detection, tachometer detection, fuel gauge detection, and water temperature gauge detection.

[0019] Specifically, indicator light testing verifies whether various indicator lights on the dashboard can correctly light up and turn off when receiving corresponding control signals, and whether their color, brightness, and response time meet design requirements; speedometer testing verifies whether the correspondence between the speedometer pointer or digital display value and the vehicle speed signal input to the tested instrument meets standard specifications; tachometer testing verifies the linearity and accuracy between the engine tachometer display value and the input speed signal; fuel gauge testing verifies the correspondence between the fuel level displayed on the fuel gauge and the actual remaining fuel, paying particular attention to the accuracy of the low fuel warning area; and coolant temperature gauge testing verifies the correspondence between the coolant temperature gauge display value and the input coolant temperature sensor signal.

[0020] When a user submits a test task to the cloud server via the client, the cloud server receives the task and first parses it. The server's internal task parser scans the text and extracts two types of information: the model of the instrument under test and the test items, using predefined keyword matching rules or natural language processing models. If the user does not explicitly specify, it may default to all items. For example, a task may explicitly require only speedometer and fuel gauge testing, in which case the parsing process will extract only these two items.

[0021] Secondly, the cloud server matches the corresponding basic test cases from the cloud test case library based on the instrument under test model and test items. The matching process involves the cloud server querying and comparing the parsed instrument under test model and test items in the cloud test case library to find the most suitable test case record. The basic test cases include at least signal type, parameter range, test timing and judgment criteria. The signal types include at least high and low level signals, PWM wave signals, frequency signals, resistance signals and voltage signals.

[0022] Specifically, signal type refers to the type of electrical signal that needs to be output to the instrument under test; high and low level signals are binary discrete signals, usually used to simulate switch inputs; PWM wave signals are pulse width modulation signals, which transmit analog information by changing the duty cycle, and are often used to control LED brightness or drive stepper motors; frequency signals are square wave or sine wave signals whose frequency changes with time, and are the most common output form of vehicle speed sensors and tachometer sensors; resistance signals transmit information by changing the resistance value, and are the standard output form of fuel level sensors, water temperature sensors, and temperature sensors; voltage signals are continuous analog voltage values, used to simulate the output of some linear sensors.

[0023] The parameter range is the minimum and maximum values ​​that the test needs to cover for each signal type; the test sequence specifies the order of test points or different test items, the waiting time, and the signal switching time sequence; the judgment criterion is the standard for judging whether the response of the instrument under test is qualified.

[0024] The system extracts the instrument model and test items, connects to a cloud-based test case library via a cloud server for matching, and executes an SQL query based on the extracted instrument model and test items. If the query is successful, the cloud server obtains a basic test case uniquely corresponding to that model and item, including signal type, parameter range, test timing, and decision criteria. SQL, or Structured Query Language, is a standard programming language used to manage relational databases. For example, the signal type might be: frequency signal; parameter range: minimum 10Hz, maximum 2000Hz, step base 20Hz; test timing: starting from the minimum value, increasing in 20Hz steps to the maximum value, with a stable output time of 2 seconds at each frequency point and an interval of 0.5 seconds; decision criteria: displayed vehicle speed = frequency × 0.1, with an allowable error of ±2km / h or ±3%, whichever is greater.

[0025] Finally, the cloud server inputs the basic test cases into the pre-trained test strategy optimization model. The test strategy optimization model aims to maximize the anomaly detection rate and minimize the test time. It iteratively optimizes the parameter range and test timing in the basic test cases, outputs the optimized parameter step size sequence and test point density, and generates the initial programmable instruction sequence based on the optimized parameter step size sequence and test point density.

[0026] Specifically, iterative optimization employs algorithms such as reinforcement learning to simulate or utilize historical data, trying various combinations of parameter step sizes and test point densities, and evaluating the merits of each combination based on a preset reward function to find the optimal or near-optimal. The parameter step size sequence contains the incremental value for each step between the minimum and maximum values ​​of the parameter range. The test point density is the number of test points set within a unit interval on the parameter axis, where high-density areas correspond to fine-grained testing and low-density areas correspond to fast scanning.

[0027] After obtaining the basic test cases, the cloud server uses them as input data to feed into a pre-trained test strategy optimization model. Internally, the model learns from historical test data, aiming to maximize the anomaly detection rate and minimize test time. It analyzes and iteratively optimizes the parameter range and original test timing in the basic test cases. After iteration, the model outputs the optimized parameter step sequence and the corresponding test point density distribution. Finally, the model optimizes the step sequence by combining the specific parameter values ​​of each test point within the parameter range with information such as signal type and output duration from the basic test cases to generate an initial sequence of programmable instructions for subsequent steps.

[0028] In step S10 of the method provided in this application embodiment, the process of constructing the test strategy optimization model includes: Collect historical test datasets, which include multiple sets of instrument models, test items, programmable control command parameters, response data and corresponding fault tags. The programmable control command parameters include at least voltage value, frequency value, resistance value, duty cycle and pulse count. The response data includes at least the instrument display value, indicator light status, CAN message content and analog output value. Based on reinforcement learning, an initial test strategy optimization model is constructed. Using the instrument models and test items in the historical test dataset as the state space, the corresponding programmable instruction parameters as the action space, and the comprehensive reward function as the optimization objective, a pre-trained test strategy optimization model is obtained through multiple rounds of iterative training. The construction process of the comprehensive reward function includes: The ratio of the number of anomalies detected in the current round of testing to the total number of test points is used as the anomaly detection rate. Get the total test time for the current round of testing; Based on the preset first weight coefficient and second weight coefficient, the comprehensive reward function is calculated according to the following function: the comprehensive reward function is equal to the anomaly detection rate multiplied by the first weight coefficient minus the total test time multiplied by the second weight coefficient, wherein both the first weight coefficient and the second weight coefficient are positive values.

[0029] In this embodiment, historical test datasets are first collected. These datasets include multiple sets of instrument models, test items, programmable control command parameters, response data, and corresponding fault tags. The programmable control command parameters include at least voltage, frequency, resistance, duty cycle, and pulse count. The response data includes at least the instrument display values, indicator light status, CAN message content, and analog output values. The historical test dataset is a collection of structured data generated and stored by the system or compatible system during past testing activities. The instrument model is a unique code that identifies the type of the instrument under test. The test item is the specific functional test category performed. The programmable control command parameters are the specific values ​​of the electrical signals actually output by the end-side test equipment during historical testing, used to excite the instrument under test.

[0030] Duty cycle is the ratio of the high-level time to the period in a PWM wave signal, usually expressed as a percentage; Pulse count: the number of pulses output within a specific time period, often used to simulate wheel speed sensors, etc.; Instrument display value is the angle value of a pointer instrument or the digital value displayed on an LCD screen, reflecting the instrument's reading; Indicator status is the on / off state of various indicator lights, usually represented by Boolean values; CAN message content is the data frame sent by the instrument through the CAN bus, containing the instrument's internal status, diagnostic information, etc.; Analog output value is the analog voltage or current signal output by the instrument, used for communication with other controllers; Fault tag is a mark that has been manually or automatically verified to identify whether an abnormality exists in the test and the type of abnormality.

[0031] Specifically, before building the test strategy optimization model, it is necessary to first collect a large amount of historical test datasets accumulated from previous tests of various automotive instruments using this system. For example, in the past year, speedometer, fuel gauge, and tachometer tests were performed on more than 1,000 different instrument models. For each test point in each test, the instrument model under test, the test items performed, the output programmable command parameters, and the collected response data were recorded, collecting at least 10,000 historical test data points.

[0032] The same data collection method is used to acquire the corresponding data, and fault labels are added to determine whether the test points are ultimately judged as abnormal. Fault labeling can be performed by experienced domain experts or professional labeling technicians, following principles such as 0 indicating normal and 1 indicating excessive display error. All collected and labeled data is compiled to form a historical test dataset, which will be used as data for subsequent model training.

[0033] Secondly, based on reinforcement learning, an initial test strategy optimization model is constructed. Reinforcement learning is a machine learning paradigm that utilizes an agent to learn to take actions to maximize cumulative rewards through interaction with the environment.

[0034] Specifically, the initial test strategy optimization model is first constructed. Its internal structure usually includes a policy network and a value network. The policy network is used to select actions based on the current state, and the value network is used to evaluate the merits of the current state. The structure of the model is initially obtained, but the parameters need to be optimized through subsequent training processes.

[0035] For example, a fully connected neural network is used to construct an initial test policy optimization model that includes a policy network and a value network, with both networks having the exact same neural network structure.

[0036] The input layer contains 3 neurons using a linear activation function to receive the feature vector from the state space; the input node is 3. This is followed by two hidden layers. The first hidden layer contains 64 neurons using the ReLU activation function to perform non-linear transformations and feature extraction on the input features. The second hidden layer contains 32 neurons, also using the ReLU activation function, to further abstract and compress the feature dimensions, extracting high-level decision information. The output layer contains 4 neurons using a linear activation function, outputting the cumulative discount reward value corresponding to increasing the step size, decreasing the step size, maintaining the step size, and jumping to the next test item, respectively. The cumulative discount reward value represents the expected cumulative discount reward after taking this action in a given state.

[0037] In the model training section, the initial learning rate was set to 0.001, the discount factor γ = 0.95, the initial exploration rate = 1.0, the exploration rate decay factor 0.995, the minimum exploration rate 0.01, the training batch size 32, and the maximum number of training epochs 5000. The mean squared error (MSE) loss function was used to calculate the difference between the main network's predicted value and the target network's calculated cumulative discounted reward value. During each training iteration, a starting state was randomly sampled from the historical test dataset. Random exploration was performed with a probability equal to the initial exploration rate, and the action with the highest current main network prediction value was selected with a probability of 1 - the initial exploration rate. The environment then returned an immediate reward, and the process transitioned to the next state.

[0038] Once the experience pool has accumulated enough samples, a small batch of experience is randomly sampled for training. For each sample, if the next state is the terminal state, the target cumulative discount reward value equals the immediate reward; otherwise, the target cumulative discount reward value equals the immediate reward plus the discount factor multiplied by the target network's output maximum cumulative discount reward value for the next state. The current state is input into the main network to obtain the predicted cumulative discount reward value. The predicted cumulative discount reward value for the corresponding action is extracted and compared with the target cumulative discount reward value to calculate the MSE loss. Then, the gradient is calculated using the backpropagation algorithm, and the Adam optimizer is used to update the weight parameters of the main network. Every C rounds, the parameters of the main network are completely copied to the target network to stabilize the training process. If the maximum number of training rounds (5000) is reached, or the early stopping condition is triggered, or the model's overall reward on the validation set remains stable at more than 98% of the theoretical maximum value for 50 consecutive rounds with a fluctuation range of less than 1%, the iteration stops.

[0039] During model validation, after each training epoch, the current policy is run on an independent validation set to perform a complete virtual test task, calculating three metrics: average overall reward, average anomaly detection rate, and average test time. When the average overall reward on the validation set no longer improves for 10 consecutive epochs, the current model is saved as the optimal model. Through this process, a pre-trained optimized test policy model is finally obtained.

[0040] For example, assuming the current state is instrument model A, the test item is speedometer detection, the current frequency value is 1200Hz, the deviation feature is -1.67%, and the four cumulative discount reward values ​​of the output layer after normalization are [5.2, 8.7, 6.1, 3.4], among which the cumulative discount reward value of decreasing the step size is the highest. Therefore, the agent chooses to decrease the step size action. After executing this action, the subsequent parameter step size decreases from 20Hz to 13.4Hz, the test point density increases, and a certain instrument display lag anomaly is successfully captured in the encrypted area, obtaining a high immediate reward. This experience is stored in the replay pool for subsequent training. After 500 rounds of iterative training, the model learns to actively decrease the step size to encrypt the test when the deviation feature is deteriorating, and to increase the step size to accelerate the test when the deviation feature is stable and far from the threshold, thereby achieving dual optimization of anomaly detection rate and test time.

[0041] Furthermore, using the instrument models and test items in the historical test dataset as the state space, the corresponding programmable instruction parameters as the action space, and the comprehensive reward function as the optimization objective, a pre-trained test strategy optimization model is obtained through multiple rounds of iterative training. Here, the state space is the set of all possible environmental states; the action space is the set of all possible actions that can be taken in a given state in the reinforcement learning model; and the comprehensive reward function is the core objective of the reinforcement learning model optimization, used to quantify the degree of good or bad obtained by the agent after taking a specific action in a specific state.

[0042] Specifically, after constructing the initial test strategy optimization model, it needs to be iteratively trained using historical test datasets. The initial test strategy optimization model takes the instrument models and test items in the historical dataset as part of its perceived state space, the programmable command parameters as its action space, and the maximized comprehensive reward function as the training objective.

[0043] The construction process of the comprehensive reward function includes: First, the ratio of the number of anomalies detected in the current round of testing to the total number of test points is obtained as the anomaly detection rate. The current round of testing is during the model training process, where each iteration represents a complete virtual test, generating a complete sequence of programmable instructions and simulating the execution of the sequence to obtain a virtual test result. The anomaly detection rate is the ratio of the number of anomalies successfully detected by the model in the current round of virtual testing to the total number of test points. The higher the value, the more effective the model's testing strategy is.

[0044] For each round of virtual testing, the model first obtains the number of anomalies detected in that round and the total number of test points, and calculates the ratio as the anomaly detection rate.

[0045] Secondly, obtain the total test time for the current round of testing. The total test time is the total time consumed by the virtual test process in the current round, typically calculated by summing the output duration and waiting time for each test point. A lower value indicates a more efficient testing strategy. The corresponding model obtains the total test time consumed from the start to the end of this round.

[0046] Finally, based on the preset first weight coefficient and second weight coefficient, the comprehensive reward function is calculated according to the following function: the comprehensive reward function is equal to the anomaly detection rate multiplied by the first weight coefficient minus the total test time multiplied by the second weight coefficient, where both the first weight coefficient and the second weight coefficient are positive values.

[0047] Specifically, the first and second weighting coefficients are hyperparameters used to adjust the relative importance of anomaly detection rate and total testing time in the overall reward function. A larger first weighting coefficient prioritizes anomaly detection rate, while a larger second weighting coefficient prioritizes shorter testing time. Both are positive values ​​and are dynamically set in advance by developers based on actual business needs.

[0048] For example, if more emphasis is placed on the anomaly detection rate and less on the testing time, the first weighting coefficient can be preset to 0.7 and the second weighting coefficient to 0.3.

[0049] The overall reward is calculated as follows: (Anomaly detection rate × First weight coefficient) - (Total testing time × Second weight coefficient). For example, if the anomaly detection rate in a certain round of testing is 90% and the total testing time is 100 seconds, then the overall reward is (0.9 × 0.7) - (100 × 0.3) = -29.37. A higher reward value indicates a better policy. The model undergoes thousands or even tens of thousands of iterations of training, continuously adjusting its internal neural network parameters so that the generated testing policy can obtain increasingly higher overall rewards. Finally, when the overall reward converges to a stable value, the pre-trained optimized testing policy model is obtained.

[0050] Step S10 in the method provided in this application embodiment further includes: Receive standard test specification documents uploaded by users, wherein the standard test specification documents include at least one of enterprise standards, industry standards or vehicle technical specifications; The standard test specification document is parsed using natural language processing technology to extract test items, signal types, parameter ranges, test sequences, and judgment criteria, and to generate initial test cases. The natural language processing technology includes at least word segmentation, entity recognition, and relation extraction. The initial test cases are revised and confirmed. The revised and confirmed test cases are associated with the corresponding instrument models and stored in the cloud test case library, and a revision log is recorded. The correction records in the correction log are used as training samples to iteratively update the test strategy optimization model.

[0051] In this embodiment of the application, the standard test specification document uploaded by the user is first received. The standard test specification document includes at least one of the following: enterprise standard, industry standard, or vehicle technical specification. The standard test specification document is a technical document issued by an authoritative organization that specifies the test methods for various functions and performance indicators of automotive instruments.

[0052] Natural Language Processing (NLP) enables computers to understand, parse, and generate human natural language. It is used to extract structured test information from unstructured, specification document text. Word segmentation, a technique in NLP, divides a continuous text sequence into independent word units. Entity recognition identifies named entities with specific meanings from the segmentation results. Relation extraction identifies the semantic relationships between entities. Initial test cases are structured test cases automatically generated by NLP technology after extracting information from specification documents.

[0053] Specifically, when the testing system needs to build a test case library, the user first uploads the relevant standard test specification documents to the cloud server through the client interface. Upon receiving the documents, the cloud server uses natural language processing technology to perform deep analysis. First, a combination of dictionary-based and statistical models is employed, typically using open-source word segmentation tools such as Jieba and HanLP, to segment the text. A domain-specific dictionary is pre-loaded to ensure that long sentences and paragraphs are accurately divided into independent word sequences. Then, entity recognition is performed. From the word segmentation results, key entities are located and labeled using regular expressions. For example, speedometer detection is identified as a test item, frequency signals as signal types, and 10Hz to 2000Hz as the parameter range, accurately extracting key information. Finally, through relation extraction, a syntax tree of the sentence is constructed based on syntactic analysis, and logical connections are established between discrete entities and predefined extraction rules. At the same time, general information extraction models such as UIE can be used to process complex sentence structures. Finally, these entities and their relations are integrated into structured data records, which clearly describe key fields such as test items, signal types, parameter ranges, and test actions. This automatically transforms unstructured natural language text into initial test cases that can be directly used by automated testing platforms.

[0054] Secondly, the standard test specification document is parsed using natural language processing technology to extract test items, signal types, parameter ranges, test sequences and judgment criteria, and generate initial test cases. Natural language processing technology includes at least word segmentation, entity recognition and relation extraction. Next, the initial test cases are revised and confirmed. The revised and confirmed test cases are then associated with their corresponding instrument models and stored in the cloud-based test case library, with a revision log recorded. This association storage establishes a clear link between the revised and confirmed test cases and their corresponding instrument models, and the association record is stored in the cloud-based test case library. This allows for quick matching and retrieval of the appropriate test cases when performing subsequent testing tasks for that instrument model. The revision log is a historical record of each revision operation performed on the initial test cases. The revision log typically includes the revision time, the person making the revision, the original error field, the content before revision, the content after revision, and the reason for revision.

[0055] Specifically, after generating initial test cases using NLP technology, these cases are first reviewed by a test expert or system administrator. If errors are found during the NLP parsing process, such as incorrectly extracting decision criteria, the initial test cases are then corrected, replacing erroneous fields with correct content. After all corrections are completed, a confirmation process is performed, marking the corrected test cases as accurate. Subsequently, the cloud server associates the corrected and confirmed test cases with their corresponding instrument models and inserts them into the relevant data table in the cloud test case library. Simultaneously, a correction log is generated, recording all correction information as a single log entry and stored in a dedicated log database.

[0056] Finally, the correction records in the correction log are used as training samples to iteratively update the testing strategy to optimize the model. The iterative update is the process of continuously adding newly generated training samples to the model's training dataset and retraining or fine-tuning the model periodically or in real time, so that the model's performance gradually improves as data accumulates.

[0057] Specifically, during the construction and maintenance of the cloud-based test case library, the more standard test specification documents are processed, the more correction logs are accumulated. These correction logs are used as training samples to iteratively update the natural language processing model used to parse the specification documents in the constructed test strategy optimization model.

[0058] A batch of correction records can be periodically extracted from the correction log library. For each record, the incorrectly parsed text fragment from the original document is used as input, and the manually corrected content is used as the output target, forming a set of training samples. Then, the NLP model is further trained or fine-tuned using these samples. Through multiple rounds of such iterative updates, the probability of the NLP model making the same mistakes when parsing similar documents will be significantly reduced. For example, suppose that when processing a car manufacturer's specification document, voltage signals are frequently misidentified as frequency signals. After several manual corrections, dozens of such correction records accumulate in the correction log. After fine-tuning the model using these correction records, the model learns specific vocabulary in the document, and the recognition accuracy improves from the initial 85% to 98% when processing new versions of specification documents.

[0059] In this embodiment, the test task is automatically parsed by the cloud server, and the model of the instrument under test and the test items are extracted, eliminating the ambiguity and error risk of manual interpretation and realizing the automated parsing of the test task. Then, the basic test case that completely corresponds to the current test object is located from the cloud test case library. By defining the basic test case as a structured field containing signal type, parameter range, test timing, judgment criteria, etc., the test case can be directly understood and processed by the computer program, laying the data foundation for subsequent model optimization.

[0060] S20: The cloud server sends the initial programmable instruction sequence to the end-side test device. The end-side test device outputs test signals to the instrument panel of the vehicle under test based on the initial programmable instruction sequence and simultaneously collects the response data of the instrument panel of the vehicle under test. In this embodiment, the sending process is the process by which a cloud server transmits data from the cloud to a designated end-side test device via a network communication protocol; the end-side test device is a portable hardware device physically connected to the instrument panel of the vehicle under test; the instrument panel of the vehicle under test is the instrument cluster assembly to be tested, and its physical interface is connected to the end-side test device; the test signal is a series of electrical signals generated by the end-side test device according to instructions, simulating the output of a vehicle sensor or controller, used to excite the instrument panel under test.

[0061] Synchronous acquisition involves capturing the response of the instrument under test (DUT) immediately, either simultaneously with or within a very short time delay, the output of each test signal. Response data refers to the measurable output generated by the DUT after receiving the test signal. Examples include the angular position of a pointer instrument, the value displayed on an LCD screen, the on / off status of indicator lights, and communication message data transmitted via the CAN bus.

[0062] Specifically, the cloud server generates initial programmable commands and sends them over the network to the end-side test device connected to the instrument under test. Upon receiving the command sequence, the end-side test device's internal microcontroller parses the commands and, according to the sequence's timing, controls its internal signal generation circuit to output test signals to the corresponding input pins of the vehicle's instrument panel. Simultaneously, the end-side device's synchronous acquisition function is activated to collect data, ultimately obtaining the response data constituting that test point.

[0063] Step S20 in the method provided in this application embodiment includes: The cloud server sends the initial sequence of programmable commands to the end-side test device; The end-side test equipment receives and parses the initial programmable instruction sequence to generate a signal output timing table, wherein the signal output timing table arranges the signal type, output value and duration of each test point in chronological order. The end-side test equipment outputs test signals to the vehicle instrument under test sequentially through its internally integrated signal output module according to the signal output timing table. The signal output module includes at least a high / low level output module, a voltage output module, a PWM wave output module, a frequency output module, and a resistance output module. While outputting test signals, the end-side test equipment collects the response data of the vehicle instrument panel under test in real time through the internally integrated detection module. The detection module includes at least a voltage detection module, a CAN communication module, and an image acquisition module. The response data includes at least the instrument pointer position, the LCD screen display content, the indicator light status, and the CAN bus message. The end-side testing equipment associates and stores the collected response data with the signal type and output value of the current test point, forming a timestamped test data record.

[0064] In this embodiment, the cloud server first sends the initial programmable instruction sequence to the end-side test device. The sending is the process by which the cloud server transmits data from the cloud to the designated end-side test device through a network communication protocol. Considering the real-time and reliability of the test, a long connection method is usually adopted, and a data verification and retransmission mechanism is added to ensure the complete arrival of the instruction sequence. The end-side test device is a portable hardware device that is physically connected to the instrument panel of the vehicle under test. It is usually composed of a microcontroller, a communication module, a signal generation circuit, a signal acquisition circuit, a power management system, etc.

[0065] Specifically, the cloud server generates an initial sequence of programmed instructions. For example, the cloud generates a JSON array containing 200 instructions, each specifying time, signal type, output value, and duration. The cloud server then distributes this initial sequence of instructions to the designated edge testing device via the network. The distribution process typically employs encrypted transmission, such as the TLS protocol, to ensure data security. The communication module of the edge testing device continuously listens for messages from the cloud. Upon receiving the instruction sequence data packet, it unpacks and verifies it. Once confirmed to be correct, it stores it in local non-volatile memory, awaiting further processing. For instance, the cloud generates a sequence of 200 frequency points for testing a vehicle speedometer, which is then completely distributed to the edge device in the laboratory within 0.5 seconds via a 5G network.

[0066] Secondly, the end-side test equipment receives and parses the initial programmable instruction sequence to generate a signal output timing table. The signal output timing table arranges the signal type, output value, and duration of each test point in chronological order. Parsing is the process by which the microcontroller of the end-side test equipment performs syntax analysis and data extraction on the received instruction sequence data. The signal output timing table is a two-dimensional table structure used to drive the signal output module, sorted by time, with each row arranged in chronological order of execution.

[0067] Specifically, after receiving the initial sequence of programmable commands, the end-side test equipment's internal microcontroller (MCU) parses the data, extracts the signal type, output value, and duration corresponding to each test point at different time offsets, and sorts them according to sequence number or absolute timestamp to generate a signal output timing table. For example, when testing the response time of the instrument to the PWM signal, the timing table ensures a precise 10ms interval between the end of the previous frequency signal and the start of the next PWM signal to avoid signal superposition interference.

[0068] Next, the end-side test equipment outputs test signals to the vehicle instrument panel under test sequentially through its internally integrated signal output modules according to the signal output timing table. These signal output modules include at least a high / low level output module, a voltage output module, a PWM wave output module, a frequency output module, and a resistance output module. The internally integrated signal output modules are hardware units on the internal circuit board of the end-side test equipment specifically designed to generate specific types of electrical signals. The high / low level output modules are circuits capable of outputting stable digital logic levels. The voltage output modules are circuits capable of outputting continuously adjustable analog voltage signals to simulate the output of linear sensors.

[0069] The PWM output module is a circuit that can output pulse width modulation signals, with independently adjustable frequency and duty cycle, used to simulate signals requiring dimming or speed control; the frequency output module is a circuit that can output adjustable frequency square wave or sine wave signals, used to simulate pulse sequences generated by vehicle speed sensors, tachometers, etc.; the resistor output module is a circuit that can output programmable resistance values, usually implemented through digital potentiometers or resistor networks, used to simulate the output of variable resistors such as fuel level sensors and coolant temperature sensors.

[0070] Specifically, after generating the signal output timing table, the MCU of the end-side test equipment activates the corresponding internally integrated signal output modules sequentially according to the time order in the timing table, outputting all types of electrical signals required for instrument testing in the timing table. For example, the first row of the timing table requires the output of a PWM wave with a duty cycle of 30% and a duration of 1000ms. The MCU then sends configuration commands, such as frequency = 1kHz and duty cycle = 30%, to the PWM wave output module via its internal bus, and then activates the module's output pin. After 1000ms, the MCU switches to the frequency output module according to the next row of the timing table, configuring it to output a 100Hz square wave for 2000ms. This process continues sequentially until all test points in the timing table have been completed. Throughout the entire process, the MCU is responsible for precise time management and module switching, ensuring that the timing accuracy of the signal output reaches the millisecond level.

[0071] Simultaneously, while outputting test signals, the end-side test equipment collects the response data of the vehicle's instrument panel in real time through its internally integrated detection module. The detection module includes at least a voltage detection module, a CAN communication module, and an image acquisition module. The response data includes at least the instrument pointer position, LCD screen display content, indicator light status, and CAN bus messages.

[0072] Specifically, the detection module is a hardware unit integrated within the end-side testing equipment, used to sense and measure the output signal of the instrument under test; the voltage detection module is a circuit used to measure the analog voltage or current signal output by the instrument under test, such as measuring the backlight brightness control voltage fed back by the instrument; the CAN communication module is a circuit unit containing a CAN controller and a transceiver, used to communicate with the instrument's bus interface, receive CAN messages sent by the instrument, and obtain the instrument's internal status, fault codes, diagnostic information, etc.

[0073] The image acquisition module uses a miniature camera or image sensor to capture real-time visual information such as pointer position, LCD screen digits, and indicator light status on / off by pointing it at the instrument's display panel. Combined with image recognition algorithms, the image can be converted into structured data. The pointer position is obtained from the image acquisition and angle detection algorithms, showing the pointer's deflection angle relative to zero or the scale value it points to. The LCD screen display content is extracted from the image using optical character recognition technology, showing the numbers, text, or icons displayed on the LCD screen. The indicator light status is determined by image analysis to indicate the ON / OFF state of each indicator light. The CAN bus message is a data frame conforming to a specific protocol sent by the instrument via the CAN bus, containing rich internal information.

[0074] While the end-side test equipment outputs test signals according to the timing schedule, the MCU triggers the internally integrated detection module to perform real-time data acquisition. Specifically, the voltage detection module is connected to the analog output pin of the instrument to measure and record the voltage value. The CAN communication module listens to the CAN bus, captures all messages sent by the instrument, and filters out messages relevant to the current test.

[0075] The image acquisition module captures images of the dashboard using a miniature camera. The MCU or cloud-based system then extracts the pointer position, LCD display content, and indicator light status from these images. For example, when outputting a 1200Hz test signal, the image acquisition module captures an image of the dashboard and identifies the pointer as pointing to 118 km / h; the CAN communication module captures a data frame with ID 0×123 from the instrument cluster, which contains a vehicle speed of 117 km / h; and the voltage detection module may measure a backlight voltage of 2.5V.

[0076] Finally, the end-side test equipment associates and stores the collected response data with the signal type and output value of the current test point to form a timestamped test data record. That is, the MCU of the end-side test equipment associates the scattered response data: pointer position 118, CAN vehicle speed 117, backlight voltage 2.5 with the signal type of the current test point: frequency, output value: 1200Hz, and adds the current system timestamp to form a timestamped test data record.

[0077] In this embodiment, the initial programmable command sequence sent from the cloud is parsed by the end-side testing equipment to generate a signal output timing table, avoiding timing conflicts and signal interference, and ensuring the controllability and repeatability of the testing process. By integrating high / low level output modules, voltage output modules, PWM wave output modules, frequency output modules, and resistance output modules within the end-side testing equipment, the convenience and automation of the testing are improved. By integrating voltage detection modules, CAN communication modules, and image acquisition modules, the instrument's response performance is comprehensively captured from multiple dimensions, avoiding potential omissions due to single detection methods. The acquired response data is associated with and stored according to the signal type and output value of the current test point, and a precise timestamp is added to form a structured test data record. This allows each response data to trace back to the excitation signal, providing a solid data foundation for subsequent cloud-based evaluation and anomaly analysis.

[0078] S30: The end-side testing device uploads the response data to the cloud server in real time; In this embodiment of the application, real-time uploading is the process by which the end-side test device, after collecting response data, sends it back to the cloud server as soon as possible via the network without performing long-term storage or complex local processing.

[0079] Specifically, after the edge testing equipment completes the signal output and response data acquisition for one or more test points, the edge equipment needs to send the response data to enable subsequent intelligent analysis and dynamic adjustment in the cloud. First, the edge equipment encapsulates the response data according to a predefined data format and uploads it in real time to the designated data receiving interface of the cloud server using a wireless network. For example, after completing the test point outputting the frequency corresponding to 60km / h, the edge equipment will immediately send a data packet containing information such as timestamp T1, signal type vehicle speed frequency, output value 600Hz, acquired pointer angle 58 degrees, and CAN message to the cloud.

[0080] In this embodiment, a real-time data channel is established from the physical test site to the cloud through a real-time upload mechanism. The subsequent test strategy is dynamically adjusted based on the current test situation to realize the key data flow of adaptive dynamic testing.

[0081] S40: The cloud server evaluates the current test progress and anomaly detection status in real time based on the response data, dynamically adjusts the parameter step size and test point density of the initial programmable instruction sequence, generates an optimized programmable instruction sequence, and sends the optimized programmable instruction sequence to the end-side test device for continued execution; In this embodiment, dynamic adjustment is the process by which the cloud server modifies and optimizes the parameters of the unexecuted test instructions in real time based on the continuously received real-time response data during test execution. The parameter step size is the amount of change or increment between two adjacent test points when testing continuously changing parameters. A larger step size results in faster testing speed but lower resolution, while a smaller step size results in higher resolution but slower testing speed. The test point density is the number of test points set per unit parameter range or unit time. The higher the density, the more refined the test coverage of the area. The optimized programmable instruction sequence is a newly generated instruction sequence based on the original initial instruction sequence, after adjusting the parameter step size and test point density of subsequent parts according to the real-time evaluation results.

[0082] Specifically, once the cloud server continuously receives response data uploaded in real-time from the edge devices, it executes decision-making functions. First, it assesses the current test progress in real-time by calculating the ratio of the number of data points to the total planned number of data points. Simultaneously, it analyzes the latest response data to determine if any anomalies are present. Based on the results, it dynamically adjusts the parameter step size and test point density for the remaining untested portions. After adjustment, the cloud server reassembles the optimized instruction portion, generates a new optimized programmable instruction sequence, and immediately sends it to the edge testing equipment via the network to execute subsequent tests.

[0083] Step S40 in the method provided in this application embodiment includes: The cloud server receives the timestamped test data records uploaded in real time by the end-side test device, calculates the ratio of the number of completed test points to the total number of test points in the initial program control instruction sequence, and uses it as the current test progress value. The cloud server inputs the signal type, parameter value, and real-time deviation characteristics of the current test point into the pre-trained abnormal pattern prediction model, outputs the predicted abnormal probability value of anomalies occurring within the subsequent parameter range, and multiplies the predicted abnormal probability value by the equipment aging compensation coefficient to obtain the abnormal probability value. The adjustment coefficient is obtained by subtracting the current test progress value from 1 and then multiplying it by the anomaly probability value. The subsequent parameter step size is obtained by multiplying the preset base step size by the adjustment coefficient, and the subsequent test point density is obtained by dividing the preset base density by the adjustment coefficient. An optimized programmable instruction sequence is generated based on the subsequent parameter step size and subsequent test point density, and the optimized programmable instruction sequence is sent to the end-side test equipment to continue executing subsequent tests; The construction process of the abnormal pattern prediction model includes: Historical test data is extracted from the historical test database. The historical test data includes normal test data and abnormal test data. The abnormal test data includes signal type, parameter value, deviation characteristics, and actual fault cause after the abnormality occurred. A training sample set is constructed using the signal type, parameter value, and deviation characteristics in the historical test data as input features and the corresponding anomaly label as output label. Here, an anomaly label of 1 indicates the presence of an anomaly, and an anomaly label of 0 indicates the absence of an anomaly. An initial anomaly pattern prediction model is constructed using a neural network algorithm. The initial anomaly pattern prediction model includes an input layer, a hidden layer, and an output layer. The output layer uses a Sigmoid activation function to output the predicted anomaly probability value. The initial anomaly pattern prediction model is trained in a supervised manner using the training sample set until it is verified to converge, thus obtaining a pre-trained anomaly pattern prediction model.

[0084] In this embodiment, the cloud server first receives the timestamped test data records uploaded in real time by the end-side test device, calculates the ratio of the number of completed test points to the total number of test points in the initial programmable instruction sequence, and uses this ratio as the current test progress value. The total number of test points in the initial programmable instruction sequence is the total number of test points included in the initial programmable instruction sequence generated by the cloud before the test begins; the current test progress value is the percentage of the current test task completed.

[0085] Specifically, during test execution, the edge testing equipment uploads timestamped test data records to the cloud server in real time. Each time the cloud server's data receiving module receives a valid test data record, it immediately stores it in a temporary database and updates the number of completed test points. Assuming the initial sequence of programmable instructions contains 200 test points, and the data record for the 50th test point has been received and confirmed, the cloud server performs an update calculation, dividing the completed 50 by the total number of points (200) to obtain a current test progress value of 25%. This current test progress value is continuously updated as the test progresses and serves as an important input parameter for subsequent adjustment coefficient calculations.

[0086] When the test progress value is 90%, there are only 10% of test points remaining on the cloud server. Even if a high risk of anomalies is predicted later, it is not advisable to reduce the step size too much, otherwise too many additional test points will be added, resulting in a serious overrun of test time. Conversely, when the progress value is 10%, there is a lot of room left, and the step size can be adjusted more aggressively according to the probability of anomalies.

[0087] Secondly, the cloud server inputs the signal type, parameter value, and real-time deviation characteristics of the current test point into the pre-trained anomaly pattern prediction model, and outputs the predicted anomaly probability value for anomalies occurring within the subsequent parameter range. The predicted anomaly probability value is multiplied by the equipment aging compensation coefficient to obtain the anomaly probability value. The real-time deviation characteristics are extracted from the response data of the current test point and reflect the degree to which the instrument response deviates from expectations. The anomaly pattern prediction model is a neural network model trained with historical data, taking signal type, parameter value, and deviation characteristics as inputs, and outputting probability values ​​between 0 and 1.

[0088] The subsequent parameter range is a range of parameters following the parameter value of the current test point, and its length can be dynamically set according to the test item. The predicted anomaly probability value is the original probability output by the model, ranging from [0,1]. The larger the value, the higher the probability of anomalies occurring in the subsequent range. The equipment aging compensation coefficient is a decimal between 0 and 1, used to correct the impact of accuracy drift caused by long-term use of the end-side test equipment on anomaly prediction. The anomaly probability value is the final anomaly probability after aging compensation correction, used for the calculation of subsequent adjustment coefficients.

[0089] After receiving the data from the current test point, the cloud server extracts the corresponding signal type and parameter values, and calculates the real-time deviation characteristics. For example, if the measured vehicle speed is 118 km / h and the theoretical speed is 120 km / h, the absolute deviation is -2 km / h and the relative deviation is -1.67%. This data is then fed into a pre-trained anomaly pattern prediction model. After forward propagation, the model outputs a predicted anomaly probability value, representing the probability that an anomaly will occur in the parameter range following the parameter value at the current test point.

[0090] During long-term use, the signal output and acquisition modules of end-side testing equipment may experience systematic deviations between their actual output values ​​and theoretical settings due to component aging, mechanical wear, and other factors. This is known as equipment aging drift. Therefore, the equipment aging compensation coefficient of the current end-side testing equipment is multiplied by the predicted anomaly probability value to obtain the anomaly probability value. Generally, the more severe the equipment aging, the smaller the compensation coefficient, resulting in less compensation for the predicted anomaly probability and avoiding false alarms caused by equipment errors; conversely, the less severe the equipment aging, the larger the compensation.

[0091] Assuming that the voltage detection module of the end-side device of the aging equipment drifts, resulting in a systematically low response data, the model outputs a relatively high predicted anomaly probability of 0.75, and the aging compensation coefficient is 0.9952. The corrected anomaly probability value is 0.75 × 0.9952 ≈ 0.75.

[0092] Furthermore, subtracting the current test progress value from 1 and multiplying it by the anomaly probability value yields the adjustment coefficient. This adjustment coefficient dynamically adjusts the subsequent parameter step size and test point density. For example, 1 minus the current test progress value multiplied by the anomaly probability value = (1-0.25)×0.75≈0.56.

[0093] The adjustment coefficient can adaptively control the aggressiveness of subsequent testing strategies. For example, in the early stages of testing, the probability of anomalies is not high, and because the remaining proportion is large, the adjustment coefficient will not be too small, allowing for a certain amount of strategy adjustment space. In the later stages of testing, even if the probability of anomalies is high, the adjustment coefficient will be suppressed because the remaining proportion is small, preventing over-adjustment from causing the test to fail to be completed on time.

[0094] Furthermore, the subsequent parameter step size is obtained by multiplying the preset base step size by the adjustment coefficient, and the subsequent test point density is obtained by dividing the preset base density by the adjustment coefficient. Here, the preset base step size is the pre-set parameter step size value; the subsequent parameter step size is the actual step size that will be used for the untested portion after dynamic adjustment; the subsequent test point density is the actual density that will be used for the untested portion after dynamic adjustment; and the preset base density is the pre-set test point density value, which is the number of test points within a unit parameter interval.

[0095] The preset base step size is usually derived from the default step size in the basic test cases or manually set according to the characteristics of the test item. For example, for speedometer testing, the preset base step size could be 10Hz; for fuel gauge testing, it could be 10Ω. Subsequent parameter step size = preset base step size × adjustment coefficient. Since the adjustment coefficient is usually between 0 and 1, the subsequent step size is usually less than or equal to the preset base step size. The smaller the adjustment coefficient, the smaller the step size and the denser the test points; the larger the adjustment coefficient, the larger the step size and the sparser the test points. Subsequent test point density = preset base density / adjustment coefficient. Since the adjustment coefficient ≤ 1, the subsequent density is usually greater than or equal to the preset base density. The smaller the adjustment coefficient, the higher the density and the denser the test points; the larger the adjustment coefficient, the lower the density and the sparser the test points.

[0096] For example, assuming a preset base step size of 20Hz, meaning one test point is set every 20Hz, and a preset base density of 0.5 test points per 10Hz, then the subsequent parameter step size is 20 × 0.56 ≈ 11.2Hz; the subsequent test point density is 0.5 / 10 ÷ 0.56 ≈ 0.89 points / 10Hz, meaning one test point every 11.2Hz, consistent with the calculation result of the step size.

[0097] Furthermore, an optimized programmable instruction sequence is generated based on the subsequent parameter step size and subsequent test point density, and then sent to the end-side test equipment to continue executing subsequent tests. The optimized programmable instruction sequence is a newly generated instruction sequence based on the initial programmable instruction sequence, according to the dynamically adjusted step size and density, and only contains the subsequent untested parts. Sending is done by the cloud server transmitting the newly generated optimized instruction sequence to the end-side test equipment via the network. Continuing to execute subsequent tests means that after receiving the optimized instruction sequence, the end-side test equipment stops executing the original remaining instructions and instead continues to output signals and collect data according to the new instruction sequence.

[0098] Specifically, after calculating the subsequent parameter step size and test point density, the cloud server, based on the parameter value of the last completed test point and the upper limit of the test parameter range, generates test point instructions starting from the parameter value of the last test point plus one subsequent parameter step size, increasing at intervals of subsequent parameter step sizes until approaching the upper limit of the test parameter range. The generated test point instructions are packaged into an optimized programmable instruction sequence. The cloud server sends the optimized sequence to the edge testing device. Upon receiving the sequence, the edge device replaces the old instruction sequence that has not yet been executed locally and continues to execute subsequent tests, outputting signals and collecting data according to the new, finer step size.

[0099] The construction process of the abnormal pattern prediction model includes: First, historical test data is extracted from the historical test database. The historical test data includes normal test data and abnormal test data. Abnormal test data includes signal type, parameter value, deviation characteristics, and actual fault cause after the abnormality occurred. The historical test database is a large database stored in the cloud that accumulates complete data of all past test tasks. It includes information such as instrument model, test items, programmable instruction sequence, response data, judgment results, and manually annotated fault causes for each test.

[0100] Normal test data consists of test records where the response data of all test points meets the judgment criteria and is marked as qualified; abnormal test data consists of test records where the response data of at least one test point does not meet the judgment criteria and is marked as unqualified; input features are independent variables used to train the model; abnormal labels are target variables used for supervised learning.

[0101] Specifically, when building an anomaly pattern prediction model, the first step is to prepare training data. The cloud server extracts a large amount of historical test data from the historical test database, including both normal and anomaly test data.

[0102] Secondly, for each test record, the signal type and parameter value are extracted, and the deviation feature is calculated. This is then used as input features, with the anomaly label corresponding to the test point as the output target, to construct a training sample set. A label of 1 indicates the presence of an anomaly, meaning the test point or test interval is confirmed as unqualified; a label of 0 indicates the absence of an anomaly. For example, a sample can be represented as: Input (signal type = resistance, parameter value = 350Ω, deviation feature = +0.25 divisions), output (anomaly label = 1). At least 1000 sample data points are trained, and these sample data are then integrated to obtain the training sample data.

[0103] Furthermore, an initial anomaly pattern prediction model is constructed using a neural network algorithm. This model includes an input layer, a hidden layer, and an output layer. The output layer uses a sigmoid activation function to output the predicted anomaly probability value. The neural network algorithm is a machine learning algorithm that mimics the structure and function of biological neural networks. It consists of multiple neuron layers and can fit complex nonlinear function relationships by learning from a large amount of data.

[0104] The Sigmoid activation function is a commonly used activation function that maps any real-valued input to the (0,1) interval and outputs a probability value. Each training sample in supervised training contains input features and corresponding true output labels. By comparing the difference between the predicted output and the true label, the loss function is calculated, and then the backpropagation algorithm is used to update the model parameters so that the prediction gradually approaches the true value. Validation convergence is considered to have reached convergence when the loss function value on the validation set no longer decreases significantly or begins to increase. This is used to determine whether to stop training. A pre-trained anomaly pattern prediction model is a neural network model with good generalization ability after supervised training and validation convergence. It can be used to predict the anomaly probability of new input test point data in real time.

[0105] Specifically, the initial anomaly pattern prediction model includes an input layer, hidden layers, and an output layer. The input layer is the first layer of the neural network, responsible for receiving input features. The number of neurons in this layer equals the dimension of the input features, for example, 1 / 2. The hidden layers are intermediate layers between the input and output layers, and there can be one or more. The neurons in the hidden layers perform non-linear transformations on the input features, extracting higher-level feature representations. The number of hidden layers and the number of neurons in each layer are hyperparameters in the model design. The output layer is the last layer of the neural network, responsible for outputting the final prediction result, and includes one neuron to output the anomaly probability value.

[0106] For example, a fully connected neural network is used to construct the initial anomaly pattern prediction model, which includes an input layer, a hidden layer, and an output layer. The input layer contains 8 neurons with a linear activation function to receive feature vectors and capture non-linear relationships; the hidden layer contains 32 neurons with a ReLU activation function to perform non-linear transformation and feature extraction on the input features; the output layer contains 1 neuron with a Sigmoid activation function to compress the linear combination result of the hidden layer output to between 0 and 1, directly outputting the predicted anomaly probability value.

[0107] Finally, the initial anomaly pattern prediction model is trained in a supervised manner using the training sample set until it is verified to converge, thus obtaining the pre-trained anomaly pattern prediction model.

[0108] Specifically, during training, a batch of samples is input into the model each time, the cross-entropy loss between the predicted probability and the true label is calculated, and then the model's weights and bias parameters are updated using the backpropagation algorithm. This process is repeated, and the model performance is evaluated on the validation set after each training epoch. When the loss value on the validation set no longer decreases for several consecutive epochs, the model is considered to have converged in validation, and training is stopped.

[0109] For example, 80% of the training sample set is used as the training set, and 20% as the validation set. The initial learning rate is set to 0.001, the Adam optimizer is used, and the maximum number of training epochs is 200. The batch size is 64. Internally, the model first adjusts the weights and biases through forward propagation, then obtains the predicted anomaly probability value using the Sigmoid activation function. Subsequently, it calculates the output layer error using backpropagation and the binary cross-entropy loss function, calculates the hidden layer error and weight gradient using the chain rule, and finally updates the weight matrix and biases using the Adam optimizer based on the calculated gradients. After each training epoch, the loss value is calculated on the validation set. If the validation loss does not decrease for 20 consecutive epochs, training stops. The criteria for stopping iterations are: reaching the maximum number of training epochs (200), triggering the early stopping condition, or the validation set loss value being below 0.01 for 10 consecutive epochs and the validation accuracy exceeding 98%.

[0110] After each training epoch, the current model is run on the validation set to calculate the validation loss and validation accuracy. Validation accuracy is defined as follows: a predicted anomaly probability greater than 0.5 is considered an anomaly, while a probability consistent with the true label is considered normal. When the loss value on the validation set no longer decreases for 10 consecutive epochs, the current model is saved as the initial anomaly pattern prediction model.

[0111] In step S40 of the method provided in this application embodiment, the calculation process of the equipment aging compensation coefficient includes: The cloud server obtains the device identifier of the current edge-side test device and the cumulative usage time of the current edge-side test device; The cloud server retrieves the historical calibration records of the current end-side test device from the cloud device archive based on the device identifier. The historical calibration records include the time points of each calibration, the usage time during calibration, and the calibration results. The calibration results include the relative deviation between the measured value of the current end-side test device under the standard signal input and the standard value. The relative deviation is equal to the difference between the measured value and the standard value divided by the standard value. The cloud server constructs the accuracy drift curve of the current end-side test equipment based on the usage time and corresponding calibration results of each calibration in the historical calibration record using a curve fitting method. The cloud server queries the accuracy drift curve based on the cumulative usage time of the current edge test equipment to obtain the corresponding current relative deviation value. Calculate the sum of the absolute value of the current relative deviation and 1 to obtain the denominator for aging compensation; Divide 1 by the aging compensation denominator to obtain the equipment aging compensation coefficient.

[0112] In this embodiment of the application, the cloud server first obtains the device identifier and the cumulative usage time of the current edge test device. The device identifier is a unique code or serial number that identifies the edge test device and is usually fixed in the non-volatile memory of the device during production. The cumulative usage time is the total working time of the current edge test device from the first time it was put into use until the current moment when it actually performs the test task.

[0113] When the cloud server needs to calculate the aging compensation factor for an edge test device currently performing a test task, it first sends a request to the device via the communication link, or obtains the device identifier and cumulative usage time from the device's most recently reported status information. The cumulative usage time does not include standby, hibernation, or shutdown time and is typically accumulated by the device's built-in clock each time a test task is executed. The device identifier is used to uniquely locate the device's information file in the cloud database, and the cumulative usage time serves as the x-axis of the accuracy drift curve for subsequent queries. For example, an edge test device with device identifier S has a cumulative usage time of 1560 hours.

[0114] Secondly, the cloud server retrieves the historical calibration records of the current end-side test equipment from the cloud device archive based on the device identifier. The historical calibration records include the time point of each calibration, the usage time during calibration, and the calibration results. The calibration results include the relative deviation between the measured value of the current end-side test equipment under the standard signal input and the standard value. The relative deviation is equal to the difference between the measured value and the standard value divided by the standard value.

[0115] Historical calibration records are data records generated when the equipment has undergone metrological calibration in the past; calibration time point is the date and time of performing the calibration operation; usage duration during calibration is the cumulative number of hours the equipment has been used up to that calibration time; calibration results typically include the relative deviation between the test values ​​of multiple signal channels of the equipment and the standard values ​​under standard signal input; standard signal input is the standard, known quantity electrical signal input from the high-precision calibration source to the end-side test equipment during the calibration process.

[0116] The measured value is the actual value measured by the internal detection module of the end-side testing equipment after receiving the standard signal input; the relative deviation is an indicator that quantifies the degree to which the measurement accuracy of the equipment deviates from the standard. A positive relative deviation indicates that the measured value of the equipment is too high, and a negative relative deviation indicates that it is too low.

[0117] Specifically, the cloud server initiates a query request to the cloud device archive based on the device identifier. The database returns all historical calibration records for the device, obtaining the time point of each calibration, the usage duration during calibration, and the calibration results. The calibration results are typically the relative deviations between the test values ​​of multiple signal channels of the device and the standard values ​​under standard signal input. The relative deviation is calculated as (measured value - standard value) / standard value. For example, if the standard value for a certain signal channel is 5V, and the device's measured value is 4.9V, then the relative deviation is (4.9 - 5) / 5 = -0.02.

[0118] Furthermore, the cloud server constructs the accuracy drift curve of the current end-side test equipment using a curve fitting method based on the usage time and corresponding calibration results of each calibration in the historical calibration record. The curve fitting method finds the continuous curve that best represents the changing trend of discrete data points. The choice of method depends on the theoretical model of the equipment aging law or the distribution characteristics of the data points.

[0119] The accuracy drift curve is a continuous curve with cumulative usage time as the x-axis and relative deviation as the y-axis, describing how the measurement accuracy of an equipment changes with increasing usage time. The curve can be monotonically increasing, monotonically decreasing, or increasing then decreasing, depending on the actual physical process of equipment aging.

[0120] Specifically, to predict the relative deviation of a device over any usage duration, the cloud server uses a curve fitting method to process the data points. Assuming that device aging typically exhibits a linear or approximately linear trend, a linear regression method can be chosen. Let the relative deviation be y and the usage duration be x, fitting a straight line: Relative deviation y = a × x + b. Here, a and b are the coefficients of the linear equation, a represents the slope, and b represents the intercept, which can be determined through the linear fitting process.

[0121] According to the curve, the relative deviation is approximately +0.14% at 0, -0.0026 at 1000, and -0.0066 at 2000. If the data points exhibit a non-linear trend, such as accelerated aging, a quadratic polynomial or exponential function can be chosen for fitting, for example, a relative deviation of... Where c, d, and e are the coefficients of the equation, which can be calculated based on the fitting process.

[0122] Furthermore, the cloud server queries the accuracy drift curve based on the current cumulative usage time of the end-side test equipment to obtain the corresponding current relative deviation value. The query involves substituting the known horizontal axis value into the function expression of the accuracy drift curve to calculate the corresponding vertical axis value. The current relative deviation value is the relative deviation of the equipment's measurement accuracy under the current cumulative usage time, predicted based on the accuracy drift curve. It reflects the direction and degree of the systematic deviation of the equipment's measured value relative to the standard value under the current aging state.

[0123] Based on the function expression of the accuracy drift curve, the cumulative usage time of the current end-side test equipment is substituted to calculate the current relative deviation value. For example, if the least squares method is used to calculate a = -0.000004 / hour and b = +0.0014, then the fitted accuracy drift curve is relative deviation = -0.0004 × x + 0.14. If the cumulative usage time is 1560 hours, then the current relative deviation value = -0.0004 × 1560 + 0.14 = -0.00484, indicating that the measured value is on average lower than the standard value by about 0.00484.

[0124] Further, the sum of the absolute value of the current relative deviation and 1 is calculated to obtain the aging compensation denominator, which is an intermediate variable used to calculate the aging compensation coefficient. Specifically, the absolute value of the current relative deviation is calculated, and then the sum of the absolute value of the current relative deviation and 1 is calculated to obtain the aging compensation denominator. When the equipment has almost no aging and the relative deviation is close to 0, the aging compensation denominator is close to 1, the compensation coefficient is close to 1, and it hardly weakens the anomaly probability value; when the equipment is severely aged and the absolute value of the relative deviation reaches 1, it moderately weakens the anomaly probability value; when the absolute value of the relative deviation reaches 50%, it significantly weakens the anomaly probability value.

[0125] The aging compensation denominator = 1 + |current relative deviation value|. Since the absolute value of the current relative deviation value is ≥ 0, the aging compensation denominator is ≥ 1. For example, if the absolute value of the current relative deviation value = |-0.00484| = 0.00484, then the aging compensation denominator = 1 + 0.00484 = 1.00484.

[0126] Finally, 1 is divided by the aging compensation denominator to obtain the equipment aging compensation coefficient. Specifically, the equipment aging compensation coefficient = 1 / aging compensation denominator = 1 / (1 + |current relative deviation value|), with a value range of (0,1). When the equipment has no current relative deviation, the equipment aging compensation coefficient is 1; when the equipment has a deviation, the equipment aging compensation coefficient is less than 1; the larger the relative deviation, the smaller the equipment aging compensation coefficient. For example, the equipment aging compensation coefficient = 1 / 1.00484 ≈ 0.9952.

[0127] In this embodiment, a pre-trained anomaly pattern prediction model is used to predict the probability of anomalies occurring within subsequent parameter intervals at the current test point, and the subsequent step size and density are adjusted to effectively capture critical anomalies. The current test progress value is incorporated into the adjustment coefficient calculation, enabling the adjustment strategy to adaptively balance remaining test time and potential anomaly risks, improving overall test efficiency. Subsequently, a neural network is used for supervised training to construct the anomaly pattern prediction model, identifying abnormal response patterns of the automotive instrument panel under various stimuli. Furthermore, an equipment aging compensation coefficient is used to correct the predicted anomaly probability, eliminating errors caused by end-side equipment accuracy drift.

[0128] S50: The cloud server compares the final collected response data with the standard responses in the historical data to generate a test report.

[0129] In this embodiment of the application, the final collected response data is a complete response dataset that the cloud server receives from the end device after the entire test task is completed, covering all test points; the standard response in the historical data is the response data for a specific model of instrument under a specific test item, stored in the cloud test case library or a dedicated database; the comparison is to compare the measured response data with the standard response data item by item.

[0130] Specifically, once the test process is complete and the cloud server possesses the final collected response data, it retrieves the standard response from the historical data corresponding to that test task from the database. Subsequently, the cloud server compares the actual speed displayed on the instrument panel at each test point with the theoretical speed. Finally, the cloud server aggregates all information, generates a test report, and stores it in the cloud or pushes it to the user.

[0131] Step S50 in the method provided in this application embodiment includes: The cloud server acquires all the final collected response data, compares the all response data with the corresponding standard response data in the cloud test case library point by point, and calculates the absolute error and relative error of each test point. Based on whether the absolute error and the relative error exceed the preset pass threshold, each test point is determined to be a pass test point or a fail test point, and all test points determined to be fail are recorded to form a list of fail test points. For each non-conforming test point in the list of non-conforming test points, the cloud server extracts the signal type, output value, and response data corresponding to the non-conforming test point, and associates them with the abnormal probability value generated by the abnormal pattern prediction model during the test. The pass / fail judgment results, the list of non-compliant test points, the anomaly probability values ​​corresponding to the non-compliant test points, and the final optimized programmable instruction sequence are summarized to generate a structured test report.

[0132] In this embodiment of the application, firstly, the cloud server acquires all the final collected response data, compares all the response data with the corresponding standard response data in the cloud test case library point by point, and calculates the absolute error and relative error of each test point. The standard response data is the theoretically correct response value corresponding to the current test item stored in the cloud test case library. Point-by-point comparison is the process of comparing the actual response data of each test point with the corresponding standard response data.

[0133] Absolute error is the absolute difference between the measured value and the standard value; relative error is the ratio of the absolute error to the standard value, usually expressed as a percentage. Relative error eliminates the dimensions, making it easier to compare test items with different ranges.

[0134] Specifically, after the entire testing process is completed, the cloud server retrieves all the final collected response data from the database. Assume this test executes 200 test points, each corresponding to a frequency input value and its measured displayed vehicle speed. Simultaneously, the cloud server retrieves the standard response data corresponding to this test task from the cloud test case library, i.e., the theoretical vehicle speed value corresponding to each frequency point. Subsequently, the cloud server performs a point-by-point comparison from the first to the 200th test point: for each test point, it extracts the measured vehicle speed value and the standard vehicle speed value, calculating the absolute error = |measured vehicle speed value - standard vehicle speed value|. The relative error = (|measured vehicle speed value - standard vehicle speed value| / standard vehicle speed value) × 100%. For example, with a target speed of 100 km / h and a measured speed of 97 km / h, the absolute error is 3 km / h, and the relative error is 3%.

[0135] Secondly, based on whether the absolute error and relative error exceed the preset pass / fail threshold, each test point is determined to be a pass / fail test point or a fail / fail test point. All test points determined to be fail / fail are recorded to form a fail / fail test point list. The preset pass / fail threshold is a pre-set upper limit of error used to determine whether a test point is pass / fail. A pass / fail test point is a test point whose absolute error and relative error are both within the preset pass / fail threshold range. A fail / fail test point is a test point whose absolute error or relative error exceeds the preset pass / fail threshold range at least once. The fail / fail test point list records the information of all test points determined to be fail / fail in the order of testing.

[0136] Specifically, the first step is to set a preset acceptable threshold. This threshold is usually determined based on industry standards, company specifications, or the technical requirements of a specific vehicle model. For speedometers, the acceptable threshold may be specified as an absolute error not exceeding ±2 km / h and a relative error not exceeding ±3%; for fuel gauges, the acceptable threshold may be that the difference between the displayed fuel level and the theoretical fuel level does not exceed ±1 / 8 of a division.

[0137] The absolute and relative errors for each test point are calculated. The cloud server judges the results based on preset pass / fail thresholds. If both are within the limits, the test point is considered passable; if either the relative or absolute error exceeds the limit, or both, the test point is considered failable. Assume the pass / fail thresholds for speedometer testing are: absolute error ≤ 2 km / h and relative error ≤ 3%. For the 50th test point, the absolute error is 1.2 km / h and the relative error is 1.1%, both within the limits, so the test point is considered passable. For the 120th test point, the absolute error is 2.5 km / h (exceeding 2 km / h) and the relative error is 2.1% (not exceeding 3%). Because the absolute error exceeds the limit, the test point is considered failable. The cloud server records all failable test points, extracting information such as test point number, frequency value, measured speed, standard speed, absolute error, and relative error to form a list of failable test points.

[0138] Furthermore, for each non-conforming test point in the list of non-conforming test points, the cloud server extracts the signal type, output value, and response data corresponding to the non-conforming test point, and associates them with the anomaly probability value generated by the anomaly pattern prediction model during the test.

[0139] Specifically, for each non-compliant test point in the list of non-compliant test points, the cloud server extracts the corresponding signal type, output value, and response data from the original test data records. Simultaneously, it searches for the anomaly probability value corresponding to that test point from the anomaly probability records stored in real-time during the test. Since the cloud calculates and stores the anomaly probability value at each test point, it can find the anomaly probability value output by the model when the test point is executed, assumed to be 0.78 (78%). The cloud server then associates the corresponding anomaly probability value with the record for that non-compliant test point.

[0140] Finally, the pass / fail judgment results, the list of non-pass test points, the anomaly probability values ​​corresponding to the non-pass test points, and the final optimized programmable instruction sequence are summarized and organized according to a structured template to generate a structured test report. For example, the report can be in JSON format, with the root node containing a test summary and a list of non-pass points, and each point containing fields, the optimized instruction sequence, and other main parts.

[0141] In this embodiment, by calculating the absolute and relative errors point by point, the system determines whether the test exceeds the acceptable threshold. The generated report accurately locates each non-compliant test point and its specific deviation value, providing direct quantitative evidence for instrument calibration and fault repair. All non-compliant test points are centrally recorded in a non-compliant test point list, improving troubleshooting efficiency. Anomaly probability values ​​are associated with each non-compliant test point, enhancing the report's interpretability. The inclusion of the final optimized programmable control instruction sequence in the report ensures the reproducibility of the test results. Finally, standardized summaries are generated to produce a structured test report, facilitating subsequent archiving and quality trend analysis, and enabling efficient reuse of test results.

[0142] The embodiments of this application, through the above specific implementation methods, achieve the following technical effects: In this embodiment, the test task is first automatically parsed by the cloud server, and the model of the instrument under test and the test items are extracted, eliminating the ambiguity and error risk of manual interpretation and realizing the automated parsing of the test task. Then, the basic test cases that completely correspond to the current test object are located from the cloud test case library. By defining the basic test cases as structured fields containing signal type, parameter range, test sequence, judgment criteria, etc., the test cases can be directly understood and processed by the computer program, laying the data foundation for subsequent model optimization.

[0143] Secondly, the initial programmable command sequence sent from the cloud is parsed and a signal output timing table is generated using the end-side testing equipment. This avoids timing conflicts and signal interference, ensuring the controllability and repeatability of the testing process. By integrating high / low level output modules, voltage output modules, PWM wave output modules, frequency output modules, and resistance output modules into the end-side testing equipment, the convenience and automation of the testing are improved. By integrating voltage detection modules, CAN communication modules, and image acquisition modules, the instrument's response performance is comprehensively captured from multiple dimensions, avoiding potential omissions due to single detection methods. The acquired response data is associated with and stored according to the signal type and output value of the current test point, and precise timestamps are added to form structured test data records. This allows each response data to trace back to the excitation signal, providing a solid data foundation for subsequent cloud-based evaluation and anomaly analysis.

[0144] Furthermore, by establishing a real-time data channel from the physical test site to the cloud through a real-time upload mechanism, subsequent test strategies can be dynamically adjusted based on the current test situation, realizing a key data flow for adaptive dynamic testing.

[0145] Furthermore, a pre-trained anomaly pattern prediction model is used to predict the probability of anomalies occurring within subsequent parameter ranges at the current test point, and the subsequent step size and density are adjusted to effectively capture critical anomalies. The current test progress value is incorporated into the adjustment coefficient calculation, allowing the adjustment strategy to adaptively balance remaining test time and potential anomaly risks, improving overall test efficiency. Subsequently, a neural network is used for supervised training to construct an anomaly pattern prediction model that identifies abnormal response patterns of automotive instruments under various stimuli. In addition, an equipment aging compensation coefficient is used to correct the predicted anomaly probability, eliminating errors caused by end-side equipment accuracy drift.

[0146] Finally, by calculating the absolute and relative errors point by point, the system determines whether the test results exceed the acceptable threshold. The generated report accurately identifies each non-compliant test point and its specific deviation value, providing direct quantitative evidence for instrument calibration and fault repair. All non-compliant test points are centrally recorded in a non-compliant test point list, improving troubleshooting efficiency. Anomaly probability values ​​are associated with each non-compliant test point, enhancing the report's interpretability. The inclusion of the final optimized programmable control instruction sequence in the report ensures the reproducibility of the test results. Finally, standardized summaries are generated to produce a structured test report, facilitating subsequent archiving and quality trend analysis, and enabling efficient reuse of test results.

[0147] Example 2, as Figure 2 As shown, based on the same inventive concept as the end-to-cloud collaborative automotive instrument programmable control testing method provided in Embodiment 1, this embodiment of the invention also provides an end-to-cloud collaborative automotive instrument programmable control testing system, the system comprising: The programmable instruction generation module 11 is used to receive test tasks submitted by users on the cloud server, call basic test cases from the cloud test case library according to the test tasks, and input the basic test cases into the pre-trained test strategy optimization model to generate an initial programmable instruction sequence, wherein the test strategy optimization model takes maximizing the anomaly detection rate and minimizing the test time as optimization objectives. The response data acquisition module 12 is used for the cloud server to send the initial programmable instruction sequence to the end-side test device, and the end-side test device to output test signals to the instrument panel of the vehicle under test based on the initial programmable instruction sequence, and simultaneously collect the response data of the instrument panel of the vehicle under test. The response data upload module 13 is used by the end-side test device to upload the response data to the cloud server in real time. The programmable instruction optimization module 14 is used by the cloud server to evaluate the current test progress and anomaly detection in real time based on the response data, dynamically adjust the parameter step size and test point density of the initial programmable instruction sequence in the subsequent steps, generate an optimized programmable instruction sequence, and send the optimized programmable instruction sequence to the end-side test device for continued execution. The comparison test module 15 is used by the cloud server to compare the final collected response data with the standard response in the historical data and generate a test report.

[0148] In one embodiment, the programmable instruction generation module 11 is used for: The cloud server parses the test task and extracts the instrument model and test items from the test task. The test items include at least indicator light detection, speedometer detection, tachometer detection, fuel gauge detection, and water temperature gauge detection. The cloud server matches corresponding basic test cases from the cloud test case library according to the model of the instrument under test and the test item. The basic test cases include at least signal type, parameter range, test sequence and judgment criteria. The signal type includes at least high and low level signals, PWM wave signals, frequency signals, resistance signals and voltage signals. The cloud server inputs the basic test cases into a pre-trained test strategy optimization model. The test strategy optimization model aims to maximize the anomaly detection rate and minimize the test time. It iteratively optimizes the parameter range and test timing in the basic test cases, outputs the optimized parameter step sequence and test point density, and generates an initial program control instruction sequence based on the optimized parameter step sequence and test point density.

[0149] The process of constructing the test strategy optimization model includes: Collect historical test datasets, which include multiple sets of instrument models, test items, programmable control command parameters, response data and corresponding fault tags. The programmable control command parameters include at least voltage value, frequency value, resistance value, duty cycle and pulse count. The response data includes at least the instrument display value, indicator light status, CAN message content and analog output value. Based on reinforcement learning, an initial test strategy optimization model is constructed. Using the instrument models and test items in the historical test dataset as the state space, the corresponding programmable instruction parameters as the action space, and the comprehensive reward function as the optimization objective, a pre-trained test strategy optimization model is obtained through multiple rounds of iterative training. The construction process of the comprehensive reward function includes: The ratio of the number of anomalies detected in the current round of testing to the total number of test points is used as the anomaly detection rate. Get the total test time for the current round of testing; Based on the preset first weight coefficient and second weight coefficient, the comprehensive reward function is calculated according to the following function: the comprehensive reward function is equal to the anomaly detection rate multiplied by the first weight coefficient minus the total test time multiplied by the second weight coefficient, wherein both the first weight coefficient and the second weight coefficient are positive values.

[0150] The process of building the cloud-based test case library includes: Receive standard test specification documents uploaded by users, wherein the standard test specification documents include at least one of enterprise standards, industry standards or vehicle technical specifications; The standard test specification document is parsed using natural language processing technology to extract test items, signal types, parameter ranges, test sequences, and judgment criteria, and to generate initial test cases. The natural language processing technology includes at least word segmentation, entity recognition, and relation extraction. The initial test cases are revised and confirmed. The revised and confirmed test cases are associated with the corresponding instrument models and stored in the cloud test case library, and a revision log is recorded. The correction records in the correction log are used as training samples to iteratively update the test strategy optimization model.

[0151] In one embodiment, the response data acquisition module 12 is used for: The cloud server sends the initial sequence of programmable commands to the end-side test device; The end-side test equipment receives and parses the initial programmable instruction sequence to generate a signal output timing table, wherein the signal output timing table arranges the signal type, output value and duration of each test point in chronological order. The end-side test equipment outputs test signals to the vehicle instrument under test sequentially through its internally integrated signal output module according to the signal output timing table. The signal output module includes at least a high / low level output module, a voltage output module, a PWM wave output module, a frequency output module, and a resistance output module. While outputting test signals, the end-side test equipment collects the response data of the vehicle instrument panel under test in real time through the internally integrated detection module. The detection module includes at least a voltage detection module, a CAN communication module, and an image acquisition module. The response data includes at least the instrument pointer position, the LCD screen display content, the indicator light status, and the CAN bus message. The end-side testing equipment associates and stores the collected response data with the signal type and output value of the current test point, forming a timestamped test data record.

[0152] In one embodiment, the programmable instruction optimization module 14 is used for: The cloud server receives the timestamped test data records uploaded in real time by the end-side test device, calculates the ratio of the number of completed test points to the total number of test points in the initial program control instruction sequence, and uses it as the current test progress value. The cloud server inputs the signal type, parameter value, and real-time deviation characteristics of the current test point into the pre-trained abnormal pattern prediction model, outputs the predicted abnormal probability value of anomalies occurring within the subsequent parameter range, and multiplies the predicted abnormal probability value by the equipment aging compensation coefficient to obtain the abnormal probability value. The adjustment coefficient is obtained by subtracting the current test progress value from 1 and then multiplying it by the anomaly probability value. The subsequent parameter step size is obtained by multiplying the preset base step size by the adjustment coefficient, and the subsequent test point density is obtained by dividing the preset base density by the adjustment coefficient. An optimized programmable instruction sequence is generated based on the subsequent parameter step size and subsequent test point density, and the optimized programmable instruction sequence is sent to the end-side test equipment to continue executing subsequent tests; The construction process of the abnormal pattern prediction model includes: Historical test data is extracted from the historical test database. The historical test data includes normal test data and abnormal test data. The abnormal test data includes signal type, parameter value, deviation characteristics, and actual fault cause after the abnormality occurred. A training sample set is constructed using the signal type, parameter value, and deviation characteristics in the historical test data as input features and the corresponding anomaly label as output label. Here, an anomaly label of 1 indicates the presence of an anomaly, and an anomaly label of 0 indicates the absence of an anomaly. An initial anomaly pattern prediction model is constructed using a neural network algorithm. The initial anomaly pattern prediction model includes an input layer, a hidden layer, and an output layer. The output layer uses a Sigmoid activation function to output the predicted anomaly probability value. The initial anomaly pattern prediction model is trained in a supervised manner using the training sample set until it is verified to converge, thus obtaining a pre-trained anomaly pattern prediction model.

[0153] The calculation process for the equipment aging compensation coefficient includes: The cloud server obtains the device identifier of the current edge-side test device and the cumulative usage time of the current edge-side test device; The cloud server retrieves the historical calibration records of the current end-side test device from the cloud device archive based on the device identifier. The historical calibration records include the time points of each calibration, the usage time during calibration, and the calibration results. The calibration results include the relative deviation between the measured value of the current end-side test device under the standard signal input and the standard value. The relative deviation is equal to the difference between the measured value and the standard value divided by the standard value. The cloud server constructs the accuracy drift curve of the current end-side test equipment based on the usage time and corresponding calibration results of each calibration in the historical calibration record using a curve fitting method. The cloud server queries the accuracy drift curve based on the cumulative usage time of the current edge test equipment to obtain the corresponding current relative deviation value. Calculate the sum of the absolute value of the current relative deviation and 1 to obtain the denominator for aging compensation; Divide 1 by the aging compensation denominator to obtain the equipment aging compensation coefficient.

[0154] In one embodiment, the comparison test module 15 is used for: The cloud server acquires all the final collected response data, compares the all response data with the corresponding standard response data in the cloud test case library point by point, and calculates the absolute error and relative error of each test point. Based on whether the absolute error and the relative error exceed the preset pass threshold, each test point is determined to be a pass test point or a fail test point, and all test points determined to be fail are recorded to form a list of fail test points. For each non-conforming test point in the list of non-conforming test points, the cloud server extracts the signal type, output value, and response data corresponding to the non-conforming test point, and associates them with the abnormal probability value generated by the abnormal pattern prediction model during the test. The pass / fail judgment results, the list of non-compliant test points, the anomaly probability values ​​corresponding to the non-compliant test points, and the final optimized programmable instruction sequence are summarized to generate a structured test report.

[0155] Compared to existing technologies, this application first automatically parses the test task and extracts the instrument model and test items through a cloud server, eliminating the ambiguity and error risk of manual interpretation and realizing the automated parsing of the test task. Then, it locates the basic test cases that completely correspond to the current test object from the cloud test case library. By defining the basic test cases as structured fields containing signal type, parameter range, test sequence, and judgment criteria, the test cases can be directly understood and processed by the computer program, laying a data foundation for subsequent model optimization.

[0156] Secondly, the initial programmable command sequence sent from the cloud is parsed and a signal output timing table is generated using the end-side testing equipment. This avoids timing conflicts and signal interference, ensuring the controllability and repeatability of the testing process. By integrating high / low level output modules, voltage output modules, PWM wave output modules, frequency output modules, and resistance output modules into the end-side testing equipment, the convenience and automation of the testing are improved. By integrating voltage detection modules, CAN communication modules, and image acquisition modules, the instrument's response performance is comprehensively captured from multiple dimensions, avoiding potential omissions due to single detection methods. The acquired response data is associated with and stored according to the signal type and output value of the current test point, and precise timestamps are added to form structured test data records. This allows each response data to trace back to the excitation signal, providing a solid data foundation for subsequent cloud-based evaluation and anomaly analysis.

[0157] Furthermore, by establishing a real-time data channel from the physical test site to the cloud through a real-time upload mechanism, subsequent test strategies can be dynamically adjusted based on the current test situation, realizing a key data flow for adaptive dynamic testing.

[0158] Furthermore, a pre-trained anomaly pattern prediction model is used to predict the probability of anomalies occurring within subsequent parameter ranges at the current test point, and the subsequent step size and density are adjusted to effectively capture critical anomalies. The current test progress value is incorporated into the adjustment coefficient calculation, allowing the adjustment strategy to adaptively balance remaining test time and potential anomaly risks, improving overall test efficiency. Subsequently, a neural network is used for supervised training to construct an anomaly pattern prediction model that identifies abnormal response patterns of automotive instruments under various stimuli. In addition, an equipment aging compensation coefficient is used to correct the predicted anomaly probability, eliminating errors caused by end-side equipment accuracy drift.

[0159] Finally, by calculating the absolute and relative errors point by point, the system determines whether the test results exceed the acceptable threshold. The generated report accurately identifies each non-compliant test point and its specific deviation value, providing direct quantitative evidence for instrument calibration and fault repair. All non-compliant test points are centrally recorded in a non-compliant test point list, improving troubleshooting efficiency. Anomaly probability values ​​are associated with each non-compliant test point, enhancing the report's interpretability. The inclusion of the final optimized programmable control instruction sequence in the report ensures the reproducibility of the test results. Finally, standardized summaries are generated to produce a structured test report, facilitating subsequent archiving and quality trend analysis, and enabling efficient reuse of test results.

[0160] Example 3, as Figure 3 As shown, this embodiment of the invention provides a portable device for the programmable testing of automotive instruments using edge-cloud collaboration. The device includes: a memory 210 for storing computer software program 211; and a processor 220 for reading and executing the computer software program, thereby realizing a programmable testing method for automotive instruments using edge-cloud collaboration.

[0161] In embodiments of the present invention, such as Figure 4 , Figure 5 As shown in the figure, this application provides a prototype schematic diagram of a portable device for programmable testing of automotive instruments that enables edge-cloud collaboration.

[0162] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

Claims

1. A method for testing the programmable control of automotive instruments using edge-cloud collaboration, characterized in that, The method includes: The cloud server receives the test task submitted by the user, calls the basic test cases from the cloud test case library according to the test task, and inputs the basic test cases into the pre-trained test strategy optimization model to generate an initial program control instruction sequence. The test strategy optimization model aims to maximize the anomaly detection rate and minimize the test time. The cloud server sends the initial programmable instruction sequence to the end-side testing device. The end-side testing device outputs test signals to the instrument panel of the vehicle under test based on the initial programmable instruction sequence and simultaneously collects the response data of the instrument panel of the vehicle under test. The end-side testing device uploads the response data to the cloud server in real time. The cloud server evaluates the current test progress and anomaly detection in real time based on the response data, dynamically adjusts the parameter step size and test point density of the initial programmable instruction sequence, generates an optimized programmable instruction sequence, and sends the optimized programmable instruction sequence to the end-side test device for continued execution. The cloud server compares the final collected response data with the standard responses in historical data to generate a test report.

2. The method for programmable control testing of automotive instruments with end-to-cloud collaboration according to claim 1, characterized in that, The cloud server receives the test task submitted by the user, retrieves basic test cases from the cloud test case library according to the test task, and inputs the basic test cases into the pre-trained test strategy optimization model to generate an initial sequence of programmable instructions, including: The cloud server parses the test task and extracts the instrument model and test items from the test task. The test items include at least indicator light detection, speedometer detection, tachometer detection, fuel gauge detection, and water temperature gauge detection. The cloud server matches corresponding basic test cases from the cloud test case library according to the model of the instrument under test and the test item. The basic test cases include at least signal type, parameter range, test sequence and judgment criteria. The signal type includes at least high and low level signals, PWM wave signals, frequency signals, resistance signals and voltage signals. The cloud server inputs the basic test cases into a pre-trained test strategy optimization model. The test strategy optimization model aims to maximize the anomaly detection rate and minimize the test time. It iteratively optimizes the parameter range and test timing in the basic test cases, outputs the optimized parameter step sequence and test point density, and generates an initial program control instruction sequence based on the optimized parameter step sequence and test point density.

3. The method for programmable control testing of automotive instruments with end-to-cloud collaboration according to claim 2, characterized in that, The process of building a test strategy optimization model includes: Collect historical test datasets, which include multiple sets of instrument models, test items, programmable control command parameters, response data and corresponding fault tags. The programmable control command parameters include at least voltage value, frequency value, resistance value, duty cycle and pulse count. The response data includes at least the instrument display value, indicator light status, CAN message content and analog output value. Based on reinforcement learning, an initial test strategy optimization model is constructed. Using the instrument models and test items in the historical test dataset as the state space, the corresponding programmable instruction parameters as the action space, and the comprehensive reward function as the optimization objective, a pre-trained test strategy optimization model is obtained through multiple rounds of iterative training. The construction process of the comprehensive reward function includes: The ratio of the number of anomalies detected in the current round of testing to the total number of test points is used as the anomaly detection rate. Get the total test time for the current round of testing; Based on the preset first weight coefficient and second weight coefficient, the comprehensive reward function is calculated according to the following function: the comprehensive reward function is equal to the anomaly detection rate multiplied by the first weight coefficient minus the total test time multiplied by the second weight coefficient, wherein both the first weight coefficient and the second weight coefficient are positive values.

4. The method for programmable control testing of automotive instruments with end-to-cloud collaboration according to claim 2, characterized in that, The process of building a cloud-based test case library includes: Receive standard test specification documents uploaded by users, wherein the standard test specification documents include at least one of enterprise standards, industry standards or vehicle technical specifications; The standard test specification document is parsed using natural language processing technology to extract test items, signal types, parameter ranges, test sequences, and judgment criteria, and to generate initial test cases. The natural language processing technology includes at least word segmentation, entity recognition, and relation extraction. The initial test cases are revised and confirmed. The revised and confirmed test cases are associated with the corresponding instrument models and stored in the cloud test case library, and a revision log is recorded. The correction records in the correction log are used as training samples to iteratively update the test strategy optimization model.

5. The method for programmable control testing of automotive instruments with end-to-cloud collaboration according to claim 1, characterized in that, The cloud server sends the initial programmable command sequence to the end-side testing device. The end-side testing device outputs test signals to the instrument cluster of the vehicle under test based on the initial programmable command sequence, and simultaneously collects the response data of the instrument cluster, including: The cloud server sends the initial sequence of programmable commands to the end-side test device; The end-side test equipment receives and parses the initial programmable instruction sequence to generate a signal output timing table, wherein the signal output timing table arranges the signal type, output value and duration of each test point in chronological order. The end-side test equipment outputs test signals to the vehicle instrument under test sequentially through its internally integrated signal output module according to the signal output timing table. The signal output module includes at least a high / low level output module, a voltage output module, a PWM wave output module, a frequency output module, and a resistance output module. While outputting test signals, the end-side test equipment collects the response data of the vehicle instrument panel under test in real time through the internally integrated detection module. The detection module includes at least a voltage detection module, a CAN communication module, and an image acquisition module. The response data includes at least the instrument pointer position, the LCD screen display content, the indicator light status, and the CAN bus message. The end-side testing equipment associates and stores the collected response data with the signal type and output value of the current test point, forming a timestamped test data record.

6. The method for programmable control testing of automotive instruments using edge-cloud collaboration according to claim 1, characterized in that, The cloud server evaluates the current test progress and anomaly detection status in real time based on the response data, dynamically adjusts the parameter step size and test point density of the initial programmable instruction sequence, generates an optimized programmable instruction sequence, and sends the optimized programmable instruction sequence to the end-side test equipment for continued execution, including: The cloud server receives the timestamped test data records uploaded in real time by the end-side test device, calculates the ratio of the number of completed test points to the total number of test points in the initial program control instruction sequence, and uses it as the current test progress value. The cloud server inputs the signal type, parameter value, and real-time deviation characteristics of the current test point into the pre-trained abnormal pattern prediction model, outputs the predicted abnormal probability value of anomalies occurring within the subsequent parameter range, and multiplies the predicted abnormal probability value by the equipment aging compensation coefficient to obtain the abnormal probability value. The adjustment coefficient is obtained by subtracting the current test progress value from 1 and then multiplying it by the anomaly probability value. The subsequent parameter step size is obtained by multiplying the preset base step size by the adjustment coefficient, and the subsequent test point density is obtained by dividing the preset base density by the adjustment coefficient. An optimized programmable instruction sequence is generated based on the subsequent parameter step size and subsequent test point density, and the optimized programmable instruction sequence is sent to the end-side test equipment to continue executing subsequent tests; The construction process of the abnormal pattern prediction model includes: Historical test data is extracted from the historical test database. The historical test data includes normal test data and abnormal test data. The abnormal test data includes signal type, parameter value, deviation characteristics, and actual fault cause after the abnormality occurred. A training sample set is constructed using the signal type, parameter value, and deviation characteristics in the historical test data as input features and the corresponding anomaly label as output label. Here, an anomaly label of 1 indicates the presence of an anomaly, and an anomaly label of 0 indicates the absence of an anomaly. An initial anomaly pattern prediction model is constructed using a neural network algorithm. The initial anomaly pattern prediction model includes an input layer, a hidden layer, and an output layer. The output layer uses a Sigmoid activation function to output the predicted anomaly probability value. The initial anomaly pattern prediction model is trained in a supervised manner using the training sample set until it is verified to converge, thus obtaining a pre-trained anomaly pattern prediction model.

7. The method for programmable control testing of automotive instruments with end-to-cloud collaboration according to claim 6, characterized in that, The calculation process for the equipment aging compensation coefficient includes: The cloud server obtains the device identifier of the current edge-side test device and the cumulative usage time of the current edge-side test device; The cloud server retrieves the historical calibration records of the current end-side test device from the cloud device archive based on the device identifier. The historical calibration records include the time points of each calibration, the usage time during calibration, and the calibration results. The calibration results include the relative deviation between the measured value of the current end-side test device under the standard signal input and the standard value. The relative deviation is equal to the difference between the measured value and the standard value divided by the standard value. The cloud server constructs the accuracy drift curve of the current end-side test equipment based on the usage time and corresponding calibration results of each calibration in the historical calibration record using a curve fitting method. The cloud server queries the accuracy drift curve based on the cumulative usage time of the current edge test equipment to obtain the corresponding current relative deviation value. Calculate the sum of the absolute value of the current relative deviation and 1 to obtain the denominator for aging compensation; Divide 1 by the aging compensation denominator to obtain the equipment aging compensation coefficient.

8. The method for programmable control testing of automotive instruments with end-to-cloud collaboration according to claim 1, characterized in that, The cloud server compares the final collected response data with the standard responses in historical data to generate a test report, including: The cloud server acquires all the final collected response data, compares the all response data with the corresponding standard response data in the cloud test case library point by point, and calculates the absolute error and relative error of each test point. Based on whether the absolute error and the relative error exceed the preset pass threshold, each test point is determined to be a pass test point or a fail test point, and all test points determined to be fail are recorded to form a list of fail test points. For each non-conforming test point in the list of non-conforming test points, the cloud server extracts the signal type, output value, and response data corresponding to the non-conforming test point, and associates them with the abnormal probability value generated by the abnormal pattern prediction model during the test. The pass / fail judgment results, the list of non-compliant test points, the anomaly probability values ​​corresponding to the non-compliant test points, and the final optimized programmable instruction sequence are summarized to generate a structured test report.

9. A cloud-edge collaborative automotive instrument control testing system, characterized in that, A method for performing programmable control testing of an automotive instrument cluster with end-to-cloud collaboration as described in any one of claims 1-8, the system comprising: The programmable instruction generation module is used to receive test tasks submitted by users on the cloud server, call basic test cases from the cloud test case library according to the test tasks, and input the basic test cases into a pre-trained test strategy optimization model to generate an initial programmable instruction sequence. The test strategy optimization model aims to maximize the anomaly detection rate and minimize the test time. The response data acquisition module is used for the cloud server to send the initial programmable instruction sequence to the end-side test device, and the end-side test device to output test signals to the instrument panel of the vehicle under test based on the initial programmable instruction sequence, and simultaneously collect the response data of the instrument panel of the vehicle under test. A response data upload module is used by the end-side test device to upload the response data to the cloud server in real time. The programmable instruction optimization module is used by the cloud server to evaluate the current test progress and anomaly detection in real time based on the response data, dynamically adjust the parameter step size and test point density of the initial programmable instruction sequence in the subsequent steps, generate an optimized programmable instruction sequence, and send the optimized programmable instruction sequence to the end-side test device for continued execution. The comparison test module is used by the cloud server to compare the final collected response data with the standard responses in historical data and generate a test report.

10. A portable device for testing the programmable control of automotive instruments using edge-cloud collaboration, characterized in that, include: Memory, used to store computer software programs; A processor is used to read and execute the computer software program, thereby implementing the end-to-cloud collaborative automotive instrument programmable test method as described in any one of claims 1-8.

Images