A tightening control method and device for a final assembly system

Abstract

The application discloses a tightening control method and device applicable to a general assembly system, and relates to the technical field of tightening tool state monitoring; in view of the problem that a fixed threshold value is difficult to accurately identify a deterioration state, tightening timing data and working condition information when a wireless tightening tool executes a target tightening task are acquired, a first feature extraction network and a second feature extraction network based on joint training of health state paired samples are used to respectively determine working condition reference features and full-time sequence features; the full-time sequence features are projected and separated along the working condition reference direction to obtain working condition correlation features, and state residual features are constructed according to direction information and amplitude information of the residual features; the deviation degree is determined based on a health reference state model, and a deterioration state is confirmed in combination with a continuous task judgment rule to output deterioration early warning information. Thus, the interference of working condition changes on deterioration judgment can be reduced, and the accuracy and stability of early deterioration monitoring of the wireless tightening tool can be improved.

Description

Technical Field

[0001] This application relates to the field of tightening tool condition monitoring technology in automobile assembly production lines, and in particular to a tightening control method and device that can be used in the assembly system. Background Technology

[0002] In automotive assembly lines, wireless tightening tools are typically installed at different workstations. These tools tighten bolts at target connection points based on the specific tightening task for each vehicle. During mixed-line production, the target torque, number of bolts, tightening duration, and load conditions vary across different vehicle models and workstations. This causes the real-time torque, motor current, motor speed, tightening angle, wireless transmission delay, and battery voltage collected by the wireless tightening tools to change depending on the tightening conditions.

[0003] Existing methods for detecting the condition of tightening tools typically determine the tool's condition directly based on the difference between operational data and fixed health characteristics or fixed thresholds. However, changes in normal operation caused by vehicle model, target torque, number of bolts, or workstation switching may overlap with changes in operation caused by early deterioration of the drive motor, torque sensor, battery management unit, or ratchet head. Fixed health characteristics or fixed thresholds are insufficient to distinguish between normal changes in operating conditions and changes in the condition of tool components, making the assessment of the deterioration condition of wireless tightening tools susceptible to interference from changes in operating conditions. Summary of the Invention

[0004] This application provides a method and device for monitoring the degradation of wireless tightening tools in a final assembly system, in order to reduce the interference of changes in tightening conditions on the judgment of the degradation status of wireless tightening tools.

[0005] Firstly, this application provides a method for monitoring the degradation of wireless tightening tools in a final assembly system. This method can be executed by a tightening control device. The tightening control device acquires tightening timing data generated when the wireless tightening tool performs a target tightening task, as well as the corresponding operating condition information. Based on the operating condition information, it determines operating condition baseline features through a first feature extraction network, and determines full-scale timing features based on the tightening timing data through a second feature extraction network jointly trained with the first feature extraction network based on health state paired samples. It determines the projection component of the full-scale timing features onto the operating condition baseline direction represented by the operating condition baseline features, uses the projection component as the operating condition associated feature, and obtains initial residual features based on the difference between the full-scale timing features and the operating condition associated features. It constructs state residual features based on the direction and amplitude information of the initial residual features. It determines the degree of deviation of the state residual features from the health baseline state model, determines the degradation state of the wireless tightening tool based on the degree of deviation, and outputs degradation warning information corresponding to the degradation state. The health baseline state model is established based on the state residual features obtained when the wireless tightening tool is in a healthy state.

[0006] In the above method, the tightening control device first acquires the tightening timing data and operating condition information corresponding to the target tightening task. Second, by using two feature extraction networks jointly trained with paired samples of healthy states, it determines the operating condition baseline features and full-scale timing features with comparable directional relationships. Then, it determines the operating condition correlation features of the full-scale timing features in the operating condition baseline direction and constructs state residual features using the directional and amplitude information of the remaining features. Third, it determines the degree of deviation of the state residual features from the healthy baseline state model. Finally, it confirms the deterioration state and outputs deterioration warning information based on multiple target tightening tasks continuously executed by the same wireless tightening tool. Because the state judgment simultaneously retains the residual direction changes and residual intensity changes that cannot be explained by the current operating condition baseline direction, it can reduce the interference of normal operating condition changes on the tool state judgment and avoid the loss of deterioration intensity information caused by normalizing only the residual direction.

[0007] In one possible design, the moment corresponding to the tightening start trigger signal is taken as the zero point of time. Real-time torque data, motor current data, motor speed data, tightening angle data, wireless transmission delay data, and battery voltage data are acquired. A unified relative time axis is established based on the sampling timestamps corresponding to each data point. Data from different sampling moments are mapped to this unified relative time axis to obtain the original time series matrix. The vehicle model code, target torque, number of bolts, and workstation number are acquired, and the task identifier of the target tightening task is used to bind each working condition information to the original time series matrix. Through this design, data from different acquisition sources can maintain a correspondence with the same target tightening task and its working condition information.

[0008] In one possible design, median filtering is applied to the original time series matrix. Based on the historical data range of the corresponding workstation and vehicle model, the data in each dimension of the filtered time series matrix are normalized. Vehicle model codes and workstation numbers are categorized, and target torque and bolt quantity are numerically normalized or range-coded. The results of these codes are then combined to obtain the working condition coding features. This design reduces the impact of local sampling disturbances and differences in the dimensions of different physical quantities on subsequent feature extraction, while preserving the numerical relationships between continuous working condition parameters.

[0009] In one possible design, multiple pairs of health state samples are acquired. Each pair of health state samples includes the working condition encoding features and normalized temporal feature matrix corresponding to the same target tightening task. Each pair of health state samples is processed by a first feature extraction network and a second feature extraction network, respectively. Based on the alignment constraints between the outputs of the two networks for the same sample and the working condition discrimination constraints between the outputs of networks corresponding to different working condition indices, the parameters of the two networks are jointly updated. After the jointly trained model is validated, the current working condition encoding features are input into the first feature extraction network to obtain the working condition baseline features. Through this design, the working condition baseline features and the full temporal features not only have the same feature dimension, but also form a comparable directional relationship under the pair of health state samples, while avoiding the convergence of features corresponding to different working condition indices to the same output.

[0010] In one possible design, it is determined whether the current operating condition index is within the trained encoding range of the first feature extraction network. If it is within the trained encoding range, the operating condition benchmark feature corresponding to the current operating condition index is added or updated in the operating condition benchmark feature library. If it is outside the trained encoding range, the health status pair samples corresponding to the current operating condition index are first obtained, and the two feature extraction networks are updated and verified before the corresponding operating condition benchmark feature is generated and stored. This design avoids the direct use of network outputs that have not been trained with samples from new vehicle models, new workstations, or new parameter ranges as operating condition benchmark features.

[0011] In one possible design, multiple temporal convolutional processing units with different time receptive ranges are used to extract temporal changes at different time scales. The temporal responses of each feature channel are summarized along the time dimension, and the channel response coefficients are determined based on the summary results, and the outputs of the corresponding feature channels are adjusted. Residual fusion, temporal dimension aggregation, and feature mapping are performed on the adjusted features at different time scales to obtain the full temporal features. This design can simultaneously preserve both local changes during the tightening process and changes spanning a longer time range.

[0012] In one possible design, the eigenvalues ​​at the same position in the full time-series characteristics and the operating condition reference characteristics are multiplied and summed to obtain the directional coincidence. The scale of the operating condition reference direction is determined based on the sum of squares of the eigenvalues ​​in the operating condition reference characteristics and the positive stability quantity. The projection coefficient is determined based on the ratio between the directional coincidence and the scale. The operating condition reference characteristics are then proportionally adjusted using the projection coefficient to obtain the operating condition associated characteristics. The initial residual characteristics are obtained by subtracting the corresponding eigenvalues ​​in the operating condition associated characteristics from each eigenvalue in the full time-series characteristics. The state residual characteristics are obtained based on the direction and magnitude information of the initial residual characteristics. Through this design, the state residual characteristics simultaneously include the residual change direction and residual change magnitude.

[0013] In one possible design, multiple state residual features obtained during the normal execution of multiple target tightening tasks by the same wireless tightening tool are acquired to form a healthy state sample set. This healthy state sample set is then divided into a model fitting sample set and a threshold calibration sample set. A health probability distribution including multiple Gaussian components is established based on the model fitting sample set. The target Gaussian component to which the current state residual feature most likely belongs is determined, and the Mahalanobis distance is determined based on the mean deviation of the current state residual feature relative to the target Gaussian component, the health fluctuation scale of each dimension, and the correlation between different dimensions. A preset deviation threshold is determined based on the statistical distribution of the Mahalanobis distance corresponding to the threshold calibration sample set, and the deterioration state is confirmed based on the Mahalanobis distance of multiple consecutive target tightening tasks by the same wireless tightening tool. This design allows model parameter fitting and threshold calibration to use mutually distinguishable healthy samples, reducing the possibility of a single disturbance triggering a deterioration state.

[0014] In one possible design, the state residual features are input into a fault component classifier trained using historical fault samples to obtain matching probabilities for multiple candidate deteriorated components. Based on these matching probabilities, candidate deteriorated components are determined, and deterioration warning information is generated and sent to the workstation display device and the production line maintenance system. The raw data, intermediate features, deviation degree, deterioration state, and deterioration warning information are associated and stored according to task identifiers. This design ensures that the warning result corresponds to the data processing process that generated it.

[0015] Secondly, this application provides a wireless tightening tool deterioration monitoring device for a final assembly system. The device includes a data acquisition module, a feature determination module, a feature separation module, a deterioration judgment module, and an early warning output module. These modules interact to implement the wireless tightening tool deterioration monitoring method described in the first aspect.

[0016] The technical effects that can be achieved in the second aspect mentioned above are the same as those that can be achieved in the corresponding design in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0018] Figure 1 This is a schematic diagram illustrating an application scenario of an automobile assembly system provided in an embodiment of this application.

[0019] Figure 2 This is a schematic diagram of a wireless tightening tool degradation monitoring software architecture provided in an embodiment of this application.

[0020] Figure 3A flowchart illustrating a method for monitoring the degradation of a wireless tightening tool, as provided in an embodiment of this application.

[0021] Figure 4 This is a schematic diagram of a unified time axis mapping of multi-source tightening data provided in an embodiment of this application.

[0022] Figure 5 This is a schematic diagram of joint training of a dual-branch feature space provided in an embodiment of this application.

[0023] Figure 6 This is a schematic diagram of a working condition benchmark feature extraction and feature library management provided for an embodiment of this application.

[0024] Figure 7 This is a schematic diagram of a multi-scale full-time feature extraction method provided in an embodiment of this application.

[0025] Figure 8 This is a schematic diagram of the projection separation of working condition related features provided in an embodiment of this application.

[0026] Figure 9 This is a schematic diagram illustrating a health baseline status assessment and degradation early warning system provided in an embodiment of this application.

[0027] Figure 10 This is a schematic diagram of the structure of a wireless tightening tool deterioration monitoring device provided in an embodiment of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the embodiments of this application will be further described in detail below with reference to the accompanying drawings. The following embodiments are used to illustrate this application, and not to limit this application.

[0029] In the embodiments of this application, "at least one" means one or more, and "more than one" means two or more. "And / or" is used to describe the association relationship of associated objects, and can indicate the existence of one associated object alone, the existence of multiple associated objects at the same time, or any combination of multiple associated objects.

[0030] The target tightening task in this embodiment can correspond to one tightening cycle performed by a wireless tightening tool on a target connection position. The number of bolts can represent the total number of target tightening cycles that the current vehicle needs to perform at the corresponding workstation. For the implementation method of merging multiple tightening cycles at the same workstation into one task, the data can be divided into multiple sub-sequences according to each tightening start trigger signal, and then the following feature processing process can be performed separately, or the features of each sub-sequence can be aggregated according to the execution order.

[0031] The tightening control device in this application refers to a workstation control computer, edge computing device, or server capable of acquiring data from wireless tightening tools and production line systems and executing the following degradation monitoring method. The wireless tightening tool degradation monitoring device can be deployed on the tightening control device in the form of software modules, hardware circuits, or a combination of software and hardware.

[0032] Figure 1 This is a schematic diagram illustrating an application scenario of an automobile assembly system provided in an embodiment of this application. (Reference) Figure 1 The automotive assembly system may include wireless tightening tools, workstation data acquisition and control equipment, production line data service equipment, and operation and maintenance management terminals.

[0033] Wireless tightening tools are installed at the tightening station on the automotive assembly line and may include a wireless torque wrench body, a torque sensor, an angle sensor, a drive motor, a battery management unit, a wireless communication module, and a ratchet actuator. The torque sensor collects real-time torque data, the angle sensor collects tightening angle data, the drive motor controller provides motor current and speed data, the battery management unit provides battery voltage data, and the wireless communication module provides transmission timestamps.

[0034] Workstation data acquisition and control equipment may include a fixed barcode scanner, photoelectric switch, wireless receiver, workstation industrial control computer, and workstation display device. The fixed barcode scanner is used to collect vehicle model barcodes or vehicle identification information, the wireless receiver is used to receive operating data sent by the wireless tightening tool, and the workstation industrial control computer can act as a tightening control device to perform data alignment, feature extraction, feature separation, and status judgment processes.

[0035] Production line data service equipment may include a Tightening Data Management System (DMS) server and a Manufacturing Execution System (MES) server. The DMS server stores historical tightening data, model parameters, operating condition baseline feature library, health status samples, and degradation warning information. The MES server provides the workstation control computer with vehicle model, target torque, bolt quantity, and workstation task configuration. The maintenance management terminal centrally displays the degradation status and degradation warning information of the wireless tightening tools.

[0036] Figure 2 This is a schematic diagram of a wireless tightening tool degradation monitoring software architecture provided in an embodiment of this application. (Reference) Figure 2The software architecture can include an edge layer, a service layer, and an application layer. The edge layer can be deployed on the industrial control computer at the workstation, including a multi-source data acquisition and adaptation unit, a time series preprocessing unit, a dual-branch feature extraction unit, and a feature separation and degradation judgment unit; the service layer can be deployed on the DMS data server and the MES system server, including a model training and update unit, a working condition benchmark feature library management unit, a health model management unit, and a data lifecycle management unit; the application layer can be deployed on the workstation display device and the operation and maintenance management terminal, including a degradation early warning display unit, a candidate component location unit, and a data query and traceability unit.

[0037] Figure 3 A flowchart illustrating a method for monitoring the degradation of a wireless tightening tool, as provided in an embodiment of this application. (Reference) Figure 3 The method includes: acquiring tightening timing data and operating condition information corresponding to the target tightening task; preprocessing the tightening timing data and operating condition information; determining the operating condition baseline features through a first feature extraction network; determining the full timing features through a second feature extraction network; projecting and separating to construct state residual features; determining the degree of deviation of the state residual features from the health baseline state model; confirming the deterioration state based on multiple target tightening tasks continuously executed by the same wireless tightening tool, and outputting deterioration warning information.

[0038] The baseline features characterize the current operating conditions within the aligned feature space formed by joint training, representing the baseline orientation. The full-time features characterize the multi-dimensional temporal changes of the current target tightening task. The operating condition correlation features are the projected components distributed along the baseline orientation in the full-time features. The state residual features consist of the residual orientation features and residual strength after separating the operating condition correlation features. They may include changes in tool component state, operational changes not explained by the current baseline orientation, and sampling disturbances; therefore, they need to be combined with the health baseline state model and continuous judgment rules to determine the degradation state.

[0039] During data acquisition, the workstation control computer uses the moment corresponding to the tightening start trigger signal of the wireless tightening tool as the zero point of the current target tightening task. Starting from the zero point, it acquires real-time torque data, motor current data, motor speed data, tightening angle data, wireless transmission delay data, and battery voltage data. The wireless transmission delay data can be determined based on the difference between the transmission timestamp corresponding to the wireless tightening tool sending data and the reception timestamp corresponding to the wireless receiver receiving data at the workstation.

[0040] Figure 4This is a schematic diagram illustrating a unified timeline mapping for multi-source tightening data provided in this application embodiment. The workstation industrial control computer establishes a unified relative timeline according to a preset sampling time interval, using the zero point of time as the first unified sampling moment. Subsequently, each preset sampling time interval forms the next unified sampling moment, until the end time of the current target tightening task is covered. The preset sampling time interval can be determined based on the highest effective change frequency among the six types of data, the sampling period provided by each acquisition unit, and the processing capability of the workstation industrial control computer, and remains consistent during model training and real-time inference phases.

[0041] For data with a sampling frequency higher than the unified sampling frequency, the sampled value closest to the unified sampling time can be selected, or the average, median, maximum or change of multiple original sampled values ​​between two adjacent unified sampling times can be calculated. For data with a sampling frequency lower than the unified sampling frequency, the nearest value preservation or linear interpolation method can be used to supplement it. For wireless transmission delay generated by data packets, the delay value corresponding to a data packet can be filled into the unified sampling interval covered by the data packet.

[0042] For example, if a tightening cycle lasts two seconds and the uniform sampling interval is set to ten milliseconds, then the uniform relative time axis starts from time zero, and a sampling position is set every ten milliseconds until the two seconds are over. If torque and current data are collected every millisecond, the median value of the ten samples within each ten millisecond interval can be placed in the corresponding sampling position; if battery voltage is collected every 100 milliseconds, the most recent value can be used between two battery voltage samples; if wireless transmission latency is generated as a data packet every 50 milliseconds, the latency value of the data packet can be filled into the corresponding five uniform sampling positions.

[0043] After time mapping, the industrial control computer at the workstation forms a multi-row, six-column original time series matrix according to a unified sampling time. Each row corresponds to a unified sampling time, and each column corresponds to a fixed type of operating data. The six columns are, in order, real-time torque, motor current, motor speed, tightening angle, wireless transmission delay, and battery voltage. The units for each column can be N·m, amperes, revolutions per minute, degrees, milliseconds, and volts, in that order. When the sequence lengths of different tightening tasks are different, the variable-length sequence can be retained and aggregated in the time dimension at the end of the time series feature extraction network. Alternatively, data exceeding the preset length can be truncated, and the portion below the preset length can be supplemented to a preset value and a valid position marker can be generated.

[0044] The workstation control computer also acquires the vehicle model code, target torque, number of bolts, and workstation number. The vehicle model code and workstation number are category information, the target torque is a continuous numerical value, and the number of bolts is an ordered numerical value. The workstation control computer uses the task identifier of the target tightening task to associate and store the above working condition information with the corresponding original time series matrix to avoid data mismatch between adjacent vehicles or adjacent tightening cycles.

[0045] In filtering, the industrial control computer at the workstation can use a sliding window containing five consecutive sampling points. For the current sampling point at the center of the window, select the five values ​​of the current sampling point and the two sampling points before and after it, sort the five values ​​in order of size, and replace the current sampling point with the value in the middle position. If there are fewer than five sampling points at the beginning or end of the sequence, the boundary values ​​can be copied, the sequence can be expanded by mirroring, or the window can be reduced.

[0046] For example, in a torque sequence, five consecutive sampled values ​​are 29 N·m, 31 N·m, 120 N·m, 30 N·m, and 32 N·m, where 120 N·m may be due to transient disturbances. After sorting these five values, the median value is 31 N·m, so 31 N·m is used to replace the 120 N·m at the center of the window. This processing can suppress isolated spikes while preserving the overall trend of torque increase or decrease.

[0047] In the normalization process, the industrial control computer at each workstation determines the historical minimum and maximum values ​​for the current workstation and the current vehicle model for each column of data, and uses the difference between the two as the historical range of change for that column of data. To avoid the inability to perform scale conversion due to the historical maximum and historical minimum values ​​being the same, a minimum normalization scale consistent with the unit of the physical quantity in that column can be pre-set for each column, and the minimum normalization scale is used when the historical range of change is less than the minimum normalization scale.

[0048] For each filtered sampled value, the increase in the sampled value relative to the historical minimum is first determined. Then, this increase is compared with the normalization scale of the corresponding column, and the resulting ratio is used as the normalization result. The same historical range, minimum normalization scale, and out-of-range handling methods are used in both the training and real-time inference phases. When the current sampled value exceeds the historical range, a normalization result greater than one or less than zero can be retained to preserve the degree of out-of-range behavior; alternatively, it can be truncated to a preset range to limit the impact of extreme outliers on the network input.

[0049] For example, if a certain vehicle model's historical healthy torque range at a certain workstation is 20 N·m to 60 N·m, and the current torque after filtering is 30 N·m, then the current torque is 10 N·m higher than the historical minimum, and the historical range is 40 N·m. Therefore, the normalization result is 0.2. As another example, if the historical maximum and minimum values ​​of a certain stable battery voltage column are both 24 volts, and the preset minimum normalization scale is 0.1 volts, then 0.1 volts is used as the normalization scale for this column to avoid division by zero.

[0050] For operating condition information, vehicle model codes and workstation numbers can be encoded using fixed-dimensional one-hot encoding, embedded encoding, or encoding for unknown category positions. Target torque can be numerically normalized according to the allowable torque range of the wireless tightening tool, or it can be divided into multiple continuous torque intervals and encoded using interval encoding. The number of bolts can be numerically normalized according to the minimum and maximum allowable numbers, or encoded according to ordered intervals such as single bolts, few bolts, and many bolts. All codes are combined in a fixed order to obtain the operating condition coding characteristics.

[0051] For example, a specific tightening task corresponds to vehicle model A, workstation 3, target torque of 45 N·m, and number of bolts of 4. Vehicle model A and workstation 3 are converted into codes for their corresponding categories and positions. If the allowable torque range for the wireless tightening tool is 10 N·m to 110 N·m, then 45 N·m corresponds to a normalized value of 0.35 within this range. If the allowable number of bolts is 1 to 9, then the number 4 corresponds to a normalized value of 0.375. Combining these results in the order of vehicle model, workstation, target torque, and number of bolts yields the working condition coding characteristics of the target tightening task.

[0052] Figure 5 This diagram illustrates a dual-branch feature space joint training method provided in this application embodiment. Each pair of health state samples includes the working condition encoding features and the normalized temporal feature matrix corresponding to the same normal target tightening task. The working condition encoding features are input into the first feature extraction network to obtain the working condition branch output; the normalized temporal feature matrix is ​​input into the second feature extraction network to obtain the temporal branch output. Both branch outputs use the same feature dimension.

[0053] To enable directional comparison between the two branch outputs, the overall length of each branch output is first determined during training. Then, each feature value in the branch output is divided by the sum of the overall length and the positive stability factor, ensuring that the two branch outputs primarily retain directional information. The positive stability factor is a sufficiently small positive number to prevent instability when one branch output approaches all zeros.

[0054] Pairwise alignment constraints are used to handle paired samples of the same health state. For the operational branch output and temporal branch output of the same sample, the feature values ​​at the same position are multiplied and summed to determine the degree of consistency between the two output directions; the closer the two directions are, the higher the degree of consistency and the smaller the alignment loss. The alignment loss of all paired samples in a training batch is averaged to obtain the pairwise alignment result of that training batch.

[0055] For example, the first three normalized feature values ​​of the working condition branch output for the same normal tightening task are 0.6, 0.6, and 0.5, while the first three feature values ​​of the time-series branch output are 0.58, 0.62, and 0.51. Since their directions are similar, the corresponding alignment loss is relatively small. However, if the time-series branch output is -0.6, 0.5, and -0.5, the directions differ significantly, and the alignment loss increases accordingly. During training, the parameters of both networks are adjusted to reduce this difference.

[0056] The working condition classification constraint is used to prevent all working condition outputs from converging in the same direction. The working condition branch output and the time-series branch output are input into a shared working condition classification head, which outputs the predicted probability that the current sample belongs to each working condition index. For the working condition index to which the sample truly belongs, the higher the predicted probability, the smaller the classification loss; the lower the predicted probability, the larger the classification loss. During training, the classification losses of both branches are calculated simultaneously and averaged.

[0057] For example, the training samples cover vehicle model A in the 40 N·m working condition at station one, vehicle model A in the 60 N·m working condition at station one, vehicle model B in the 40 N·m working condition at station two, and vehicle model B in the 60 N·m working condition at station two. If a sample truly belongs to vehicle model A in the 40 N·m working condition at station one, and the working condition classification head gives a 90% prediction probability for that working condition, the classification loss is small; however, if the sample is mainly predicted to be vehicle model B in the 60 N·m working condition at station two, the classification loss increases.

[0058] Different operating condition interval constraints are used to maintain the distinguishability between different operating condition indices. For any two samples with different operating condition indices in a training batch, the Euclidean distance between their normalized operating condition branch outputs is calculated. When this distance is greater than or equal to the preset interval, no interval loss is added; when the distance is less than the preset interval, the interval loss is increased by the square of the insufficient distance. This avoids excessive clustering of operating condition benchmark features corresponding to different operating condition indices.

[0059] For example, if the preset interval is 0.5, and the output distance between two samples under different working conditions is 0.7, no interval penalty is applied to that sample pair; if the output distance is 0.2, the distance is still 0.3 short, and the training process increases the interval loss according to this degree of insufficiency, so that the output of the two working conditions gradually widens.

[0060] During training, the pairwise alignment results, working condition classification results, and different working condition interval results are combined according to preset non-negative coefficients to form the overall training objective. After each training batch is completed, the parameters of the first feature extraction network, the second feature extraction network, and the working condition classification head are updated simultaneously according to the overall training objective. Each constraint coefficient can be determined through validation samples to ensure that the outputs of the two branches for the same working condition remain directionally similar, while maintaining distinguishability between different working conditions.

[0061] In one optional instance, the coefficients of the pairwise alignment constraint and the working condition classification constraint are both set to 1, and the coefficient of the different working condition interval constraint is set to 0.2. If the two branch directions of the same working condition are still unstable on the verification sample, the coefficient of the pairwise alignment constraint can be increased; if different working conditions are easily confused, the coefficients of the working condition classification constraint or the different working condition interval constraint can be increased. The above values ​​are only one feasible parameter, and the actual parameters can be adjusted according to the sample size and the number of working conditions.

[0062] The health status paired samples can be divided into training samples, validation samples, and test samples. Training samples are used to update network parameters, validation samples are used to determine constraint coefficients, the timing of stopping training, and the model version, and test samples are used to independently evaluate the alignment of the two branches and the ability to classify work conditions after training. Each sample set should avoid containing data from the same tightening task; when dividing by time, earlier time periods can be used as training samples, and later time periods as validation and test samples to check the model's applicability to subsequent data.

[0063] In one optional training instance, 800 pairs of health status samples covering four working condition indices are collected, with 70% used for training, 15% for validation, and 15% for testing. Each training batch contains 604 samples, and the parameters are updated using an adaptive moment estimation optimization method with an initial learning rate of 0.001. Training is stopped when the validation loss does not decrease for ten consecutive training epochs, and the model version with the lowest validation loss and the working condition classification accuracy that meets the preset requirements is selected.

[0064] To avoid the dominance of training by work conditions with a large number of samples, samples can be drawn from each work condition index in approximately equal quantities when constructing training batches. For work conditions with fewer samples, the sampling can be increased, repeated sampling can be used, or the weight of that work condition in the model release decision can be reduced. Before the model is released, the pairwise orientation consistency under each work condition index can be checked separately to avoid masking the problem of insufficient alignment of individual work conditions by simply averaging the results of all samples.

[0065] In one optional network architecture example, the input dimension of the work condition coding features is 32. The first feature extraction network consists of two one-dimensional convolutional layers, one max-pooling layer, and one fully connected layer. The two convolutional layers have 16 and 32 output channels, respectively, and both have a kernel length of 3. The max-pooling window length is 2, and the fully connected layer outputs 604-dimensional work condition branch features. This architecture provides a specific method for processing fixed-dimensional work condition coding features.

[0066] When the current working condition index is within the coverage of the training samples, the workstation control computer inputs the current working condition coding features into the first feature extraction network that has passed the verification, obtains the working condition benchmark features, and stores or reads the working condition benchmark features from the working condition benchmark feature library according to the working condition index formed by the vehicle model code, target torque range, bolt quantity range and workstation number.

[0067] When a new vehicle type, new workstation type, or target torque and bolt quantity exceeding the training value range are encountered that are not covered by the training samples, the original model output is not directly used as the benchmark feature for the working condition. The server first collects paired samples of health status confirmed by quality records and tool status under the new working condition, and uses the new working condition samples together with the original working condition samples for network updates. After successful verification, a new model version and corresponding benchmark features for the working condition are generated.

[0068] For example, the original model only covers models A and B, and a new model C is added to the production line. When model C is first deployed, the system can continue to collect data for the tightening task, but it will mark the corresponding state as a new operating condition to be modeled, and will not directly generate a formal degradation judgment based on the output of the original model. After collecting healthy samples of model C under each target torque range, the server updates the two networks to confirm that the original performance of models A and B has not significantly decreased and that model C meets the alignment and classification requirements before releasing a new model version.

[0069] During the full-scale temporal feature determination process, the normalized temporal feature matrix is ​​input into the second feature extraction network. Figure 7 This is a schematic diagram of a multi-scale full-time feature extraction method provided in an embodiment of this application. The second feature extraction network may include multiple residual temporal convolutional blocks, each of which includes a dilated convolution processing unit, a channel response adjustment unit, and residual connections.

[0070] In one example of an optional network architecture, the second feature extraction network comprises four residual temporal convolutional blocks. The dilated convolutional kernels have a length of 3. These four residual temporal convolutional blocks sequentially span adjacent sampling ranges, two sampling intervals, four sampling intervals, and eight sampling intervals. The number of output channels for each residual temporal convolutional block is 32, 604, 604, and 604, respectively. Each residual temporal convolutional block includes two layers of dilated convolutions, followed by a linear rectified activation function.

[0071] The dilated convolution processing unit selects multiple time locations from the input features for convolution processing according to corresponding time intervals. The residual temporal convolution blocks with shorter time intervals mainly extract the local changes in torque, current, and rotational speed between adjacent sampling points; the residual temporal convolution blocks with longer time intervals mainly extract the longer process changes from the start of tightening, bonding, torque rise to near the target torque.

[0072] For example, with a uniform sampling interval of ten milliseconds and a kernel length of 3, convolutional units spanning adjacent sampling ranges can simultaneously observe local changes within approximately 30 milliseconds; convolutional units spanning eight sampling intervals can span a time range of approximately 170 milliseconds to identify slower changes such as the duration of current rise, the process of speed decrease, or the time required for torque to reach its plateau.

[0073] The channel response adjustment unit first averages the response of each feature channel at all time positions to obtain the overall response level of the channel; then, it compresses the number of channels to a preset ratio of the original number through the first mapping layer, and after nonlinear activation, it restores the original number of channels by the second mapping layer; finally, it obtains the response coefficient of each channel through the response function with a value between 0 and 1.

[0074] The response coefficient of each channel is multiplied by the characteristic value of that channel at each time position to improve the response of channels that are more representative of the current tightening task and reduce the response of channels with less information or that are more affected by noise. If the number of channels in the residual branch is different from that in the convolution branch, the number of channels in the residual branch can be adjusted by pointwise convolution, and then the outputs of the two branches are added together according to their positions.

[0075] For example, a residual temporal convolutional block forms 604 feature channels. Among them, the channels related to current rise time and torque slope have relatively high overall responses in the current task, and their response coefficients are 0.8 and 0.9 after channel response adjustment. A certain channel affected by occasional wireless delay has a response coefficient of 0.2. After adjustment, the first two channels retain more response, while the response of the latter channel is reduced accordingly.

[0076] After processing the four residual temporal convolutional blocks, the average or maximum value of each output channel along the time dimension can be calculated to convert time-series data of different lengths into a fixed number of channel features. Then, a fully connected mapping layer is used to map the channel features to the same dimension as the benchmark features, resulting in the full set of time-series features. The first and second feature extraction networks are deployed using the same jointly trained model version to maintain the directional comparison relationship between the two outputs.

[0077] In the process of separating operating condition-related features Figure 8The relationship between full-quantity timing features, operating condition reference direction, operating condition correlation features, and initial residual features is illustrated. The workstation control computer first multiplies the feature values ​​at the same position in the full-quantity timing features with the operating condition reference features, and then adds all the products together. The result is used to characterize the degree of directional overlap of the full-quantity timing features in the operating condition reference direction.

[0078] The workstation control computer then squares and sums each feature value in the working condition reference feature to obtain the self-scale of the working condition reference direction. To avoid the scale being too small when the working condition reference feature is close to all zeros, a preset positive stability factor is added to this scale. The projection coefficient is obtained by dividing the degree of directional coincidence by the self-scale after adding the positive stability factor.

[0079] By using projection coefficients to scale up or down each feature value in the operating condition reference feature by the same proportion, the projection component of the full time series feature in the operating condition reference direction is obtained, and this projection component is used as the operating condition associated feature. The initial residual feature is obtained by subtracting the feature value at the same position in the operating condition associated feature from the feature value at each position in the full time series feature.

[0080] For example, for ease of explanation, the full time series features are simplified to two eigenvalues, 3 and 4, and the load condition reference features are simplified to 1 and 0. Multiplying the values ​​at the same locations and summing them yields 3. The load condition reference feature itself has a scale of 1 and a projection coefficient of 3. Therefore, the load condition associated features are 3 and 0. After subtracting each item from the full time series features, the initial residual features are 0 and 4. This example indicates that changes in the first direction can be explained by the current load condition reference direction, while changes in the second direction are not explained by this load condition reference direction.

[0081] The industrial control computer at the workstation squares each eigenvalue in the initial residual characteristics and sums them. Then, it takes the square root of the sum of the squares to obtain the residual strength. The residual strength reflects the overall amplitude of the initial residual characteristics. Dividing each eigenvalue in the initial residual characteristics by the sum of the residual strength and the positive stability quantity yields the residual direction characteristics. The residual direction characteristics reflect the relative distribution of changes not explained by the operating condition reference direction across the characteristic dimensions.

[0082] Arrange all eigenvalues ​​in the residual directional feature in their original order and add a residual strength at the end to obtain the state residual feature. If both the full time series feature and the operating condition baseline feature are 604-dimensional, then the residual directional feature is 604-dimensional. After adding a residual strength, the state residual feature becomes 605-dimensional.

[0083] Continuing with the simplified example above, the initial residual features are 0 and 4, the residual strength is 4, and the residual direction features are close to 0 and 1. The state residual features can be constructed in the order of 0, 1, and 4. If the initial residual features of another target tightening task are 0 and 0.4, then its residual direction is still close to 0 and 1, but the residual strength is 0.4. Thus, although the residual directions of the two tasks are the same, the state residual features can still distinguish cases where the residual amplitudes differ by a factor of ten.

[0084] In a real-world scenario, when switching from a low-torque model to a high-torque model, the torque, current, and tightening duration may increase overall. If these changes are primarily distributed along the reference direction corresponding to the current high-torque operating condition, a significant portion of them will enter the operating condition-related characteristics, and the initial residual characteristics will not increase significantly in tandem simply due to the increase in the torque target.

[0085] In another scenario, as friction in the drive motor bearings gradually increases, the motor current may continuously rise at the same target torque and speed, while the time required to reach the target torque also increases. This change cannot be fully explained by the operating condition baseline; the current and time-related dimensions in the initial residual characteristics gradually increase, and the residual strength also increases with continuous tasks, thus providing information for subsequent degradation assessment.

[0086] The projection separation described above is used to separate the feature components distributed along the current operating condition baseline direction in the full time series features, but it does not mean that it can eliminate all unmodeled operating condition changes. State residual features may still include changes in ambient temperature, differences in tool installation, sampling disturbances, or unmodeled operating conditions, so it is also necessary to combine the health baseline state model and continuous task judgment rules.

[0087] During the health baseline state model establishment phase, the residual characteristics of the same wireless tightening tool are obtained when it normally performs multiple target tightening tasks after passing factory acceptance, forming a health state sample set. Normal tasks can be confirmed based on factory acceptance results, production line quality records, tool maintenance records, and fault-free operation records within a preset observation period.

[0088] The health status sample set is divided into a model fitting sample set and a threshold calibration sample set, with no identical state residual features in the two sets. The model fitting sample set is used to determine the health probability distribution, while the threshold calibration sample set is used to determine deviations from the threshold. This avoids underestimating the actual range of health fluctuations by using the same batch of samples to both fit the model and determine the threshold.

[0089] The health probability distribution can be represented using a Gaussian mixture model containing multiple Gaussian components. Each Gaussian component characterizes a local clustering state of the health state residual features, such as a health sub-distribution formed under different battery charge ranges, different tool temperature stages, or different slight assembly differences. Each Gaussian component includes at least its weight in the healthy samples, the mean of each feature dimension, and the covariance relationship between the feature dimensions.

[0090] Multiple Gaussian components can be fitted using the expectation-maximization method. First, the weights, mean, and covariance of each Gaussian component are initialized. Then, for each model-fitted sample, the relative probability of that sample being generated by each Gaussian component is calculated, and each relative probability is converted into an attribution proportion. Next, based on the attribution proportion of all samples to each Gaussian component, the weights, mean, and covariance of that Gaussian component are updated. The attribution proportion calculation and parameter update are repeated until the change in the model's overall explanatory power of the samples in consecutive iterations is less than a preset stopping point, or the preset maximum number of iterations is reached.

[0091] The number of Gaussian components can be preset or determined through validation comparisons. For example, candidate models containing 2, 3, 4, and 5 Gaussian components can be built, and the model can be selected based on the average log-likelihood of the validation samples, the information criterion, or the subsequent false alarm rate. If the healthy state samples mainly form 3 clustered regions, 3 Gaussian components can be used; if the sample size is insufficient, it is not advisable to set too many components to avoid each component containing only a small number of samples.

[0092] During the real-time judgment phase, for each current state residual feature, its probability density is determined for each Gaussian component. This probability density is then multiplied by the weight of that Gaussian component to obtain the unnormalized assignment of that Gaussian component to the current state residual feature. Dividing the unnormalized assignment of a Gaussian component by the sum of the unnormalized assignments of all Gaussian components yields the probability that the current state residual feature belongs to that Gaussian component.

[0093] The Gaussian component with the highest assignment probability is identified as the target Gaussian component. If the assignment probabilities of two Gaussian components are similar, the component with the higher assignment probability can be selected. Alternatively, the deviation from the two components can be calculated separately and the smaller value can be taken. The same selection method should be used in both the training and real-time inference phases.

[0094] To determine the degree of deviation, first calculate the difference between the current state residual features and the mean of the target Gaussian components in each feature dimension. Then, using the covariance matrix of the target Gaussian components, scale correction is applied to the differences in each dimension according to the corresponding health fluctuation range and the correlation between different dimensions. Dimensions with smaller health fluctuations contribute more to the degree of deviation when equal numerical differences occur; conversely, dimensions with larger health fluctuations contribute less. The square root of the scale-corrected combined difference is then taken to obtain the Mahalanobis distance.

[0095] To ensure the stable inversion of the covariance matrix, the same regularization value can be added to each diagonal position of the covariance matrix. The regularization value is greater than 0 and can be determined based on the size of the healthy sample and the smallest eigenvalue of the covariance matrix. A regularization value that is too small may still lead to instability in the inversion, while a value that is too large may weaken the differences in health fluctuations across dimensions; therefore, it is advisable to validate the sample selection.

[0096] For example, the state residual feature can be simplified to two dimensions. The two differences between the current sample and the mean of the target Gaussian component are 0.1 and 0.2, respectively; the health variances of the two dimensions are 0.04 and 0.25, respectively, and the correlation between the two dimensions is not considered for the time being. The first difference corresponds to 0.5 standard deviations after conversion according to the health standard deviation, and the second difference corresponds to 0.4 standard deviations. Adding the squares of the two scale-corrected results and taking the square root, we get a Mahalanobis distance of approximately 0.64. This example illustrates that the same numerical difference has a higher anomalous contribution in the dimension with smaller health fluctuations.

[0097] For each state residual feature in the threshold calibration sample set, the above-described target Gaussian component selection and Mahalanobis distance determination process is performed, and the samples are sorted in ascending order of Mahalanobis distance. The Mahalanobis distance corresponding to a preset quantile position can be used as a preset deviation threshold. For example, when the threshold calibration sample set contains 200 samples, the distance of samples located near the 95th percentile after sorting can be used as the threshold. The specific quantile position can be determined based on the allowable false alarm rate and fault response requirements.

[0098] The workstation control computer records the deviation degree according to the time sequence of the target tightening tasks performed by the same wireless tightening tool. When the Mahalanobis distance corresponding to a preset number of consecutive target tightening tasks exceeds a preset deviation threshold, the wireless tightening tool is determined to be in a degraded state. In one example, the preset number is five. If the deviation degree of the five times is, in order, exceeding, not exceeding, exceeding, and exceeding the threshold, the degraded state is not confirmed; only when all five deviations exceed the threshold is confirmation made.

[0099] In another optional approach, an early warning observation state can be set. When at least three out of the last five tasks exceed the preset deviation threshold, but the five-times consecutive rule has not yet been met, the tool is placed in observation state and the frequency of subsequent sampling and recording is increased; a formal degradation warning is then output when the consecutive judgment rule is met. This observation state is an optional auxiliary process and does not replace the continuous confirmation rule.

[0100] To avoid contaminating the health model with degraded samples, the residual features corresponding to the tightening task of a target that is determined to be in a degraded state are not added to the health state sample set. New healthy samples can be confirmed based on production line quality records, tool maintenance results, and preset health sample acceptance criteria. Updates to the health baseline state model are completed on the server and distributed after version verification.

[0101] After confirming the deterioration state, the current state residual features or the combined state residual features corresponding to multiple consecutive target tightening tasks of the same wireless tightening tool can be input into the faulty component classifier. Historical fault samples include the state residual features corresponding to a preset number of target tightening tasks before the fault occurred, as well as faulty component tags confirmed based on maintenance records, component replacement records, or disassembly and inspection results.

[0102] In one optional instance, the faulty component classifier is a random forest classifier. During training, multiple decision trees are constructed, and each decision tree randomly selects a portion from the historical fault samples and state residual feature dimensions for split learning. During real-time inference, each decision tree outputs candidate component categories, and the matching probability of candidate degraded components such as drive motor, torque sensor, battery management unit, ratchet head, and wireless communication module is determined based on the tree voting ratio obtained for each category.

[0103] For example, a random forest containing 200 decision trees can be trained using 300 sets of historical fault samples confirmed by maintenance. Each set of samples consists of the state residual features of five consecutive tightening tasks before the fault occurred, the average value of the five residual strengths, and the trend of change. The training data can be grouped according to the tool number to avoid overestimating the classification performance by having adjacent samples of the same tool simultaneously enter the training and validation sets.

[0104] Early degradation of the drive motor can be characterized by a continuously increasing residual in the current-related channel, a prolonged decrease in motor speed under the same target torque, and a gradual increase in residual strength. Torque sensor anomalies can be manifested as a shift in the correspondence between the torque-related channel and the current and speed-related channels. Battery management unit anomalies can be characterized by an increased battery voltage drop during load conditions, synchronized with current changes. Ratchet head wear can be characterized by a change in the stage relationship between the tightening angle and the torque increase process. Wireless communication module anomalies can be characterized by an increase in the correlation between the wireless transmission delay channel and its number of transitions. These characteristics are merely illustrative examples for training samples; the actual classification criteria are determined by historical fault samples.

[0105] The component with the highest matching probability can be identified as a candidate deteriorated component. When the highest matching probability is lower than a preset confirmation threshold, the component's confirmation result is output. When the matching probabilities of two or more components all exceed the corresponding thresholds, multiple candidate deteriorated components can be output simultaneously, and maintenance personnel can be prompted to check them in order of matching probability.

[0106] Figure 9 This diagram illustrates a health baseline status assessment and degradation early warning system provided in an embodiment of this application. The workstation control computer generates degradation early warning information based on the tool number, workstation number, candidate degraded components, degradation status, and early warning generation time. This information is then sent to the workstation display device and the production line maintenance system. The degradation early warning information indicates that the wireless tightening tool is showing a degradation trend, but does not limit the tool to having already experienced functional failure.

[0107] The workstation control computer associates the original timing matrix, operating condition information, normalized timing feature matrix, operating condition baseline features, full timing features, operating condition correlation features, state residual features, deviation degree, deterioration state, candidate deteriorated parts, and deterioration warning information according to the task identifier of the target tightening task, and stores them to the DMS data server.

[0108] In Application Example 1, the same wireless tightening tool switches from a 40 N·m task for vehicle model A to a 70 N·m task for vehicle model B. The initial torque, current, and tightening duration all increase significantly. However, the 70 N·m condition for vehicle model B retains the corresponding baseline characteristics, and most of the changes are identified as condition-related features. The residual characteristics remain within the coverage of the healthy baseline state model, therefore no degradation warning is output. This example illustrates that normal changes caused by the condition switch will not directly trigger a warning simply because the initial data amplitude increases.

[0109] In application example two, the wireless tightening tool continuously performs tasks with the same vehicle model, workstation, and target torque. As the friction of the drive motor bearing increases, the residual error related to the motor current gradually strengthens, the time for the motor speed to decrease is prolonged, and the residual intensity continuously increases from near the healthy range. The deviation of the current five consecutive tasks all exceeds the preset deviation threshold, thus confirming a deterioration state; the fault component classifier gives the drive motor the highest matching probability, and the deterioration warning message prompts to prioritize checking the drive motor.

[0110] In application example 3, a brief congestion occurred in the production line's wireless network, causing a sudden increase in the wireless transmission delay for a target tightening task. However, the torque, current, speed, and angle changes remained normal. The deviation of this task may have exceeded the threshold, but the subsequent four tasks returned to normal. Therefore, the continuous confirmation rule was not met, and no formal degradation warning was output. If the wireless transmission delay remains abnormal in multiple consecutive tasks, and the fault component classifier has the highest matching probability for the wireless communication module, then a candidate degradation warning for the wireless communication module is output.

[0111] In application example 4, a new model C and a new target torque range are added to the production line. Since this operating condition index exceeds the range of the trained encoding, the system does not directly use the original model output for formal state judgment. Instead, it collects quality-verified healthy tightening samples, updates the two feature extraction networks and completes verification, and then establishes a corresponding record for model C in the operating condition baseline feature library. After the model update, the verification results for models A and B should also be rechecked to avoid performance degradation of the original operating conditions due to training under the new operating conditions.

[0112] In Application Example 5, a target task involves four bolt tightening cycles on the same vehicle at one workstation. The workstation's industrial control computer divides the data into four sub-sequences based on the four tightening start trigger signals, obtaining four state residual features for each. Then, according to the bolt execution order, the average of the four state residual features is calculated, and the maximum residual strength is taken, or the features are sequentially concatenated before task-level aggregation. If the residual strength of one bolt cycle is abnormal while the other three are normal, the maximum residual strength and the corresponding sub-sequence identifier can be retained to allow maintenance personnel to trace the specific connection location.

[0113] Figure 10 This is a schematic diagram of a wireless tightening tool degradation monitoring device provided in an embodiment of this application. The device includes a data acquisition module, a feature determination module, a feature separation module, a degradation judgment module, and an early warning output module.

[0114] The data acquisition module acquires the tightening timing data and operating condition information corresponding to the target tightening task. The feature determination module determines the operating condition baseline features through a first feature extraction network and the full timing features through a second feature extraction network. The feature separation module determines the operating condition associated features and constructs state residual features based on the direction and amplitude information of the remaining features. The degradation judgment module determines the degree of deviation of the state residual features from the healthy baseline state model and determines the degradation state. The early warning output module generates and outputs degradation early warning information.

[0115] The above modules can be implemented using software functional modules or using a hardware structure that combines a processor, memory, and communication interface. The tightening control device may include at least one processor and at least one memory. The memory stores a computer program, which the processor reads and executes to implement the above method steps. The communication interface is used for data interaction with at least one of the following: a wireless tightening tool, an MES system server, a DMS data server, a workstation display device, and a production line maintenance system.

[0116] Those skilled in the art will understand that all or part of the steps in the above embodiments can be implemented by computer program instructions and related hardware. The computer program can be stored in a computer-readable storage medium, and when the processor executes the computer program, it implements all or part of the steps in the above method embodiments.

[0117] The above description is merely a specific embodiment of this application and is not intended to limit this application. Those skilled in the art can adjust or substitute the above embodiments without departing from the technical concept of this application, and all such adjustments or substitutions should be included within the protection scope of this application.

Claims

1. A tightening control method applicable to a final assembly system, characterized in that, The method, applied to tightening control equipment, includes: Acquire the tightening timing data generated when the wireless tightening tool performs the target tightening task and the working condition information corresponding to the target tightening task. The tightening timing data is used to characterize the timing operation changes of the wireless tightening tool in the target tightening task, and the working condition information is used to characterize the working condition conditions corresponding to the target tightening task. Based on the working condition information, the working condition baseline features are determined by the first feature extraction network, and based on the tightening time series data, the full time series features are determined by the second feature extraction network jointly trained with the first feature extraction network based on health status paired samples. The projection component of the full time series feature in the working condition reference direction represented by the working condition reference feature is determined, the projection component is used as the working condition associated feature, and the initial residual feature is obtained according to the difference between the full time series feature and the working condition associated feature. The state residual features are constructed based on the direction and magnitude information of the initial residual features; Determine the degree of deviation of the state residual characteristics from the health baseline state model, determine the deterioration state of the wireless tightening tool based on the degree of deviation, and output deterioration warning information corresponding to the deterioration state; The health baseline state model is established based on the state residual characteristics obtained when the wireless tightening tool is in a healthy state.

2. The method as described in claim 1, characterized in that, The acquisition of tightening timing data generated when the wireless tightening tool performs the target tightening task and the corresponding working condition information of the target tightening task includes: Using the moment corresponding to the tightening start trigger signal of the wireless tightening tool as the zero point of time, real-time torque data, motor current data, motor speed data, tightening angle data, wireless transmission delay data, and battery voltage data are acquired respectively. A unified relative time axis is established based on the sampling timestamps corresponding to each data point. Data points with different sampling times are mapped to the corresponding sampling times on the unified relative time axis to obtain the original time series matrix. Obtain the vehicle model code, target torque, number of bolts, and workstation number, and use the task identifier of the target tightening task to bind the vehicle model code, target torque, number of bolts, and workstation number to the original time series matrix.

3. The method as described in claim 2, characterized in that, Before determining the operating condition baseline characteristics and the full time series characteristics, the method further includes: The median filtering of each dimension of the original time series matrix is ​​performed according to a preset sliding window to obtain the filtered time series matrix. Based on the historical data range of the workstation and vehicle model corresponding to the target tightening task, the data of each dimension in the filtered time series matrix are normalized to obtain the normalized time series feature matrix. The vehicle model code and the workstation number are categorized and coded, the target torque and the number of bolts are numerically normalized or range-coded, and the coding results are combined in a preset order to obtain the working condition coding features.

4. The method as described in claim 3, characterized in that, The joint training of the first feature extraction network and the second feature extraction network, as well as the determination of the working condition benchmark features, include: Multiple health status pairs are obtained. Each health status pair includes the working condition coding features and normalized time series feature matrix corresponding to the same target tightening task. Each pair of health status samples is processed by the first feature extraction network and the second feature extraction network respectively. Based on the alignment constraint between the outputs of the two networks for the same sample and the condition discrimination constraint between the outputs of the networks corresponding to different condition indices, the parameters of the two networks are jointly updated. After the jointly trained model is validated, the current working condition encoded features are input into the first feature extraction network to obtain the working condition baseline features. It is then determined whether the current working condition index is within the trained encoding range of the first feature extraction network. If so, the working condition baseline features are stored in the working condition baseline feature library of the corresponding working condition index. If not, the two networks are updated using the health status pair samples corresponding to the current working condition index, and then the working condition baseline features are generated and stored.

5. The method as described in claim 4, characterized in that, The process of determining full temporal features through a second feature extraction network includes: The normalized temporal feature matrix is ​​input into the second feature extraction network, and the temporal operation changes of the wireless tightening tool at different time scales are extracted by multiple temporal convolution processing units with different time perception ranges. The temporal responses of each feature channel are summarized along the time dimension. Based on the summary results, the channel response coefficients corresponding to each feature channel are determined, and the output of the corresponding feature channel is adjusted using the channel response coefficients. The adjusted features at different time scales are subjected to residual fusion and time dimension aggregation, and feature mapping is performed according to the feature dimensions of the operating condition baseline features to obtain the full time series features.

6. The method as described in claim 5, characterized in that, The process of determining the operating condition correlation features and the state residual features includes: Multiply and sum the feature values ​​at the same position in the full time series feature and the operating condition reference feature respectively to obtain the directional coincidence amount; sum the squares of each feature value in the operating condition reference feature and add the positive stability amount to obtain the scale amount of the operating condition reference direction; determine the projection coefficient according to the ratio between the directional coincidence amount and the scale amount. The projection coefficients are used to proportionally adjust each feature value in the working condition reference feature to obtain the projection component. The projection component is used as the working condition associated feature. The feature value at the corresponding position in the working condition associated feature is subtracted from each feature value in the full time series feature to obtain the initial residual feature. The residual strength is determined by the square root of the sum of the squares of the feature values ​​in the initial residual feature. Each feature value in the initial residual feature is divided by the sum of the residual strength and another positive stable quantity to obtain the residual direction feature. The residual direction feature and the residual strength are combined in a preset order to obtain the state residual feature.

7. The method as described in claim 6, characterized in that, The process of establishing the health baseline state model includes: After the same wireless tightening tool passes factory acceptance, multiple state residual features are obtained during the normal execution of multiple target tightening tasks to form a health state sample set, and the health state sample set is divided into a model fitting sample set and a threshold calibration sample set. Based on the state residual features of the model fitting sample set, a health probability distribution including multiple Gaussian components is established, and the weight, mean vector and covariance matrix corresponding to each Gaussian component are determined. The health baseline state model is established based on the health probability distribution and the model parameters of each Gaussian component.

8. The method as described in claim 7, characterized in that, Determining the degree of deviation and the state of degradation includes: Determine the probability of the current state residual feature belonging to the multiple Gaussian components, and determine the Gaussian component with the highest probability of belonging as the target Gaussian component. The difference between the current state residual feature and the mean vector of the target Gaussian component is determined, and the difference is scaled according to the health fluctuation range of each feature dimension and the correlation between different feature dimensions using the regularized covariance matrix to obtain the Mahalanobis distance, and the Mahalanobis distance is used as the degree of deviation. A preset deviation threshold is determined based on the statistical distribution of the deviation degree corresponding to the threshold calibration sample set, and a preset number of deviation degrees are continuously acquired according to the time sequence of the same wireless tightening tool performing the target tightening task. When all preset number of deviation degrees exceed the preset deviation threshold, it is determined that the wireless tightening tool is in a deteriorated state.

9. The method as described in claim 8, characterized in that, After determining that the wireless tightening tool is in a deteriorated state, the method further includes: The state residual features are input into a fault component classifier trained using historical fault samples to obtain the matching probability of each of the multiple candidate deteriorated components. The candidate deteriorated components are determined based on each matching probability. The candidate deteriorated components include at least one of a drive motor, a torque sensor, a battery management unit, a ratchet head, and a wireless communication module. The deterioration warning information is generated based on the tool number of the wireless tightening tool, the workstation number, the candidate deteriorated parts, and the deterioration status, and the deterioration warning information is sent to the workstation display device and the production line operation and maintenance system. The tightening timing data, the operating condition information, the operating condition baseline characteristics, the full timing characteristics, the state residual characteristics, the deviation degree, the degradation state, and the degradation early warning information are associated and stored according to the task identifier.

10. A tightening control device for use in a final assembly system, for implementing the method of claim 1, characterized in that, include: The data acquisition module is used to acquire the tightening timing data generated when the wireless tightening tool performs the target tightening task, as well as the working condition information corresponding to the target tightening task; The feature determination module is used to determine the benchmark features of the working condition through the first feature extraction network, and to determine the full-quantity time-series features through the second feature extraction network jointly trained with the first feature extraction network based on health status paired samples. The feature separation module is used to determine the working condition related features of the full time series features in the working condition reference direction represented by the working condition reference features, and to construct state residual features based on the direction information and amplitude information of the difference between the full time series features and the working condition related features. The degradation judgment module is used to determine the degree of deviation of the state residual characteristics from the health baseline state model, and to determine the degradation state of the wireless tightening tool based on the degree of deviation. The early warning output module is used to generate and output corresponding degradation early warning information based on the degradation status.

Images