Method and apparatus for detecting material cancellation, and electronic device

Abstract

The application discloses a material verification detection method and device and electronic equipment. The method comprises the following steps: obtaining material verification time series data and work order data, wherein the material verification time series data represents the number of verified materials within a preset time window, and the work order data is used to represent work order information corresponding to multiple business types within the preset time window; performing feature fusion on the material verification time series data and the work order data to obtain fusion data; processing the fusion data through a deep learning model to obtain an output sequence; and determining a material verification detection result through the output sequence. The application solves the technical problem that, in the related art, when the consistency verification of the serial code of a terminal device is performed, the abnormal detection of the device cannot be performed when the device does not support the serial code, resulting in incomplete verification.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and more specifically, to a method, apparatus, and electronic device for detecting material write-offs. Background Technology

[0002] In the field of operator network maintenance, installation and maintenance personnel play a crucial role in the entire ecosystem, encompassing various network devices such as optical modems, routers, IPTV, and smart home devices. Their work scope includes installation, configuration, debugging, maintenance, and sales assistance. To ensure the accuracy, transparency, and rational use of resources in the installation and maintenance process, the verification of installation and maintenance materials has become a critical task. This verification process involves recording, managing, and verifying the materials used by installation and maintenance personnel to ensure consistency with actual work needs, while avoiding waste and misuse of resources.

[0003] However, during the material verification process, several anomalies often occur, including but not limited to duplicate verification, omissions, mismatched specifications, and even fraudulent verification. Currently, the integrated dispatch system utilizes serial number scanning, identification, and persistent storage technologies to verify the consistency of serial numbers on some terminal devices. If an inconsistency is detected, the system will trigger a work order alarm or fail the verification to prevent inconsistencies and omissions. Material management personnel also use the system to set experience-based thresholds based on statistical data to report abnormal material usage.

[0004] However, the relevant technologies perform consistency verification on the serial numbers of terminal devices. When the devices do not support the serial numbers, they cannot be anomaly detected, resulting in incomplete verification.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This application provides a method, apparatus, and electronic device for detecting material verification, which at least solves the technical problem in related technologies where consistency verification is performed on the serial numbers of terminal devices, but when the devices do not support serial numbers, abnormal detection of these devices cannot be performed, resulting in incomplete verification.

[0007] According to one aspect of the embodiments of this application, a method for detecting material write-offs is provided, comprising: acquiring material write-off time series data and work order data, wherein the material write-off time series data represents the quantity of materials written off within a preset time window, and the work order data represents work order information corresponding to multiple business types within the preset time window; performing feature fusion on the material write-off time series data and work order data to obtain fused data; processing the fused data through a deep learning model to obtain an output sequence, wherein the deep learning model includes a multi-head attention mechanism and a Fourier transform attention mechanism, the multi-head attention mechanism and the Fourier transform attention mechanism being used to extract key features from the fused data; and determining the detection result of the material write-offs through the output sequence.

[0008] Optionally, after obtaining the material reimbursement time series data and work order data, the method further includes: processing the material reimbursement time series data in at least one of the following ways: soft reset, peak noise, length replacement, and smooth replacement, wherein soft reset is used to generate random noise based on the material reimbursement time series data; peak noise is used to add a sharp data noise to the material reimbursement time series data; length replacement is used to stretch or shrink the material reimbursement time series data in the time dimension according to a preset ratio; and smooth replacement is used to replace all data in the first time period of the material reimbursement time series data with data at any point in the first time period.

[0009] Optionally, the material reimbursement time series data and work order data are feature-fused to obtain fused data, including: processing the second weight, material reimbursement time series data, third weight, work order data, and second bias value using a first activation function to obtain a second intermediate variable; determining the first intermediate variable based on the first weight, second intermediate variable, and first bias value; processing the fourth weight, first intermediate variable, third bias value, fifth weight, and fourth bias value using a second activation function to obtain residual data; obtaining the first feature data by standardizing the material reimbursement time series data and residual data; and processing the first feature data using a third activation function to obtain fused data.

[0010] Optionally, the first feature data is processed by a third activation function to obtain fused data, including: processing the first feature data, the sixth weight, and the fifth bias value by a third activation function to obtain a third intermediate variable; and processing the third intermediate variable, the seventh weight, and the sixth bias value by a third activation function to obtain fused data.

[0011] Optionally, the fused data is processed using a deep learning model, including: processing the fused data using a multi-head attention mechanism to obtain a first result; and processing the fused data using a Fourier transform attention mechanism to obtain a second result.

[0012] Optionally, the fused data is processed using a multi-head attention mechanism to obtain a first result, including: normalizing the sixth bias value, the first data, and the second data using a fourth activation function to obtain a fourth intermediate variable, wherein the first data is determined by the data in the first position of the fused data and the eighth weight, and the second data is determined by the data in the second position of the fused data and the ninth weight; determining the attention variables corresponding to the first data, the second data, and the third data based on the fourth intermediate variable and the third data, wherein the third data is determined by the data in the third position of the fused data and the tenth weight; and determining the first result based on the attention variables and the eleventh weight.

[0013] Optionally, the fused data is processed using a Fourier transform attention mechanism to obtain a second result, including: processing the data at the first position, the data at the second position, and the data at the third position in the fused data using Fourier transform to obtain a first transform value, a second transform value, and a third transform value; processing the variance between the first transform value, the second transform value, and the data at the first position in the fused data according to a fifth activation function to obtain a fifth intermediate variable; and obtaining the second result by performing an inverse Fourier transform on the fifth intermediate variable and the third transform value.

[0014] Optionally, the fused data is processed by a deep learning model to obtain an output sequence, including: determining the twelfth weight corresponding to the first result and determining the thirteenth weight corresponding to the second result; and determining the output sequence based on the first result, the twelfth weight, the second result and the thirteenth weight.

[0015] Optionally, after processing the fused data using a deep learning model, the method further includes: processing the fourteenth weight, the output sequence, and the seventh bias value using the sixth activation function to obtain the sixth intermediate variable; and processing the sixth intermediate variable, the fifteenth weight, and the eighth bias value using the sixth activation function to obtain the second output sequence.

[0016] Optionally, the detection result of material verification is determined by the output sequence, including: when the value in the second output sequence obtained after processing the output sequence is the first value, the detection result of material verification is determined to be normal; when the value in the second output sequence is the second value, the detection result of material verification is determined to be abnormal.

[0017] According to another aspect of the embodiments of this application, a material verification detection device is also provided, comprising: an acquisition module for acquiring material verification time series data and work order data, wherein the material verification time series data represents the quantity of materials verified within a preset time window, and the work order data represents work order information corresponding to multiple business types within the preset time window; a fusion module for performing feature fusion on the material verification time series data and the work order data to obtain fused data; a processing module for processing the fused data through a deep learning model to obtain an output sequence, wherein the deep learning model includes a multi-head attention mechanism and a Fourier transform attention mechanism, the multi-head attention mechanism and the Fourier transform attention mechanism being used to extract key features from the fused data; and a determination module for determining the detection result of material verification based on the output sequence.

[0018] According to another aspect of the embodiments of this application, an electronic device is also provided, including: a memory for storing program instructions; and a processor connected to the memory for executing program instructions to perform the following functions: acquiring material reimbursement time series data and work order data, wherein the material reimbursement time series data represents the quantity of materials reimbursed within a preset time window, and the work order data represents work order information corresponding to multiple business types within the preset time window; fusing the material reimbursement time series data and work order data to obtain fused data; processing the fused data through a deep learning model to obtain an output sequence, wherein the deep learning model includes a multi-head attention mechanism and a Fourier transform attention mechanism, the multi-head attention mechanism and the Fourier transform attention mechanism being used to extract key features from the fused data; and determining the detection result of material reimbursement through the output sequence.

[0019] According to another aspect of the embodiments of this application, a non-volatile storage medium is also provided, the non-volatile storage medium including a stored computer program, wherein the device containing the non-volatile storage medium executes the above-mentioned material verification detection method by running the computer program.

[0020] In this embodiment, material reimbursement time series data and work order data are acquired. The material reimbursement time series data represents the quantity of materials reimbursed within a preset time window, and the work order data represents the work order information corresponding to multiple business types within the preset time window. The material reimbursement time series data and work order data are fused to obtain fused data. The fused data is then processed by a deep learning model to obtain an output sequence. The deep learning model includes a multi-head attention mechanism and a Fourier transform attention mechanism, which are used to extract key features from the fused data. This achieves the goal of determining the detection result of material reimbursement based on the output sequence, thereby realizing the technical effect of comprehensively detecting and verifying material reimbursement. This solves the technical problem in related technologies where consistency verification is performed on the serial codes of terminal devices, but when the devices do not support serial codes, abnormal detection of these devices cannot be performed, resulting in incomplete verification. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0022] Figure 1 This is a hardware structure block diagram of a computer terminal for implementing a detection method for material verification according to an embodiment of this application;

[0023] Figure 2 This is a flowchart of a material verification detection method according to an embodiment of this application;

[0024] Figure 3a This is a schematic diagram of a material verification and detection model according to an embodiment of this application;

[0025] Figure 3b This is a schematic diagram illustrating four replacement strategies according to embodiments of this application;

[0026] Figure 3c This is a flowchart of an anomaly detection process for the verification of installation and maintenance materials according to an embodiment of this application;

[0027] Figure 3d This is a schematic diagram of an abnormal verification detection result according to an embodiment of this application;

[0028] Figure 4 This is a structural diagram of a material verification detection device according to an embodiment of this application. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0030] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0031] First, some nouns or terms that appear in the explanation of the embodiments of this application shall be interpreted as follows:

[0032] A gated residual network (GRN) is a neural network implemented using gated units and residual units, which integrates multidimensional variable data while avoiding neural network degradation.

[0033] Layer Normalization (LN) is a normalization method in neural networks. Compared with batch normalization, it has the advantages of not depending on batch size, being able to learn online, and having features that are less likely to be lost.

[0034] A multilayer perceptron (MLP) is an artificial neural network consisting of an input layer, hidden layers, an output layer, and activation functions.

[0035] Installation and maintenance material verification refers to the process by which installation and maintenance personnel submit material usage information and verify user business information in the system when using materials provided by the system.

[0036] In the relevant technologies, several problems exist in the process of verifying and reimbursing installation and maintenance materials: First, auxiliary materials such as optical fibers, network cables, pigtails, multimode connectors, RJ45 connectors, and some new terminal devices do not support serial number consistency detection for the aforementioned systems. Failure to detect these anomalies may lead to unnecessary high-cost material expenditures. Second, user services and material types are diverse, and the conditions vary significantly across different work teams, rural areas, and urban areas. Control methods based on experience thresholds or material modeling are insufficient for achieving fine-grained, accurate, and rapid management, resulting in oversights, omissions, and misjudgments of anomalies. Finally, on-site construction involves complex conditions. For example, errors in material system terminal types, such as failed material serial number associations, as well as material damage or loss, cannot be automatically detected and resolved, requiring significant manual intervention.

[0037] To address the aforementioned issues, this application provides a method for detecting material write-offs. By inputting work order information related to installation and maintenance personnel's write-offs, installation / relocation, repair, and collaborative write-off work orders (or covariate data or work order data), along with time-series data of the write-off materials, a trained Transformer model is used to extract latent information from the data. This enables real-time monitoring of write-off anomalies, more accurately identifying instances of waste or misuse of materials by installation and maintenance personnel, detecting write-off anomalies, and sending alarms to management for verification and processing, thereby achieving the goals of resource conservation and cost reduction. The aforementioned material write-off detection method can be implemented in… Figure 1 The computer terminals shown are explained below.

[0038] The material verification detection method provided in this application can be executed on a mobile terminal, computer terminal, or similar computing device. Figure 1 A hardware block diagram of a computer terminal for implementing a material verification detection method is shown. Figure 1 As shown, the computer terminal 10 may include one or more processors (shown as 102a, 102b, ..., 102n in the figure) (the processor may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission module 106 for communication functions. In addition, it may also include: a display, a keyboard, a cursor control device, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, and a BUS bus. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 10 may also include... Figure 1 The more or fewer components shown, or having the same Figure 1The different configurations shown.

[0039] It should be noted that the aforementioned one or more processors and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be implemented wholly or partially as software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be wholly or partially integrated into any other element in the computer terminal 10. As involved in the embodiments of this application, the data processing circuits serve as processor control (e.g., selection of a variable resistor termination path connected to an interface).

[0040] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the material verification detection method in this embodiment. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the aforementioned material verification detection method. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0041] The transmission module 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission module 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission module 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0042] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 10.

[0043] It should be noted here that, in some optional embodiments, the above... Figure 1 The computer terminal shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 1This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned computer terminal.

[0044] Under the above operating environment, this application provides an embodiment of a material verification detection method. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0045] Figure 2 This is a flowchart of a material verification detection method according to an embodiment of this application, such as... Figure 2 As shown, the method includes the following steps:

[0046] Step S202: Obtain material reimbursement time series data and work order data. The material reimbursement time series data represents the quantity of materials reimbursed within a preset time window, and the work order data represents the work order information corresponding to multiple business types within the preset time window.

[0047] In step S202 above, the material reimbursement time series data and work order data are two types of key data sequences input into the input layer. The material reimbursement time series data records the quantity of specific materials reimbursed by installation and maintenance within a certain time window, and generates abnormal training data through a random outlier (value) replacement device. The aforementioned work order data can also be called covariate sequence data. The covariate data covers work order information corresponding to different service types (including broadband, cameras, etc.) of installation and maintenance within the same time period, such as the number of fault work orders and installation work orders. It also includes user service information, such as speed, quantity of various products, and fault reasons (such as network cables, routers, etc.), work order type (installation, relocation, fault, construction area attribute), the region where the construction is located, whether the construction community is a newly built community, etc., work order service type (broadband, ITV, smart home camera, ETV, router, networking), number of service terminals (mainly for ETV), and statistics are performed according to a certain time window and dimension.

[0048] Step S204: Feature fusion is performed on the material verification time series data and work order data to obtain fused data.

[0049] In step S204 above, feature fusion can be performed in the feature fusion extraction layer. This layer employs a gated residual network, primarily used to fuse covariate sequence data such as material verification time series data and work order data. The linear embedding layer within the feature fusion extraction layer extracts feature information from the fused data using a linear multilayer perceptron, laying the foundation for the next processing step.

[0050] Step S206: The fused data is processed by a deep learning model to obtain an output sequence. The deep learning model includes a multi-head attention mechanism and a Fourier transform attention mechanism, which are used to extract key features from the fused data.

[0051] In step S206 above, the deep learning model can be, for example, a Transformer model, which includes an encoder and a decoder. It incorporates techniques such as multi-head attention, Fourier transform attention, multilayer linear perceptron, layer normalization, and activation functions to extract hidden features from sequence features. This stage of processing provides important feature representations for subsequent anomaly detection.

[0052] Step S208: Determine the test results of material verification by outputting the sequence.

[0053] In step S208 above, the output sequence is processed by using a multilayer perceptron (MLP) combined with a softmax function to obtain sequence data about normal and abnormal labels, so as to complete the judgment and labeling of abnormal cancellation situations.

[0054] The steps in steps S202 to S208 above can be achieved through... Figure 3a The material verification and detection model is implemented, specifically, in Figure 3a The model consists of four key parts: an input layer, a feature fusion and extraction layer, a Transformer body, and an output layer. This reconciliation detection model aims to effectively detect and determine abnormal reconciliation situations.

[0055] The aforementioned material reimbursement detection model enables more accurate analysis of installation and maintenance reimbursement data. It fully leverages the fusion characteristics of material reimbursement time-series data and work order data, thereby improving the accuracy and efficiency of detecting abnormal reimbursement situations. The material reimbursement detection model is shown in the formula, where, Y represents whether the write-off is abnormal at the current time t, f is the function used to determine whether the write-off is abnormal, and Y represents the value of f. t-l:t Let s be the material value written off in the previous l time windows before the current time t, and s be a static variable. t-l:t This refers to the work order data that was processed within the previous l time windows at the current time t.

[0056]

[0057] Through the above steps S202 to S208, the purpose of determining the material verification result based on the output sequence is achieved, thus realizing the technical effect of comprehensively verifying material verification. This solves the technical problem in related technologies where consistency verification is performed on the serial codes of terminal devices, but when the devices do not support serial codes, anomaly detection cannot be performed on these devices, leading to incomplete verification. A detailed explanation follows.

[0058] In step S202 of the above-mentioned material verification detection method, after obtaining the material verification time series data and work order data, the method further includes: processing the material verification time series data by at least one of the following methods: soft reset, peak noise, length replacement, and smooth replacement, wherein soft reset is used to generate random noise based on the material verification time series data; peak noise is used to add a sharp data noise based on the material verification time series data; length replacement is used to stretch or shrink the material verification time series data in the time dimension according to a preset ratio; and smooth replacement is used to replace all data in the first time period of the material verification time series data with data at any point in the first time period.

[0059] In the embodiments of this application Figure 3a In this context, random outlier replacement plays a crucial role in reconciliation sequence data. Its main purpose is to address the data imbalance caused by the small proportion of outliers. By introducing random outlier replacement, self-supervised learning of sequence data can be achieved.

[0060] This application employs a random time window approach to replace random positions in the data under a certain error probability. Specifically, four replacement strategies can be adopted, such as... Figure 3bThe diagram illustrates soft reset, peak noise, length replacement, and smooth replacement. In soft reset, the data within a random window is slightly adjusted to simulate anomalies. In some embodiments of this application, a segment of random noise is added to the original sequence and marked as an anomaly. Peak noise adds noisy data to the random window to simulate data anomalies. In some embodiments of this application, a sharp data noise is added to the original sequence and marked as an anomaly. Length replacement involves modifying the data length within the random window to create an anomalous data window. In some embodiments of this application, the data of the original sequence can be stretched or shrunk proportionally over time. Smooth replacement replaces the data within the random window with stable anomalous data. In some embodiments of this application, a segment of the original time series data can be replaced based on a specific point in the time series. These replacement strategies can simulate anomalies to a certain extent, thereby enriching the training dataset and enhancing the model's ability to identify anomalies. It is important to emphasize that the random outlier replacer only works during the training phase and is not applied during testing and deployment.

[0061] By introducing a random outlier replacement, the problem of data imbalance can be effectively addressed, improving the self-supervised learning effect of the material verification and detection model. It also provides more reliable and robust training data for the material verification and detection model, thereby further enhancing the model's performance and credibility.

[0062] In step S204 of the above-mentioned material verification detection method, the material verification time series data and work order data are feature-fused to obtain fused data. Specifically, this includes the following steps: processing the second weight, material verification time series data, third weight, work order data, and second bias value using a first activation function to obtain a second intermediate variable; determining the first intermediate variable based on the first weight, second intermediate variable, and first bias value; processing the fourth weight, first intermediate variable, third bias value, fifth weight, and fourth bias value using a second activation function to obtain residual data; obtaining the first feature data by standardizing the material verification time series data and residual data; and processing the first feature data using a third activation function to obtain fused data.

[0063] In the above steps, the first feature data is processed by the third activation function to obtain fused data. Specifically, the steps are as follows: the first feature data, the sixth weight, and the fifth bias value are processed by the third activation function to obtain the third intermediate variable; the third intermediate variable, the seventh weight, and the sixth bias value are processed by the third activation function to obtain fused data.

[0064] In some embodiments of this application, the above factors need to be taken into consideration because the current status of each installation and maintenance team and the workload of installation and troubleshooting during material verification vary. The feature fusion extraction layer consists of a GRN and a linear embedding layer, aiming to achieve data fusion of covariate data (i.e., the aforementioned work order data) and material verification time series data while maintaining high-quality features. GRN plays a key role in the feature fusion process, and its main function is to ensure that features are not lost while fusing data information. By using GRN, covariate data and material verification time series data can be effectively fused, realizing comprehensive analysis of multidimensional data.

[0065] On the other hand, the linear embedding layer is composed of an MLP (Multi-Level Processing) module, whose main task is to extract key features from the material reimbursement time series data. This layer is designed to capture the inherent patterns in the material reimbursement time series data. Through MLP processing, useful information in the data can be effectively extracted, providing a solid foundation for subsequent processing and analysis.

[0066] In summary, the structure of the feature fusion extraction layer aims to fully leverage the advantages of GRN and the linear embedding layer to achieve the organic fusion of covariate data and material verification time series data, thereby improving the accuracy of the neural network. GRN ensures the quality of the fusion, while the linear embedding layer focuses on feature extraction from the material verification time series data, together constructing a powerful feature fusion extraction layer.

[0067] The formula for GRN is as follows:

[0068] GRN(a,c)=LayerNorm(a+GLU(η1)),

[0069] η1=W1η2+b1

[0070] η2=ELU(W2a+W3c+b2)

[0071] GLU(n1)=sigmod(W4n1+b3)⊙(W5n1+b4)

[0072] Wherein, ELU represents the first activation function, W2 represents the second weight, a represents the material write-off time series data, W3 represents the third weight, c represents the work order data, b2 represents the second bias value, η2 represents the second intermediate variable, W1 represents the first weight, b1 represents the first bias value, η1 represents the first intermediate variable, sigmoid represents the second activation function, W4 represents the fourth weight, b3 represents the third bias value, W5 represents the fifth weight, b4 represents the fourth bias value, GLU is the residual linear unit, GLU(η1) represents the residual data, LayerNorm represents the function for standardization, and GRN(a,c) represents the first feature data.

[0073] The aforementioned first feature data is input into an MLP, which consists of input, hidden states, output, and activation functions. The following explanation uses the formula for an MLP with only one hidden layer:

[0074] mlp(H)=relu(W7H+b6)

[0075] H = relu(W6X + b5)

[0076] relu(y) = max(0, y)

[0077] In the above formula, ReLU represents the third activation function, which determines the maximum value between 0 and the input value y; W6 represents the sixth weight; X replaces GRN(a, c) to represent the first feature data; b5 represents the fifth bias value; H represents the third intermediate variable; W7 represents the seventh weight; b6 represents the sixth bias value; and mlp(H) represents the fused data. It should be noted that the number of layers in an MLP can be set according to actual needs. The above formula is only an example and does not limit the number of layers in an MLP.

[0078] In step S206 of the above-mentioned material verification detection method, the fused data is processed by a deep learning model, specifically including the following steps: the fused data is processed by a multi-head attention mechanism to obtain a first result; the fused data is processed by a Fourier transform attention mechanism to obtain a second result.

[0079] In the above steps, the fused data is processed through a multi-head attention mechanism to obtain the first result, specifically including the following steps: The sixth bias value, the first data, and the second data are normalized using the fourth activation function to obtain the fourth intermediate variable. The first data is determined by the data in the first position of the fused data and the eighth weight, and the second data is determined by the data in the second position of the fused data and the ninth weight. Based on the fourth intermediate variable and the third data, the attention variables corresponding to the first, second, and third data are determined. The third data is determined by the data in the third position of the fused data and the tenth weight. Based on the attention variables and the eleventh weight, the first result is determined.

[0080] In the above steps, the fused data is processed through a Fourier transform attention mechanism to obtain a second result. Specifically, the steps include: processing the data at the first position, the data at the second position, and the data at the third position in the fused data using Fourier transform to obtain a first transform value, a second transform value, and a third transform value; processing the variance between the first transform value, the second transform value, and the data at the first position in the fused data using a fifth activation function to obtain a fifth intermediate variable; and performing an inverse Fourier transform on the fifth intermediate variable and the third transform value to obtain the second result.

[0081] In this embodiment of the application, the deep learning model is a Transformer architecture. The Transformer body accurately mines the potential information in the verification data through an encoder and decoder based on an attention mechanism, providing key support for the deep processing and modeling of the data.

[0082] The Transformer body of this application consists of L encoders and decoders connected in series, aiming to achieve efficient abstraction and feature extraction of verification data. Each encoder is composed of a multi-head attention mechanism, layer normalization (LN), and a multilayer perceptron (MLP). This design enables the encoder to fully capture the correlation information and patterns in the data.

[0083] The application of multi-head attention mechanisms in the encoder allows the model to simultaneously consider data dependencies at different locations, thereby better establishing the internal connections within the data. Layer normalization helps maintain the model's stability and convergence, thus improving training performance. The introduction of multilayer perceptrons further enriches the encoder's capabilities, enabling it to extract features at different levels of abstraction.

[0084] The Transformer body's structural design fully leverages the advantages of attention mechanisms, standardization, and multilayer perceptrons, enabling the model to more accurately capture key information in reimbursement data, thus providing strong support for subsequent anomaly detection and classification. Through the concatenation of L encoders and decoders, the Transformer body in this application achieves high-level data representation and feature extraction, resulting in significant benefits in reimbursement data analysis. The MultiHead(Q, K, V) formula for the multi-head attention mechanism is shown below:

[0085]

[0086]

[0087] MultiHead(Q,K,V)=[H1,…,H h W H

[0088]

[0089] In the above formula, Softmax represents the fourth activation function, Attention(Q, K, y) is the general formula for the attention mechanism, Q is the query, which in this embodiment refers to the data in the first position of the fused data, K is the key, which in this embodiment refers to the data in the second position of the fused data, and V is the value, which in this embodiment refers to the data in the third position of the fused data. B represents the sixth bias value. This refers to the first data mentioned above. This represents the eighth weight mentioned above, which is the weight of the data in the first position of the fused data at the h-th attention head. This refers to the second data mentioned above. This represents the ninth weight mentioned above, which is the weight of the data in the second position of the fused data at the h-th attention head, d. attn Let A(Q,K) represent the variance of the first and second data points. A(Q,K) is the normalization function for the attention mechanism; here, scaled dot product attention is used. In this application, This represents the fourth intermediate variable mentioned above. This refers to the third data mentioned above. This represents the tenth weight mentioned above, which is the weight of the data in the third position in the fused data at the h-th attention head. This represents the attentional variable mentioned above, that is, the attentional variable corresponding to the h-th attention head, or H. h W indicates HHere, represents the trainable weight parameters, i.e., the eleventh weight mentioned above; H represents the attention head; h represents the number of attention heads; and MultiHead(Q,K,V) represents the first result mentioned above. Unlike the original Transformer, this embodiment does not use positional encoding technology, but instead implements it through bias B, where B is a learnable relative positional encoding parameter.

[0090] To improve the periodic processing performance of the model, this application introduces a Fourier transform attention mechanism. The Fourier transform attention mechanism is expressed as follows:

[0091]

[0092] In the above formula, For Fourier transform attention mechanism, This represents the first transformation value mentioned above. This represents the second transformation value mentioned above. This represents the third transformation value mentioned above. Let d represent the Fourier transform and the inverse Fourier transform, respectively, σ represent the fifth activation function mentioned above, and d q This represents the variance between the data points that are in the first position in the fused data. This represents the fifth intermediate variable mentioned above. This indicates the second result mentioned above.

[0093] The Discrete Fourier Transform formula used in this embodiment is as follows:

[0094]

[0095] Where N is the sequence length, x n The original sequence, i.e., the sequence corresponding to the fused data, is X. k This is a sequence of Fourier transforms. The formula for the inverse discrete Fourier transform is shown below.

[0096]

[0097] The derivation of the Discrete Fourier Transform is as follows: Periodic functions can be expanded using Fourier series. After introducing the δ(t) function, more general periodic functions can also be expanded because they can be sampled using the δ(t) function, becoming finite-length discrete sequences, which can be considered periodic signals. Based on the continuous signal x(t), sampled N times according to the sampling time Ts, δ(t-nT)... s By periodically extending these N values, a periodic discrete signal x[n] can be obtained, with a period T = N*Ts and a frequency f = 2pi / T. Within one period T, its expression is:

[0098]

[0099]

[0100] The Fourier series of the discrete periodic signal can be obtained as:

[0101]

[0102] According to the sifting property of the δ(t) function, there is

[0103]

[0104] Let

[0105] X[kω]·T s =X[k]

[0106] We obtain:

[0107]

[0108] The above process is the Fourier transform of the discrete periodic signal. Theoretically, the frequency spectrum of a discrete periodic signal has infinitely many components, but due to its periodicity (after expanding it into trigonometric functions, i.e., sin and cos), generally only the principal value interval 0<=k<N-1 is taken for research.

[0109] In step S206 of the above material write-off detection method, processing the fused data through the deep learning model to obtain an output sequence specifically includes the following steps: determining the twelfth weight corresponding to the first result, and determining the thirteenth weight corresponding to the second result; determining the output sequence according to the first result, the twelfth weight, the second result and the thirteenth weight.

[0110] In the embodiment of the present application, the output of the attention mechanism is obtained by connecting MultiHead(Q,K,V) with a trainable weight vector as the output, so as to obtain the output sequence.

[0111] After processing the fused data through the deep learning model in step S206 of the above material write-off detection method, the method further comprises: processing the fourteenth weight, the output sequence and the seventh bias value through the sixth activation function to obtain a sixth intermediate variable; processing the sixth intermediate variable, the fifteenth weight and the eighth bias value through the sixth activation function to obtain a second output sequence.

[0112] In the embodiment of the present application, the output layer is composed of an MLP and a Softmax function.

[0113] mlp(D)=Softmax(W 15 D+b8)

[0114] D = Softmax(W 14 Y+b7)

[0115] In the above formula, Softmax represents the sixth activation function of the output layer, W 14 This represents the fourteenth weight mentioned above, Y represents the output sequence, b7 represents the seventh bias value, D represents the sixth intermediate variable, and W... 15 Here, b represents the fifteenth weight, b8 represents the eighth bias value, and mlp(D) represents the second output sequence. It should be noted that the number of MLP layers can be set according to actual needs; the above formula is merely an example and does not limit the number of MLP layers in the output layer.

[0116] In step S208 of the above-mentioned material verification detection method, the detection result of material verification is determined by the output sequence, specifically including the following steps: when the value in the second output sequence obtained after processing the output sequence is the first value, the detection result of material verification is determined to be normal; when the value in the second output sequence is the second value, the detection result of material verification is determined to be abnormal.

[0117] In this embodiment of the application, when the value in the second output sequence is the first value (i.e., 0), it indicates that the material verification test result is normal; when the data in the second output sequence is the second value (i.e., 1), it indicates that the material verification test result is abnormal.

[0118] To better understand the anomaly detection process for the aforementioned material verification, the following is combined with... Figure 3c The above process will be explained.

[0119] The installation and maintenance material verification anomaly detection process based on self-supervised learning Transformer is as follows: Figure 3c As shown, it mainly includes three parts: data preparation, model building, and anomaly detection. Specifically, it can be divided into 6 steps: extracting material verification data, data augmentation, generating abnormal data, training the model, deployment and operation, and verification anomaly early warning.

[0120] 1) Material reimbursement data extraction: Extract reimbursement data for optical modems, network cables, routers, and cameras from the material reimbursement report database. This data is time series data of material reimbursement.

[0121] 2) Data augmentation: Extract obstacles, installations, and sales assistance work orders from the scheduling system, along with the reasons for work order responses, the work team, region, and material warehouse of the installation and maintenance personnel, and perform clustering and perspective analysis, which is equivalent to obtaining work order data.

[0122] 3) Generate abnormal data. First, manually check the historical data and mark the abnormal situations. Then, use a random replacement method to generate abnormal data.

[0123] 4) Train the model: Use GPU to train the deep learning model and fine-tune the Transformer parameters. Select the model with good overall accuracy and precision and save it.

[0124] 5) Deploy deep learning models in the production environment and use real-time data from the material system directly without random replacement.

[0125] 6) Based on the material affiliation, push abnormal information to the material administrator and control personnel, and verify and provide feedback on material usage information.

[0126] Specifically:

[0127] Step 1: Extract verification data

[0128] Extract the reimbursement list from the reimbursement data. The reimbursement list includes material type, associated work order number, reimbursement time, and reimbursement installation and maintenance personnel. Pivot the reimbursement list data according to time windows to generate material reimbursement time series data, which can be represented as material type (index), installation and maintenance personnel (index), start time (index), and material consumption (value). Simultaneously, based on the installation and maintenance personnel's affiliation, material reimbursement data can also be generated for work teams, material warehouses, regions, and the entire area. For example, a time window of 4 hours (Category 6 network cable, personnel A, 2023-02-04 12:00:00, 3) or (router, work team B, 2023-03-06 12:00:00, 1).

[0129] Step 2: Data Augmentation

[0130] The purpose of data augmentation is to correlate covariates of the reconciliation data (the aforementioned work order data) in order to further detect reconciliation anomalies.

[0131] First, based on the aforementioned installation and maintenance personnel information, we can extract the time windows, various work order information for installation and maintenance personnel, and static information. Work order information includes installation work orders (broadband, IPTV, bullet cameras, PTZ cameras, landlines, ETV, etc.), fault work orders (faulty optical modems, faulty network cables, faulty fiber optic cables, faulty set-top boxes, etc.), quality improvement work orders (old network cables, insufficient gateway capacity, insufficient router capacity, etc.), and sales assistance work orders (routers, cameras, etc.), along with their material quantities. We cluster these based on the time windows and correlate the work order quantity with the material consumption time series data for each installation and maintenance personnel. Static information includes the number of maintenance users in each installation and maintenance personnel's assigned area, sourced from the installation and maintenance internet platform. This can be represented as installation and maintenance personnel (index), start time (index), quantity (value) of various installation work orders, quantity (value) of fault work orders, quantity (value) of quality improvement work orders, quantity (value) of sales assistance work orders, and hour (value). We can also categorize the information by the organizational structure of the installation and maintenance personnel, such as work teams, material warehouses, districts / counties, and the entire region.

[0132] Step 3: Generate abnormal data

[0133] Based on the daily manual checks by the materials administrator and the manual feedback from various installation and maintenance personnel, data windows showing excessive material usage are marked as abnormal. (This step can be skipped depending on the actual situation.)

[0134] Because the aforementioned outlier data constitutes a small portion of the total data, it leads to data imbalance and makes training difficult. Therefore, the random outlier replacer described in Section 1 of this method is used to replace the material consumption data sequence, generating outlier data. Different methods can be applied multiple times to different positions within the same data sequence, serving as different input data.

[0135] Step 4: Training the model

[0136] Normalization: Next, the numerical data in X are normalized according to dimension Z using a formula:

[0137]

[0138] The loss function is set to binary cross-entropy, as shown in the formula below. To train the Transformer: set the total number of iterations (iter), and adjust the parameters to train the data. Use the ADAM method for error backpropagation to update and improve the Transformer model weights. Adjustable parameters include:

[0139]

[0140] Where t0 is the initial time, t 0+l The time when the sequence ends. To determine if it matches the exception label, a t The outlier score is either 0 or 1, ∑||W|| is the sum of all weight L2 norms, and weight_decaly is an adjustable weight decay.

[0141] The following explains the parameters present in the model:

[0142] 1) Transformer parameter L, which determines the number of encoders and decoders connected in series in the Transformer body, and affects the number of model parameters and training difficulty.

[0143] 2) The learning rate is the step size for training the neural network. The learning rate is usually set between 0.001 and 0.01.

[0144] 3) Batch size, usually set to 16-64, determines the amount of data to be trained on a neural network at one time.

[0145] 4) The number of attention heads, usually set to 2-4, determines the number of attention heads that the Transformer can compute in parallel.

[0146] 5) The sequence encoder has a length of l. Based on the previous l time series data of X, it calculates whether there is an anomaly in the write-off at the current time t.

[0147] Accuracy is calculated on the test set, parameters are fine-tuned multiple times, and appropriate model weights are selected and saved for prediction.

[0148] Step 5: Deployment and Operation

[0149] Deploy the trained model in the production environment. At this time, simultaneously retrieve data from the reconciliation report library and the installation and maintenance scheduling system, and perform correlation according to the data augmentation method in step two. Do not label data anomalies or use a random anomaly replacement.

[0150] Step Six: Anomaly Detection

[0151] If the Transformer model detects an anomaly in the materials of a certain installation and maintenance personnel, team, or area, it will send the anomaly information to the superiors according to the organizational structure for verification and handling.

[0152] The above process is illustrated through the following examples: Taking City A as an example, and focusing on installation and maintenance teams, the process involves training data by extracting six months of network cable and prefabricated fiber optic installation and repair data (covariates), followed by testing with one month's data. The anomaly detection process is detailed below. A total of 830,000 data entries related to material reimbursement and usage were collected, covering 12 regional marketing centers / district offices, 124 installation and maintenance teams (80 urban and 44 rural areas), and 52 installation and maintenance material databases. This included 100,000 data entries for network cables and prefabricated fiber optic cables, totaling 760,000 and 120,000 meters respectively. There were 183,000 broadband installation and relocation work orders, 23,000 ITV installation work orders, 180,000 fault work orders, and 41,000 work orders related to network cables. Specific steps included:

[0153] (1) Extracting the reimbursement data: Extract the material reimbursement data list from the integrated scheduling system and view it according to the time window, which is 1 day. The following table is the material reimbursement list.

[0154]

[0155] (2) Data augmentation: Based on the installation and maintenance personnel in the write-off list, the work orders of the installation and maintenance personnel are extracted from the integrated dispatch system, and pivoted according to time, and connected with the write-off pivot table according to time. The table below shows the data after data augmentation.

[0156]

[0157]

[0158] (3) Generate abnormal data. First, based on the manual verification results of the material administrator, some data is marked as abnormal. For example, on April 4, 2023, the second shift had a large consumption of network cables, while the number of various work orders was small. According to manual feedback, the network cable reimbursement was confirmed to be abnormal. Therefore, the data at this time point is marked as abnormal as 1. At the same time, the random replacement device of this application is used to replace the reimbursement sequence data to generate abnormal data.

[0159] (4) Model training environment: Python 3.9, PyTorch 2.0, CUDA 11.8, CPU Intel 13700KF, GPU NVIDIA RTX 4090. The model was built, and after multiple parameter adjustments, the iteration count was set to 3000, the number of multi-heads to 2, the learning rate to 0.0009, the weight_declay to 0.00026, and the Transformer parameter L to 2. The model was then trained using 5 months of data and validated using 1 month of data. Specific performance metrics are shown in the table below:

[0160] Algorithm Name This application transformer transformer GRU SVM accuracy 0.8721 0.8429 0.8112 0.8321 F1 0.4205 0.3018 0.1551 0.0191

[0161] (5) Deploy the model, save the training model, and deploy it in the intranet environment.

[0162] (6) Anomaly detection: All work teams in City A were checked for reconciliation issues. Once an anomaly was detected, the reconciliation status of the work team was automatically sent to the material administrators in each area via email for verification. This method detected 208 anomalies in network cable and fiber optic cable reconciliation in July, of which 198 were confirmed to be anomalies by manual verification, with an accuracy rate of 85.57%. Figure 3d This is a schematic diagram of the results of the verification anomaly detection.

[0163] The material verification detection method provided in this application, based on existing serial code verification, uses installation and maintenance work order information and historical material usage as algorithm inputs to perform anomaly detection on auxiliary materials and terminals in installation and maintenance verification. This allows for early warning and control of the use of auxiliary materials and terminals, preventing excessive material usage by installation and maintenance personnel, ultimately achieving the goal of saving resources and reducing costs. This patent achieves accurate verification anomaly detection by designing an anomaly detection Transformer, introducing a Gated Residual Network (GRN), and using a random replacement data generation method.

[0164] It has the following advantages: 1. By combining an improved attention mechanism with Fourier transform, linear embedding, encoder, and decoder, it extracts potential information and periodic features within a certain time window, effectively avoiding misjudgments of reconciliation anomalies caused by non-standard installation and maintenance operations. 2. Addressing the problem of complex relationships in material usage due to numerous related factors that cannot be directly calculated, a gated residual network is introduced to integrate information on material reconciliation status and related variables such as rural / urban areas, work order status, and user business status, enabling more accurate detection of reconciliation anomalies. 3. Addressing the data imbalance problem where reconciliation anomalies account for a small percentage of the total data, a self-supervised deep learning process is constructed by designing a random replacement anomaly generation method. This solves the data imbalance problem to some extent and can more accurately detect material loss and material reconciliation anomalies caused by upstream system anomalies.

[0165] Figure 4 This is a structural diagram of a material verification and detection device according to an embodiment of this application, as shown below. Figure 4 As shown, the device includes:

[0166] The acquisition module 42 is used to acquire material reimbursement time series data and work order data. The material reimbursement time series data represents the quantity of materials reimbursed within a preset time window, and the work order data represents the work order information corresponding to multiple business types within the preset time window.

[0167] The fusion module 44 is used to perform feature fusion on the material reimbursement time series data and work order data to obtain fused data;

[0168] Processing module 46 is used to process the fused data through a deep learning model to obtain an output sequence. The deep learning model includes a multi-head attention mechanism and a Fourier transform attention mechanism, which are used to extract key features from the fused data.

[0169] Module 48 is used to determine the test results of material verification through the output sequence.

[0170] In the aforementioned material verification detection device, the acquisition module 42, fusion module 44, processing module 46, and determination module 48 achieve the goal of determining the material verification detection result based on the output sequence. This realizes the technical effect of comprehensively detecting and verifying material verification, and solves the technical problem in related technologies where consistency verification is performed on the serial codes of terminal devices, but when the devices do not support serial codes, abnormal detection of these devices cannot be performed, resulting in incomplete verification. A detailed explanation follows.

[0171] In the acquisition module of the aforementioned material verification detection device, after acquiring the material verification time series data and work order data, the acquisition module processes the material verification time series data in at least one of the following ways: soft reset, peak noise, length replacement, and smooth replacement. Soft reset is used to generate random noise based on the material verification time series data; peak noise is used to add a sharp data noise to the material verification time series data; length replacement is used to stretch or shrink the material verification time series data in the time dimension according to a preset ratio; and smooth replacement is used to replace all data in the first time period of the material verification time series data with data from any point in the first time period.

[0172] In the fusion module of the aforementioned material verification detection device, the fusion module is used to process the second weight, material verification time series data, third weight, work order data, and second bias value through a first activation function to obtain a second intermediate variable; determine the first intermediate variable based on the first weight, second intermediate variable, and first bias value; process the fourth weight, first intermediate variable, third bias value, fifth weight, and fourth bias value through a second activation function to obtain residual data; obtain the first feature data by standardizing the material verification time series data and residual data; and obtain fused data by processing the first feature data through a third activation function.

[0173] In the fusion module of the aforementioned material verification detection device, the fusion module is also used to process the first feature data, the sixth weight, and the fifth bias value through the third activation function to obtain the third intermediate variable; and to process the third intermediate variable, the seventh weight, and the sixth bias value through the third activation function to obtain fused data.

[0174] In the processing module of the aforementioned material verification detection device, the processing module is used to process the fused data through a multi-head attention mechanism to obtain a first result; and to process the fused data through a Fourier transform attention mechanism to obtain a second result.

[0175] In the processing module of the aforementioned material verification detection device, the processing module is further used to normalize the sixth bias value, the first data, and the second data using the fourth activation function to obtain a fourth intermediate variable. The first data is determined by the data in the first position of the fused data and the eighth weight, and the second data is determined by the data in the second position of the fused data and the ninth weight. Based on the fourth intermediate variable and the third data, attention variables corresponding to the first data, the second data, and the third data are determined. The third data is determined by the data in the third position of the fused data and the tenth weight. Based on the attention variables and the eleventh weight, the first result is determined.

[0176] In the processing module of the aforementioned material verification detection device, the processing module is further used to process the data in the first position, the data in the second position, and the data in the third position of the fused data respectively through Fourier transform to obtain a first transform value, a second transform value, and a third transform value; to process the variance between the first transform value, the second transform value, and the data in the first position of the fused data according to the fifth activation function to obtain a fifth intermediate variable; and to obtain a second result by performing an inverse Fourier transform on the fifth intermediate variable and the third transform value.

[0177] In the processing module of the aforementioned material verification detection device, the processing module is further used to determine the twelfth weight corresponding to the first result and the thirteenth weight corresponding to the second result; and to determine the output sequence based on the first result, the twelfth weight, the second result and the thirteenth weight.

[0178] In the processing module of the aforementioned material verification detection device, the processing module is also used to process the fourteenth weight, the output sequence and the seventh bias value through the sixth activation function to obtain the sixth intermediate variable; and to process the sixth intermediate variable, the fifteenth weight and the eighth bias value through the sixth activation function to obtain the second output sequence.

[0179] In the determination module of the aforementioned material verification detection device, the determination module is used to determine that the material verification detection result is normal when the value in the second output sequence obtained after processing the output sequence is the first value; and to determine that the material verification detection result is abnormal when the value in the second output sequence is the second value.

[0180] It should be noted that, Figure 4 The material verification detection device shown is used to perform... Figure 2 The material verification detection method shown above also applies to the material verification detection device, and will not be repeated here.

[0181] This application also provides an electronic device, including: a memory for storing program instructions; and a processor connected to the memory for executing program instructions to perform the following functions: acquiring material reimbursement time series data and work order data, wherein the material reimbursement time series data represents the quantity of materials reimbursed within a preset time window, and the work order data represents work order information corresponding to multiple business types within the preset time window; fusing the material reimbursement time series data and work order data to obtain fused data; processing the fused data using a deep learning model to obtain an output sequence, wherein the deep learning model includes a multi-head attention mechanism and a Fourier transform attention mechanism, which are used to extract key features from the fused data; and determining the detection result of material reimbursement through the output sequence.

[0182] The aforementioned electronic devices achieve the goal of determining the detection results of material verification based on the output sequence, thereby realizing the technical effect of comprehensively detecting and verifying material verification. This solves the technical problem in related technologies where consistency verification is performed on the serial codes of terminal devices, but when the devices do not support serial codes, abnormal detection of these devices cannot be performed, resulting in incomplete verification.

[0183] It should be noted that the aforementioned electronic equipment is used to perform Figure 2 The material verification detection method shown above also applies to this electronic device, and will not be repeated here.

[0184] This application embodiment also provides a non-volatile storage medium, which includes a stored computer program. The device containing the non-volatile storage medium executes the following material verification detection method by running the computer program: acquiring material verification time-series data and work order data, wherein the material verification time-series data represents the quantity of materials verified within a preset time window, and the work order data represents work order information corresponding to multiple business types within the preset time window; fusing the material verification time-series data and work order data to obtain fused data; processing the fused data using a deep learning model to obtain an output sequence, wherein the deep learning model includes a multi-head attention mechanism and a Fourier transform attention mechanism, which are used to extract key features from the fused data; and determining the detection result of the material verification based on the output sequence.

[0185] By using the aforementioned non-volatile storage medium, the purpose of determining the detection result of material verification based on the output sequence is achieved, thereby realizing the technical effect of comprehensively detecting and verifying material verification. This solves the technical problem in related technologies where consistency verification is performed on the serial codes of terminal devices, but when the devices do not support serial codes, abnormal detection of these devices cannot be performed, resulting in incomplete verification.

[0186] It should be noted that the aforementioned non-volatile storage media is used for execution. Figure 2 The material write-off detection method shown above also applies to this non-volatile storage medium, and will not be repeated here.

[0187] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0188] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0189] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0190] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0191] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0192] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0193] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for detecting material write-offs, characterized in that, include: Obtain material reimbursement time series data and work order data, wherein the material reimbursement time series data represents the quantity of materials reimbursed within a preset time window, and the work order data is used to represent work order information corresponding to multiple business types within the preset time window; The material reimbursement time series data and the work order data are fused using feature fusion to obtain fused data; The fused data is processed by a deep learning model to obtain an output sequence. The deep learning model includes a multi-head attention mechanism and a Fourier transform attention mechanism, which are used to extract key features from the fused data. The detection results of material verification are determined by the output sequence; The fused data is processed using a deep learning model, including: processing the fused data using the multi-head attention mechanism to obtain a first result; and processing the fused data using the Fourier transform attention mechanism to obtain a second result. The multi-head attention mechanism is used to process the fused data to obtain a first result, including: normalizing the sixth bias value, the first data, and the second data using a fourth activation function to obtain a fourth intermediate variable, wherein the first data is determined by the data in the first position of the fused data and the eighth weight, and the second data is determined by the data in the second position of the fused data and the ninth weight; determining the attention variables corresponding to the first data, the second data, and the third data based on the fourth intermediate variable and the third data, wherein the third data is determined by the data in the third position of the fused data and the tenth weight; and determining the first result based on the attention variables and the eleventh weight. The second result is obtained by processing the fused data through the Fourier transform attention mechanism, including: processing the data at the first position, the data at the second position, and the data at the third position in the fused data through Fourier transform to obtain a first transform value, a second transform value, and a third transform value; processing the variance between the first transform value, the second transform value, and the data at the first position in the fused data according to a fifth activation function to obtain a fifth intermediate variable; and obtaining the second result by performing an inverse Fourier transform on the fifth intermediate variable and the third transform value. The fused data is processed by a deep learning model to obtain an output sequence, including: determining a twelfth weight corresponding to the first result and a thirteenth weight corresponding to the second result; and determining the output sequence based on the first result, the twelfth weight, the second result, and the thirteenth weight.

2. The method according to claim 1, characterized in that, After obtaining the material reimbursement time series data and work order data, the method further includes: The material write-off time series data shall be processed by at least one of the following methods: soft reset, peak noise, length replacement, and smooth replacement, wherein, The soft reset is used to generate random noise based on the material write-off time series data; The peak noise is used to add a sharp data noise to the material write-off time series data; The length replacement is used to stretch or shrink the time dimension according to a preset ratio based on the material write-off time series data; The smooth replacement is used to replace all data within a first time period in the material verification time series data with data from any point within the first time period.

3. The method according to claim 1, characterized in that, The material reimbursement time series data and the work order data are fused using feature extraction to obtain fused data, including: The second intermediate variable is obtained by processing the second weight, the material reimbursement time series data, the third weight, the work order data, and the second bias value through the first activation function; The first intermediate variable is determined based on the first weight, the second intermediate variable, and the first bias value; The residual data is obtained by processing the fourth weight, the first intermediate variable, the third bias value, the fifth weight, and the fourth bias value through the second activation function; After standardizing the material write-off time series data and the residual data, the first feature data is obtained; The first feature data is processed by a third activation function to obtain the fused data.

4. The method according to claim 3, characterized in that, The first feature data is processed by a third activation function to obtain the fused data, including: The third intermediate variable is obtained by processing the first feature data, the sixth weight, and the fifth bias value through the third activation function. The fused data is obtained by processing the third intermediate variable, the seventh weight, and the sixth bias value through the third activation function.

5. The method according to claim 1, characterized in that, After processing the fused data using a deep learning model, the method further includes: The sixth intermediate variable is obtained by processing the fourteenth weight, the output sequence, and the seventh bias value through the sixth activation function; The sixth intermediate variable, the fifteenth weight, and the eighth bias value are processed by the sixth activation function to obtain the second output sequence.

6. The method according to claim 5, characterized in that, The detection results of material verification are determined by the output sequence, including: When the value in the second output sequence obtained after processing the output sequence is the first value, the detection result of the material verification is determined to be normal. When the value in the second output sequence is the second value, the detection result of the material verification is determined to be abnormal.

7. A detection device for material verification, characterized in that, include: The acquisition module is used to acquire material reimbursement time series data and work order data. The material reimbursement time series data represents the quantity of materials reimbursed within a preset time window, and the work order data represents work order information corresponding to multiple business types within the preset time window. The fusion module is used to perform feature fusion of the material reimbursement time series data and the work order data to obtain fused data; The processing module is used to process the fused data through a deep learning model to obtain an output sequence. The deep learning model includes a multi-head attention mechanism and a Fourier transform attention mechanism, which are used to extract key features from the fused data. A determination module is used to determine the detection result of material verification based on the output sequence; The processing module is further configured to process the fused data using the multi-head attention mechanism to obtain a first result; and to process the fused data using the Fourier transform attention mechanism to obtain a second result; The processing module is further configured to normalize the sixth bias value, the first data, and the second data using a fourth activation function to obtain a fourth intermediate variable, wherein the first data is determined by the data in the first position of the fused data and the eighth weight, and the second data is determined by the data in the second position of the fused data and the ninth weight; based on the fourth intermediate variable and the third data, to determine the attention variables corresponding to the first data, the second data, and the third data, wherein the third data is determined by the data in the third position of the fused data and the tenth weight; and based on the attention variables and the eleventh weight, to determine the first result; The processing module is further configured to process the data at the first position, the data at the second position, and the data at the third position in the fused data using Fourier transform to obtain a first transform value, a second transform value, and a third transform value; process the variance between the first transform value, the second transform value, and the data at the first position in the fused data according to a fifth activation function to obtain a fifth intermediate variable; and obtain the second result by performing an inverse Fourier transform on the fifth intermediate variable and the third transform value. The processing module is further configured to determine the twelfth weight corresponding to the first result and the thirteenth weight corresponding to the second result; and to determine the output sequence based on the first result, the twelfth weight, the second result and the thirteenth weight.

8. An electronic device, characterized in that, include: Memory, used to store program instructions; A processor, connected to the memory, is configured to execute program instructions to perform the following functions: acquiring material reimbursement time-series data and work order data, wherein the material reimbursement time-series data represents the quantity of materials reimbursed within a preset time window, and the work order data represents work order information corresponding to multiple business types within the preset time window; fusing the material reimbursement time-series data and the work order data to obtain fused data; and processing the fused data using a deep learning model to obtain an output sequence, wherein the deep learning model includes a multi-head attention mechanism and a Fourier transform attention mechanism, and the multi-head attention mechanism and the Fourier transform attention mechanism are used ... The process involves extracting key features from the fused data; determining the detection result of material write-off through the output sequence; and processing the fused data using a deep learning model, including: processing the fused data through the multi-head attention mechanism to obtain a first result; processing the fused data through the Fourier transform attention mechanism to obtain a second result; and processing the fused data through the multi-head attention mechanism to obtain the first result, including: normalizing the sixth bias value, the first data, and the second data through the fourth activation function to obtain a fourth intermediate variable, wherein the first data is determined by the data in the first position of the fused data and the eighth weight. The second data is determined by the data in the second position of the fused data and the ninth weight; based on the fourth intermediate variable and the third data, attention variables corresponding to the first data, the second data, and the third data are determined, wherein the third data is determined by the data in the third position of the fused data and the tenth weight; based on the attention variable and the eleventh weight, the first result is determined; the fused data is processed through the Fourier transform attention mechanism to obtain the second result, including: processing the data in the first position, the data in the second position, and the data in the third position of the fused data through Fourier transform respectively. The data is processed to obtain a first transform value, a second transform value, and a third transform value; the variance between the first transform value, the second transform value, and the data at the first position in the fused data is processed according to a fifth activation function to obtain a fifth intermediate variable; the second result is obtained by performing an inverse Fourier transform on the fifth intermediate variable and the third transform value; the fused data is processed by a deep learning model to obtain an output sequence, including: determining a twelfth weight corresponding to the first result and determining a thirteenth weight corresponding to the second result; and determining the output sequence based on the first result, the twelfth weight, the second result, and the thirteenth weight.

9. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored computer program, wherein the device containing the non-volatile storage medium executes the material write-off detection method according to any one of claims 1 to 6 by running the computer program.

Images