Vehicle-mounted electrical performance intelligent test system based on deep learning

Abstract

The invention relates to the technical field of vehicle-mounted electronic product testing, and discloses a vehicle-mounted electrical performance intelligent testing system based on deep learning, which comprises a feature construction module, an intelligent decision module, an instruction generation module, a cooperative control module, a parameter synchronous detection module, an anomaly analysis module and a model optimization feedback module. The feature construction module collects vehicle-mounted electrical performance data and external environment perception data to perform feature extraction to obtain a fusion feature vector set; the intelligent decision-making module generates a detection strategy configuration parameter through a parameter priority adaptive algorithm; the instruction generation module generates a BUS box instruction set; the cooperative control module automatically builds a test environment; the parameter synchronization detection module executes multi-parameter synchronization verification; the anomaly analysis module generates an intelligent anomaly analysis report; the model optimization feedback module optimizes the decision ability of the model. According to the invention, automatic construction of a test environment, multi-parameter synchronous verification and intelligent anomaly analysis can be realized, and the test efficiency and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of automotive electronic product testing technology, and more specifically, to an intelligent testing system for automotive electrical performance based on deep learning. Background Technology

[0002] With the increasing level of automotive electronics, the functions of in-vehicle electronic products are becoming increasingly complex, requiring comprehensive testing of multiple electrical performance parameters such as voltage, current, pulse width modulation signals, and bus messages. Traditional in-vehicle electrical performance testing methods mainly rely on manually setting up test circuits and configuring test parameters. Testers manually connect the test equipment to each pin of the product under test according to the product specifications, and set the detection thresholds and sampling parameters for each parameter channel one by one. This method suffers from problems such as large circuit setup errors, low efficiency, and difficulty in ensuring the consistency of the test environment.

[0003] Existing automotive electrical performance testing systems typically employ a fixed parameter testing sequence and independent parameter verification methods, failing to dynamically adjust testing strategies based on the actual state of the product. For multiple parameters with interrelationships, traditional methods use time-sharing testing to verify each parameter individually, failing to capture the temporal correlation characteristics between parameters and resulting in low efficiency in multi-parameter verification. Furthermore, when abnormal parameters are detected, traditional methods can only provide a simple judgment that the parameter exceeds the limit, lacking the ability to analyze the underlying causes of the anomaly. Testers must rely on experience for fault diagnosis, demonstrating weak anomaly analysis capabilities.

[0004] Therefore, there is a need for an on-board electrical performance testing system that can automatically build the test environment, intelligently generate testing strategies based on product status, support multi-parameter synchronous verification, and have intelligent anomaly analysis capabilities, in order to solve the technical problems of large test circuit construction errors, low efficiency of multi-parameter verification, and weak anomaly analysis capabilities in traditional testing methods. Summary of the Invention

[0005] This invention provides a deep learning-based intelligent testing system for vehicle electrical performance, which solves the technical problems of large errors in test circuit construction, low efficiency of multi-parameter verification, and weak anomaly analysis capabilities in related technologies.

[0006] This invention provides a deep learning-based intelligent testing system for vehicle electrical performance, comprising: The feature construction module is used to collect on-board electrical performance data and external environment perception data, perform multi-dimensional feature extraction, and obtain a fused feature vector set. The intelligent decision-making module is used to build an intelligent decision-making model. It inputs the fused feature vector set into the decision-making model and generates detection strategy configuration parameters through a parameter priority adaptive algorithm. The instruction generation module is used to obtain the detection strategy configuration parameters and to obtain the BUS box automated construction instruction set and detection control instruction set through the instruction generation method. The collaborative control module is used to send automated setup instruction sets and detection control instruction sets to the BUS box, automatically build the test environment and configure the detection parameters to obtain the completed test environment; The parameter synchronization detection module is used to perform detection on the BUS box according to the multi-parameter synchronization verification strategy based on the established test environment, and obtain the detection result dataset. The anomaly analysis module is used to generate intelligent anomaly analysis reports based on the detection result dataset and external environmental perception data. The model optimization feedback module is used to feed back the intelligent anomaly analysis report to the intelligent decision-making model to optimize the model's decision-making capabilities.

[0007] In a preferred embodiment, the intelligent decision-making model includes: The system comprises a feature extraction subnetwork, a feature fusion subnetwork, and a decision output subnetwork; the feature extraction subnetwork includes a visual feature extraction branch and a temperature feature extraction branch. The feature fusion subnetwork adopts an attention mechanism fusion method. After linear transformation and bias processing of visual depth features and temperature depth features, attention weights are obtained by normalization through the softmax function. The visual depth features and temperature depth features are multiplied by their corresponding attention weights and then added together to obtain the fused depth features. The decision output subnetwork contains three output branches, which respectively output parameter priority sequence, multi-parameter synchronous verification combination configuration, and dynamic detection threshold configuration.

[0008] In a preferred embodiment, the parameter priority adaptive algorithm includes: The failure risk assessment subprocess and the parameter priority dynamic calculation subprocess; the failure risk assessment subprocess selects a preset number of similar cases based on the fused feature vector set and the failure state feature library, and calculates the failure risk probability by counting the frequency of anomalies in each parameter channel. The parameter priority dynamic calculation subprocess calculates the detection priority based on the failure risk probability and parameter safety level. The detection priority is obtained by multiplying the basic priority by the first weighting coefficient and adding the failure risk probability by the second weighting coefficient. The parameter channels are sorted according to the detection priority to obtain the parameter priority sequence.

[0009] In a preferred embodiment, the multi-parameter synchronous verification combination configuration includes: The intelligent decision-making model establishes a parameter correlation matrix, where the matrix elements represent the correlation strength between parameters; the model adjusts the parameter correlation matrix based on the fusion feature vector of the current product. Extract parameter pairs whose association strength exceeds the association threshold from the adjusted parameter association matrix; use a greedy clustering algorithm to divide the synchronous verification group, select the parameter pair with the highest association strength as the initial member, and add the parameter to the group if the association strength between a parameter and any member in the group exceeds the association threshold. Repeat this process until all associated parameters are divided to form a multi-parameter synchronous verification combination configuration.

[0010] In a preferred embodiment, the instruction generation module includes: The parameter detection order is determined according to the parameter priority sequence. For each parameter to be detected, the corresponding detection channel number and product pin number are determined according to the parameter type, and the coordinates of the relay position to be closed are calculated. For synchronous verification parameter groups, the relay positions corresponding to all parameters within the group are determined simultaneously to generate synchronous closing instructions; the relay matrix automatic construction instructions include instruction header, relay row number, relay column number, action type, and delay time fields.

[0011] In a preferred embodiment, the collaborative control module includes: Establishing a communication connection with the BUS box involves four stages: device discovery, connection request, authentication, and parameter negotiation. During the device discovery phase, a device query message is sent via broadcast, and the BUS box returns a device information response; during the connection request phase, a TCP connection request is sent to the BUS box; during the authentication phase, an authentication message containing the device identifier is sent; during the parameter negotiation phase, both parties negotiate communication parameters; after the connection is established, heartbeat data packets are sent periodically, and if the heartbeat times out for a preset number of consecutive times, the connection is determined to be disconnected and re-established.

[0012] In a preferred embodiment, the parameter synchronization detection module includes: The synchronous verification control unit takes out a synchronous verification group from the multi-parameter synchronous verification combination configuration and determines the list of parameters that need to be synchronously verified and the corresponding detection channels within the synchronous verification group; The synchronous triggering subunit sends a synchronous trigger signal to all detection channels in the synchronous verification group simultaneously; the synchronous acquisition subunit drives the synchronous acquisition clock, and each detection channel synchronously samples at the rising edge of the synchronous acquisition clock to obtain sampling data with consistent time; each detection channel performs individual judgment on the sampled data. The synchronization judgment subunit performs a comprehensive judgment based on the configured synchronization judgment logic. If the synchronization judgment logic is a logical AND judgment, it checks whether all individual judgment results are normal. If the synchronization judgment logic is a weighted judgment, it performs a weighted calculation on the individual judgment results according to the importance weight of each channel. If the weighted sum exceeds the judgment threshold, the comprehensive judgment result is abnormal.

[0013] In a preferred embodiment, the anomaly analysis module includes: Anomaly feature identification is performed on visual image data. The visual image of the current product is compared with the standard visual image. The difference is calculated using the image difference method. Areas with large differences are identified as potential anomaly areas. Anomaly identification is performed on the temperature distribution data. The temperature distribution of the current product is compared with that of a normal product. Locations where the temperature difference exceeds a preset temperature difference threshold are identified as temperature anomaly points. The system performs correlation analysis on visual anomalies, temperature anomalies, and electrical performance anomalies. It queries the historical fault analysis database to find correlation patterns and performs pattern matching between the current product's anomaly combinations and historical fault cases. It then uses the Euclidean distance similarity calculation method to select a preset number of historical fault cases as reference cases and extracts the root cause of the fault from these reference cases.

[0014] In a preferred embodiment, the anomaly analysis module further includes: Calculate the abnormal deviation of each abnormal parameter. For abnormal parameters that exceed the limit, if the parameter value exceeds the upper limit threshold, the abnormal deviation is the actual value of the parameter minus the upper limit threshold and then divided by the upper limit threshold. If the parameter value is lower than the lower limit threshold, the abnormal deviation is the lower limit threshold minus the actual value of the parameter and then divided by the lower limit threshold. Anomaly severity scores are calculated based on the degree of deviation and the safety level of the parameters. For synchronous verification groups where the results of multi-parameter synchronous verification are abnormal, the anomaly severity score is the sum of the severity scores of each abnormal parameter within the group. Based on the anomaly severity scores, product anomalies are classified into levels, and thresholds for minor and severe anomalies are set for grading.

[0015] In a preferred embodiment, the model optimization feedback module includes: Construct failure state learning samples, extract visual and temperature features from the fused feature vector set as input features, and extract abnormal state information from the intelligent anomaly analysis report as output labels; Update the failure state feature library, insert the learning samples one by one into the failure state feature library, perform deduplication check before insertion, and if the similarity exceeds the deduplication threshold, it is determined to be a duplicate sample and will not be inserted. Optimize the parameter priority algorithm, calculate the priority effectiveness score by statistically analyzing the detection order and anomaly detection of each parameter, and adjust the weight parameters if the priority effectiveness score is lower than the effectiveness threshold. Optimize the multi-parameter synchronous verification strategy by updating the parameter correlation matrix based on the actual data from this test, and then re-divide the synchronous verification groups using the updated parameter correlation matrix.

[0016] The beneficial effects of this invention are as follows: A deep learning intelligent decision-making model is constructed through the intelligent decision-making module. The parameter priority adaptive algorithm dynamically generates parameter priority sequences and multi-parameter synchronous verification combination configurations based on the product's fusion feature vector, realizing the intelligent generation of detection strategies. The instruction generation module converts the detection strategies into control instructions executable by the BUS box. The collaborative control module automatically controls the relay matrix to build test circuits and configure detection parameters, realizing the automated construction of the test environment. This solves the problem of large errors in test circuit construction in traditional methods and improves the accuracy and consistency of test environment construction. The parameter synchronization detection module employs a multi-parameter synchronous verification mechanism to synchronously trigger, collect, and comprehensively judge related parameters, achieving temporal consistency verification of multiple parameters and improving the efficiency of multi-parameter verification. The anomaly analysis module combines visual image data and temperature distribution data to perform correlation analysis on electrical performance anomalies, infers the underlying causes of anomalies based on pattern matching of historical fault cases, assesses the severity of anomalies, and generates intelligent anomaly analysis reports, thus enhancing anomaly analysis capabilities. The model optimization feedback module feeds the detection results back to the intelligent decision model for continuous optimization, enabling the test system to learn and improve its capabilities. Attached Figure Description

[0017] Figure 1 This is a block diagram of an intelligent testing system for vehicle electrical performance based on deep learning, according to the present invention. Figure 2 This is a flowchart of an intelligent testing system for vehicle electrical performance based on deep learning, according to the present invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses an intelligent testing system for vehicle electrical performance based on deep learning, such as... Figures 1 to 2 As shown, it includes the following steps: The feature construction module is used to collect on-board electrical performance data and external environment perception data, perform multi-dimensional feature extraction, and obtain a fused feature vector set. The data acquisition module collects data from the vehicle-mounted electronic products under test. A multi-dimensional feature extraction algorithm is used to extract features from the collected data, resulting in a fused feature vector set, which provides a data foundation for subsequent deep learning-based intelligent decision-making. The data acquisition module includes an electrical performance data acquisition unit and an environmental perception data acquisition unit; these two units work together to achieve comprehensive data acquisition.

[0020] S11, Collect vehicle electrical performance dataset; Based on the electrical performance data acquisition unit in the data acquisition module, electrical performance data of the vehicle electronic product under test is collected to obtain vehicle electrical performance dataset D1. The vehicle electrical performance dataset D1 includes voltage data, current data, pulse width modulation data, digital input data, analog input data, and CAN / LIN bus message data.

[0021] The electrical performance data acquisition unit adopts a multi-channel parallel acquisition architecture, with each channel operating independently without interference. Voltage data is acquired through the voltage acquisition channel, which includes a voltage divider circuit and an analog-to-digital converter (ADC). The voltage divider circuit adjusts the voltage to be measured to the input range of the ADC. The ADC has a resolution of 16 bits, and the acquisition frequency is determined based on the voltage change characteristics of the product under test. For products with voltage change frequencies below 1kHz, a sampling frequency of 1kHz is used; for products with voltage change frequencies above 1kHz, a sampling frequency of 10kHz is used to ensure compliance with the Nyquist sampling theorem. Current data is acquired through the current acquisition channel, which uses a precision sampling resistor and a differential amplifier circuit to convert the current signal into a voltage signal before acquisition. The sampling resistor value is selected according to the range of the current to be measured; a 0.1-ohm sampling resistor is used for small current measurements, and a 0.01-ohm sampling resistor is used for large current measurements. Pulse width modulation (PWM) data is acquired through the PWM acquisition channel, which uses a high-speed counter and edge capture circuitry to record the period, duty cycle, and level change timing of the PWM signal. The counter clock frequency is 100MHz, and the time resolution reaches the 10ns level. Digital input data is acquired through the digital signal acquisition channel, which uses a Schmitt trigger for level shaping and records the high and low level states and change timing of the digital signal. Analog input data is acquired through the analog signal acquisition channel, which uses a follower and anti-aliasing filter to record the voltage values ​​and change curves of the analog signal. CAN / LIN bus message data is acquired through the bus message acquisition channel, which integrates CAN and LIN bus transceivers, parses the frame header, data field, and check field of the bus message, and supports CAN2.0A, CAN2.0B, and LIN2.0 protocols.

[0022] S12, Collect the external environment perception dataset; Based on the environmental perception data acquisition unit in the data acquisition module, the external environment status of the vehicle electronic product under test is perceived and collected to obtain the external environment perception dataset D2. The external environment perception dataset D2 includes visual image data and temperature distribution data.

[0023] Visual image data is acquired using an industrial camera to capture the appearance of the product under test. The industrial camera employs a CMOS image sensor, and the pixel resolution is determined based on the size and detail requirements of the product under test. For small products under 100mm in size, a 2-megapixel industrial camera is used, while for large products over 100mm, a 5-megapixel industrial camera is used. The industrial camera is equipped with an adjustable focal length lens, with a focal length range of 8mm to 50mm and an aperture range of F1.4 to F16, supporting both autofocus and manual focus modes. The shooting angle of the industrial camera covers the front, sides, and key details of the product under test, capturing visual features such as the product's appearance, indicator light status, and display screen content. To ensure image quality, a ring-shaped LED light source provides uniform illumination with a color temperature of 5500K (Kelvin) and an illuminance of 1000 lux. Visual image data is acquired after the product under test has been powered on and stabilized. The stabilization time is determined based on the product characteristics, typically 30 to 60 seconds after power-on.

[0024] Temperature distribution data is collected from the surface of the product under test using a thermal imager. The thermal imager employs an uncooled infrared detector with an operating wavelength of 8μm to 14μm, a temperature measurement range of -20℃ to 150℃, and a measurement accuracy of ±2℃. The infrared resolution of the thermal imager is determined based on the temperature distribution characteristics of the product under test. A 160×120 pixel thermal imager is used for products with uniform temperature distribution, while a 320×240 pixel thermal imager is used for products with uneven temperature distribution. The thermal imager is equipped with an infrared lens with a field of view of 25° and a spatial resolution of 2.5mrad. Temperature distribution data is collected after the product under test has reached a thermal steady state after being powered on. The criterion for determining thermal steady state is a temperature change of less than 1℃ within 5 consecutive minutes.

[0025] S13, Extract electrical performance features; Based on the vehicle electrical performance dataset D1, an electrical performance feature extraction algorithm is used to extract electrical performance features, resulting in the electrical performance feature vector F_elec. The electrical performance feature extraction algorithm employs appropriate feature extraction methods for different types of electrical performance data to ensure that the extracted features comprehensively reflect the electrical performance status of the product.

[0026] For voltage data, statistical feature extraction methods were used to extract the voltage mean, voltage standard deviation, peak-to-peak voltage, and voltage rate of change. The voltage mean was calculated by summing the values ​​of all collected voltage sampling points and dividing by the total number of sampling points. The voltage standard deviation was calculated by squarening the difference between each voltage sampling point value and the voltage mean, summing the results, dividing by the total number of sampling points, and then taking the square root. The peak-to-peak voltage was calculated by finding the maximum and minimum values ​​among all voltage sampling points; the voltage rate of change was calculated by dividing the difference between two adjacent voltage sampling points by the sampling time interval, averaging all instantaneous voltage rate of change.

[0027] For current data, the same statistical feature extraction method as for voltage data is used to extract the mean current, standard deviation current, peak current, and rate of change of current. The peak current is calculated by finding the maximum value among all current sampling points.

[0028] For pulse width modulation (PWM) data, frequency domain analysis is used for feature extraction. A Fourier transform is performed on the PWM signal to convert the time-domain signal into a frequency-domain signal, obtaining the spectral distribution. The dominant frequency component is extracted from the spectral distribution, corresponding to the frequency point with the largest spectral amplitude. The duty cycle of the PWM signal is calculated by dividing the high-level duration within a complete cycle by the total cycle time. Harmonic components are extracted, representing the spectral amplitude at integer multiples of the dominant frequency component.

[0029] For CAN / LIN bus message data, a bus message parsing method is used to extract message features. This involves parsing the frame header information to extract the message identifier and message type; parsing the data field to extract the data content carried by the message; statistically analyzing message periodic features, which is achieved by recording the transmission timestamps of messages with the same identifier, calculating the time difference between two adjacent transmission timestamps, and averaging all time differences to obtain the message period; and statistically analyzing data field change features, which involves comparing the data field content of two adjacent messages and recording the position and magnitude of the changed data bytes.

[0030] The extracted voltage features, current features, pulse width modulation features, and bus message features are arranged and combined in order of feature type to obtain the electrical performance feature vector F_elec.

[0031] S14, Extracting External Environmental Perception Features: Based on the external environmental perception dataset D2, an environmental perception feature extraction algorithm is used to extract perception features, resulting in a perception feature vector F_sense. This algorithm combines computer vision and temperature analysis techniques to extract the product's appearance and thermal characteristics from visual images and temperature distributions.

[0032] For visual image data, image feature extraction methods are used to extract product appearance features. The visual images undergo preprocessing, including grayscale conversion, denoising, and contrast enhancement. Grayscale conversion transforms color images into grayscale to reduce computational complexity. Denoising uses median filtering to remove image noise, with the kernel size determined based on the noise level. Contrast enhancement employs histogram equalization to enhance image contrast.

[0033] The product appearance and morphology features are extracted from the preprocessed visual image. These features include the product outline shape, surface texture, and structural integrity. The product outline shape extraction method involves using an edge detection algorithm, specifically the Canny edge detection algorithm. This algorithm performs Gaussian filtering to smooth the image, then calculates the image gradient magnitude and direction, followed by non-maximum suppression to preserve the edge backbone. Finally, double threshold detection and edge connection are used to obtain the complete edges. The extracted edges are then fitted to obtain the product outline shape. The product surface texture extraction method uses the gray-level co-occurrence matrix (GLCM) method to extract texture features. After calculating the GLCM of the image, texture statistics such as contrast, correlation, energy, and homogeneity are extracted.

[0034] The indicator light status features are extracted from the visual image. These features include indicator light brightness, indicator light color, and indicator light blinking frequency. The brightness is extracted by locating the indicator light area in the visual image and calculating the average grayscale value of all pixels within that area. The color is extracted by extracting the red, green, and blue (RGB) components of the indicator light area from the color visual image and determining the color based on the proportional relationship between these components. The blinking frequency is extracted by continuously acquiring multiple frames of visual images and statistically analyzing the periodicity of the indicator light brightness changes over time.

[0035] The visual image is used to extract display screen content features, including the clarity, completeness, and accuracy of the displayed content. The method for extracting display content clarity is to calculate the gradient magnitude of the image in the display screen area; a larger gradient magnitude indicates higher image clarity. The method for extracting display content completeness is to use a character recognition algorithm to identify the characters or graphics displayed on the screen and determine whether the characters or graphics are displayed completely. The method for extracting display content accuracy is to compare the identified display content with the expected display content to determine whether the display content is correct.

[0036] For temperature distribution data, temperature features are extracted using temperature feature extraction methods. The temperature distribution data undergoes preprocessing, including outlier removal and temperature data smoothing. Outlier removal involves identifying and removing temperature points whose values ​​exceed a reasonable temperature range; the reasonable temperature range is determined based on the operating temperature range of the product under test. Temperature data smoothing uses a moving average method to smooth the temperature data.

[0037] The preprocessed temperature distribution data is analyzed to extract the mean temperature. The mean temperature is calculated by summing the values ​​of all temperature sampling points in the temperature distribution data and then dividing by the total number of sampling points. The temperature gradient is extracted by calculating the temperature difference between adjacent temperature sampling points and dividing by the spatial distance. The temperature gradient reflects the unevenness of the temperature distribution on the product surface. Hotspot locations are extracted by identifying the highest temperature sampling points in the temperature distribution data and recording their spatial coordinates. Finally, temperature change trends are extracted by continuously collecting temperature distribution data multiple times and calculating the rate of temperature change at the same sampling point over time. A positive rate of temperature change indicates an upward temperature trend, while a negative rate of temperature change indicates a downward temperature trend.

[0038] The extracted product appearance features, indicator light status features, display screen content features, and temperature features are arranged and combined in order of feature type to obtain the perception feature vector F_sense.

[0039] S15, integrate electrical performance features and sensing features; based on the electrical performance feature vector F_elec and the sensing feature vector F_sense, use the feature fusion method to perform feature fusion to obtain the fused feature vector set F.

[0040] The feature fusion method employs feature concatenation, combining the electrical performance feature vector F_elec and the sensing feature vector F_sense according to their feature dimensions. Before feature concatenation, both the electrical performance and sensing feature vectors are normalized to eliminate differences in units and numerical ranges between different features. The normalization process uses min-max normalization, subtracting the minimum value of each feature dimension from its value, and then dividing by the difference between the maximum and minimum values ​​of that feature dimension to obtain the normalized feature value. The normalized feature value ranges from 0 to 1.

[0041] The normalized electrical performance feature vector and the perception feature vector are concatenated, and then connected in dimensional order to obtain a fused feature vector. The dimension of the fused feature vector is equal to the sum of the dimensions of the electrical performance feature vector and the perception feature vector. This fused feature vector is used as a fused feature vector set F, which contains both the electrical performance information of the product under test and the perception information of the external environment, providing comprehensive data input for subsequent deep learning intelligent decision-making.

[0042] The intelligent decision-making module is used to build an intelligent decision-making model. It inputs the fused feature vector set into the decision-making model and generates detection strategy configuration parameters through a parameter priority adaptive algorithm. Based on the fused feature vector set F obtained from the feature construction module, a deep learning intelligent decision-making model is used to analyze the fused feature vector set to obtain the detection strategy configuration parameters C, providing an intelligent decision-making basis for the subsequent automated construction and detection of BUS boxes. The deep learning intelligent decision-making model intelligently generates a detection strategy adapted to the current product state by analyzing the product's electrical performance characteristics and environmental perception characteristics.

[0043] S21, Construct a deep learning intelligent decision-making model; Based on deep learning technology, construct a deep learning intelligent decision-making model M, which adopts a joint learning architecture of visual features and temperature features. The network structure of the deep learning intelligent decision-making model M includes three main parts: a feature extraction subnetwork, a feature fusion subnetwork, and a decision output subnetwork.

[0044] The feature extraction subnetwork includes a visual feature extraction branch and a temperature feature extraction branch. The visual feature extraction branch uses a convolutional neural network structure, containing four convolutional layers and two pooling layers. The first convolutional layer contains 32 kernels (3×3), with a stride of 1 and padding of 1, and uses the ReLU activation function. The first pooling layer uses max pooling, with a kernel size of 2×2 and a stride of 2. The second convolutional layer contains 64 kernels with the same parameters as the first convolutional layer. The second pooling layer also has the same parameters as the first pooling layer. The third and fourth convolutional layers contain 128 and 256 kernels, respectively. The temperature feature extraction branch uses a fully connected neural network structure, containing three fully connected layers. The first fully connected layer has 128 neurons, the second has 64 neurons, and the third has 32 neurons, all using the ReLU activation function.

[0045] The feature fusion subnetwork employs an attention mechanism fusion method. The implementation of the attention mechanism includes two steps: attention weight calculation and feature weighted fusion. Attention weight calculation involves linearly transforming and biasing the visual depth features and temperature depth features, then normalizing them using a softmax function to obtain the attention weights for the visual and temperature features. Feature weighted fusion involves multiplying the visual depth features by their corresponding attention weights, multiplying the temperature depth features by their corresponding attention weights, and then summing the two to obtain the fused depth features.

[0046] The decision output subnetwork consists of two fully connected layers and three output branches. The first fully connected layer has 128 neurons, and the second fully connected layer has 64 neurons. The three output branches output parameter priority sequences, multi-parameter synchronous verification combination configurations, and dynamic detection threshold configurations, respectively, and the output branches use linear activation functions.

[0047] S22, Training the Deep Learning Intelligent Decision Model: The deep learning intelligent decision model M is trained based on historical product test data, enabling the model to learn the mapping relationship between product features and detection strategies. The historical product test data includes the fused feature vectors of historical products, the actual abnormal states of historical products, and the optimal detection strategies for historical products. The historical dataset contains no fewer than 1000 product samples, covering both normal and abnormal products, with abnormal products accounting for 30%.

[0048] The deep learning intelligent decision-making model is trained using supervised learning, with the fused feature vectors of historical products used as model input and the optimal detection strategies of historical products used as supervisory labels for model output. The training objective function employs the mean squared error loss function, calculating the average of the squared differences between the predicted output and the true label for all samples. The Adam optimization algorithm is used to update model parameters during training, with a learning rate of 0.001, momentum parameters β1 set to 0.9, and β2 set to 0.999.

[0049] The training process employs batch training with a batch size of 32. The training and validation sets are divided in an 8:2 ratio. The training iterations are set to 100 epochs. Training is stopped early when the validation set loss value no longer decreases for 10 consecutive epochs to prevent overfitting. A learning rate decay strategy is used during training, reducing the learning rate to 0.5 times its original value every 20 epochs.

[0050] S23, Execute the parameter priority adaptive algorithm; based on the fused feature vector set F obtained from the feature construction module, input it into the trained deep learning intelligent decision model M, and the model executes the parameter priority adaptive algorithm to dynamically generate a parameter priority sequence. The parameter priority adaptive algorithm includes two sub-processes: failure risk assessment and dynamic calculation of parameter priorities.

[0051] The failure risk assessment subprocess evaluates the failure risk probability of each parameter channel based on the visual and temperature features in the fused feature vector set F, combined with the failure state feature library built within the deep learning model. The failure state feature library stores feature data from no fewer than 500 failure cases. The failure risk assessment employs an Euclidean distance similarity matching method to calculate the Euclidean distance between the current feature vector and the feature vectors of each failure case. The top 10 failure cases with the smallest Euclidean distance are selected as similar cases, and the frequency of anomalies in each parameter channel within these 10 cases is counted. The failure risk probability is equal to the number of times that parameter channel exhibits anomalies in the 10 similar cases divided by 10.

[0052] The parameter priority dynamic calculation sub-process calculates the detection priority based on the failure risk probability and parameter safety level. Parameter safety levels are preset: the base priority for voltage, current, and temperature parameters is set to 0.8; the base priority for PWM and bus message parameters is set to 0.6; and the base priority for brightness and display content parameters is set to 0.4. The detection priority is obtained by multiplying the base priority by 0.4 and then adding the failure risk probability multiplied by 0.6. All parameter channels are then sorted according to their detection priority from high to low to obtain the parameter priority sequence.

[0053] S24, Generate a multi-parameter synchronous verification configuration; Based on the fused feature vector set F and the deep learning intelligent decision model M, generate a multi-parameter synchronous verification configuration to achieve synchronous verification of associated parameters. An internal parameter association matrix is ​​established within the deep learning intelligent decision model, where matrix elements represent the association strength between parameter i and parameter j. The association strength is obtained through historical data statistics, calculated as: the number of times parameters i and j are simultaneously abnormal divided by the number of times parameter i is abnormal alone plus the number of times parameter j is abnormal alone minus the number of times they are simultaneously abnormal.

[0054] The model adjusts the parameter association matrix based on the fused feature vector of the current product. The adjustment method is as follows: multiply the original association strength by 1, add the adjustment coefficient multiplied by the similarity of the corresponding feature of the parameter. The adjustment coefficient is set to 0.2. The association threshold is set to 0.6, and parameter pairs with an association strength exceeding 0.6 are extracted from the adjusted parameter association matrix.

[0055] The synchronous verification group partitioning employs a greedy clustering algorithm. The parameter pair with the highest correlation strength is selected as the initial member of the first synchronous verification group. Then, other parameters are iterated over; if the correlation strength between a parameter and any member in the group exceeds 0.6, the parameter is added to the group. This process is repeated until no new parameters are added. The above process is repeated for the remaining ungrouped parameters until all correlated parameters are partitioned. The member parameter lists of all synchronous verification groups are combined to form a multi-parameter synchronous verification combination configuration.

[0056] S25, Generate dynamic detection threshold configuration; based on the feature distribution of the fused feature vector set, a dynamic threshold generation algorithm is used to generate the detection threshold for each parameter channel. The dynamic threshold generation algorithm extracts operating condition features from the fused feature vector set, including average temperature, average current, and average PWM duty cycle. Product data with similar operating condition features are retrieved from historical test data. The similarity is judged using the Euclidean distance method, and the top 50 products with the smallest distance are selected as similar operating condition samples.

[0057] The normal value distribution of each parameter channel in similar operating condition samples is statistically analyzed, and the parameter mean and standard deviation are calculated. For continuous parameters, the detection threshold is set using a 3-standard-deviation criterion: the upper threshold is the mean plus 3 times the standard deviation, and the lower threshold is the mean minus 3 times the standard deviation. For voltage parameters, if the calculated threshold range exceeds the operating voltage range specified in the product specifications, the range specified in the specifications shall prevail. For current parameters, the upper threshold is increased by a 10% safety margin. For temperature parameters, the upper threshold is set based on the product's highest operating temperature, and the lower threshold is set based on the ambient temperature.

[0058] For digital signal parameters, the detection thresholds adopt the expected logical state form, which is obtained from the product specifications. For bus message parameters, the detection thresholds include message period thresholds and message content verification rules. The message period thresholds are set according to the bus protocol specification, and the message content verification rules are defined according to the product communication protocol. The detection thresholds for each parameter channel are packaged into a dynamic detection threshold configuration.

[0059] S26, Output detection strategy configuration parameters; Based on the parameter priority sequence generated in step S23, the multi-parameter synchronous verification combination configuration generated in step S24, and the dynamic detection threshold configuration generated in step S25, these three are packaged into detection strategy configuration parameters. These parameters provide a complete intelligent decision-making basis for subsequent automated BUS box construction and detection. The detection strategy configuration parameters are encapsulated in JSON format for easy parsing and transmission by the collaborative control module.

[0060] The instruction generation module is used to obtain the detection strategy configuration parameters and to obtain the BUS box automated construction instruction set and detection control instruction set through the instruction generation method. Based on the detection strategy configuration parameters obtained from the intelligent decision-making module, an instruction generation algorithm is used to generate a BUS box instruction set. This instruction set drives the BUS box hardware unit to achieve automated assembly and detection. The instruction generation algorithm converts the detection strategy configuration parameters into control instructions executable by the BUS box hardware.

[0061] S31 generates automatic relay matrix setup instructions. Based on the parameter priority sequence and multi-parameter synchronous verification combination configuration in the detection strategy configuration parameters, the automatic relay matrix setup instructions are generated using a relay path planning algorithm. The BUS box integrates a 16×16 relay matrix. The relays are single-pole single-throw type, with a contact capacity of 5A and a mechanical life of 10 million cycles. The row lines of the relay matrix connect to the detection channel output of the BUS box, and the column lines connect to the pin inputs of the product under test.

[0062] The relay path planning algorithm determines the parameter detection order based on the parameter priority sequence. For each parameter to be detected, the corresponding detection channel number and product pin number are determined according to the parameter type, and the coordinates of the relay position to be closed are calculated. For the synchronous verification parameter group, the relay positions corresponding to all parameters in the group are determined simultaneously, and a synchronous closing command is generated. The relay switching timing is set as follows: the relay mechanical response time is 10ms, the circuit settling time is 50ms, and the time interval between two adjacent switching actions is set to 60ms.

[0063] The automatic relay matrix setup instructions use hexadecimal encoding format. Each instruction includes a header, relay row number, relay column number, action type, and delay time. A header of 0xA1 indicates a relay control instruction; an action type of 0x01 indicates closing; and 0x00 indicates opening. The delay time is in milliseconds.

[0064] S32 generates a multi-parameter synchronization detection command; based on the multi-parameter synchronization verification combination configuration in the detection strategy configuration parameters, a multi-parameter synchronization detection command is generated. Each synchronization verification parameter group is traversed, and the detection channel identifiers corresponding to all parameters within the group are extracted. The synchronization detection command includes three parts: synchronization trigger time, channel identifier list, and synchronization judgment logic.

[0065] The synchronization triggering time adopts a hardware clock synchronization mechanism. The BUS box integrates a high-precision clock source with a clock accuracy of 1μs. The synchronization determination logic is set according to the parameter correlation strength. Parameter groups with a correlation strength greater than 0.8 use logical AND determination, while parameter groups with a correlation strength between 0.6 and 0.8 use weighted determination. The weighting coefficient is obtained by normalizing the correlation strength.

[0066] The multi-parameter synchronization detection command uses a hexadecimal encoding format. The command includes fields such as command header, synchronization group number, number of channels, channel list, and decision logic. The command header 0xB1 indicates a synchronization detection command, and the decision logic is 0x01 for logical AND, 0x02 for logical OR, and 0x03 for weighted decision.

[0067] S33, Generate detection parameter configuration instructions; based on the dynamic detection threshold configuration in the detection strategy configuration parameters, generate detection parameter configuration instructions. Iterate through the detection thresholds of each parameter channel, extracting the threshold type and threshold value. For continuous value parameters, the threshold parameters include an upper threshold and a lower threshold, with the threshold value represented in floating-point format. For digital signal parameters, the threshold parameters represent the expected logical state, using Boolean values. For bus message parameters, the threshold parameters include the message period range and the message content mask.

[0068] The detection parameter configuration command also includes sampling parameter configuration. The sampling frequency is set according to the parameter type: 1kHz for voltage and current parameters, 10kHz for PWM parameters, and twice the bus baud rate for bus message parameters. The sampling duration is uniformly set to 1 second. The sampling trigger condition is set to software trigger mode.

[0069] The detection parameter configuration instructions use hexadecimal encoding format. The instructions include fields such as instruction header, channel number, threshold type, upper threshold, lower threshold, and sampling frequency. The instruction header 0xC1 indicates a parameter configuration instruction, the threshold type 0x01 indicates a continuous value, 0x02 indicates a logical value, and 0x03 indicates a message type.

[0070] S34, Package and generate the BUS box instruction set; based on the relay matrix automatic construction instruction generated in step S31, the multi-parameter synchronous detection instruction generated in step S32, and the detection parameter configuration instruction generated in step S33, package them in the order of instruction execution to obtain the complete BUS box instruction set. The instruction execution order is as follows: execute the parameter configuration instruction to configure the detection parameters, then execute the relay control instruction to build the test circuit, and finally execute the synchronous detection instruction to start the synchronous detection.

[0071] The data structure of the BUS box instruction set includes an instruction set header, parameter configuration instruction groups, relay control instruction groups, synchronization detection instruction groups, and an instruction set checksum. The instruction set header includes the version number, total number of instructions, product identifier, and timestamp. The instruction set checksum is calculated using the CRC16 algorithm to ensure the integrity and correctness of instruction transmission.

[0072] The collaborative control module is used to send automated setup instruction sets and detection control instruction sets to the BUS box, automatically build the test environment and configure the detection parameters to obtain the completed test environment; Based on the instruction set generated by the instruction generation module, the instruction set is sent to the BUS box hardware unit by the collaborative control module to obtain the BUS box ready status and complete the automated construction of the test environment.

[0073] S41, Establish a communication connection between the collaborative control module and the BUS box; the collaborative control module establishes a communication connection with the BUS box hardware unit through an Ethernet interface. The Ethernet interface uses the TCP / IP protocol, with a communication rate of 100Mbps, and supports full-duplex communication. The communication connection establishment process includes four stages: device discovery, connection request, authentication, and parameter negotiation.

[0074] During the device discovery phase, the collaborative control module broadcasts a device query message. Upon receiving the query message, the BUS box returns a device information response, including the device model, serial number, and IP address. During the connection request phase, the collaborative control module sends a TCP connection request to the BUS box's IP address. During the authentication phase, the collaborative control module sends an authentication message containing the device identifier, and the BUS box verifies the identifier's validity. During the parameter negotiation phase, both parties negotiate communication parameters, including packet size, timeout, and retransmission count. The packet size is set to 1024 bytes, the timeout to 5 seconds, and the retransmission count to 3. After the connection is established, the collaborative control module sends a heartbeat packet every 10 seconds. The BUS box returns a heartbeat response. If three consecutive heartbeat timeouts occur, the connection is considered broken and re-established.

[0075] S42, Issue BUS Box Command Set; Based on the established communication connection, the collaborative control module issues the BUS box command set to the BUS box hardware unit. Command issuance uses a packet transmission mechanism, with each data packet being 1024 bytes in size. The collaborative control module divides the command set into packets according to the packet size; each data packet contains a header, sequence number, data content, and CRC checksum. The header is 0xAA55, and the sequence number increments from 0.

[0076] The collaborative control module sends data packets sequentially according to their sequence numbers, waiting for a 5-second timeout after each packet transmission from the BUS box. Upon receiving a data packet, the BUS box performs a CRC check; if the check passes, it returns an acknowledgment (0xACK); otherwise, it returns a negative acknowledgment (0xNAK). If a timeout occurs or a negative acknowledgment is received, the collaborative control module retransmits the data packet, up to a maximum of three retransmissions. After all data packets have been sent, the collaborative control module sends a transmission completion flag (0xEOT). The BUS box then reassembles and verifies the integrity of all received data packets. Upon successful verification, it returns a reception completion acknowledgment, which includes the total number of commands and the verification result.

[0077] S43, the BUS box performs automatic relay matrix setup. Based on the received BUS box instruction set, the matrix controller inside the BUS box hardware unit parses the automatic relay matrix setup instructions and controls the relay matrix to automatically build the test circuit. The matrix controller is implemented using an FPGA chip with an operating frequency of 100MHz. The matrix controller parses the relay row number, column number, action type, and delay time in the instructions to generate the relay drive timing sequence.

[0078] The relay drive circuit uses a Darlington driver chip with a drive voltage of 12V and a drive current of 100mA. For relays that need to be closed, the matrix controller outputs a high-level signal to the corresponding drive circuit. The drive circuit applies a 12V drive voltage to the relay coil, energizing it and generating a magnetic field. This magnetic field attracts the relay contacts to close, with a closing time of 10ms. For relays that need to be opened, the matrix controller outputs a low-level signal, the drive circuit stops supplying power, and the relay contacts open under the action of spring return force, with an opening time of 5ms.

[0079] The matrix controller strictly controls the relay switching sequence, with a 60ms interval between two adjacent switching actions, including 10ms of mechanical action time and 50ms of circuit stabilization time. The BUS box completes the relay switching step by step according to the instruction sequence, ultimately realizing the automatic construction of the test circuit and correctly connecting each test channel of the BUS box to the corresponding pin of the product under test.

[0080] S44, BUS box configuration detection parameters; based on the received detection parameter configuration command, the multi-channel detection unit of the BUS box configures the detection threshold and sampling parameters of each detection channel. The multi-channel detection unit includes 8 parallel detection channels: voltage detection channel, current detection channel, PWM detection channel, digital signal detection channel, analog signal detection channel, temperature detection channel, bus message detection channel, and brightness detection channel. Each detection channel includes signal conditioning circuitry, a 16-bit analog-to-digital converter, and FPGA decision logic circuitry.

[0081] The multi-channel detection unit receives detection parameter configuration instructions and distributes detection threshold parameters and sampling parameters to the corresponding detection channels according to the channel numbers. Each detection channel writes the detection threshold parameters into the threshold register of the judgment logic circuit. The threshold register is implemented using dual-port RAM with a capacity of 32 bytes. The sampling parameters are written into the sampling control register of the data acquisition circuit. The sampling control register includes a sampling frequency register, a sampling duration register, and a trigger mode register. The judgment logic circuit performs real-time anomaly judgment based on the detection threshold in the threshold register, with a judgment delay of less than 1μs.

[0082] S45, the BUS box is configured with multi-parameter synchronous verification; based on the received multi-parameter synchronous detection command, the BUS box's synchronous verification control unit configures the channel combination and synchronization logic for multi-parameter synchronous verification. The synchronous verification control unit is implemented using an FPGA chip and includes a synchronous trigger subunit, a synchronous acquisition subunit, and a synchronous judgment subunit.

[0083] The synchronization trigger subunit parses the synchronization group number and channel list in the instruction, establishes a mapping relationship between the synchronization verification group and the detection channels, and stores the mapping relationship in a lookup table. The synchronization trigger subunit configures the synchronization trigger signal, which is a hardware pulse signal with a pulse width of 1μs, and sends it simultaneously to all detection channels within the synchronization verification group.

[0084] The synchronous acquisition subunit is configured with a synchronous acquisition clock, the clock frequency of which is the highest value of the required sampling frequency for all parameters within the synchronous verification group. The synchronous acquisition clock drives multiple detection channels within the synchronous verification group to acquire data simultaneously, with a time synchronization accuracy of 1 μs. The synchronous acquisition subunit is configured with a software trigger mode for acquisition triggering.

[0085] The synchronous judgment subunit configures the comprehensive judgment method according to the judgment logic parameters in the instruction. The logical AND judgment is implemented using AND gates, the logical OR judgment using OR gates, and the weighted judgment using multiply-accumulate units. The comprehensive judgment result is output to the detection process controller in real time.

[0086] S46, return to the BUS box ready state; based on the automatic test line setup completed in step S43, the detection parameter configuration completed in step S44, and the multi-parameter synchronous verification configuration completed in step S45, the BUS box hardware unit has completed all preparations and enters the ready state. The BUS box status feedback module monitors the working status of each module in real time, with a monitoring cycle of 100ms. When the relay matrix status, detection channel configuration status, and synchronous verification configuration status are all normal, the status feedback module generates the BUS box ready state.

[0087] The data structure for the BUS box's readiness status includes fields such as status identifier, number of detection channels, number of synchronization verification groups, relay connection status, and timestamp. A status identifier of 0xRDY indicates readiness. The BUS box's readiness status is fed back to the collaborative control module via the communication interface. After receiving the readiness status, the collaborative control module verifies the configuration information digest to confirm that the configuration information matches the issued instruction set. Once the verification is successful, the detection process is initiated.

[0088] The parameter synchronization detection module is used to perform detection on the BUS box according to the multi-parameter synchronization verification strategy based on the established test environment, and obtain the detection result dataset. Based on the BUS box readiness status obtained from the collaborative control module, a multi-parameter synchronous verification mechanism is used to perform detection, and a detection result dataset is obtained.

[0089] S51, Initiate the detection process; based on the BUS box's ready state, the collaborative control module sends a detection start command to the BUS box. The detection start command includes fields such as a command header, detection mode, detection rounds, and detection interval. The command header is 0xSTART, the detection mode is 0x01 for single-pass detection mode and 0x02 for continuous detection mode. In single-pass detection mode, the process stops after completing one round of parameter detection; in continuous detection mode, parameter detection is performed cyclically until a stop command is received. The detection rounds specify the number of detection cycles in continuous detection mode, and the detection interval specifies the time interval between two adjacent detections, in seconds.

[0090] After the BUS box receives the detection start command, the detection process controller initializes the detection process according to the detection mode and detection parameters in the command, and executes the detection of each parameter in sequence according to the parameter priority sequence.

[0091] S52, Detection is performed according to the parameter priority sequence; based on the parameter priority sequence generated by the intelligent decision module, the BUS box performs parameter detection sequentially in descending order of parameter priority. The detection process controller retrieves the parameter to be detected with the highest priority from the parameter priority sequence, determines the detection channel corresponding to the parameter, and sends a detection trigger signal to the detection channel.

[0092] Upon receiving a trigger signal, the detection channel activates the signal conditioning circuit and data acquisition circuit. The signal conditioning circuit amplifies, filters, and performs level conversion on the raw signal output by the product under test. The data acquisition circuit uses a 16-bit analog-to-digital converter for sampling and quantization; the sampling frequency and duration are determined based on the sampling parameters configured by the collaborative control module. The data acquisition circuit outputs digitized sampled data, and the judgment logic circuit compares the sampled data with the detection threshold configured by the collaborative control module to determine whether the parameter is abnormal.

[0093] For voltage and current parameters, the judgment logic is as follows: if the sampled data exceeds the upper threshold or falls below the lower threshold, it is judged as abnormal, and the abnormality type is recorded as "over-limit abnormality". For PWM parameters, the frequency and duty cycle of the sampled data are extracted; if the frequency or duty cycle exceeds the threshold range, it is judged as abnormal. For digital signal parameters, the logic level state of the sampled data is extracted; if the logic level state does not match the expected state, it is judged as abnormal. For bus message parameters, the bus message in the sampled data is parsed, and the message format and content are verified; if the message format is incorrect or the content does not conform to the protocol specification, it is judged as abnormal.

[0094] The detection channel packages the judgment result, sampling data, and anomaly information into a detection result for that parameter. The detection result includes the parameter identifier, detection timestamp, judgment result, sampling data, and anomaly information. The detection process controller receives the detection result for that parameter, stores it in the detection result cache, and then continues to retrieve the next parameter to be detected from the parameter priority sequence, repeating the above detection process until all parameters have been detected.

[0095] S53, Perform multi-parameter synchronous verification; Based on the multi-parameter synchronous verification combination configuration generated by the intelligent decision module, the synchronous verification control unit of the BUS box performs multi-parameter synchronous verification to achieve synchronous detection and comprehensive judgment of related parameters. The synchronous verification control unit takes a synchronous verification group from the multi-parameter synchronous verification combination configuration and determines the list of parameters that need to be synchronously verified and the corresponding detection channels within the group.

[0096] The synchronous trigger subunit simultaneously sends a synchronous trigger signal to all detection channels within the group. This synchronous trigger signal uses a hardware pulse signal to ensure that all detection channels start detection simultaneously. The synchronous acquisition subunit drives a synchronous acquisition clock, and the data acquisition circuits of each detection channel sample synchronously on the rising edge of the synchronous acquisition clock, with a time synchronization accuracy of 1μs, resulting in time-consistent sampled data. The judgment logic circuit of each detection channel performs individual judgment on the sampled data, obtaining an individual judgment result for each channel.

[0097] The synchronization judgment subunit receives the individual judgment results from all detection channels within the synchronization verification group and performs a comprehensive judgment based on the synchronization judgment logic configured by the collaborative control module. If the synchronization judgment logic is an AND judgment, it checks whether all individual judgment results are normal. If all are normal, the comprehensive judgment result is normal; otherwise, the comprehensive judgment result is abnormal. If the synchronization judgment logic is an OR judgment, it checks whether any individual judgment results are abnormal. If any are abnormal, the comprehensive judgment result is abnormal. If the synchronization judgment logic is a weighted judgment, it weights the individual judgment results according to the importance weight of each channel, assigning a value of 1 to abnormal judgment results and 0 to normal judgment results, calculating the weighted sum. If the weighted sum exceeds the judgment threshold, the comprehensive judgment result is abnormal.

[0098] The synchronous judgment subunit outputs the comprehensive judgment result of the synchronous verification group. This result includes the synchronous verification group identifier, the comprehensive judgment conclusion, the individual judgment result for each channel, and the detection timestamp. The detection process controller receives the comprehensive judgment result and stores it in the detection result cache. The synchronous verification control unit continues to retrieve the next synchronous verification group from the multi-parameter synchronous verification combination configuration, repeating the above synchronous verification process until all synchronous verification groups are verified. Through the multi-parameter synchronous verification mechanism, the BUS box can synchronously collect data from multiple related parameters and perform comprehensive judgment, effectively determining the product's functional status.

[0099] S54, Summarize the detection results; Based on the detection results of each parameter completed in step S52 and the multi-parameter synchronous verification results completed in step S53, the detection process controller of the BUS box summarizes all detection results to obtain a detection result dataset. The detection process controller reads the detection results of each parameter and the comprehensive judgment results of each synchronous verification group from the detection result cache, sorts the detection results according to the detection timestamp, and generates a time-series detection result record.

[0100] The detection results for each parameter are compiled, including parameter identifier, detection timestamp, judgment result, anomaly type, sampled data, and anomaly value. The comprehensive judgment results for each synchronous verification group are also compiled, including synchronous verification group identifier, comprehensive judgment conclusion, individual judgment results for each channel, and detection timestamp. The compiled detection results for each parameter and the synchronous verification results are then packaged, and metadata such as detection task identifier, detection round, and detection completion time are added to form a complete detection result dataset R.

[0101] The detection result dataset is encapsulated in JSON format and fed back to the collaborative control module through the status feedback module and communication interface of the BUS box. The collaborative control module receives the detection result dataset and passes it to the subsequent anomaly analysis module for intelligent analysis.

[0102] The anomaly analysis module is used to generate intelligent anomaly analysis reports based on the detection result dataset and external environmental perception data. Based on the detection result dataset obtained from the parameter synchronization detection module, the fused feature vector set obtained from the feature construction module, and the external environment perception dataset, an intelligent anomaly analysis report is generated using an intelligent anomaly analysis algorithm, providing testers with a comprehensive product status assessment.

[0103] S61, Summarize Abnormal Results: Based on the detection result dataset, an abnormal result summarization method is used to summarize the abnormal states of each channel and the combined judgment results of multi-parameter synchronous verification. The detection results of each parameter in the detection result dataset are traversed, and parameter records with abnormal judgment results are extracted. The number of abnormal parameters is counted, and a list of abnormal parameter identifiers is recorded. Abnormal parameters are classified and statistically analyzed according to their abnormality type, including over-limit anomalies, frequency anomalies, duty cycle anomalies, logic anomalies, and message anomalies. The frequency of occurrence of each abnormality type is counted, and an anomaly type distribution statistical result is generated.

[0104] The system iterates through the comprehensive judgment results of each synchronous verification group in the detection result dataset, extracts the records of synchronous verification groups whose comprehensive judgment conclusion is abnormal, counts the number of abnormal synchronous verification groups, and records the identifier list of abnormal synchronous verification groups and the specific abnormal channels within each group. The number of abnormal parameters, the list of abnormal parameter identifiers, the distribution of abnormal types, the number of abnormal synchronous verification groups, and the details of the abnormal synchronous verification groups are packaged into a summary of abnormal results.

[0105] S62, perform visual and temperature anomaly correlation analysis; based on visual image data and temperature distribution data in the external environment perception dataset, combined with electrical performance anomaly information in the detection result dataset, use correlation analysis methods to identify the underlying causes of product anomalies.

[0106] Anomaly identification is performed on visual image data by comparing the current product's visual image with a standard visual image of a normal product. Image differencing is used to calculate the differences between the two images, and areas with significant differences are identified as potential anomaly areas. Anomaly features extracted include appearance distortion, color anomalies, abnormal indicator light status, and abnormal display content. Anomaly identification is also performed on temperature distribution data by comparing the current product's temperature distribution with that of a normal product. Temperature difference distribution is calculated, and locations with temperature differences exceeding 5°C are identified as temperature anomaly points. Temperature anomaly features extracted include localized overheating, abnormal temperature gradients, and hotspot location shifts.

[0107] This study performs correlation analysis on visual anomalies, temperature anomalies, and electrical performance anomalies. It queries a historical fault analysis database containing at least 300 product fault cases to find correlation patterns between specific electrical performance anomalies and visual / temperature anomalies. The current product's electrical performance anomalies, visual anomalies, and temperature anomalies are then matched with these historical fault cases using Euclidean distance similarity calculation. The top five historical fault cases with the highest similarity are selected as reference cases, from which the root cause of the fault is extracted as an inference of the underlying cause of the current product's anomalies. The visual anomaly features, temperature anomaly features, correlation patterns, and inferred root causes obtained from the correlation analysis are packaged into a visual and temperature anomaly correlation analysis result.

[0108] S63, Assess the Severity of Anomalies: Based on the anomaly values ​​in the detection result dataset and the multi-parameter synchronous verification results, an anomaly severity assessment method is used to evaluate the severity of product anomalies. The anomaly deviation rate for each anomaly parameter is calculated. For out-of-limit anomalies, if the parameter value exceeds the upper threshold, the anomaly deviation rate is the actual parameter value minus the upper threshold and then divided by the upper threshold. If the parameter value is below the lower threshold, the anomaly deviation rate is the lower threshold minus the actual parameter value and then divided by the lower threshold. For logical anomalies, the anomaly deviation rate is set to a fixed value of 1.0. For message anomalies, the anomaly deviation rate is the number of error bits in the message divided by the total number of bits.

[0109] An anomaly severity score is calculated based on the anomaly deviation and the parameter's safety level. The anomaly severity score is the anomaly deviation multiplied by the parameter safety level coefficient. The safety level coefficient for voltage, current, and temperature parameters is set to 2.0, for PWM and bus message parameters to 1.5, and for other parameters to 1.0. For a synchronization verification group where multiple parameters show anomalies, the anomaly severity score is the sum of the severity scores of each anomalous parameter within the group.

[0110] Product anomalies are categorized into levels based on anomaly severity scores, with a threshold of 0.2 for minor anomalies and 0.5 for severe anomalies. An anomaly severity score below 0.2 is considered a minor anomaly, between 0.2 and 0.5 a moderate anomaly, and above 0.5 a severe anomaly. The anomaly deviation, severity score, and level of each anomaly parameter are packaged together to form the anomaly severity assessment result.

[0111] S64, Generate Intelligent Anomaly Analysis Report; Based on the anomaly result summary information obtained in step S61, the visual and temperature anomaly correlation analysis results obtained in step S62, and the anomaly severity assessment results obtained in step S63, an intelligent anomaly analysis report is generated. The report framework includes six parts: report header, detection overview, anomaly details, correlation analysis, severity assessment, and handling recommendations.

[0112] The report header includes basic information such as the product identification, testing time, testing personnel, and report number. The testing overview section includes the total number of tested parameters, the number of abnormal parameters, the number of normal parameters, and the pass rate (the pass rate is calculated as the number of normal parameters divided by the total number of tested parameters, then multiplied by 100%). The anomaly details section lists detailed information for each abnormal parameter, including parameter identifier, anomaly type, anomaly value, testing timestamp, and anomaly deviation. The correlation analysis section includes visual anomaly characteristics, temperature anomaly characteristics, correlation patterns, and inferences about underlying causes. The severity assessment section includes an anomaly severity score and anomaly level. The handling recommendations section provides targeted handling suggestions based on the anomaly level: minor anomalies are recommended for continued observation, moderate anomalies for repair and inspection, and severe anomalies for immediate shutdown.

[0113] The content is organized and formatted according to the report framework to generate a structured intelligent anomaly analysis report. The report is packaged in JSON format for easy reading by testers and systematic storage. After the report is generated, it is output to the test terminal or stored in the database through the collaborative control module.

[0114] The model optimization feedback module is used to feed back the intelligent anomaly analysis report to the deep learning intelligent decision-making model to optimize the model's decision-making capabilities. Based on the intelligent anomaly analysis report generated by the anomaly analysis module, the fused feature vector set obtained by the feature construction module, the detection strategy configuration parameters obtained by the intelligent decision-making module, and the detection result dataset obtained by the parameter synchronization detection module, the deep learning intelligent decision-making model is optimized using the model optimization method to obtain the optimized deep learning intelligent decision-making model, thereby realizing the continuous learning and capability improvement of the model.

[0115] S71. Constructing Failure State Learning Samples: Based on the fused feature vector set and the intelligent anomaly analysis report, construct failure state learning samples. Extract visual and temperature features from the fused feature vector set. Visual features include product appearance and shape characteristics, indicator light status characteristics, and display screen content characteristics. Temperature features include average temperature, temperature gradient, and hotspot locations. Extract anomaly state information from the intelligent anomaly analysis report. Anomaly state information includes anomaly parameter identifiers, anomaly type, anomaly severity, and underlying causes of the anomaly.

[0116] Visual and temperature features are used as input features for the learning samples, and abnormal state information is used as the output label. The input features are normalized, and the output labels are one-hot encoded, converting the abnormal state information into a numerical vector form. The input features and output labels are combined to form a failure state learning sample. For products that pass inspection, visual and temperature features are extracted as input, and the normal state label is used as output to construct a normal state learning sample. The failure state learning samples and normal state learning samples obtained in this inspection are packaged together, and metadata such as sample source identifier and collection time are added to form the learning sample set for this inspection.

[0117] S72, Update the failure state feature library; Based on the learning sample set constructed in step S71, update the failure state feature library within the deep learning intelligent decision-making model. The failure state feature library stores the failure state features and corresponding anomaly information of historical products. The data structure includes feature vectors, anomaly labels, and sample element information.

[0118] Each learning sample in the learning sample set is inserted into the failure state feature library one by one. Before insertion, the learning samples are deduplicated, and the feature similarity between the new learning sample and the existing samples in the feature library is calculated using the cosine similarity method, which calculates the dot product of the two vectors divided by the product of their magnitudes. If the similarity exceeds the deduplication threshold of 0.95, the sample is considered a duplicate and is not inserted. For non-duplicate new learning samples, they are added to the failure state feature library. The feature library uses a KD-tree index structure to accelerate sample retrieval and supports fast similarity queries. The capacity limit of the feature library is set to 10,000 samples. When the feature library capacity reaches the limit, the oldest sample is removed using a first-in, first-out (FIFO) strategy. By continuously updating the failure state feature library, the deep learning intelligent decision-making model can continuously learn new product failure state features.

[0119] S73, Optimize the parameter priority algorithm; based on the detection result dataset and the parameter priority sequence in the detection strategy configuration parameters, evaluate the effectiveness of the parameter priority sequence and optimize the priority calculation algorithm. Statistically analyze the detection order and anomaly detection of each parameter, and calculate the priority effectiveness score. The calculation method is: sum the priority rankings of all abnormal parameters, divide by the total number of abnormal parameters, and then take the reciprocal. A higher effectiveness score indicates a more reasonable priority setting. If the priority effectiveness score is lower than the effectiveness threshold of 0.5, the current parameter priority sequence is deemed insufficiently effective, and the priority algorithm needs to be optimized.

[0120] The optimization method is as follows: Analyze parameters with unreasonable priority settings, extract the characteristics and actual failure risks of these parameters, and adjust the weight parameters in the parameter priority algorithm. The weight parameters include the basic priority weight and the failure risk weight, with initial values ​​of 0.4 and 0.6, respectively. Adjust the weight parameters based on the optimization results, with an adjustment step size of 0.05. Recalculate the parameter priority sequence using the adjusted weight parameters. If the effectiveness score of the new priority sequence is higher than the original sequence, then the parameter priority algorithm is updated using the new weight parameters.

[0121] S74, Optimize the multi-parameter synchronous verification strategy; based on the judgment results of multi-parameter synchronous verification combinations in the detection result dataset, evaluate the effectiveness of the synchronous verification combination and optimize the parameter correlation analysis algorithm. Statistically analyze the comprehensive judgment results of each synchronous verification group and the individual judgment results of each parameter within the group. For synchronous verification groups with an abnormal comprehensive judgment result, check whether multiple parameters within the group are indeed simultaneously abnormal. Calculate the synchronous verification effectiveness score, calculated as: the number of simultaneously abnormal parameter pairs divided by the total number of synchronous verification groups. A higher effectiveness score indicates a more effective synchronous verification strategy.

[0122] The optimization method is as follows: Adjust the synchronous verification groups with insufficient effectiveness, and remove parameters with weak correlation within the groups. Based on the actual data from this test, analyze the actual correlation performance between parameters and update the parameter correlation matrix. Use the updated parameter correlation matrix to re-divide the synchronous verification groups and generate an optimized multi-parameter synchronous verification combination configuration.

[0123] S75, Optimize the dynamic threshold generation algorithm; optimize the dynamic detection threshold generation algorithm based on the matching degree between the detection result dataset and the actual anomalies. Evaluate the accuracy of the dynamic detection threshold. For parameters judged as abnormal, manually verify whether the parameter is truly abnormal, and count the number of false positives and false negatives. Calculate the threshold accuracy by dividing the number of correct judgments by the total number of judgments and then multiplying by 100%. If the accuracy is lower than the accuracy threshold of 95%, the dynamic detection threshold is considered improperly set, and the threshold generation algorithm needs to be optimized.

[0124] The optimization method is as follows: Analyze the parameters of false positives and false negatives, extract the actual values ​​of these parameters, and set thresholds. For false positive parameters, increase the tolerance coefficient of the threshold, adjusting the step size to 0.1. For false negative parameters, decrease the tolerance coefficient of the threshold, adjusting the step size to 0.1. Regenerate the dynamic detection threshold using the adjusted tolerance coefficient, apply the new threshold to historical detection data for regression testing, and evaluate the accuracy of the new threshold. If the accuracy of the new threshold improves, update the dynamic threshold generation algorithm using the adjusted tolerance coefficient.

[0125] S76, output the optimized deep learning intelligent decision model; based on the failure state feature library updated in step S72, the parameter priority algorithm optimized in step S73, the multi-parameter synchronous verification strategy optimized in step S74, and the dynamic threshold generation algorithm optimized in step S75, integrate and update the optimized deep learning intelligent decision model to obtain the optimized deep learning intelligent decision model.

[0126] The model update process is as follows: The updated failure state feature library replaces the old feature library within the model; the weight parameters of the optimized parameter prioritization algorithm are updated to the model's priority calculation unit; the optimized parameter correlation matrix is ​​updated to the model's synchronization verification unit; and the tolerance coefficient of the optimized threshold generation algorithm is updated to the model's threshold generation unit. The updated model is then tested and verified using a verification dataset containing 100 product samples to evaluate its decision accuracy and stability. If the model's decision accuracy reaches 95% or higher, the optimized model is deployed to the actual test system, replacing the original model. If the model's performance does not meet the requirements, the update is rolled back, and the original model continues to be used.

[0127] The optimized deep learning intelligent decision-making model is used in subsequent detection tasks, continuously receiving new detection data and anomaly feedback, constantly learning and optimizing, forming a closed-loop continuous improvement mechanism, which continuously improves the detection capability of the vehicle electrical performance intelligent testing system.

[0128] This invention focuses on the application of intelligent testing of electrical performance of vehicle controller products, taking a certain model of vehicle LED lighting controller as the actual application object. This controller needs to perform comprehensive testing of multiple parameters such as voltage, current, PWM signal, digital signal, bus message and temperature. The intelligent vehicle electrical performance testing system based on deep learning, as described in this invention, was used for testing. The system collects electrical performance data from the controller and environmental perception data. An industrial camera captures the controller's appearance and LED indicator status, while a thermal imager captures the surface temperature distribution of the controller, resulting in a fused feature vector set. After analyzing the fused features using a deep learning intelligent decision model, local hotspots in the controller's temperature distribution and fluctuations in LED brightness are identified. Based on this, a parameter priority sequence is dynamically generated, prioritizing current and temperature parameters. Simultaneously, LED brightness and PWM signal parameters are grouped into a synchronous verification group. The system automatically generates a BUS box instruction set, and the relay matrix inside the BUS box automatically constructs the test circuit according to the instructions. After configuring detection thresholds, multi-parameter synchronous verification is performed. During the testing process, abnormal fluctuations in the current parameter and temperatures exceeding the set thresholds are detected. The synchronous verification group detects simultaneous anomalies in both LED brightness and PWM duty cycle. The system generates an intelligent anomaly analysis report. Through visual and temperature anomaly correlation analysis, it identifies abnormal heating in the controller's internal power module, causing current fluctuations and unstable brightness. The severity of the anomaly is assessed as moderate, and repair and inspection are recommended. The detection results are fed back to the deep learning model for optimization. The model updates its failure state feature library and optimizes its parameter priority algorithm, enabling subsequent testing of similar products to more accurately identify power module anomalies. The entire testing process achieves automated test environment setup, simultaneous verification of multiple parameters, and intelligent anomaly analysis, effectively improving testing efficiency and accuracy.

[0129] Tables 1 and 2 show examples of data obtained during the actual testing process: Table 1. Example of electrical performance test data for vehicle-mounted LED lighting controllers

[0130] Table 2. Examples of external environment perception data for vehicle-mounted LED lighting controllers

[0131] The intelligent testing system for vehicle electrical performance based on deep learning provided by this invention realizes the intelligent and automated testing of vehicle electronic products, and solves the technical problems of large error in test circuit construction, low efficiency of multi-parameter verification and weak anomaly analysis capability in traditional testing methods, thus providing an effective technical means for quality assurance of vehicle electronic products.

[0132] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A deep learning-based intelligent testing system for vehicle electrical performance, characterized in that, include: The feature construction module is used to collect on-board electrical performance data and external environment perception data, perform multi-dimensional feature extraction, and obtain a fused feature vector set. The intelligent decision-making module is used to build an intelligent decision-making model. It inputs the fused feature vector set into the decision-making model and generates detection strategy configuration parameters through a parameter priority adaptive algorithm. The instruction generation module is used to obtain the detection strategy configuration parameters and to obtain the BUS box automated construction instruction set and detection control instruction set through the instruction generation method. The collaborative control module is used to send automated setup instruction sets and detection control instruction sets to the BUS box, automatically build the test environment and configure the detection parameters to obtain the completed test environment; The parameter synchronization detection module is used to perform detection on the BUS box according to the multi-parameter synchronization verification strategy based on the established test environment, and obtain the detection result dataset. The anomaly analysis module is used to generate intelligent anomaly analysis reports based on the detection result dataset and external environmental perception data. The model optimization feedback module is used to feed back the intelligent anomaly analysis report to the intelligent decision-making model to optimize the model's decision-making capabilities.

2. The intelligent testing system for vehicle electrical performance based on deep learning according to claim 1, characterized in that, The intelligent decision-making model includes: The system comprises a feature extraction subnetwork, a feature fusion subnetwork, and a decision output subnetwork; the feature extraction subnetwork includes a visual feature extraction branch and a temperature feature extraction branch. The feature fusion subnetwork adopts an attention mechanism fusion method. After linear transformation and bias processing of visual depth features and temperature depth features, attention weights are obtained by normalization through the softmax function. The visual depth features and temperature depth features are multiplied by their corresponding attention weights and then added together to obtain the fused depth features. The decision output subnetwork contains three output branches, which respectively output parameter priority sequence, multi-parameter synchronous verification combination configuration, and dynamic detection threshold configuration.

3. The intelligent testing system for vehicle electrical performance based on deep learning according to claim 1, characterized in that, The parameter priority adaptive algorithm includes: The failure risk assessment subprocess and the parameter priority dynamic calculation subprocess; the failure risk assessment subprocess selects a preset number of similar cases based on the fused feature vector set and the failure state feature library, and calculates the failure risk probability by counting the frequency of anomalies in each parameter channel. The parameter priority dynamic calculation subprocess calculates the detection priority based on the failure risk probability and parameter safety level. The detection priority is obtained by multiplying the basic priority by the first weighting coefficient and adding the failure risk probability by the second weighting coefficient. The parameter channels are sorted according to the detection priority to obtain the parameter priority sequence.

4. The intelligent testing system for vehicle electrical performance based on deep learning according to claim 2, characterized in that, The multi-parameter synchronous verification combination configuration includes: The intelligent decision-making model establishes a parameter correlation matrix, where the matrix elements represent the correlation strength between parameters; the model adjusts the parameter correlation matrix based on the fusion feature vector of the current product. Extract parameter pairs whose association strength exceeds the association threshold from the adjusted parameter association matrix; use a greedy clustering algorithm to divide the synchronous verification group, select the parameter pair with the highest association strength as the initial member, and add the parameter to the group if the association strength between a parameter and any member in the group exceeds the association threshold. Repeat this process until all associated parameters are divided to form a multi-parameter synchronous verification combination configuration.

5. The intelligent testing system for vehicle electrical performance based on deep learning according to claim 1, characterized in that, The instruction generation module includes: The parameter detection order is determined according to the parameter priority sequence. For each parameter to be detected, the corresponding detection channel number and product pin number are determined according to the parameter type, and the coordinates of the relay position to be closed are calculated. For synchronous verification parameter groups, the relay positions corresponding to all parameters within the group are determined simultaneously to generate synchronous closing instructions; the relay matrix automatic construction instructions include instruction header, relay row number, relay column number, action type, and delay time fields.

6. The intelligent testing system for vehicle electrical performance based on deep learning according to claim 1, characterized in that, The collaborative control module includes: Establishing a communication connection with the BUS box involves four stages: device discovery, connection request, authentication, and parameter negotiation. During the device discovery phase, a device query message is sent via broadcast, and the BUS box returns a device information response; during the connection request phase, a TCP connection request is sent to the BUS box; during the authentication phase, an authentication message containing the device identifier is sent; during the parameter negotiation phase, both parties negotiate communication parameters; after the connection is established, heartbeat data packets are sent periodically, and if the heartbeat times out for a preset number of consecutive times, the connection is determined to be disconnected and re-established.

7. The intelligent testing system for vehicle electrical performance based on deep learning according to claim 1, characterized in that, The parameter synchronization detection module includes: The synchronous verification control unit takes out a synchronous verification group from the multi-parameter synchronous verification combination configuration and determines the list of parameters that need to be synchronously verified and the corresponding detection channels within the synchronous verification group; The synchronous triggering subunit sends a synchronous trigger signal to all detection channels in the synchronous verification group simultaneously; the synchronous acquisition subunit drives the synchronous acquisition clock, and each detection channel synchronously samples at the rising edge of the synchronous acquisition clock to obtain sampling data with consistent time; each detection channel performs individual judgment on the sampled data. The synchronization judgment subunit performs a comprehensive judgment based on the configured synchronization judgment logic. If the synchronization judgment logic is a logical AND judgment, it checks whether all individual judgment results are normal. If the synchronization judgment logic is a weighted judgment, it performs a weighted calculation on the individual judgment results according to the importance weight of each channel. If the weighted sum exceeds the judgment threshold, the comprehensive judgment result is abnormal.

8. The intelligent testing system for vehicle electrical performance based on deep learning according to claim 1, characterized in that, The anomaly analysis module includes: Anomaly feature identification is performed on visual image data. The visual image of the current product is compared with the standard visual image. The difference is calculated using the image difference method. Areas with large differences are identified as potential anomaly areas. Anomaly identification is performed on the temperature distribution data. The temperature distribution of the current product is compared with that of a normal product. Locations where the temperature difference exceeds a preset temperature difference threshold are identified as temperature anomaly points. The system performs correlation analysis on visual anomalies, temperature anomalies, and electrical performance anomalies. It queries the historical fault analysis database to find correlation patterns and performs pattern matching between the current product's anomaly combinations and historical fault cases. It then uses the Euclidean distance similarity calculation method to select a preset number of historical fault cases as reference cases and extracts the root cause of the fault from these reference cases.

9. The intelligent testing system for vehicle electrical performance based on deep learning according to claim 1, characterized in that, The anomaly analysis module also includes: Calculate the abnormal deviation of each abnormal parameter. For abnormal parameters that exceed the limit, if the parameter value exceeds the upper limit threshold, the abnormal deviation is the actual value of the parameter minus the upper limit threshold and then divided by the upper limit threshold. If the parameter value is lower than the lower limit threshold, the abnormal deviation is the lower limit threshold minus the actual value of the parameter and then divided by the lower limit threshold. Anomaly severity scores are calculated based on the degree of deviation and the safety level of the parameters. For synchronous verification groups where the results of multi-parameter synchronous verification are abnormal, the anomaly severity score is the sum of the severity scores of each abnormal parameter within the group. Based on the anomaly severity scores, product anomalies are classified into levels, and thresholds for minor and severe anomalies are set for grading.

10. The intelligent testing system for vehicle electrical performance based on deep learning according to claim 1, characterized in that, The model optimization feedback module includes: Construct failure state learning samples, extract visual and temperature features from the fused feature vector set as input features, and extract abnormal state information from the intelligent anomaly analysis report as output labels; Update the failure state feature library, insert the learning samples one by one into the failure state feature library, perform deduplication check before insertion, and if the similarity exceeds the deduplication threshold, it is determined to be a duplicate sample and will not be inserted. Optimize the parameter priority algorithm, calculate the priority effectiveness score by statistically analyzing the detection order and anomaly detection of each parameter, and adjust the weight parameters if the priority effectiveness score is lower than the effectiveness threshold. Optimize the multi-parameter synchronous verification strategy by updating the parameter correlation matrix based on the actual data from this test, and then re-divide the synchronous verification groups using the updated parameter correlation matrix.

Images