Work order review methods, devices, electronic equipment and computer-readable storage media

Abstract

This application discloses a work order review method, apparatus, electronic device, and computer-readable storage medium. The method includes: acquiring information filled in several first work order items from a work order to be reviewed; comparing the reference information and the filled information of each first work order item to obtain a first work order review result; wherein the reference information of each first work order item is predicted using a target large language model, the first work order review result is used to characterize whether the filled information of each first work order item is compliant, and the target large language model is obtained by adjusting the network parameters of each training layer of a pre-trained initial large language model using sample work orders, and each training layer is a network layer added to each target network layer of the initial large language model. Through the above method, this application can improve the accuracy of work order review.

Description

Technical Field

[0001] This application relates to the field of work order review technology, and in particular to a work order review method, apparatus, electronic device and computer-readable storage medium. Background Technology

[0002] Currently, the compliance review of the information filled in the work order items is mostly done manually. Manual review has low accuracy and low efficiency.

[0003] For example, regarding the review of after-sales service orders for power batteries, after the power battery service station completes its after-sales work, it submits the after-sales service order for the battery fault in the GSS system. Subsequently, engineers manually review the compliance of the information filled in the work order items of the after-sales service order for the battery fault. The review accuracy is low and the review efficiency is low. Summary of the Invention

[0004] This application provides at least one work order review method, apparatus, electronic device, and computer-readable storage medium.

[0005] The first aspect of this application provides a work order review method, which includes: obtaining the information filled in for several first work order items in the work order to be reviewed; comparing the reference information and the information filled in for each first work order item to obtain the first work order review result; wherein, the reference information for each first work order item is predicted using a target large language model, the first work order review result is used to characterize whether the information filled in for each first work order item is compliant, the target large language model is obtained by adjusting the network parameters of each training network layer of a pre-trained initial large language model using sample work orders, and each training network layer is a network layer added to each target network layer of the initial large language model.

[0006] Therefore, the target large language model is obtained by adjusting the network parameters of each training layer of the pre-trained initial large language model using sample work orders. Thus, the target large language model has the ability to predict the compliance information for each first work order item, or in other words, it can accurately predict the compliance information for each first work order item. Therefore, the reference information for each first work order item in the work order to be reviewed, predicted using the target large language model, is the compliance information, or can be considered a theoretical compliance template. Therefore, comparing the reference information and the submitted information for each first work order item is comparing the compliance information with the actual submitted information for each first work order item, thereby determining or verifying the compliance of the submitted information for each first work order item in the work order to be reviewed, and achieving accurate review of the work order to be reviewed.

[0007] Among them, the work order to be reviewed is a battery fault after-sales work order. The work order to be reviewed also includes target fault-related data, which includes at least one of the following: target fault description information, target fault handling result; the step of predicting the reference information of each first work order item using the target large language model includes: using the target large language model to predict the reference information of each first work order item based on the target fault-related data.

[0008] Therefore, the reference information for each first work order item is predicted based on the target fault-related data using the target large language model. Thus, the reference information for each first work order item is the compliance information for each first work order item, or in other words, it can be regarded as a theoretical compliance reporting template.

[0009] The first work order items include first-type work order items and second-type work order items. The correlation between first-type work order items and target fault-related data is greater than that between second-type work order items and target fault-related data. Based on the target fault-related data, the target large language model is used to predict the reference information of each first work order item, including: inputting the target fault-related data into the target large language model to obtain the reference information of the first-type work order items output by the target large language model; and / or, inputting the target fault-related data into the target large language model to obtain the target intermediate information output by the target large language model, and finding the reference information of the second-type work order items from the correlation information of the target intermediate information.

[0010] Therefore, for the first type of work order item that is highly correlated with the target fault-related data, the corresponding reference information can be directly predicted using the target large language model based on the target fault-related data.

[0011] For the second type of work order items with weak correlation to the target fault-related data, directly predicting the corresponding reference information based on the target fault-related data using the target large language model yields low accuracy. Therefore, we utilize the target large language model to predict both the reference information and the target intermediate information, which have strong correlations with the target fault-related data, based on the target fault-related data. This intermediate information is then used to determine the reference information for the second type of work order items, thereby improving the accuracy of the determined reference information.

[0012] The first type of work order item includes at least one of the following: fault level, fault attribute, faulty component, vehicle-level fault phenomenon, failure mode; and / or, the second type of work order item includes at least one of the following: repair code, repair time; and / or, the target intermediate information includes the component to be repaired.

[0013] Therefore, the first type of work order item, the second type of work order item, and the target intermediate information can be flexibly set.

[0014] Among them, the target large language model for predicting the reference information of the first type of work order item and the target large language model for predicting the intermediate information of the target are different or the same large language model.

[0015] Therefore, the target large language model for predicting the reference information of the first type of work order item and the target large language model for predicting the intermediate information are different large language models; thus, the prediction of the reference information and intermediate information of the first type of work order item is performed using the corresponding dedicated models to improve the prediction accuracy of the reference information and intermediate information of the first type of work order item.

[0016] The target large language model for predicting the first type of work order item and the target large language model for predicting the intermediate information of the target are the same large language model, which eliminates the need to repeatedly load multiple models, reducing inference latency and resource consumption.

[0017] The step of adjusting the network parameters of each training layer of the pre-trained initial large language model using sample work orders includes: freezing the current network parameters of the initial large language model and adding the training layer to each target network layer of the initial large language model; using the initial large language model to predict sample information of several first work order items based on sample fault-related data in the sample work orders; and adjusting the network parameters of each training layer based on the difference between the sample information of several first work order items and the real information of several first work order items.

[0018] Therefore, the network parameters, or original parameters, of the pre-trained initial large language model are kept unchanged. Trainable incremental parameters are added to each target network layer of the initial large language model. In other words, without directly modifying the original parameters of the pre-trained initial large language model, additional trainable incremental parameters are introduced to specifically adjust the trained initial large language model, resulting in the target large language model. This allows the target large language model to retain its general language understanding capabilities while learning domain knowledge related to work order review, improving its adaptability to work order review tasks and optimizing its performance in predicting compliant information submission. Furthermore, the current network parameters, or original parameters, of the pre-trained initial large language model remain unchanged and are not involved in parameter adjustments. Subsequent adjustments only involve the network parameters of the trainable network layers, avoiding the high computational cost of full parameter adjustments.

[0019] Among them, the network parameters of the network layer to be trained are the rank decomposition matrix of the network layer to be trained. The rank decomposition matrix of the network layer to be trained is obtained by decomposing the weight matrix of the network layer to be trained. At least two network layers to be trained have different matrices. The rank of the matrix of the network layer to be trained is determined based on the compatibility between the weight matrix of the network layer to be trained and the work order review task.

[0020] Therefore, the rank decomposition matrix corresponding to the network layer to be trained is obtained by decomposing the weight matrix corresponding to the network layer to be trained. If the rank of the matrix corresponding to each network layer to be trained is the same, that is, if each network layer to be trained uses a fixed matrix rank, it will lead to overfitting of the adjusted initial large language model, i.e., the target large language model, and poor model expressive ability. Based on the fit between the weight matrix of each network layer to be trained and the work order review task, the rank of the matrix corresponding to each network layer to be trained is dynamically determined, and a matching matrix rank is assigned to each network layer to be trained.

[0021] The fit between the weight matrix of the network layer to be trained and the work order review task is represented by the norm of the gradient of the weight matrix of the network layer to be trained. The steps for determining the rank of the matrix corresponding to the network layer to be trained include: obtaining the product between the norm of the network layer to be trained and the first coefficient; summing the product and the second coefficient to obtain the first sum; rounding the first sum to obtain the second sum, and obtaining the rank of the matrix corresponding to the network layer to be trained based on the second sum.

[0022] Therefore, the first coefficient is a scaling factor. By setting the first coefficient, the contribution ratio of the norm corresponding to the network layer to be trained to the matrix rank can be adjusted. The larger the first coefficient is, the more significant its impact on the matrix rank. The second coefficient is an offset term. By setting the second coefficient, the matrix rank is prevented from returning to zero when the norm corresponding to the network layer to be trained is too small, thus ensuring the minimum matrix rank.

[0023] Wherein, the second sum is an integer not greater than the first sum; and / or, based on the second sum, the matrix rank corresponding to the network layer to be trained is obtained, including: in response to the second sum being greater than or equal to a rank threshold, the rank threshold is used as the matrix rank corresponding to the network layer to be trained; in response to the second sum being less than a rank threshold, the second sum is used as the matrix rank corresponding to the network layer to be trained.

[0024] Therefore, the rounding direction of the first sum can be flexibly set.

[0025] The rank threshold is the upper limit of the matrix rank of the network layer to be trained. That is, the maximum matrix rank of the network layer to be trained can only be up to the rank threshold. Setting the upper limit of the matrix rank limits the number of trainable parameters of each network layer to avoid introducing too many parameters and ensure the lightweight characteristics. In addition, by setting the upper limit of the matrix rank, the maximum complexity of each network layer to be trained is limited, forcing each network layer to learn the most critical features with limited parameter capacity. Furthermore, the norm of the gradient of the weight matrix of some network layers to be trained may have instantaneous spikes in the early stage of training, resulting in abnormally high matrix rank calculated dynamically, causing oscillations in the training process. Therefore, by setting the upper limit of the matrix rank, the rank explosion caused by instantaneous gradient anomalies can be avoided, making the training process smoother and more predictable.

[0026] Wherein, the first coefficient of the first stage of adjustment is less than the first coefficient of the second stage of adjustment, and the first stage is earlier than the second stage; and / or, the second coefficient of the first stage of adjustment is equal to the second coefficient of the second stage of adjustment.

[0027] Therefore, in the initial adjustment phase, the network parameters of each training layer of the initial large language model are adjusted from the pre-training state, and their gradients are usually large. A small first coefficient is set to avoid assigning an excessively large matrix rank due to high gradients in the early adjustment phase, thereby avoiding oscillations in the training process due to an excessively large matrix rank. In the later adjustment phase, the gradients of the network parameters of each training layer of the initial large language model gradually decrease, and the model approaches convergence. A large first coefficient is set to increase the influence of the gradient on the matrix rank, enabling the initial large language model to learn detailed information.

[0028] The steps for determining the norm of the network layer to be trained include: taking each matrix parameter of the weight matrix of the network layer to be trained as the current parameter, obtaining the partial derivative of the work order loss with respect to the current parameter, and taking it as the partial derivative component of the current parameter; wherein, the work order loss is determined based on the difference between the sample information of several first work order items and the real work order information of several first work order items; summing the squares of the absolute values ​​of the partial derivative components corresponding to each matrix parameter to obtain the third sum; taking the square root of the third sum to obtain the norm of the network layer to be trained.

[0029] Therefore, the gradient of the network parameters of the network layer to be trained is the partial derivative of the work order loss (quantifying the difference between the model prediction and the true label) with respect to the network parameters of the network layer to be trained, reflecting the sensitivity of the network parameters of the network layer to the work order review task loss. A large gradient of the network parameters of the network layer to be trained indicates that small changes in the network parameters will significantly change the loss value, and the network layer to be trained has a large impact on the work order review task result, further indicating that the network layer to be trained is not fully adapted to the work order review task; therefore, more parameters are needed to model complex relationships, that is, a high matrix rank is required. A small gradient of the network parameters of the network layer to be trained indicates that changes in the network parameters of the network layer to be trained have a weak impact on the work order loss, further indicating that the function of the network layer to be trained is basically adapted to the work order review task.

[0030] The weight matrix corresponding to the network layer to be trained is obtained by adjusting the initial weight matrix using a third coefficient, which is the ratio of the hyperparameter to the rank of the matrix corresponding to the network layer to be trained; and / or, in the first stage of adjustment, the initial large language model adjusts the rank of the matrix corresponding to each network layer to be trained once for every first number of sample work orders processed, and in the second stage of adjustment, the initial large language model adjusts the rank of the matrix corresponding to each network layer to be trained once for every second number of sample work order data processed, wherein the first number is less than the second number, and the first stage is earlier than the second stage.

[0031] Therefore, in order to control the impact of the initial weight matrix corresponding to each network layer to be trained on the initial large language model, hyperparameters are introduced to improve the accuracy of work order review of the large language model.

[0032] In the initial adjustment phase, the network parameters of each training layer are just beginning to adapt to the work order review task from the pre-training state. The norm amplitude of each training layer is large and the direction changes drastically. If the adjustment frequency of the matrix rank is too low, the large language model may have been exploring in the wrong parameter space for a long time, making it difficult to converge or even diverge. Therefore, in the initial adjustment phase, adjusting the matrix rank of each training layer after processing the first number of sample work orders can ensure that the initial large language model converges stably. In the later adjustment phase, the initial large language model has initially converged, the norm amplitude of each training layer is small and the direction is stable, and the matrix rank configuration of each training layer has approached the optimal value. At this time, the decision space of the matrix rank configuration is reduced, eliminating the need for frequent re-evaluation and reducing unnecessary computational overhead.

[0033] The work order to be reviewed also includes information for several second work order items. The work order review method further includes: obtaining the target image associated with each second work order item; wherein the target image contains reference information for the second work order item; identifying the target image associated with each second work order item to obtain the reference information for each second work order item; comparing the reference information and the information submitted for each second work order item to obtain the second work order review result; wherein the second work order review result is used to characterize whether the information submitted for each second work order item is compliant.

[0034] Therefore, introducing image technology into work order compliance review, and using multimodal data to conduct work order compliance review, can improve the accuracy of work order review.

[0035] Specifically, the target images associated with each second work item are identified to obtain reference information for each second work item, including: extracting the region image of the area where the reference information of each second work item is located from the target images associated with each second work item; and identifying each region image to obtain the reference information for each second work item.

[0036] Therefore, by first identifying the region where the reference information of the second work order is located, and then recognizing the region image of the region where the reference information of the second work order is located, on the one hand, the recognition algorithm only needs to process the high-value region, avoiding the interference of noise and improving the recognition accuracy; on the other hand, only the region image of the region where the reference information of the second work order is located needs to be recognized, which reduces the amount of computation and improves the computational efficiency.

[0037] Among them, at least one target image associated with a second work order is a first type of image, and the reference information in the first type of image is a barcode representing the reference information; from the target images associated with each second work order, the region image of the area where the reference information of each second work order is located is extracted, including: determining the vertical projection value of each pixel row and the horizontal projection value of each pixel column in the first type of image; rotating the first type of image until the current vertical projection value and horizontal projection value of the first type of image meet the preset projection requirements, and using the current vertical projection value and horizontal projection value of the first type of image, extracting the region image corresponding to the barcode.

[0038] Therefore, in barcode recognition, if the current vertical and horizontal projection values ​​of the first type of image meet the preset projection requirements, it means that the barcode is in an ideal state or is not tilted. The barcode is arranged in a completely vertical or horizontal manner. Therefore, the current vertical and horizontal projection values ​​of the first type of image can be used to accurately locate the barcode, thereby accurately extracting the area image corresponding to the barcode.

[0039] The preset projection requirements are as follows: among the vertical projection values ​​of each pixel row, there exists a vertical projection value greater than a first projection threshold; among the horizontal projection values ​​of each pixel column, there exists a horizontal projection value greater than a second projection threshold; and / or, rotating the first type of image until the current vertical and horizontal projection values ​​of the first type of image meet the preset projection requirements, including: in response to the absence of vertical and horizontal projection values ​​greater than the projection threshold, rotating the first type of image by a preset angle, and re-executing the steps of determining the vertical projection values ​​of each pixel row and the horizontal projection values ​​of each pixel column in the first type of image and subsequent steps, until a vertical projection value greater than the projection threshold exists. The vertical projection value and horizontal projection value; after determining the vertical projection value of each pixel row and the horizontal projection value of each pixel column in the first type of image, the work order review method further includes: in response to the existence of vertical projection values ​​and horizontal projection values ​​greater than the projection threshold, according to the horizontal direction, from the vertical projection values ​​of each pixel row, finding the first and second sequential vertical projection values ​​corresponding to the vertical projection values ​​greater than the projection threshold, and using them as the upper and lower boundaries of the barcode respectively; and, according to the vertical direction, from the horizontal projection values ​​of each pixel column, finding the third and fourth sequential horizontal projection values ​​corresponding to the horizontal projection values ​​greater than the projection threshold, and using them as the left and right boundaries of the barcode respectively.

[0040] Therefore, preset projection requirements can be flexibly set.

[0041] The first type of image is rotated until the current vertical and horizontal projection values ​​of the first type of image meet the preset projection requirements. Through external intervention, the barcode is rotated and corrected from a tilted state to an ideal state, thereby enabling the barcode to be located even when it is not tilted, thus improving the positioning accuracy of the barcode.

[0042] Wherein, at least one target image associated with a second work order item is a second type of image, and the reference information in the second type of image is text information representing the reference information; wherein, extracting the region image of the area where the reference information of each second work order item is located from the target image associated with each second work order item includes: performing text detection on the second type of image using a detection model to obtain several text regions and corresponding region categories; matching the categories of the several text regions with the reference information in the second type of image, and taking the text region whose region category matches the category corresponding to the reference information in the second type of image as the region image corresponding to the reference information in the second type of image; and / or, before recognizing each region image to obtain the reference information of each second work order item, the work order review method further includes: correcting the region image.

[0043] Therefore, the detection model can detect several text regions in the second type of image and determine the region category of each text region. The region category corresponding to a text region reflects the category of reference information present in that text region. Thus, by identifying the region category corresponding to the text region, the region containing the reference information for the second work order item can be located. Furthermore, subsequent identification of the region image containing the reference information for the second work order item will ensure that the identified reference information is indeed the reference information for the second work order item, eliminating the need for further verification. This improves the accuracy and efficiency of identifying the reference information for the second work order item, thereby enhancing the accuracy and efficiency of the review process for the second work order item.

[0044] The reference information in the second type of image is textual information representing reference information. These text images often have various geometric distortions, resulting in low accuracy for direct recognition. By correcting the regional images to eliminate distortions and restore them to a standard format, optimal input conditions are created for subsequent regional image recognition, thereby improving the accuracy of the reference information for the second work order item.

[0045] The pending work orders are displayed on the display interface. For non-compliant work order items, at least one of the following steps shall be performed: adjust the display color of the non-compliant work order item and / or the information filled in the work order item; highlight the non-compliant work order item and / or the information filled in the work order item.

[0046] Therefore, by adjusting the display color of non-compliant work order items and / or the information entered in the work order items, the non-compliant parts of the work order to be reviewed can be highlighted.

[0047] By highlighting non-compliant work order items and / or the information entered in the work order items, the non-compliant parts of the work orders to be reviewed are highlighted.

[0048] The second aspect of this application provides a work order review device, which includes an acquisition module and an review module. The acquisition module is used to acquire the information filled in for several first work order items in the work order to be reviewed. The review module is used to compare the reference information and the information filled in for each first work order item to obtain the review result of the first work order. The reference information for each first work order item is predicted using a target large language model. The review result of the first work order is used to characterize whether the information filled in for each first work order item is compliant. The target large language model is obtained by adjusting the network parameters of each training network layer of the pre-trained initial large language model using sample work orders. Each training network layer is a network layer added to each target network layer of the initial large language model.

[0049] A third aspect of this application provides an electronic device including a memory and a processor. The memory stores program instructions, and the processor executes the program instructions to implement the aforementioned work order review method.

[0050] A fourth aspect of this application provides a computer-readable storage medium for storing program instructions that can be executed to implement the above-described work order review method.

[0051] The above technical solution uses sample work orders to adjust the network parameters of each training layer of the pre-trained initial large language model. Therefore, the target large language model has the ability to predict the compliance information of each first work order item, or in other words, it can accurately predict the compliance information of each first work order item. Thus, the reference information for each first work order item in the work order to be reviewed, predicted by the target large language model, is the compliance information, or can be considered a theoretical compliance template. Therefore, comparing the reference information and the submitted information for each first work order item is comparing the compliance information with the actual submitted information, thereby determining or verifying the compliance of the submitted information for each first work order item in the work order to be reviewed, and achieving accurate review of the work order to be reviewed. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating an embodiment of the work order review method provided in this application; Figure 2 This is a flowchart illustrating an embodiment of adjusting the network parameters of each training layer of a pre-trained initial large language model using sample work orders, as provided in this application. Figure 3 This is a flowchart illustrating another embodiment of the work order review method provided in this application; Figure 4 This is a schematic diagram of the structure of an embodiment of the work order review device provided in this application; Figure 5 This is a schematic diagram of the structure of an embodiment of the electronic device provided in this application; Figure 6 This is a schematic diagram of an embodiment of the computer-readable storage medium provided in this application. Detailed Implementation

[0053] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0054] In the following description, specific details such as particular system architectures, interfaces, and technologies are presented for illustrative purposes rather than for limiting purposes, in order to provide a thorough understanding of this application.

[0055] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, "many" in this document means two or more. Moreover, the term "at least one" in this document means any combination of at least two of any one or more of a plurality of objects. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0056] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the work order review method provided in this application. It should be noted that if substantially the same result is achieved, this embodiment does not necessarily reflect that outcome. Figure 1 The illustrated process sequence is limited. For example... Figure 1 As shown, this embodiment includes: Step S11: Obtain the information to be filled in for several first work order items in the work order to be reviewed.

[0057] In this embodiment, the information filled in regarding several first work order items in the work order to be reviewed is obtained.

[0058] In one implementation, the work order to be reviewed is a battery malfunction after-sales work order. The battery can be a vehicle's power battery, etc., and is not specifically limited here.

[0059] In one specific implementation, pending power battery fault after-sales service orders can be obtained from the Global Service System (GSS), a digital service order management platform specifically for the power battery industry. After the power battery service station completes its after-sales work, it submits the power battery fault after-sales service order to the GSS system, and subsequently conducts a compliance review of the power battery fault after-sales service orders submitted to the GSS system.

[0060] In one implementation, the first work order item in the work order to be reviewed includes at least one of the following: fault level, fault attribute, faulty component, vehicle-level fault phenomenon, failure mode, maintenance code, and maintenance man-hours.

[0061] Step S12: Compare the reference information and the information filled in for each item in the first work order to obtain the review result of the first work order.

[0062] In this embodiment, the reference information and the information filled in for each first work order item are compared to obtain the first work order review result. The reference information for each first work order item is predicted using the target large language model. The first work order review result is used to characterize whether the information filled in for each first work order item is compliant. The target large language model is obtained by adjusting the network parameters of each training network layer of the pre-trained initial large language model using sample work orders. Each training network layer is a network layer added to each target network layer of the initial large language model.

[0063] The pre-trained initial large language model learns from massive amounts of general text data during the pre-training stage. However, this data may not contain professional knowledge, terminology, or compliance rules in the field of work order review. Therefore, the pre-trained initial large language model cannot understand the business logic, industry standards, or specific reporting requirements in the work order. Directly using the pre-trained initial large language model to predict the reference information of each first work order item results in low prediction accuracy. By adjusting the network parameters of the pre-trained initial large language model using sample work orders, domain knowledge of the work order review domain can be internalized into its parameters. This allows the adjusted initial large language model to learn domain knowledge of the work order review domain, thereby improving the adaptability of the adjusted initial large language model to the work order review task and optimizing its performance in the work order review task. In other words, by adjusting the network parameters of the pre-trained initial large language model using sample work orders, the adjusted initial large language model can predict the compliance information of each first work order item, or in other words, it can accurately predict the compliance information of each first work order item.

[0064] The target large language model is obtained by adjusting the network parameters of each training layer of the pre-trained initial large language model using sample work orders. Each training layer is a network layer added to the target network layer of the initial large language model. Adding training layers to each target network layer (transformer layer) of the initial large language model means adding trainable incremental parameters to each target network layer. In other words, without directly modifying the original parameters of the pre-trained initial large language model, additional trainable incremental parameters are introduced to specifically adjust the trained initial large language model, resulting in the target large language model. On the one hand, this allows the target large language model to retain general language understanding capabilities; on the other hand, it enables the target large language model to learn domain knowledge in the work order review field, improving its adaptability to work order review tasks and optimizing its performance in predicting compliant information submission—that is, improving the target large language model's ability to predict compliant information submission. In addition, the network parameters of the network layers to be trained are adjusted, that is, the parameters of the added trainable incremental parameters are adjusted to avoid the high computational cost of full parameter adjustment.

[0065] The target large language model is obtained by adjusting the network parameters of each training layer of the pre-trained initial large language model using sample work orders. Therefore, the target large language model has the ability to predict the compliance information of each first work order item, or in other words, it can accurately predict the compliance information of each first work order item. Therefore, the reference information of each first work order item in the work order to be reviewed, predicted by the target large language model, is the compliance information, or in other words, it can be regarded as a theoretical compliance template.

[0066] The reference information for each first work order item is predicted using the target large language model based on target fault-related data. Therefore, the reference information for each first work order item is the compliant reporting information for that first work order item, or it can be regarded as a theoretical compliant reporting template. Thus, comparing the reference information and the reported information for each first work order item is comparing the compliant reporting information with the actual reported information for each first work order item. This allows us to determine or verify the compliance of the reported information for each first work order item in the work order to be reviewed, thereby achieving accurate review of the work order to be reviewed.

[0067] In one implementation, if the reference information and the information filled in for the first work order are inconsistent, the information filled in for the first work order is determined to be non-compliant; if the reference information and the information filled in for the first work order are consistent, the information filled in for the first work order is determined to be compliant.

[0068] In one implementation, after adjusting the network parameters of each training network layer of the pre-trained initial large language model using sample work orders to obtain the target large language model, the target large language model is encapsulated, and then the encapsulated target large language model is deployed on the service cluster.

[0069] In one implementation, the work order to be reviewed is a battery fault after-sales work order. The sample work orders used to adjust the network parameters of each network layer to be trained in the pre-trained initial large language model can be historical work orders obtained through the GSS system.

[0070] There is no limit to the number of sample work orders used to adjust the network parameters of each training layer of the pre-trained initial large language model. For example, the number of sample work orders is 51822.

[0071] In one implementation, after adjusting the network parameters of each training layer of the pre-trained initial large language model using sample work orders to obtain the target large language model, the target large language model is tested using a test set.

[0072] There is no limit to the number of sample work orders in the test set. For example, the number of sample work orders in the test set is 292.

[0073] In one specific implementation, the work order to be reviewed is a battery fault after-sales work order, and the sample work orders in the test set can be historical work orders obtained through the GSS system.

[0074] In one embodiment, before adjusting the network parameters of each training layer of the pre-trained initial large language model using sample work orders to obtain the target large language model, the work order information of several first work order items in the sample work orders will be cleaned and standardized in sequence.

[0075] The cleaning of work order information for several first work order items in the sample work orders can include removing outliers, removing null values, removing meaningless values, etc., and is not limited here.

[0076] In one embodiment, the work order to be reviewed is a battery fault after-sales work order. The work order to be reviewed also includes target fault-related data, which includes at least one of the following: target fault description information and target fault handling results. The step of predicting the reference information of each first work order item using a target large language model can be: using the target large language model based on the target fault-related data to predict the reference information of each first work order item. The target large language model is obtained by adjusting the network parameters of each training layer of the pre-trained initial large language model using sample work orders. Therefore, the target large language model has the ability to predict the compliance reporting information of each first work order item, or in other words, it can accurately predict the compliance reporting information of each first work order item. Therefore, the reference information about each first work order item in the work order to be reviewed, predicted by using the target large language model based on the target fault-related data, is compliance reporting information, or in other words, it can be regarded as a theoretical compliance reporting template.

[0077] For example, taking the target fault-related data, including the target fault description information and the target fault handling result, as an example: The target fault description information and the target fault handling result can be specifically as follows: The vehicle instrument panel reports a fault, and the vehicle's power is limited; Our maintenance personnel arrive at the site, connect to the host computer, and the host computer has no fault; Feedback is sent to the external service, and the vehicle instrument panel reports a fault (B110) while the vehicle is in motion. The whole vehicle analysis is that the battery swapping connector is overheating. After the fault occurs, the whole vehicle's power is limited. The investigation found that there were burn marks on the battery swapping base connector and the vehicle's multi-in-one plug connection end. The battery swapping base connector needs to be replaced, and the multi-in-one connector on the whole vehicle side needs to be replaced simultaneously; After replacing the battery swapping base connector, the vehicle returns to normal, and this fault does not reappear while driving.

[0078] In one embodiment, several first work order items include first-type work order items and second-type work order items. The correlation between first-type work order items and target fault-related data is greater than that between second-type work order items and target fault-related data. Reference information for each first work order item is predicted based on the target fault-related data using a target large-scale language model. Specifically, the target fault-related data is input into the target large-scale language model to obtain the reference information for the first-type work order items output by the target large-scale language model. For first-type work order items with a high correlation to the target fault-related data, their corresponding reference information can be directly predicted using the target large-scale language model based on the target fault-related data.

[0079] In one specific implementation, the first type of work order item includes at least one of the following: fault level, fault attribute, faulty component, vehicle-level fault phenomenon, and failure mode.

[0080] For example, taking a power battery failure after-sales work order for a vehicle as an example: In the power battery failure after-sales work order for a vehicle, the information to be filled in for several first-class work order items can be as follows: Fault level - general fault, fault attribute - not originally C0, faulty component - Pack / wiring harness assembly / low voltage wiring harness, vehicle-level fault phenomenon - unable to power on, failure mode - detachment / loosening.

[0081] In one embodiment, several first work order items include first-type work order items and second-type work order items. The correlation between first-type work order items and target fault-related data is greater than that between second-type work order items and target fault-related data. Reference information for each first-type work order item is predicted based on the target fault-related data using a target large-scale language model. Specifically, the target fault-related data is input into the target large-scale language model to obtain target intermediate information output by the model, and reference information for second-type work order items is retrieved from the correlation information of the target intermediate information. For second-type work order items with weak correlation to the target fault-related data, directly predicting their corresponding reference information using the target large-scale language model based on the target fault-related data results in low accuracy. Therefore, the target large-scale language model is used to predict target intermediate information with strong correlation to both the reference information of second-type work order items and the target fault-related data, thereby using the target intermediate information to determine the reference information of second-type work order items and improving the accuracy of the determined reference information for second-type work order items.

[0082] In one specific implementation, the second type of work order item includes at least one of the following: maintenance code, maintenance hours.

[0083] In one specific implementation, the target intermediate information includes the components that need repair.

[0084] For example, taking a vehicle power battery fault after-sales work order as an example: the target intermediate information can be: Components to be repaired - Packing strap_16mm*0.8mm·PET plastic steel packing strap (for warehouse shipment)###White sticker###High voltage wiring harness between boxes. In the vehicle power battery fault after-sales work order, the information to be filled in for several second-type work order items can be as follows: Repair code - B0147, Repair time - 1 (1 indicates that the repair time is 1 hour, and the repair time is the standard working time for the repair action corresponding to the repair code); Repair code - B0175, Repair time - 1; Repair code - B0155, Repair time - 1.

[0085] In one specific implementation, the target large language model for predicting the reference information of the first type of work order item and the target large language model for predicting the target intermediate information are different large language models. The prediction of the reference information and target intermediate information of the first type of work order item is performed using corresponding dedicated models, thereby improving the prediction accuracy of the reference information and target intermediate information of the first type of work order item.

[0086] Specifically, the target large language model for predicting the reference information of the first type of work order item is the first large language model, which is obtained by adjusting the network parameters of each training layer of the pre-trained initial large language model using the first sample work order. The first sample work order is a battery fault after-sales work order, and the work order information of the first type of work order item in the first sample work order is used as the real information of the first type of work order item. The target large language model for predicting the target intermediate information is the second large language model, which is obtained by adjusting the network parameters of each training layer of the pre-trained initial large language model using the second sample work order. The second sample work order is a battery fault after-sales work order, and the second sample work order is marked with real intermediate information.

[0087] In other specific implementations, the target large language model for predicting the first type of work order item and the target large language model for predicting the target intermediate information are the same large language model. Since the target large language model for predicting the first type of work order item and the target large language model for predicting the target intermediate information are the same large language model, there is no need to repeatedly load multiple models, reducing inference latency and resource consumption.

[0088] Specifically, the target large language model for predicting the reference information of the first type of work order item and the target large language model for predicting the intermediate information of the target are the same large language model. It is obtained by adjusting the network parameters of each training network layer of the pre-trained initial large language model using sample work orders. The sample work order is a battery fault after-sales work order. The work order information of the first type of work order item in the sample work order is used as the real information of the first type of work order item. In addition, the sample work order is marked with real intermediate information.

[0089] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of adjusting the network parameters of each training layer of a pre-trained initial large language model using sample work orders, as provided in this application. It should be noted that if substantially the same result is obtained, this embodiment does not necessarily reflect that outcome. Figure 2 The illustrated process sequence is limited. For example... Figure 2 As shown, this embodiment includes: Step S21: Freeze the current network parameters of the initial large language model and add the network layer to be trained to each target network layer of the initial large language model.

[0090] In this embodiment, the current network parameters of the initial large language model are frozen, and the network layers to be trained are added to each target network layer of the initial large language model. Specifically, the current network parameters of the pre-trained initial large language model are maintained. W Alternatively, the original parameters remain unchanged, and a trainable network layer is added to each target network layer (transformer layer) of the initial large language model. That is, a trainable incremental parameter Δ is added to each target network layer of the initial large language model. W In other words, without directly modifying the original parameters of the pre-trained initial large language model, additional trainable incremental parameters are introduced to specifically adjust the trained initial large language model, thus obtaining the target large language model. On the one hand, this allows the target large language model to retain its general language understanding capabilities; on the other hand, it enables the target large language model to learn domain knowledge in the work order review field, improving its adaptability to work order review tasks and optimizing its performance in predicting compliant information submission—that is, enhancing its ability to predict compliant information submission. Furthermore, the current network parameters, or original parameters, of the pre-trained initial large language model remain unchanged and are not involved in parameter adjustment. Subsequent adjustments only affect the network parameters of the layers to be trained; that is, only the added trainable incremental parameters are adjusted, avoiding the high computational cost of full parameter adjustment.

[0091] The initial large language model after adding the network layers to be trained to each target network layer can be approximately represented as:

[0092] in, h This represents the feature output of each target network layer after being added to the network layer to be trained; W 0 represents the current network parameters of the initial large language model, including W. q W k W v ;Δ W This represents the network parameters of the target network layer corresponding to the network layer to be trained, specifically the incremental parameter matrix, or it can also be called the weight matrix.

[0093] Step S22: Using the initial large language model, based on the sample fault-related data in the sample work orders, predict the sample information of several first work order items.

[0094] In this embodiment, the initial large language model is used to predict sample information of several first work order items based on sample fault-related data in the sample work orders.

[0095] Step S23: Based on the difference between the sample information of several first work orders and the real information of several first work orders, adjust the network parameters of each network layer to be trained.

[0096] In this embodiment, the network parameters of each training network layer are adjusted based on the differences between sample information and real information of several first work order items. By adjusting the network parameters of each training network layer using these differences, domain knowledge of the work order review domain can be internalized into its parameters. This allows the model to learn domain knowledge related to work order review, thereby improving the adaptability of the target large language model obtained through training convergence to the work order review task and optimizing its performance in predicting compliant information. In other words, it enhances the target large language model's ability to predict compliant information, thus improving the accuracy of work order review.

[0097] In one embodiment, the network parameters of the network layers to be trained are the rank decomposition matrices corresponding to the network layers to be trained. Based on the differences between the sample information and the true information of several first work items, the network parameters of each network layer to be trained are adjusted. Specifically, the rank decomposition matrices of each network layer to be trained are adjusted based on the differences between the sample information and the true information of several first work items. By performing low-rank decomposition on the network parameters of each network layer to be trained to obtain the rank decomposition matrix, and using the rank decomposition matrix corresponding to each network layer to be trained as the network parameters of each network layer, the number of network parameters of each network layer to be trained can be significantly reduced. Fewer parameters mean faster update speed, shorter iteration cycle, and faster fitting of the pre-trained initial large language model to work item data.

[0098] Wherein, with the network parameters of each network layer to be trained being the rank decomposition matrix corresponding to each network layer to be trained, the initial large language model can be approximately represented as:

[0099] Where h represents the features output by each target network layer after being added to the network layer to be trained; W0 represents the current network parameters of the initial large language model, including Wq, Wk, and Wv; ΔW represents the network parameters of the network layer to be trained corresponding to each target network layer, specifically the incremental parameter matrix, or weight matrix; BA represents the low-rank adaptive weight matrix of ΔW, that is, the two rank decomposition matrices corresponding to each network layer to be trained. The rank decomposition matrix B has a dimension of d×r and is initialized to all zeros, and the rank decomposition matrix A has a dimension of r×d and is initialized using a Gaussian function. r represents the rank of the matrix, which is much smaller than d.

[0100] It should be noted that if the incremental parameter matrix ΔW is a full-rank matrix (r=d), its parameter count is the number of all elements in the matrix, and the parameter count = d². The incremental parameter matrix ΔW is decomposed into two rank decomposition matrices B∈Rd×r and A∈Rr×d, and the parameter count = 2dr. When r is much smaller than d, the parameter count decreases from d² to 2dr, and the parameter count is greatly reduced, achieving the effect of approximating the full-rank d² with a parameter count of 2dr.

[0101] In one embodiment, the network parameters of the network layer to be trained are the rank decomposition matrix corresponding to the network layer to be trained. The rank decomposition matrix corresponding to the network layer to be trained is obtained by decomposing the weight matrix corresponding to the network layer to be trained. At least two network layers to be trained have different matrix ranks. The matrix rank corresponding to the network layer to be trained is determined based on the compatibility between the weight matrix of the network layer to be trained and the work order review task.

[0102] The rank decomposition matrix corresponding to each network layer to be trained is obtained by decomposing the weight matrix corresponding to each network layer to be trained. In other words, the rank decomposition matrix corresponding to each network layer to be trained is obtained by decomposing the weight matrix using the rank of the matrix corresponding to the network layer to be trained. If the rank of the matrices corresponding to each network layer to be trained is the same, that is, if each network layer to be trained uses a fixed rank, it will lead to overfitting of the adjusted initial large language model, i.e., the target large language model, resulting in poor model expressive ability. Therefore, based on the fit between the weight matrix of each network layer to be trained and the work order review task, the rank of the matrix corresponding to each network layer to be trained is dynamically determined, and a matching rank is assigned to each network layer. That is, an improved low-rank adaptation technique is used to adjust the network parameters of each network layer to be trained in the pre-trained initial large language model.

[0103] It should be noted that each network layer to be trained uses a fixed matrix rank, which means that the matrix rank of each network layer to be trained is the same.

[0104] In one specific implementation, the fit between the weight matrix of the network layer to be trained and the work order review task is represented by the norm of the gradient of the weight matrix of the network layer to be trained. The steps for determining the rank of the matrix corresponding to the network layer to be trained are as follows: obtain the product between the norm of the network layer to be trained and the first coefficient; sum the product with the second coefficient to obtain the first sum; round the first sum to obtain the second sum, and obtain the rank of the matrix corresponding to the network layer to be trained based on the second sum. The first coefficient is a scaling factor. By setting the first coefficient, the contribution ratio of the norm of the network layer to the rank of the matrix can be adjusted. The larger the first coefficient, the more significant the impact on the rank of the matrix. The second coefficient is an offset term. By setting the second coefficient, the rank of the matrix is ​​prevented from returning to zero when the norm of the network layer to be trained is too small, ensuring the minimum rank of the matrix.

[0105] There are no restrictions on the magnitudes of the first and second coefficients; they can be set according to actual usage needs. For example, the first coefficient can be 0.05, 0.1, etc., and the second coefficient can be 2, 4, etc.

[0106] In one specific implementation, the second sum is an integer not greater than the first sum. That is, when the first sum is not an integer, the first sum is rounded down to obtain the second sum.

[0107] In one specific implementation, the matrix rank corresponding to the network layer to be trained is obtained based on the second sum. Specifically: in response to the second sum being greater than or equal to a rank threshold, the rank threshold is used as the matrix rank corresponding to the network layer to be trained; in response to the second sum being less than the rank threshold, the second sum is used as the matrix rank corresponding to the network layer to be trained. The rank threshold is the upper limit of the matrix rank of the network layer to be trained. That is, the maximum matrix rank of the network layer to be trained can only be the rank threshold. Setting the upper limit of the matrix rank limits the number of trainable parameters for each network layer to be trained, avoiding the introduction of too many parameters and ensuring lightweight characteristics. In addition, by setting the upper limit of the matrix rank, the maximum complexity of each network layer to be trained is limited, forcing each network layer to learn the most critical features using a limited parameter capacity. Furthermore, the norm of the gradient of the weight matrix of some network layers to be trained may have instantaneous spikes in the early stage of training, leading to the dynamic calculation of abnormally high matrix ranks and causing oscillations in the training process. Therefore, by setting the upper limit of the matrix rank, rank explosion caused by instantaneous gradient anomalies can be avoided, making the training process smoother and more predictable.

[0108] The rank threshold is not limited in size and can be set according to actual usage needs. For example, the rank threshold can be 64.

[0109] In one specific implementation, the first coefficient of the first stage of adjustment is smaller than the first coefficient of the second stage of adjustment, and the first stage precedes the second stage. That is, the first coefficient in the initial adjustment phase is smaller than the first coefficient in the later adjustment phase. In the initial adjustment phase, the network parameters of each training layer of the initial large language model are adjusted from the pre-training state, and their gradients are usually large. A small first coefficient is set to avoid assigning an excessively large matrix rank due to high gradients in the initial adjustment phase, thereby preventing training oscillations caused by an excessively large matrix rank. In the later adjustment phase, the gradients of the network parameters of each training layer of the initial large language model gradually decrease, and the model approaches convergence. A large first coefficient is set to increase the influence of the gradient on the matrix rank, enabling the initial large language model to learn detailed information.

[0110] The first coefficient for the first stage of adjustment can be 0.05, and the first coefficient for the second stage of adjustment can be 0.1.

[0111] In one specific embodiment, the second coefficient of the first adjustment stage is equal to the second coefficient of the second adjustment stage. The second coefficients of both the first and second adjustments can be 4.

[0112] In one specific implementation, the formula for calculating the rank of the matrix corresponding to each network layer to be trained is as follows:

[0113] in, Indicates the first l The rank of the matrix corresponding to the layer of the network to be trained; This represents the rank threshold, which is the upper limit of the matrix rank. Indicates the first l The norm of the network layer to be trained; α Indicates the first coefficient; β This indicates the second coefficient.

[0114] In one specific implementation, the step of determining the norm corresponding to the network layer to be trained is as follows: Each matrix parameter of the weight matrix of the network layer to be trained is taken as a current parameter, and the partial derivative of the work order loss with respect to the current parameter is obtained as the partial derivative component of the current parameter; wherein, the work order loss is determined based on the difference between the sample information of several first work order items and the actual work order information of several first work order items; the squares of the absolute values ​​of the partial derivative components corresponding to each matrix parameter are summed to obtain a third sum; the square root of the third sum is taken to obtain the norm corresponding to the network layer to be trained. The specific formula is as follows:

[0115] in, Indicates the first l The norm of the network layer to be trained; Indicates the first l The gradient of the network parameters—weight matrix—of the network layer to be trained is a matrix, so it can also be called the weight gradient matrix. The elements in the i-th row and j-th column of the weight gradient matrix represent the partial derivatives of the work order loss with respect to the matrix parameters in the i-th row and j-th column of the weight matrix; m and n represent the dimensions of the weight gradient matrix. This indicates the third total.

[0116] It should be noted that the gradient of the network parameters of the network layer to be trained is the partial derivative of the work order loss (quantifying the difference between the model prediction and the true label) with respect to the network parameters of the network layer to be trained, reflecting the sensitivity of the network parameters of the network layer to the work order review task loss.

[0117] A large gradient in the network parameters of the layer to be trained indicates that even small changes in the network parameters significantly alter the loss value. This suggests that the layer has a significant impact on the work order review task, further indicating that it is not fully adapted to the task. Therefore, more parameters are needed to model complex relationships, requiring a high matrix rank, as a high matrix rank brings a large number of parameters. Conversely, a small gradient in the network parameters of the layer to be trained indicates that changes in the network parameters have a weak impact on the work order loss. This suggests that the layer's functionality is largely adapted to the work order review task. This is because the underlying target network layer primarily performs basic semantic recognition, the patterns are universal, and the pre-trained knowledge can be directly reused, requiring fewer parameters to be tuned. Therefore, only a lower matrix rank is needed.

[0118] The weight gradient matrix is ​​a matrix containing numerous partial derivative components. A single gradient component only reflects local sensitivity and cannot directly measure the importance of the entire layer. The norm quantifies the total energy of all gradient components in the weight gradient matrix, reflecting the overall sensitivity of the layer's parameters.

[0119] Overfitting refers to the initial large language model excessively memorizing details (including noise) from the training data, leading to a significant drop in performance on new data. (For example, in simple tasks, the actual required matrix rank might be 4, but fixing it to 8 introduces redundant parameters. These redundant parameters might learn random noise in the training data, causing the initial large language model to incorrectly associate irrelevant fields in new work orders.) For simple tasks, network layers handle basic semantics, which can be covered with a small number of parameters. Redundant parameters easily introduce noise, so only a lower matrix rank is needed. For complex tasks, network layers require comprehensive reasoning and diverse logical combinations, necessitating more parameters to capture feature interactions and domain rules. Therefore, a high matrix rank is required to represent complex feature interactions.

[0120] In one specific implementation, the weight matrix corresponding to the network layer to be trained is obtained by adjusting the initial weight matrix using a third coefficient, which is the ratio of the hyperparameter to the rank of the matrix corresponding to the network layer to be trained. In other words, to control the impact of the initial weight matrix corresponding to each network layer to be trained on the initial large language model, hyperparameters are introduced to improve the prediction accuracy of the compliant reporting information of the large language model.

[0121] The initial large language model can be approximated as:

[0122] in, h This represents the feature output of each target network layer after being added to the network layer to be trained; W 0 represents the current network parameters of the initial large language model, including W. q W k W v ;ΔW γ represents the network parameters of the target network layer corresponding to the network layer to be trained—the initial weight matrix; γ / r represents the third coefficient corresponding to each network layer to be trained, where γ represents the hyperparameter, r represents the matrix rank corresponding to each network layer to be trained, and γ can be 15, 16, etc., without limitation here.

[0123] It should be noted that, in the formula, the larger the third coefficient, the larger the weight matrix corresponding to each network layer to be trained, and the easier it is for the initial large language model to overfit in the work order review task; the smaller the third coefficient, the smaller the weight matrix corresponding to each network layer to be trained, the less obvious the adjustment effect on the initial large language model, and the initial large language model underfits.

[0124] In other specific implementations, the initial weight matrix corresponding to the network layer to be trained can also be directly used as the weight matrix of the network layer to be trained, which is not limited here.

[0125] In one specific implementation, during the first stage of adjustment, the initial large language model adjusts the matrix rank of each network layer to be trained once every first number of sample work orders processed. During the second stage of adjustment, the initial large language model adjusts the matrix rank of each network layer to be trained once every second number of sample work orders processed. The first number is less than the second number, and the first stage precedes the second stage. In the initial adjustment phase, the network parameters of each network layer to be trained are just beginning to adapt to the work order review task from the pre-training state. The norm amplitude of each network layer to be trained is large, and the direction changes drastically. If the adjustment frequency of the matrix rank is too low, the large language model may have been exploring in the wrong parameter space for a long time, making convergence difficult or even causing divergence. Therefore, adjusting the matrix rank of each network layer to be trained once every first number of sample work orders processed in the initial adjustment phase enables the initial large language model to converge stably. In the later stages of adjustment, the initial large language model has initially converged, the norm magnitude of each network layer to be trained is small and the direction is stable, and the matrix rank configuration of each network to be trained has tended to be optimal. At this time, the decision space of the matrix rank configuration is reduced, and there is no need to frequently re-evaluate, thus reducing unnecessary computational overhead.

[0126] When adjusting the initial large language model, the computational complexity of the initial large model in the Transformer architecture has a quadratic relationship with the length of the text sequence. Longer text sequences lead to increased computation time and resources, affecting training efficiency. Furthermore, longer text sequences contain a large amount of redundant information, impacting the generalization ability of the initial large language model. Therefore, in one implementation, a preset proportional quantile is used as the maximum truncation length (MAX LENGTH) of the initial large language model input. By truncating longer text, the initial large language model is not only forced to focus on more important parts and avoid being disturbed by irrelevant content, but also the waste of GPU resources is reduced, improving the training efficiency and performance of the initial large language model.

[0127] The preset ratio can be 95%, etc., and is not limited here.

[0128] Please see Figure 3 , Figure 3 This is a flowchart illustrating another embodiment of the work order review method provided in this application. It should be noted that if substantially the same result is achieved, this embodiment does not necessarily reflect that outcome. Figure 3 The illustrated process sequence is limited. For example... Figure 3 As shown, the work order to be reviewed also includes information for several second work order items. In this embodiment, these include: Step S31: Obtain the target image associated with each second work order item.

[0129] In this embodiment, target images associated with each second work order item are obtained; wherein, the target images contain reference information for the second work order item. The information entered for the second work order item is manually input, which may contain subjective description biases or information omissions. However, the target images associated with the second work order item contain the reference information for that second work order item; therefore, the target images associated with the second work order item objectively record the reference information for that second work order item. Therefore, by subsequently comparing the entered information for the second work order item with the reference information obtained from recognizing the target images associated with the second work order item, it can be determined whether the entered information for the second work order item is compliant. That is, introducing image technology into work order compliance review, and using multimodal data fusion for work order compliance review, can improve the accuracy of work order review.

[0130] Step S32: Identify the target images associated with each second work order item to obtain reference information for each second work order item.

[0131] In this embodiment, the target images associated with each second work order are identified to obtain reference information for each second work order.

[0132] In one embodiment, the target image associated with each second work order can be directly identified to obtain reference information for each second work order.

[0133] In other embodiments, the target images associated with each second work order item are identified separately to obtain reference information for each second work order item. Specifically, this involves extracting the region image of the area where the reference information of each second work order item is located from the target images associated with each second work order item; and identifying each region image to obtain the reference information for each second work order item. The target images associated with the second work order item may contain irrelevant text, shadows, or other noise, and direct full-image recognition would severely interfere with the recognition results. Therefore, by first identifying the region where the reference information of the second work order item is located, and then recognizing the region image of that region, the recognition algorithm only needs to process high-value regions, avoiding noise interference and improving recognition accuracy. Furthermore, by only recognizing the region image of the area where the reference information of the second work order item is located, the computational load is reduced, and computational efficiency is improved.

[0134] In one specific implementation, the target image associated with at least one second work order is a first-type image, and the reference information in the first-type image is a barcode representing the reference information. In this case, the region image of the area where the reference information of each second work order is located is extracted from the target images associated with each second work order. Specifically, this involves: determining the vertical projection value of each pixel row and the horizontal projection value of each pixel column in the first-type image; rotating the first-type image until the current vertical and horizontal projection values ​​of the first-type image meet preset projection requirements; and using the current vertical and horizontal projection values ​​of the first-type image, extracting the region image corresponding to the barcode. In barcode recognition, if the current vertical and horizontal projection values ​​of the first-type image meet the preset projection requirements, it indicates that the barcode is in an ideal state or is not tilted, and the barcode is arranged vertically or horizontally. Therefore, the current vertical and horizontal projection values ​​of the first-type image can be used to accurately locate the barcode, thereby accurately extracting the region image corresponding to the barcode.

[0135] The formulas for calculating the vertical projection value of each pixel row and the horizontal projection value of each pixel column in the first type of image are as follows:

[0136]

[0137] in, I (x,y) represents the gray value at coordinates (x,y); I (x, y+1) represents the gray value at coordinates (x, y+1); I (x+1,y) represents the gray value at coordinates (x+1,y); P x(y) represents the horizontal projection value of the x-th pixel column, which is the sum of the absolute differences of all columns between the pixel row with y coordinate and the pixel row with y+1 coordinate; P y (x) represents the vertical projection value of the y-th pixel row, which is the sum of the absolute differences between all rows between the pixel column with horizontal coordinate x and the pixel column with horizontal coordinate x+1.

[0138] In one specific implementation, the preset projection requirements are: in each pixel row, there exists a vertical projection value greater than a first projection threshold; and in each pixel column, there exists a horizontal projection value greater than a second projection threshold. Ideally, the barcode consists of bars and spaces arranged perfectly vertically or horizontally. Therefore, the projected image of the barcode will show obvious peaks in certain specific rows and / or columns. The presence of vertical projection values ​​greater than the first projection threshold in each pixel row and horizontal projection values ​​greater than the second projection threshold in each pixel column indicates that the projected image of the barcode will show obvious peaks in certain specific rows and columns. This signifies that the barcode is perfectly vertically or horizontally arranged, meaning the barcode is in an ideal state or is not tilted. The first and second projection thresholds are not limited in magnitude and can be set according to actual usage needs.

[0139] However, once a tilt occurs, the difference operation according to the above formula will no longer be calculated for a single bar or space, but will involve mixed information spanning multiple bars or spaces. For example, if a barcode is tilted at 45°, pixels that originally belonged to different bars or spaces will be treated as continuous during projection calculation, resulting in... P x (y) and P y The result of (x) is no longer a clear spike. Therefore, to determine that the first type of image is tilted, it is necessary to rotate the first type of image until the current vertical projection value and horizontal projection value of the first type of image show obvious peaks in certain specific rows and / or columns. In this way, through external intervention, the barcode is rotated and corrected from the tilted state to the ideal state, so that the barcode can be located even when it is not tilted, thus improving the positioning accuracy of the barcode.

[0140] In one specific embodiment, the first type of image is rotated until the current vertical projection value and horizontal projection value of the first type of image meet the preset projection requirements. Specifically, in response to the absence of vertical projection values ​​and horizontal projection values ​​greater than the projection threshold, the first type of image is rotated by a preset angle, and the steps of determining the vertical projection value of each pixel row and the horizontal projection value of each pixel column in the first type of image and subsequent steps are re-executed until there are vertical projection values ​​and horizontal projection values ​​greater than the projection threshold. After determining the vertical projection values ​​of each pixel row and the horizontal projection value of each pixel column in the first type of image, in response to the presence of vertical projection values ​​and horizontal projection values ​​greater than the projection threshold, according to the horizontal direction, the first and second ordered vertical projection values ​​corresponding to the vertical projection values ​​greater than the projection threshold are found from the vertical projection values ​​of each pixel row, and these are respectively used as the upper and lower boundaries of the barcode. And, according to the vertical direction, the third and fourth ordered horizontal projection values ​​corresponding to the horizontal projection values ​​greater than the projection threshold are found from the horizontal projection values ​​of each pixel column, and these are respectively used as the left and right boundaries of the barcode.

[0141] After determining the vertical projection values ​​of each pixel row and the horizontal projection values ​​of each pixel column in the first type of image, it is first determined whether the vertical projection values ​​of each pixel row and the horizontal projection values ​​of each pixel column in the first type of image meet the preset projection requirements. If the vertical projection values ​​of each pixel row and the horizontal projection values ​​of each pixel column in the first type of image meet the preset projection requirements, it means that the barcode is currently in an ideal state or is not tilted. The barcode is currently arranged in a completely vertical or horizontal manner. Therefore, the vertical projection values ​​and horizontal projection values ​​of the first type of image can be directly used to accurately locate the barcode, thereby accurately extracting the area image corresponding to the barcode without rotating the first type of image.

[0142] If the vertical projection values ​​of each pixel row and the horizontal projection values ​​of each pixel column in the first determined image do not meet the preset projection requirements, it means that the barcode is currently tilted. Therefore, the first image needs to be rotated until the current vertical and horizontal projection values ​​of the first image meet the preset projection requirements. Through external intervention, the barcode can be rotated and corrected from the tilted state to the ideal state, thereby enabling the barcode to be located even when it is not tilted, thus improving the accuracy of barcode positioning.

[0143] The process involves rotating the first type of image in steps of a preset angle. After each rotation, it is determined whether the current vertical and horizontal projection values ​​of the first type of image meet the preset projection requirements. This process continues until the current vertical and horizontal projection values ​​of the first type of image meet the preset projection requirements. At this point, the rotation of the first type of image is stopped, and subsequent steps are executed. Alternatively, the rotation of the first type of image is stopped until the rotation angle covers all possible angles. In this case, the barcode is considered invalid, and subsequent steps are not executed.

[0144] The preset angle can be 15°, and the rotation angle of the first type of image can cover 0° to 345°, with no blind spots.

[0145] In one specific implementation, the target image associated with at least one second work order is a first type of image, and the reference information in the first type of image is a barcode representing the reference information. After extracting the region image corresponding to the barcode, the ZXing algorithm is used to decode the region image corresponding to the barcode, and the reference information represented by the barcode can be parsed out.

[0146] In one specific implementation, the target image associated with at least one second work order is a second type of image, and the reference information in the second type of image is text information representing the reference information. In this case, the region image of the area where the reference information of each second work order is located is extracted from the target images associated with each second work order. Specifically, a detection model is used to perform text detection on the second type of image to obtain several text regions and their corresponding region categories. The several text regions are then matched with the reference information in the second type of image for category selection, and the text regions whose region categories match the categories corresponding to the reference information in the second type of image are taken as the region images corresponding to the reference information in the second type of image. The detection model can detect several text regions in the second type of image and determine the region category of each text region. The region category corresponding to the text region reflects the category of the reference information present in the text region. Therefore, by using the region category corresponding to the text region, the region where the reference information of the second work order is located can be locked. In addition, it should be emphasized that the subsequent process involves identifying the region image where the reference information for the second work order item is located. Therefore, the identified reference information will necessarily be the reference information for the second work order item, without the need for further confirmation of whether the identified reference information is indeed the reference information for the second work order item. This improves the accuracy and efficiency of identifying the reference information for the second work order item, thereby improving the accuracy and efficiency of the review of the second work order item.

[0147] The detection model can be the YOLOv10 model. Additionally, the second work order item for the associated target image being a second-class image can be the BMU version number, HVB serial number, MBMU version number, SBMU version number, IMM encoding, VIN code, CSC version number, etc., without limitation.

[0148] In one specific implementation, the target image associated with at least one second work order is a second type of image. The reference information in the second type of image is text information representing the reference information. Before recognizing each region image to obtain the reference information for each second work order, the second type of image is also corrected. The reference information in the second type of image is text information representing the reference information. These text images often have various geometric distortions, resulting in low accuracy for direct recognition. By correcting the region images, the distortions are eliminated, restoring the region images to a standard format, creating optimal input conditions for subsequent region image recognition, thereby improving the accuracy of the reference information for the recognized second work order.

[0149] In one specific implementation, the target image associated with at least one second work order item is a second type of image, and the reference information in the second type of image is text information representing the reference information. Specifically, the Text Recognition STVRv2 model can be used to identify the region image corresponding to the second work order item to obtain the reference information of the second work order item.

[0150] In one specific implementation, the work order to be reviewed is a battery fault after-sales work order. The second work order item of the work order to be reviewed is the BMU version number. The detection model is trained using the BMU version number detection dataset. The detection dataset is the battery management system interface diagram in the GSS system work order attachment, which is labeled using the Labelme tool.

[0151] There is no limit to the number of images in the BMU version number detection dataset; for example, 468 images.

[0152] In one specific implementation, the converged detection model can also be validated using a BMU version number validation dataset. There is no limit to the number of images in the BMU version number validation dataset; for example, 58 images.

[0153] In one specific implementation, the converged detection model can also be tested using a BMU version number test dataset. There is no limit to the number of images in the BMU version number test dataset; for example, 58 images.

[0154] In one specific implementation, after training the detection model using the BMU version number detection dataset, the detection model is validated using the BMU version number verification dataset and tested using the BMU version number test dataset; then, the detection model is encapsulated and deployed on the service cluster.

[0155] Step S33: Compare the reference information and the information filled in for the second work order to obtain the review result of the second work order.

[0156] In this embodiment, the reference information and the submitted information of the second work order item are compared to obtain the second work order review result; wherein, the second work order review result is used to characterize whether the submitted information of each second work order item is compliant. The reference information of the second work order item is obtained by recognizing the target image associated with the second work order item, therefore, the reference information of the second work order item is the objectively recorded reference information of the target image associated with the second work order item. Therefore, by comparing the submitted information of the second work order item with the reference information obtained by recognizing the target image associated with the second work order item, it can be determined whether the submitted information of the second work order item is compliant. That is, by introducing image technology into the work order compliance review and by fusing multimodal data for work order compliance review, the accuracy of work order review can be improved.

[0157] In one implementation, the work order to be reviewed is displayed on the display interface. For non-compliant work order items, the display color of the non-compliant work order items and / or the information filled in the work order items is adjusted. By adjusting the display color of the non-compliant work order items and / or the information filled in the work order items, the non-compliant parts of the work order to be reviewed are highlighted.

[0158] The adjusted display color is not limited; for example, the adjusted display color is red. For instance, taking a battery-related after-sales service work order awaiting review as an example, the non-compliant work order items and their corresponding information will be highlighted in red on the GSS system interface.

[0159] In one implementation, the work order to be reviewed is displayed on the interface. For non-compliant work order items, the non-compliant work order items and / or the information filled in for the work order items are highlighted. By highlighting the non-compliant work order items and / or the information filled in for the work order items, the non-compliant parts of the work order to be reviewed are highlighted.

[0160] Please see Figure 4 , Figure 4 This is a schematic diagram of an embodiment of the work order review device provided in this application. The work order review device 40 includes an acquisition module 41 and an review module 42; the acquisition module 41 is used to acquire the information filled in for several first work order items in the work order to be reviewed; the review module 42 is used to compare the reference information and the information filled in for each first work order item to obtain the review result of the first work order; wherein, the reference information for each first work order item is predicted using a target large language model, the review result of the first work order is used to characterize whether the information filled in for each first work order item is compliant, the target large language model is obtained by adjusting the network parameters of each training network layer of the pre-trained initial large language model using sample work orders, and each training network layer is a network layer added to each target network layer of the initial large language model.

[0161] Among them, the above-mentioned work orders to be reviewed are battery fault after-sales work orders. The work orders to be reviewed also include target fault-related data, which includes at least one of the following: target fault description information and target fault handling results. The review module 42 is used to predict the reference information of each first work order item using the target large language model, including: using the target large language model to predict the reference information of each first work order item based on the target fault-related data.

[0162] The aforementioned first work order items include first-type work order items and second-type work order items. The correlation between the first-type work order items and the target fault-related data is greater than that between the second-type work order items and the target fault-related data. The review module 42 is used to predict the reference information of each first work order item based on the target fault-related data using the target large language model, including: inputting the target fault-related data into the target large language model to obtain the reference information of the first-type work order items output by the target large language model; and / or, inputting the target fault-related data into the target large language model to obtain the target intermediate information output by the target large language model, and finding the reference information of the second-type work order items from the correlation information of the target intermediate information.

[0163] The first type of work order item mentioned above includes at least one of the following: fault level, fault attribute, faulty component, vehicle-level fault phenomenon, failure mode; and / or, the second type of work order item mentioned above includes at least one of the following: repair code, repair time; and / or, the target intermediate information mentioned above includes the component to be repaired.

[0164] Among them, the target large language model for predicting the reference information of the first type of work order item and the target large language model for predicting the intermediate information of the target are different or the same large language model.

[0165] The work order review device 40 also includes an adjustment module 43. The adjustment module 43 is used to adjust the network parameters of each training network layer of the pre-trained initial large language model using sample work orders. The adjustment module 43 includes: freezing the current network parameters of the initial large language model and adding the training network layer to each target network layer of the initial large language model; using the initial large language model to predict sample information of several first work order items based on sample fault-related data in the sample work orders; and adjusting the network parameters of each training network layer based on the difference between the sample information of several first work order items and the real information of several first work order items.

[0166] Among them, the network parameters of the network layer to be trained are the rank decomposition matrix of the network layer to be trained. The rank decomposition matrix of the network layer to be trained is obtained by decomposing the weight matrix of the network layer to be trained. At least two network layers to be trained have different matrix ranks. The matrix rank of the network layer to be trained is determined based on the compatibility between the weight matrix of the network layer to be trained and the work order review task.

[0167] The adaptation between the weight matrix of the network layer to be trained and the work order review task is represented by the norm of the gradient of the weight matrix of the network layer to be trained. The adjustment module 43 is used to determine the matrix rank corresponding to the network layer to be trained, including: obtaining the product between the norm of the network layer to be trained and the first coefficient; summing the product and the second coefficient to obtain the first sum; rounding the first sum to obtain the second sum, and obtaining the matrix rank corresponding to the network layer to be trained based on the second sum.

[0168] Wherein, the second sum is an integer not greater than the first sum; and / or, the adjustment module 43 is used to obtain the matrix rank corresponding to the network layer to be trained based on the second sum, including: in response to the second sum being greater than or equal to the rank threshold, using the rank threshold as the matrix rank corresponding to the network layer to be trained; in response to the second sum being less than the rank threshold, using the second sum as the matrix rank corresponding to the network layer to be trained.

[0169] Wherein, the first coefficient of the first stage of the above adjustment is less than the first coefficient of the second stage of the adjustment, and the first stage is earlier than the second stage; and / or, the second coefficient of the first stage of the above adjustment is equal to the second coefficient of the second stage of the adjustment.

[0170] The adjustment module 43 is used to determine the norm of the network layer to be trained, including: taking each matrix parameter of the weight matrix of the network layer to be trained as the current parameter, obtaining the partial derivative of the work order loss with respect to the current parameter, and taking it as the partial derivative component of the current parameter; wherein, the work order loss is determined based on the difference between the sample information of several first work order items and the real work order information of several first work order items; summing the squares of the absolute values ​​of the partial derivative components corresponding to each matrix parameter to obtain the third sum; taking the square root of the third sum to obtain the norm of the network layer to be trained.

[0171] The weight matrix corresponding to the network layer to be trained is obtained by adjusting the initial weight matrix using a third coefficient, which is the ratio of the hyperparameter to the rank of the matrix corresponding to the network layer to be trained; and / or, in the first stage of adjustment, the initial large language model adjusts the rank of the matrix corresponding to each network layer to be trained once for every first number of sample work orders processed, and in the second stage of adjustment, the initial large language model adjusts the rank of the matrix corresponding to each network layer to be trained once for every second number of sample work orders processed, wherein the first number is less than the second number, and the first stage is earlier than the second stage.

[0172] The aforementioned work order to be reviewed also includes information to be filled in for several second work order items; the review module 42 is also used to obtain the target image associated with each second work order item; wherein the target image contains reference information for the second work order item; the target image associated with each second work order item is identified to obtain the reference information for each second work order item; the reference information and the information to be filled in for the second work order item are compared to obtain the review result of the second work order; wherein the review result of the second work order is used to characterize whether the information to be filled in for each second work order item is compliant.

[0173] The review module 42 is used to identify the target images associated with each second work item to obtain reference information for each second work item, including: extracting the region image of the area where the reference information of each second work item is located from the target images associated with each second work item; and identifying each region image to obtain the reference information for each second work item.

[0174] Among them, at least one target image associated with a second work order is a first type of image, and the reference information in the first type of image is a barcode representing the reference information; the review module 42 is used to extract the region image of the area where the reference information of each second work order is located from the target image associated with each second work order, including: determining the vertical projection value of each pixel row and the horizontal projection value of each pixel column in the first type of image; rotating the first type of image until the current vertical projection value and horizontal projection value of the first type of image meet the preset projection requirements, and using the current vertical projection value and horizontal projection value of the first type of image to extract the region image corresponding to the barcode.

[0175] The aforementioned preset projection requirements are: among the vertical projection values ​​of each pixel row, there exists a vertical projection value greater than a first projection threshold; and among the horizontal projection values ​​of each pixel column, there exists a horizontal projection value greater than a second projection threshold. And / or, the review module 42 is used to rotate the first type of image until the current vertical and horizontal projection values ​​of the first type of image meet the preset projection requirements, including: in response to the absence of vertical and horizontal projection values ​​greater than the projection threshold, rotating the first type of image by a preset angle, and re-executing the steps for determining the vertical projection values ​​of each pixel row and the horizontal projection values ​​of each pixel column in the first type of image, until a value greater than the projection threshold is found. Vertical projection values ​​and horizontal projection values; the review module 42, after determining the vertical projection values ​​of each pixel row and the horizontal projection values ​​of each pixel column in the first type of image, includes: responding to the existence of vertical projection values ​​and horizontal projection values ​​greater than the projection threshold, and according to the horizontal direction, finding the first and second sequential vertical projection values ​​corresponding to the vertical projection values ​​greater than the projection threshold from the vertical projection values ​​of each pixel row, respectively serving as the upper and lower boundaries of the barcode; and, according to the vertical direction, finding the third and fourth sequential horizontal projection values ​​corresponding to the horizontal projection values ​​greater than the projection threshold from the horizontal projection values ​​of each pixel column, respectively serving as the left and right boundaries of the barcode.

[0176] Wherein, at least one target image associated with a second work order is a second type of image, and the reference information in the second type of image is text information representing the reference information; wherein, the review module 42 is used to extract the region image of the area where the reference information of each second work order is located from the target image associated with each second work order, including: using a detection model to perform text detection on the second type of image to obtain several text regions and corresponding region categories; matching the categories of the several text regions with the reference information in the second type of image respectively, and taking the text region whose region category matches the category corresponding to the reference information in the second type of image as the region image corresponding to the reference information in the second type of image; and / or, the review module 42 is used to correct the region image before recognizing each region image to obtain the reference information of each second work order.

[0177] The aforementioned work orders pending review are used to display on the display interface; the review module 42 is also used to perform at least one of the following steps on non-compliant work order items: adjust the display color of the non-compliant work order items and / or the information filled in the work order items; highlight the non-compliant work order items and / or the information filled in the work order items.

[0178] Please see Figure 5 , Figure 5This is a schematic diagram of an embodiment of the electronic device provided in this application. The electronic device 50 includes a memory 51 and a processor 52 coupled to each other. The processor 52 is used to execute program instructions stored in the memory 51 to implement the steps of any of the above-described work order review method embodiments. In a specific implementation scenario, the electronic device 50 may include, but is not limited to, a microcomputer or a server. In addition, the electronic device 50 may also include mobile devices such as laptops and tablets, which are not limited here.

[0179] Specifically, processor 52 controls itself and memory 51 to implement the steps of any of the above-described work order review method embodiments. Processor 52 can also be referred to as a CPU (Central Processing Unit). Processor 52 may be an integrated circuit chip with signal processing capabilities. Processor 52 can also be a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor. Furthermore, processor 52 can be implemented using integrated circuit chips.

[0180] Please see Figure 6 , Figure 6 This is a schematic diagram of an embodiment of the computer-readable storage medium provided in this application. The computer-readable storage medium 60 of this application embodiment stores program instructions 61. When executed, these program instructions 61 implement the methods provided by any embodiment and any non-conflicting combination of the work order review method of this application. The program instructions 61 can form a program file and be stored in the aforementioned computer-readable storage medium 60 in the form of a software product, so that a computer device (which may be a personal computer, server, or network device, etc.) can execute all or part of the steps of the methods of various embodiments of this application. The aforementioned computer-readable storage medium 60 includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0181] If the technical solution of this application involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the user through pop-up information or by asking the user to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

[0182] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A work order review method, characterized in that, The method includes: Retrieve the information filled in for several first-stage work order items from the work orders pending review; The reference information and the submitted information for each of the first work order items are compared to obtain the first work order review result. The reference information for each of the first work order items is predicted using a target large language model. The first work order review result is used to characterize whether the submitted information for each of the first work order items is compliant. The target large language model is obtained by adjusting the network parameters of each training layer of a pre-trained initial large language model using sample work orders. Each training layer is a network layer added to each target network layer of the initial large language model. The network parameters of each training layer are the rank decomposition matrix corresponding to that training layer. The rank decomposition matrix is ​​obtained by decomposing the weight matrix corresponding to that training layer. At least two training layers have different matrix ranks. The matrix rank of each training layer is determined based on the fit between the weight matrix of the training layer and the work order review task. The fit between the weight matrix of the network layer to be trained and the work order review task is represented by the norm of the gradient of the weight matrix of the network layer to be trained; the step of determining the rank of the matrix corresponding to the network layer to be trained includes: Obtain the product between the norm of the network layer to be trained and the first coefficient; Summing the product with the second coefficient yields the first sum; The first sum is rounded down to obtain the second sum, and the matrix rank corresponding to the network layer to be trained is obtained based on the second sum.

2. The method according to claim 1, characterized in that, The work order to be reviewed is a battery fault after-sales work order. The work order to be reviewed also includes target fault-related data, which includes at least one of the following: target fault description information and target fault handling results. The step of predicting the reference information for each of the first work order items using the target large language model includes: Using the target large language model based on the target fault-related data, reference information for each of the first work order items is predicted.

3. The method according to claim 2, characterized in that, The plurality of first work order items include first type work order items and second type work order items, wherein the correlation between the first type work order items and the target fault-related data is greater than the correlation between the second type work order items and the target fault-related data; the step of using the target large language model to predict reference information for each first work order item based on the target fault-related data includes: Input the target fault-related data into the target large language model to obtain reference information for the first type of work order item output by the target large language model; and / or, The target fault-related data is input into the target large language model to obtain the target intermediate information output by the target large language model, and the reference information of the second type of work order item is found from the associated information of the target intermediate information.

4. The method according to claim 3, characterized in that, The first type of work order item includes at least one of the following: fault level, fault attribute, faulty component, vehicle-level fault phenomenon, and failure mode; And / or, the second type of work order item includes at least one of the following: maintenance code, maintenance hours; And / or, the target intermediate information includes components that need repair.

5. The method according to claim 3 or 4, characterized in that, The target large language model for predicting the reference information of the first type of work order item and the target large language model for predicting the target intermediate information are different or the same large language model.

6. The method according to claim 1, characterized in that, The step of adjusting the network parameters of each training layer of the pre-trained initial large language model using sample work orders includes: Freeze the current network parameters of the initial large language model, and add the network layer to be trained to each target network layer of the initial large language model; Using the initial large language model based on the sample fault-related data in the sample work orders, the sample information of the several first work order items is predicted; Based on the difference between the sample information of the plurality of first work orders and the real information of the plurality of first work orders, the network parameters of each of the network layers to be trained are adjusted.

7. The method according to claim 1, characterized in that, The second sum is an integer not greater than the first sum; And / or, obtaining the matrix rank corresponding to the network layer to be trained based on the second sum includes: In response to the second sum being greater than or equal to the rank threshold, the rank threshold is used as the rank of the matrix corresponding to the network layer to be trained; In response to the second sum being less than the rank threshold, the second sum is taken as the rank of the matrix corresponding to the network layer to be trained.

8. The method according to claim 1, characterized in that, The first coefficient of the first stage of the adjustment is less than the first coefficient of the second stage of the adjustment, and the first stage is earlier than the second stage. And / or, the second coefficient of the first stage of the adjustment is equal to the second coefficient of the second stage of the adjustment.

9. The method according to claim 1, characterized in that, The steps for determining the norm corresponding to the network layer to be trained include: Each matrix parameter of the weight matrix of the network layer to be trained is taken as the current parameter, and the partial derivative of the work order loss with respect to the current parameter is obtained as the partial derivative component of the current parameter; wherein, the work order loss is determined based on the difference between the sample information of the plurality of first work order items and the real work order information of the plurality of first work order items. The third sum is obtained by summing the squares of the absolute values ​​of the partial derivative components corresponding to each of the matrix parameters; Taking the square root of the third sum yields the norm of the network layer to be trained.

10. The method according to claim 1, characterized in that, The weight matrix corresponding to the network layer to be trained is obtained by adjusting the initial weight matrix using a third coefficient, where the third coefficient is the ratio of the hyperparameter to the rank of the matrix corresponding to the network layer to be trained; and / or, In the first stage of the adjustment, the initial large language model adjusts the matrix rank of each of the network layers to be trained once for every first number of sample work orders processed. In the second stage of the adjustment, the initial large language model adjusts the matrix rank of each of the network layers to be trained once for every second number of sample work orders processed. The first number is less than the second number, and the first stage is earlier than the second stage.

11. The method according to claim 1, characterized in that, The work order to be reviewed also includes information for several second work order items; the method further includes: Obtain the target image associated with each of the second work order items; wherein the target image contains reference information of the second work order item; The target images associated with each of the second work orders are identified to obtain reference information for each of the second work orders. The reference information and the information filled in for the second work order item are compared to obtain the review result of the second work order; wherein, the review result of the second work order is used to characterize whether the information filled in for each item of the second work order is compliant.

12. The method according to claim 11, characterized in that, The step of identifying the target image associated with each of the second work orders to obtain reference information for each of the second work orders includes: From the target image associated with each second work order item, extract the region image of the area where the reference information of each second work order item is located; The images of each region are identified to obtain reference information for each of the second work orders.

13. The method according to claim 12, characterized in that, At least one of the target images associated with the second work order item is a first type of image, and the reference information in the first type of image is a barcode representing the reference information; The step of extracting the region image of the area where the reference information of each second work order item is located from the target image associated with each second work order item includes: Determine the vertical projection value of each pixel row and the horizontal projection value of each pixel column in the first type of image; The first type of image is rotated until the current vertical projection value and the current horizontal projection value of the first type of image meet the preset projection requirements, and the region image corresponding to the barcode is extracted using the current vertical projection value and the current horizontal projection value of the first type of image.

14. The method according to claim 13, characterized in that, The preset projection requirements are: among the vertical projection values ​​of each pixel row, there is a vertical projection value greater than the first projection threshold, and among the horizontal projection values ​​of each pixel column, there is a horizontal projection value greater than the second projection threshold. And / or, The step of rotating the first type of image until the current vertical projection value and horizontal projection value of the first type of image meet the preset projection requirements includes: In response to the absence of a vertical projection value and a horizontal projection value greater than the projection threshold, the first type of image is rotated by a preset angle, and the steps of determining the vertical projection value of each pixel row and the horizontal projection value of each pixel column in the first type of image and subsequent steps are re-executed until a vertical projection value and a horizontal projection value greater than the projection threshold are found. After determining the vertical projection value of each pixel row and the horizontal projection value of each pixel column in the first type of image, the method further includes: In response to the existence of vertical and horizontal projection values ​​greater than the projection threshold, according to the horizontal direction, from the vertical projection values ​​of each pixel row, the pixel rows corresponding to the first and second sequential vertical projection values ​​greater than the projection threshold are identified, and these are respectively used as the upper and lower boundaries of the barcode; and, According to the vertical direction, from the horizontal projection values ​​of each pixel column, find the third and fourth sequential pixel columns corresponding to the horizontal projection values ​​that are greater than the projection threshold, and use them as the left and right boundaries of the barcode, respectively.

15. The method according to claim 12, characterized in that, At least one of the target images associated with the second work order item is a second type of image, and the reference information in the second type of image is text information characterizing the reference information; wherein... The step of extracting the region image of the area where the reference information of each second work order item is located from the target image associated with each second work order item includes: The detection model is used to perform text detection on the second type of image, resulting in several text regions and their corresponding region categories; The text regions are respectively matched with the reference information in the second type of image, and the text regions whose regions match the categories of the reference information in the second type of image are used as the region images corresponding to the reference information in the second type of image. And / or, before identifying each of the said region images to obtain reference information for each of the second work order items, the method further includes: The image of the region is corrected.

16. The method according to claim 1 or 11, characterized in that, The work orders pending review are displayed on the display interface; for non-compliant work order items, at least one of the following steps is performed: Adjust the display color of non-compliant work order items and / or the information entered in the work order items; The non-compliant work order items and / or the information entered in the work order items are highlighted.

17. A work order verification device, characterized in that, The work order approval device includes: The acquisition module is used to acquire the information filled in for several first-order items in the work orders to be reviewed; The review module is used to compare the reference information and the submitted information of each first work order item to obtain the review result of the first work order. The reference information of each first work order item is predicted using a target large language model. The review result of the first work order is used to characterize whether the submitted information of each first work order item is compliant. The target large language model is obtained by adjusting the network parameters of each training layer of a pre-trained initial large language model using sample work orders. Each training layer is a network layer added to each target network layer of the initial large language model. The network parameters of each training layer are the rank decomposition matrix corresponding to the training layer. The rank decomposition matrix corresponding to the training layer is obtained by decomposing the weight matrix corresponding to the training layer. At least two training layers have different matrix ranks. The matrix rank of each training layer is determined based on the fit between the weight matrix of the training layer and the work order review task. The fit between the weight matrix of the network layer to be trained and the work order review task is represented by the norm of the gradient of the weight matrix of the network layer to be trained; the step of determining the rank of the matrix corresponding to the network layer to be trained includes: Obtain the product between the norm of the network layer to be trained and the first coefficient; Summing the product with the second coefficient yields the first sum; The first sum is rounded down to obtain the second sum, and the matrix rank corresponding to the network layer to be trained is obtained based on the second sum.

18. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory being used to store program instructions, and the processor being used to execute the program instructions to implement the work order review method as described in any one of claims 1-16.

19. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store program instructions that can be executed to implement the work order review method as described in any one of claims 1-16.

Images