Automobile wire harness electrical property detection method

Abstract

The invention relates to the technical field of automobile wire harness detection, and discloses an automobile wire harness electrical property detection method. The method comprises the following steps: acquiring a plurality of electrical parameter data points of the automobile wire harness, and dividing the operation process of the wire harness into a stable time period and a change time period by analyzing the change trend of the electrical parameter data points; in a stable time period, feature fusion processing is carried out on the collected electrical parameter data points, association information among different parameters is integrated, and fusion feature representation capable of comprehensively reflecting the stable operation state of the wire harness is generated; in the change time period, the fusion feature representation generated in the stable time period serves as a reference basis, the abnormal possibility of the electrical parameter data points is calculated, and a risk feature vector containing abnormal information is output; constructing a defect propagation model based on the risk feature vector, and predicting a defect propagation path; and according to the prediction result of the defect propagation path, a self-adaptive detection threshold is generated in combination with the current operation state of the wire harness, and accurate detection of the electrical performance of the automobile wire harness is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automobile wire harness detection, in particular to an automobile wire harness electrical property detection method. BACKGROUND

[0002] As an important part of the automobile electrical system, the automobile wire harness is responsible for transmitting power and signals, and its electrical performance is directly related to the overall operation state and safety of the automobile. With the continuous improvement of the intelligence and electrification of automobiles, the structure of the wire harness is becoming increasingly complex, the number of wires and the types of connectors are increasing significantly, and the corresponding electrical parameter monitoring requirements are becoming more and more precise. Automobile wire harness electrical property detection mainly adopts offline sampling detection or online fixed-point monitoring. Offline sampling detection needs to test the wire harness after disassembling it from the automobile, which is not only cumbersome and time-consuming, but also cannot reflect the electrical property changes of the wire harness under actual working conditions in real time, and the sampling ratio is limited, which is difficult to cover all potential quality problems, and may lead to missed detection. Although online fixed-point monitoring can realize real-time detection, it only collects and judges single data of fixed electrical parameters (such as voltage and current), and lacks dynamic analysis of parameter change process. In the actual automobile operation process, the electrical parameters of the wire harness will show dynamic change characteristics due to factors such as working condition changes (such as starting, accelerating, and braking), environmental temperature fluctuations, and vibration impact, and there are stable operation periods and rapid change periods. The existing detection method cannot effectively distinguish between these two periods. In the stable period, only single parameter data is judged in isolation, without considering the correlation between different parameters, which makes it impossible to fully capture the overall electrical property characteristics of the wire harness. In the change period, due to the lack of reference based on historical stable characteristics, it is difficult to accurately distinguish between normal fluctuations and abnormal changes, often resulting in misjudgment or lagging judgment. When a defect is detected in the wire harness, the existing technology cannot further analyze the possible propagation path of the defect, cannot predict the impact of the defect on other electrical components in advance, and can only be repaired after the defect has caused a fault, increasing the operation and maintenance cost and safety hazards of the automobile. At the same time, the detection threshold used in the existing detection method is a fixed value, which is set by manual experience and cannot be dynamically adjusted according to the actual running state of the wire harness and the parameter change trend. When facing the detection needs of wire harnesses of different vehicle models and different working conditions, the adaptability is poor, further reducing the reliability and applicability of the detection. SUMMARY

[0003] The purpose of the present application is to provide an automobile wire harness electrical property detection method to solve the problems raised in the background.

[0004] To achieve the above purpose, the present application provides an automobile wire harness electrical property detection method, which comprises: Obtaining a plurality of electrical parameter data points of an automobile wire harness; Dividing a stable time period and a change time period based on a change trend of the electrical parameter data points; In the stable time period, performing feature fusion processing on the electrical parameter data points to generate a fusion feature representation; In the change time period, calculating an abnormality possibility based on the fusion feature representation to output a risk feature vector; Constructing a defect propagation model according to the risk feature vector to predict a defect propagation path; Generating an adaptive detection threshold based on a prediction result to realize electrical detection.

[0005] Preferably, the obtaining of the plurality of electrical parameter data points of the automobile wire harness comprises: The electrical parameter data points comprise voltage values, current values and resistance values; The voltage values, the current values and the resistance values of the automobile wire harness are collected in real time by sensors to form time series data points.

[0006] Preferably, the dividing of the stable time period and the change time period based on the change trend of the electrical parameter data points comprises: A first-order difference value of the electrical parameter data point sequence is calculated, and a time point at which the first-order difference value is a positive number is selected as a growth time point; The growth time points are sorted according to the first-order difference values, a second-order difference value of the sorted sequence is calculated, and a growth time point corresponding to a maximum value of the second-order difference value is taken as a key time point; A time period before the key time point is the stable time period, and a time period after the key time point is the change time period.

[0007] Preferably, the feature fusion processing on the electrical parameter data points in the stable time period to generate the fusion feature representation comprises: Multi-dimensional features of the electrical parameter data points in the stable time period are extracted, including time dimension features and space dimension features; A dynamic weighting algorithm is used to fuse the multi-dimensional features to calculate correlation weights of different dimension features; The multi-dimensional features are weighted and averaged according to the correlation weights to obtain preliminary fusion features; Redundant information is removed through dimension reduction processing to generate a low-dimensional fusion feature representation.

[0008] Preferably, the calculating of the abnormality possibility based on the fusion feature representation in the change time period to output the risk feature vector comprises: The distance between the electrical parameter data points in the change time period and the fusion feature representation of the stable time period is compared; An abnormality probability of each electrical parameter data point is calculated according to the distance value; Aggregate the anomaly probabilities of all electrical parameter data points to form a risk feature vector.

[0009] Preferably, the constructing a defect propagation model according to the risk feature vector and predicting a defect propagation path comprises: Mapping the risk feature vector to a three-dimensional spatial coordinate of the automobile wiring harness to generate a preliminary defect concentration distribution; Dynamically correcting the preliminary defect concentration distribution based on historical defect data to obtain a spatiotemporal defect concentration distribution; Constructing a graph model according to the topological structure of the automobile wiring harness, with nodes representing key positions of the wiring harness and edges representing connection relationships; Using an attention mechanism to capture defect propagation dependency relationships between nodes and output an initial prediction of defect propagation; Combining a time series prediction model to generate a defect propagation path and diffusion trend.

[0010] Preferably, the generating an adaptive detection threshold based on the prediction result comprises: Training a normal operating condition generator to simulate normal operation data of the automobile wiring harness; Constructing a real-time discriminator to compare the prediction result with the generator data and learn the normal and abnormal boundaries; Dividing the discrimination result through a clustering algorithm to generate a multi-level adaptive detection threshold.

[0011] Preferably, the implementing electrical property detection further comprises: Dividing the detection stage according to the adaptive detection threshold, including a first stage, a second stage, and a third stage; Applying different detection strategies to electrical parameter data points in different detection stages.

[0012] Preferably, the dividing the detection stage according to the adaptive detection threshold comprises: Setting the layer with the highest node density in the defect propagation model as the first stage; Setting the layer with the most child nodes after the first stage as the second stage; Setting the layers from after the second stage to the bottom layer as the third stage.

[0013] Preferably, the applying different detection strategies to electrical parameter data points in different detection stages comprises: Using a first type of detection strategy for electrical parameter data points in the first stage; Using a second type of detection strategy for electrical parameter data points in the second stage; Using a third type of detection strategy for electrical parameter data points in the third stage.

[0014] Compared with the prior art, the present application has the following beneficial effects: In the time period division link, the method can accurately distinguish the stable time period and the change time period based on the change trend of the electrical parameter data points, breaking the traditional detection method of "one size fits all" processing mode for parameter change process. This time period processing mode makes the subsequent detection analysis more in line with the actual running characteristics of the wire harness. In the stable time period, the information of multiple electrical parameter data points is integrated through feature fusion processing, which is no longer limited to isolated judgment of a single parameter, can fully excavate the internal correlation between different parameters, and the generated fusion feature representation more comprehensively and accurately reflects the overall electrical characteristics of the wire harness in the stable running state, avoiding the feature omission problem caused by single parameter analysis, and providing a more reliable reference benchmark for subsequent abnormal judgment.

[0015] In the change time period, the abnormal possibility is calculated based on the fusion feature representation generated in the stable time period, which can effectively distinguish the normal fluctuation and abnormal change of the parameter. Since the fusion feature representation covers the multi-parameter correlation information of the wire harness in the stable running state, the degree of parameter deviation from the normal range in the change period can be more accurately identified by referring to it, and the output risk feature vector can clearly present the abnormal information, avoiding the misjudgment or lag judgment problem caused by the lack of reference benchmark in the traditional detection method, so that the detection personnel can timely master the abnormal situation of the wire harness. Based on the risk feature vector, a defect propagation model is constructed, which can further analyze the propagation rule of defects in the automobile electrical system and predict the defect propagation path. This process is no longer limited to simply detecting the existence of defects, but further explores the possible influence of defects on other electrical components, helping detection personnel to predict the fault development trend in advance, so as to take targeted preventive measures to avoid the spread of defects leading to more serious faults, and reduce the cost investment and safety hazards in the process of automobile operation and maintenance. According to the prediction result, an adaptive detection threshold is generated, which can make the detection threshold match the actual running state of the wire harness and the defect propagation situation. The traditional detection method uses a fixed threshold, which cannot adapt to the detection needs of wire harnesses under different vehicle types and different working conditions, while the adaptive detection threshold can be dynamically adjusted according to the current running characteristics and potential risks of the wire harness, ensuring the detection accuracy while improving the adaptability of the detection method, so that the detection method can be applied to different types of automobile wire harnesses and different running conditions, expanding the application range of the detection method. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The working principle diagram of the automobile wire harness electrical detection method described in the application; Figure 2 The working flowchart of the automobile wire harness electrical parameter data point acquisition method; Figure 3A workflow diagram for feature fusion processing of electrical parameter data points in a stable time period of an automobile wiring harness; Figure 4 A workflow diagram for abnormal possibility calculation and risk feature vector output in a change time period of an automobile wiring harness. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0018] Please refer to Figure 1 The present application provides an electrical detection method for an automobile wiring harness, which comprises: Obtaining electrical parameter data points such as voltage value, current value and resistance value in the running process of the automobile wiring harness to form a time series data set. Based on the change trend characteristics of the data points, the entire detection period is divided into two characteristic intervals of a stable time period and a change time period. In the stable time period, multi-dimensional feature fusion technology is used for feature extraction and dimension reduction processing of the electrical parameter data to generate representative fusion feature representation. In the change time period, the deviation degree of the current data point from the stable state feature is calculated to quantify the abnormal risk and construct a risk feature vector. A dynamic model of defect propagation is established based on the risk feature vector to predict the diffusion path of defects in the wiring harness network. Finally, the detection threshold is dynamically adjusted according to the prediction result to realize an adaptive electrical detection process.

[0019] Embodiment 1: refer to Figure 2 In the electrical detection process of the automobile wiring harness, the collection of electrical parameter data is completed by a distributed sensor network. The system deploys special sensor groups at key connection points of the engine compartment main wiring harness, the instrument panel branch wiring harness and the chassis wiring harness. Each sensor group contains three measurement modules: a Hall effect current sensor is attached to the outer insulating layer of the wire, a non-contact voltage probe is connected to the terminal, and a four-wire resistance measurement module is connected to the wiring harness connector. All sensors synchronously collect data at a period of 100 milliseconds, and transmit the data to the central processing unit through shielded twisted pair lines. The voltage measurement range covers the 0-48V working interval of the vehicle electrical system, the current measurement covers the 0-20A load range, and the resistance measurement is aimed at the 0-100Ω wire impedance range. When collecting data, millisecond-level time stamps and position codes are attached to form a three-tuple sequence containing time dimension, space dimension and parameter dimension.

[0020] The first order difference values of adjacent data points are calculated by using the sliding window method. A dynamic threshold mechanism is set: when the difference absolute value exceeds 3 times the sensor accuracy value, the time point is marked as an effective change point. For example, the voltage sensor accuracy is ±0.1V, and the voltage sequence change threshold is set to 0.3V. The effective change points of the three parameter sequences are merged according to the time axis to generate a unified key point candidate sequence. The second order difference calculation is performed on the sequence, and the local extreme value detection algorithm is used to locate the change rate sudden increase point. In specific implementation, the system scans the second order difference sequence, and when the difference values of three consecutive points meet the "low-high-low" distribution and the center point value exceeds 200% of the adjacent point, it is determined that the time is a key time. With the time as the dividing point, a 60-second continuous period is defined as a stable time period, and a 30-second period is defined as a change time period. The time period boundary adopts a rolling update strategy, and each time a new data packet arrives, the difference calculation is performed again, and when a new key time is detected and the time offset exceeds 5 seconds, the time period division result is automatically updated.

[0021] In the engine cold start test scene, the system monitors the battery voltage sequence fluctuating around the 12.3V reference value. When the ignition signal triggers, the voltage value drops to 11.8V within 300ms, and the change amount 0.5V exceeds the threshold 0.3V, which is marked as an effective change point. At the same time, the current value jumps from 0.2A to 8.5A, with a change amount of 8.3A exceeding the threshold of 0.15A; the resistance value remains unchanged at 0.05Ω. The system merges the voltage and current change points, and calculates the second order difference of the merged sequence. At 310ms after the ignition signal, the maximum difference peak is detected, and the time is determined as the key time. The system divides accordingly: 60 seconds before ignition as a stable time period, containing 600 sampling points; 30 seconds after ignition as a change time period, containing 300 sampling points. When the engine enters the idle state, the system detects a new current drop inflection point after 15 seconds, and updates the key time position and re-divides the time period.

[0022] Each sensor module has a built-in self-diagnosis function, which automatically triggers a zero-point calibration program when it detects that the signal drift exceeds 20% of the calibration value. The data transmission process uses CRC check code, and the error data packet triggers an instant retransmission mechanism. The central processing unit performs logical verification on the received data, for example, when a certain point voltage value is zero and the current value is non-zero, it is automatically marked as abnormal data and starts the redundant sensor data replacement.

[0023] The time period division result is output as a time period identification matrix, containing start timestamp, end timestamp and time period type code. The matrix is stored in a ring buffer synchronously with the original data stream, for real-time calling by subsequent processing modules. The buffer adopts a double storage area design, and the new data is written into the first storage area at the same time, and the second storage area provides read-only access for the feature extraction module, realizing data processing pipeline operation.

[0024] In the brake system wire harness monitoring case, the system detects abnormal fluctuations in the brake light circuit resistance value. The original resistance sequence fluctuates around the 2.1Ω reference value, and when three consecutive sampling points show a positive change of more than 0.15Ω, the system marks this period as a potential change period. The resistance value continues to rise to 3.0Ω within the next 0.5 seconds, and the peak value of the second-order difference appears at the 0.3 second position, which is confirmed as the critical moment. The system divides the time period accordingly and marks the period data as a high-priority detection area. When the driver releases the brake pedal, the system detects the resistance value falling process after 0.2 seconds, generates a new critical moment and updates the time period division.

[0025] When no valid change point is detected for 10 consecutive sampling periods, the stable time period range is automatically expanded; when there are more than three critical moment updates within 1 second, the burst mode is started, and the analysis window is shortened to 50% of the original value. All time period division records are written into the event log, including timestamp, critical moment coordinate, time period length, and difference characteristic value for division, forming a complete time period division trajectory.

[0026] When the charging gun is connected, the system detects that the power battery voltage rises smoothly from 400V to 420V. The first-order difference value of the voltage sequence is continuously below the threshold, and the system maintains the stable time period state. When the charging enters the constant voltage stage, the current value linearly decreases from 150A to 50A, and the system identifies the critical moment at the current change rate turning point, accurately dividing the different electrical characteristic stages of the charging process. The time period division response time is maintained within 50 milliseconds throughout the process, synchronized with the data acquisition cycle.

[0027] Example 2: see Figure 3 In the stable time period of the automobile wire harness electrical detection, the system performs multi-dimensional feature fusion processing. This process extracts twelve feature dimensions from the original data stream, including six time domain features and six spatial domain features. The time domain feature calculation is based on a sliding time window, with a window width of 30 consecutive sampling points. For voltage parameter sequences, the time domain analysis module calculates the arithmetic mean value of the data within the window as the reference feature, while the standard deviation reflects the fluctuation degree. The skewness coefficient is used to quantify the asymmetry of the data distribution, and the kurtosis coefficient detects the presence of abnormal peaks. The zero-crossing rate counts the number of signals crossing the average value, and the autocorrelation coefficient measures the correlation between the sequence. These features are updated every 5 seconds, with an exponential weighting method for smooth transition.

[0028] The spatial feature extraction considers the physical layout of the line bundle network, the difference value calculation of adjacent measurement points reflects the difference between the readings of different position sensors at the same time, and the line impedance distribution. The regional mean ratio selects the voltage ratio of the main line bundle and the branch line bundle to identify potential voltage division abnormalities. The topology distance weighted sum calculates the attenuation degree of the readings of the remote sensors according to the node distance of the line bundle connection graph. The spatial autocorrelation coefficient analyzes the parameter synchronization of different branches in the same electrical circuit. The spatial feature update period is synchronized with the time domain feature to ensure the alignment of the space-time features.

[0029] The dynamic weight distribution adopts a feature correlation analysis method, and a 12x12 feature correlation matrix is constructed to calculate the Pearson correlation coefficient between each pair of features. The eigenvalue decomposition of the correlation matrix is performed, and the feature vectors corresponding to the first three principal components are selected as the weight reference. The initial weight of the time domain feature is set to 0.6, and the initial weight of the spatial feature is set to 0.4. The weight of each dimension is dynamically adjusted according to the principal component analysis result. During the adjustment process, for the feature group with a correlation greater than 0.8, the weight distribution of the secondary features in the group is reduced; for the features with strong independence, the weight coefficient is appropriately increased. The weight update adopts a gradual adjustment strategy, and the adjustment amplitude is not more than 20% of the value of the last period.

[0030] The weighted fusion process adopts a hierarchical processing architecture: the first layer internally weights and sums the six-dimensional time domain features to generate a time domain comprehensive index. The second layer internally weights and sums the six-dimensional spatial features to generate a spatial comprehensive index. The third layer combines the two comprehensive indexes according to the maximum weight to form the preliminary fusion features. During the fusion process, an outlier filtering mechanism is set, and when the feature value of a certain dimension deviates from the historical mean value by more than three times the standard deviation, the weight of that dimension is temporarily reduced to 50% of the reference value.

[0031] The twelve-dimensional preliminary fusion features are input into the dimension reduction module, with a perplexity of 15 and a learning rate of 200. During the algorithm running process, the early exaggeration factor is dynamically adjusted, and the initial stage is set to 4, which is linearly decreased to 1 during the iteration process. The dimension reduction result outputs three-dimensional spatial coordinates, preserving the topological relationship of the original features. The three-dimensional feature points after dimension reduction are subjected to density clustering, with a neighborhood radius of 0.3 and a minimum sample number of 5. The clustering result generates a number of feature clusters, and the center point coordinates and boundary range of each cluster are calculated. Finally, the three feature clusters with the highest density are selected as the fusion feature representation of the time period, and the center point coordinates, cluster point number and distribution radius are stored.

[0032] In the new energy vehicle charging condition monitoring case, the system detects the stable characteristics of the charging pile and vehicle connection stage. Time domain analysis shows that the voltage mean value remains at 398.5V, the standard deviation is 0.8V, the skewness is 0.15, the kurtosis is 2.3, the zero-crossing rate is 0.2 times / s, and the autocorrelation coefficient is 0.92. The spatial feature shows that the voltage ratio of the main wire harness and the charging port branch is 1.02, the maximum difference value of adjacent sensors is 0.5V, the topology weighted sum is 1.15, and the spatial autocorrelation coefficient is 0.88. Dynamic weight calculation gives the time domain dominant weight 0.65 and the spatial weight 0.35. The three-dimensional feature after dimension reduction forms two main clusters, with center point coordinates (0.12, -0.05, 0.08) and (-0.03, 0.11, -0.02), and cluster radii 0.15 and 0.12, respectively. The system records this state as the reference feature representation of the charging connection stage.

[0033] Each time a new fusion feature representation is generated, its Euclidean distance from the historical reference feature is calculated. When the distance value exceeds the warning threshold for three consecutive periods, the feature relearning process is started. This process expands the sliding window size to 60 sampling points, adds a secondary feature extraction link, and recalculates the weight of each dimension. If the feature after relearning still deviates from the historical reference, a system calibration prompt is generated, suggesting checking the sensor working state or wire harness physical connection.

[0034] In the vehicle vibration test scene, the system captures the feature change caused by the loose wire harness connector. The original stable state fusion feature center point is (0.05, 0.07, -0.03), and after the vibration starts, the feature point gradually drifts to (0.18, -0.12, 0.05). The system detects that the feature drift exceeds the threshold, and automatically triggers the vibration compensation mode. This mode enhances the weight allocation of the spatial feature, especially the weight of the adjacent sensor difference feature from 0.1 to 0.25, which is more sensitive to physical connection changes. The compensated fusion feature center point is corrected to (0.08, -0.05, 0.02), which more accurately reflects the actual electrical characteristics.

[0035] The reference feature representation is stored in the cache area for real-time comparison; the historical feature sequence is stored in the solid state disk, indexed by timestamp; the long-term feature model is archived to the cloud database, supporting cross-device feature sharing. Each time the feature is updated, the system generates a feature change log, recording the feature parameters, weight distribution and dimension reduction projection information before and after adjustment, forming a complete feature evolution trajectory.

[0036] When the ambient temperature drops to -20°C, the system detects that the wiring harness resistance feature changes significantly. The original fusion feature has a resistance-related dimension weight of 0.15, and the system automatically adjusts to 0.25, enhancing the attention to temperature-sensitive features. The reduced dimension feature space presents a new cluster distribution, which is marked by the system as a low-temperature mode feature representation. When the wiring harness temperature rises after the vehicle runs for 30 minutes, the features gradually return to the standard cluster, and the system reduces the resistance feature weight accordingly, returning to the normal temperature detection mode. The feature fusion system remains stable throughout the process, without false positives or missed detections.

[0037] Example 3: see Figure 4 During the change period, the system performs anomaly likelihood calculation and defect propagation model construction. The anomaly detection module first calculates the distance measure of the current data point and the stable time period fusion feature representation. An improved Mahalanobis distance calculation method is used, considering the distribution shape and anisotropy of the feature cluster. For a detection point in a three-dimensional feature space and the feature cluster center , the distance calculation introduces the covariance matrix of the cluster :

[0038] where: represents the three-dimensional coordinates of the current detection point, is the coordinate of the feature cluster center point, is the covariance matrix of all sample points in the stable time period for this feature cluster. This distance measure can reflect the deviation of the detection point relative to the overall distribution of the feature cluster. The system sets a dynamic distance threshold, with an initial value of 3 times the feature cluster radius, and dynamically adjusts according to historical detection results. When the distance value exceeds the threshold, the anomaly probability calculation process is triggered.

[0039] The anomaly probability conversion uses a two-parameter sigmoid function to map the distance value to the [0,1] interval. The center point of the conversion function is set at 1.2 times the distance threshold, and the slope parameter is adjusted adaptively according to the density of the feature cluster. For sparse distributed feature clusters, a gentler conversion curve is used; for dense feature clusters, a steep conversion curve is used. The anomaly probability calculation result of each data point is stored together with the timestamp and location code to form an anomaly event record.

[0040] The first level of aggregation is performed in the time dimension, with a sliding window width of 15 samples. The maximum, mean, and standard deviation of the anomaly probability within the window are calculated. The second level of aggregation is performed in the spatial dimension, based on the topology of the harness. The anomaly features of neighboring detection points are weighted and averaged, with the weights proportional to the electrical coupling strength between nodes. The third level of aggregation extracts trend features by calculating the first and second order differences of the anomaly index over three consecutive time windows, capturing the dynamics of risk evolution. The final risk feature vector contains 15 dimensions, with each dimension normalized to eliminate the dimension effect.

[0041] The automotive harness system is abstracted as a directed graph structure, with nodes representing key locations such as electrical connection points, distributors, and connectors, and edges representing wire connection relationships. Node features include three parts: basic attributes (wire diameter, material, rated current), real-time state (current risk feature vector), and historical data (past fault records). Edge features include connection length, bending angle, and insulation material type. The graph structure is represented by an adjacency matrix, with matrix element values reflecting the electrical connection strength between nodes.

[0042] Each graph attention layer contains 8 independent attention heads, each calculating the propagation weight between nodes. The attention weight calculation introduces a spatial decay factor, considering the influence of line length on defect propagation.

[0043] For nodes and , the calculation of attention coefficient considers three factors: basic connection strength , risk feature similarity , and spatial decay factor , where represents the physical distance between nodes, is the decay coefficient. The model aggregates neighborhood information through three layers of graph attention networks, and finally each node obtains a risk representation containing global context.

[0044] The input sequence contains the risk feature vector of the last 10 time steps, and the temporal pattern is extracted through the causal convolution layer. The network contains four dilated convolution blocks with dilation coefficients of 1, 2, 4, and 8, respectively, with a receptive field covering the entire input sequence. Each convolution block is followed by a gated activation unit to control the information transmission strength. The output layer predicts the defect concentration distribution of the next 5 time steps, including the expected defect value and propagation direction probability of each node.

[0045] In the case of high-voltage wiring harness monitoring for new energy vehicles, the system detected a sudden increase in the anomaly probability of a node near the charging interface. Distance calculation showed that the distance of this point from the stable feature cluster reached 4.2 times the cluster radius, triggering a high-risk warning. The risk feature vector showed that the average anomaly probability of this area was 0.82, the spatial aggregation degree was 0.75, and the trend slope was 0.15. The graph attention network analysis pointed out that the anomaly was most likely to propagate along the high-voltage bus towards the battery management system, and predicted that the defect concentration of the battery end node would reach the warning value after 3 time steps. The system generated a targeted detection scheme accordingly, increasing the sampling frequency of the relevant lines to 200Hz.

[0046] After each detection period, the actual observation results were compared with the predicted values, and the loss function was calculated. When the prediction error of three consecutive periods exceeded the threshold, the model parameters were fine-tuned. The fine-tuning process used small batch gradient descent with a learning rate set to one-tenth of the initial value to prevent overfitting. Model version management used incremental storage, retaining the parameters of the last five versions, and supported fast rollback.

[0047] When the ignition coil works, it produces strong electromagnetic pulses, causing abnormal fluctuations in multiple sensor readings. The risk feature vector accurately identifies the spatio-temporal characteristics of electromagnetic interference, distinguishing it from the abnormal patterns caused by real defects. The graph attention network analyzes the discontinuity of the propagation path to determine that the anomaly is external interference rather than an internal defect of the wiring harness, avoiding false alarms. The system automatically activates the anti-interference mode, temporarily adjusts the feature weight distribution, and reduces the sensitivity to transient mutations.

[0048] The wiring harness network is projected onto a three-dimensional space according to the actual layout, with node color depth representing defect concentration and arrow direction representing the predicted propagation path. The visualization interface supports time axis dragging, showing the dynamic process of defect diffusion. Engineers can view the detailed risk features and prediction basis of any node through interactive operation, assisting in diagnostic decision-making.

[0049] The system ran continuously for 300 hours in durability testing, processing over 2 million detection data. The model parameters converged stably, and the prediction accuracy remained stable. When a new type of anomaly pattern was detected, the system adjusted the parameters within 10 cycles through online learning, adapting to the new working state. Throughout the process, resource consumption remained at a reasonable level, with CPU occupancy not exceeding 25% and memory usage stable within 4GB.

[0050] Example 4: The adaptive detection threshold generation adopts a generative adversarial network framework, containing a data generation module and a real-time discrimination module. The data generation module adopts a long short-term memory autoencoder structure, and the encoder part contains three layers of LSTM units, with 128 neurons in each layer, and the input window width is set to 30 time steps. The decoder part uses two layers of LSTM, and the data distribution of normal working conditions is constrained by the reconstruction loss. The training data set contains 200 hours of fault-free working condition records, covering 12 standard working modes such as vehicle starting, driving, and charging. The real-time discrimination module adopts a one-dimensional convolutional neural network architecture, containing four convolutional blocks: the first block has 32 5x1 convolutional kernels, the second block has 64 3x1 convolutional kernels, the third block has 128 3x1 convolutional kernels, and the fourth block has 256 1x1 convolutional kernels. Each convolutional layer is followed by a max-pooling and batch normalization operation, and the final output layer uses a Sigmoid activation function to generate an abnormal score.

[0051] In the initial stage, only pure normal data is used to train the generator, minimizing the reconstruction mean square error. In the second stage, 5% of synthetic abnormal data is injected into the training data, including pulse interference, gradual drift, and step mutation. In the third stage, the proportion of abnormalities is increased to 15%, and real fault case data is added. The discriminator training is alternated with the generator, and the discrimination accuracy is evaluated after each training round. When the validation set accuracy exceeds 95% for three consecutive times, it enters the next stage.

[0052] When the input window data passes through the fourth convolutional block, a 256-dimensional feature vector is extracted as the clustering input. The improved K-means++ algorithm is used to initialize the cluster centers, and three cluster clusters are set to correspond to the three-level detection thresholds. The clustering process introduces a feature weight mechanism, giving higher weights to feature dimensions with strong discrimination ability. After each clustering iteration, the mean of the discriminator output of the samples within the cluster is calculated, and the cluster center position is adjusted to maximize the difference between clusters. The final generated three-level threshold features are shown in Table 1.

[0053] Table 1: Adaptive detection threshold feature table.

[0054] Threshold level Feature density range Discriminator output interval Typical triggering scenarios Primary threshold High density area (0.8-1.0) 0.05-0.15 Sensor noise / transient interference Secondary threshold Medium density area (0.5-0.8) 0.15-0.35 Connector poor contact / local insulation breakdown Tertiary threshold Low density area (0.0-0.5) 0.35-0.95 Wire harness short circuit / open circuit / severe aging In the case of high-voltage wiring harness for new energy vehicles, the system identifies the battery management system master node as a defect-prone area, dividing it into the first stage detection area, which includes 1 master node and 3 direct child nodes. The second stage selects two main paths from the master node to the motor controller, covering 12 intermediate nodes. The third stage includes the remaining 46 terminal nodes distributed throughout the vehicle body. The detection parameter configuration for each stage is as follows: the first stage uses three redundant sensors for synchronous acquisition with a sampling rate of 200Hz, and data verification includes extreme value checking, slope consistency, and spatial correlation verification; the second stage uses primary and backup sensors with a sampling rate of 100Hz, implementing trend prediction and propagation pattern matching; the third stage uses a single sensor with a sampling rate of 20Hz, performing threshold monitoring and event-triggered acquisition.

[0055] In the brake system wiring harness monitoring, the system detects an abnormality in the right rear wheel speed sensor line. The discriminator output value fluctuates between 0.28-0.32, triggering a secondary threshold alarm. The system automatically promotes the relevant line node to the second stage detection mode, increasing the sampling rate from 20Hz to 100Hz. The trend analysis module detects a continuous rise in resistance value at a rate of 0.5Ω / second, and when the discriminator output breaks through 0.35, the system upgrades the node to the first stage detection, starting three-sensor redundancy verification. The diagnosis confirms that the oxidation of the connector pin causes increased contact resistance, and after replacing the component, the system returns to the basic monitoring mode within 15 seconds.

[0056] The threshold update mechanism includes two trigger conditions: periodic update, performed every 24 hours, based on retraining the discriminator and updating the cluster center using data collected on the same day; event-triggered update, initiated when a new abnormal pattern is detected, and incremental training is performed after collecting 50 consecutive abnormal samples. During the update process, a protection mechanism is set to retain the old threshold version in parallel when the new threshold deviates from the historical baseline by more than 20%, and the main version is switched after 48 hours of verification confirmation.

[0057] The system demonstrates adaptability in electric vehicle fast charging tests. When the charging power increases from 50kW to 150kW, the discriminator detects a new current ripple pattern. The initial misjudgment is an anomaly (output value 0.25), and the system initiates threshold updating after collecting 30 minutes of data. New cluster analysis identifies that this pattern belongs to the normal characteristics of high-power charging, and it is classified into the first-level threshold range (output value 0.12). After updating, the system correctly distinguishes between real insulation faults (output value 0.52) and high-power charging ripples, avoiding false positives.

[0058] The running log records the complete working state transition trajectory. In the test of a hybrid vehicle, the system records 27 stage switching events in a single day: 19 times from level one to level two, 6 times from level two to level three, and 2 times from level three to level one. The switching reason statistics show that 65% are triggered by instantaneous interference, 30% are caused by gradual fault development, and 5% are caused by sudden environmental conditions. The log data is stored as a sequence of triplets of timestamp-event type-trigger parameters, supporting later pattern analysis and system optimization.

[0059] This implementation is stable in the wet heat environment test. When the environmental humidity reaches 95%, the system detects intermittent abnormal signals at the door wire harness node. The discriminator output value fluctuates in the range of 0.18-0.25, continuously within the level one threshold range. The system maintains the third stage detection mode and only records the event without triggering the alarm. After three days, when the output value breaks through 0.35, the system upgrades to the second stage detection, confirming that the wire insulation layer is damp causing partial discharge. Throughout the process, the threshold system effectively distinguishes between environmental interference and real faults, avoiding premature triggering of maintenance alarms.

[0060] The first stage detection task occupies 40% of the processor resources, the second stage allocates 30%, the third stage allocates 20%, and the remaining 10% is used for system maintenance. When multiple nodes trigger upgrades at the same time, a priority sorting based on defect concentration is implemented. In the case of battery pack wire harness monitoring, the system once handled 4 level one alarm nodes and 7 level two alarm nodes at the same time, and the processor resources were dynamically adjusted as follows: each level one node occupied 15%, and each level two node occupied 5%, ensuring the detection needs of critical areas.

[0061] Example 5: The first-stage detection strategy implements a multi-level verification mechanism for core defect areas. In the monitoring of the main node of the new energy vehicle battery management system, the system deploys three groups of redundant current sensors, each group containing Hall effect sensors, Rogowski coils, and shunt detectors of three different principles. The data acquisition cycle is set to 5 milliseconds, and three independent channels record the charging and discharging current waveforms synchronously. The verification process is divided into three levels: the first level performs instantaneous value comparison, taking the median of the three sensor readings as the reference value, and setting a ±5% allowable deviation range; the second level analyzes the short-term trend, and the correlation test is performed on the slope of the continuous 10 sampling points, requiring the Pearson correlation coefficient to be ≥0.95; the third level verifies the law of conservation of energy, comparing the difference between the input and output current integral values, and setting the allowable error threshold to 0.3%. When the data of a certain sensor continuously exceeds the deviation range for three times, the system automatically marks it as a failed channel, switches to the backup link, and issues a calibration prompt. In the battery post loosening fault simulation test, the system detects that the difference between the readings of the three groups of sensors reaches 12% within 200 milliseconds, and the current integral deviation breaks through 0.8%, triggering a three-level linkage alarm. The second-stage detection strategy establishes a dynamic analysis model for the wire harness propagation path. In the case of monitoring the line from the vehicle body control module to the door actuator, the system installs double sensors to collect the voltage signals at the input and output ends of the door control unit. The monitoring window is set to a 100-millisecond cycle, and three key propagation features are extracted in each cycle: the signal delay time is calculated by the cross-correlation function, capturing the time lag from the controller command to the actuator response; the voltage attenuation rate measures the peak-to-peak value ratio of the output to the input; the high-frequency noise proportion is analyzed using Fourier transform to analyze the proportion of frequency components above 20 kHz. The feature update cycle is 50 milliseconds, and the system maintains a feature queue of the last 20 cycles for trend modeling. An autoregressive moving average model is established to predict the next cycle feature value, with a confidence interval of ±2 standard deviations from the predicted value. When the insulation layer of the door hinge wire harness is worn, the system detects that the voltage attenuation rate exceeds the upper limit of the prediction interval for 5 consecutive cycles, and the high-frequency noise proportion increases from 3% to 12%. The system immediately starts the path scanning mode, increases the sampling frequency to 5 kHz, and locks the wear point to the third wire position at the door seam turning point. The third-stage detection strategy uses an intelligent event trigger mechanism. The wire harness monitoring case of the vehicle entertainment system covers 46 terminal nodes, each node is configured with a single sensor to scan the voltage parameters at a basic frequency of 5 Hz. The system establishes two levels of trigger conditions: the regular trigger threshold is set to ±15% of the rated voltage, and the instantaneous jump threshold is set to a change rate of adjacent sampling points exceeding 50%. When the cigarette lighter socket is connected to a high-power device, the system detects that the 12V voltage drops to 9.6V, with an instantaneous change rate of 56%. This event activates the 10-millisecond high-speed sampling mode, recording 200 cycles continuously.The feature extraction module only calculates the mean and standard deviation, and the mean data is input into the pre-trained decision tree classifier. The model contains 5 branches: if the mean is less than 9V and the standard deviation exceeds 0.5V, it is classified as a short circuit risk; if the mean fluctuates in the 9-10V interval, it is classified as an overload state. In this incident, the system identified that the mean was stable at 9.8V and the standard deviation was 0.2V, and determined that it was a normal load change. After the device was removed, the system returned to the basic monitoring mode within 15 seconds, and no false alarms were triggered during the entire process.

[0062] During the monitoring of the electric power steering system, when the basic scanning detects a ±0.5V periodic fluctuation in the torque sensor signal, the system starts the event trigger mode. Preliminary analysis shows that the standard deviation abnormally increases to 0.38V, and the classifier output boundary confidence value. The system automatically applies for a second-level review, and transmits 10 seconds of data to the regional processing unit. The second-stage analysis identifies that the fluctuation frequency is associated with the vehicle speed by 85%, and the signal delay conforms to the normal propagation curve, but it is found that the noise component abnormally increases by 1.8dB. This result triggers the strategy linkage request to the first-level detection, and the core region starts three-sensor synchronous acquisition. Redundancy verification confirms that the impedance of the wire harness ground terminal increases from 0.1Ω to 0.8Ω, and the fault point is locked as the steering column connector oxidation. The system automatically generates a diagnostic report and marks the fault location coordinates, and permanently upgrades the related nodes to the first-level monitoring strategy, allocating exclusive communication bandwidth and computing resources.

[0063] During the 30Hz sine sweep test implemented by the electromagnetic compatibility laboratory, the third stage detects an increase in the frequency of the reverse radar line voltage fluctuations. The basic scanning mode records 8 ±10% fluctuations within 20 seconds, and the regular classifier output confidence is insufficient. The system applies for a second-level analysis to confirm that the fluctuation mode is synchronized with the vibration frequency, but the propagation characteristics conform to the normal path parameters. At this time, the first-level core region is triggered without exception, and the resource scheduling center automatically allocates an idle computing unit for in-depth verification. The three-level collaborative analysis confirms that the interference source is vibration-induced poor sensor contact, and recommends tightening the connector instead of replacing the wire harness. The entire decision-making process takes 800 milliseconds, saving unnecessary maintenance operations while maintaining the system's continuous monitoring capability.

[0064] Monitoring data from the cold region test site shows that low temperature causes periodic increases in the third stage false alarm rate. The system automatically establishes a temperature compensation model: when the ambient temperature is below -10°C, the regular trigger threshold of all terminal nodes is relaxed to ±18%, the instantaneous jump threshold is raised to 65%, and the classifier feature weight is tilted towards the mean parameter. At the same time, the self-checking period is shortened to 2 hours, and the sensor drift compensation function is enhanced. When the vehicle enters the warm workshop, the system detects that the connector temperature rises above 15°C, and automatically restores the standard parameter configuration. In the annual cycle data statistics, this dynamic adjustment mechanism maintains the winter false alarm rate at a fluctuation range of 3% comparable to that in summer.

[0065] A certain hybrid vehicle model experienced intermittent anomalies in heavy rain, with multiple nodes triggering alerts alternately in the third phase. The system prioritized resource allocation to handle the braking system nodes that directly endangered driving safety, while reducing the monitoring frequency of the motor cooling system to a minimum. An anomaly pattern propagation map was established, marking the wiper circuit as the initial source of disturbance. A two-level isolation strategy was implemented: first, the wiper control module was disconnected to eliminate interference, and second, a two-level deep detection was performed on key system nodes to confirm the integrity of the wiring harness. Under the premise of ensuring basic vehicle functions, the system identified the root cause of signal crosstalk caused by rainwater seeping into the combination switch, guiding targeted component replacement rather than comprehensive wiring harness repair.

[0066] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0067] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, alternatives, and variations can be made in the embodiments without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. A method for testing the electrical properties of automotive wiring harnesses, characterized in that, Including the following steps: Obtain multiple electrical parameter data points of the automotive wiring harness; The data points of electrical parameters are divided into stable time periods and changing time periods based on their changing trends; Within a stable time period, feature fusion processing is performed on electrical parameter data points to generate fused feature representations; Within the time period of change, the probability of anomalies is calculated based on the fused feature representation, and a risk feature vector is output. A defect propagation model is constructed based on risk feature vectors to predict defect propagation paths; An adaptive detection threshold is generated based on the prediction results to achieve electrical detection.

2. The method for testing the electrical properties of automotive wiring harnesses according to claim 1, characterized in that, The acquisition of multiple electrical parameter data points for the automotive wiring harness includes: Electrical parameter data points include voltage, current, and resistance values; The voltage, current and resistance values ​​of the vehicle's wiring harness are collected in real time by sensors to form time series data points.

3. The method for testing the electrical properties of automotive wiring harnesses according to claim 2, characterized in that, The division of stable and changing time periods based on the changing trends of electrical parameter data points includes: Calculate the first-order difference value of the electrical parameter data point sequence, and select the moments with positive first-order difference values ​​as growth moments; Sort the growth moments according to the first difference value, calculate the second difference value of the sorted sequence, and take the growth moment corresponding to the maximum value of the second difference value as the critical moment. The period before the critical moment is a stable period, while the period after the critical moment is a changing period.

4. The method for electrical testing of automotive wiring harnesses according to claim 3, characterized in that, The step of performing feature fusion processing on electrical parameter data points within a stable time period to generate a fused feature representation includes: Extract multi-dimensional features of electrical parameter data points within a stable time period, including temporal and spatial features; A dynamic weighted algorithm is used to fuse multi-dimensional features and calculate the correlation weights of features in different dimensions. The preliminary fused features are obtained by weighting the multi-dimensional features according to the relevance weights. Redundant information is removed by dimensionality reduction, generating a low-dimensional fusion feature representation.

5. The method for testing the electrical properties of automotive wiring harnesses according to claim 4, characterized in that, Within the changing time period, the probability of anomalies is calculated based on the fused feature representation, and a risk feature vector is output, including: Compare the distance between electrical parameter data points during the change period and the fused feature representations during the stable period; Calculate the probability of anomalies for each electrical parameter data point based on the distance value; The probability of anomalies in all electrical parameter data points is aggregated to form a risk feature vector.

6. The method for testing the electrical properties of automotive wiring harnesses according to claim 5, characterized in that, The step of constructing a defect propagation model based on risk feature vectors and predicting defect propagation paths includes: The risk feature vector is mapped to the three-dimensional spatial coordinates of the automotive wiring harness to generate a preliminary defect concentration distribution. The initial defect concentration distribution is dynamically corrected based on historical defect data to obtain the spatiotemporal defect concentration distribution. A graph model is constructed based on the automotive wiring harness topology, where nodes represent key locations in the wiring harness and edges represent connections. An attention mechanism is used to capture the dependency relationship of defect propagation between nodes, and an initial prediction of defect propagation is output. By combining time series prediction models, defect propagation paths and diffusion trends are generated.

7. The method for electrical testing of automotive wiring harnesses according to claim 6, characterized in that, The process of generating an adaptive detection threshold based on the prediction results includes: Train a normal operating condition generator to simulate normal operating data of automotive wiring harnesses; Build a real-time discriminator to compare prediction results with generator data and learn the boundaries between normal and abnormal; The results are segmented and judged using a clustering algorithm to generate multi-level adaptive detection thresholds.

8. The method for electrical testing of automotive wiring harnesses according to claim 7, characterized in that, The implementation of electrical detection further includes: The detection stages are divided according to the adaptive detection threshold, including the first stage, the second stage, and the third stage. Different detection strategies are applied to electrical parameter data points at different detection stages.

9. The method for electrical testing of automotive wiring harnesses according to claim 8, characterized in that, The step of dividing the detection stage according to the adaptive detection threshold includes: The layer with the highest node density in the defect propagation model is set as the first stage; The layer with the most child nodes after the first phase is designated as the second phase. The second stage is followed by the third stage at the bottom layer.

10. The method for electrical testing of automotive wiring harnesses according to claim 9, characterized in that, The application of different detection strategies to electrical parameter data points at different detection stages includes: The first type of detection strategy is used for the electrical parameter data points in the first stage; The second type of detection strategy is used for the electrical parameter data points in the second stage; A third type of detection strategy is adopted for the electrical parameter data points in the third stage.

Images