Power battery fault management and control method and system

Abstract

The invention provides a power battery fault management and control method and system.The method comprises the steps that a fault data set and a knowledge graph are obtained, the fault data set comprises various signal time-frequency graphs of a power battery and corresponding fault descriptions, and the fault descriptions comprise fault phenomena and fault keywords; the knowledge graph comprises a fault phenomenon, a fault keyword, a fault reason and a mutual relation between a management and control decision; performing multi-task reasoning training on the multi-task large model according to the fault data set to obtain an initial fault management and control model; performing multi-task reasoning optimization training on the initial fault management and control model by taking a mutual relationship in the knowledge graph as a reasoning constraint to obtain a target fault management and control model; and performing fault management and control on the power battery by using the target fault management and control model. The target fault management and control model is obtained in a mode of combining the multi-task large model and the knowledge graph, power battery fault management and control are achieved, and the precision and efficiency of power battery fault diagnosis are effectively improved.

Description

Technical Field

[0001] The present application relates to the field of battery management technology, and in particular to a power battery fault control method and system. Background Art

[0002] With the global energy crisis and the continuous improvement of environmental protection awareness, new energy vehicles, as an environmentally friendly and efficient means of transportation, have received widespread attention and rapid development. One of the core components of new energy vehicles is the power battery, whose performance directly affects the vehicle's cruising range, driving experience and safety. At present, the power battery of new energy vehicles may have a variety of faults during actual use, including but not limited to excessive battery pressure difference, battery failure to fully charge, battery cruising range reduction, vehicle loss of power, etc. These faults not only affect the normal use of the vehicle, but also may pose a potential threat to driving safety.

[0003] In the related technologies, the BMS (Battery Management System) of new energy vehicles is constantly developing, but the fault identification and early warning capabilities of its power batteries still need to be improved, and there are phenomena of untimely fault warnings or false alarms. In addition, once a power battery fails, it is also crucial to quickly and accurately locate the fault point and take corresponding treatment measures. However, the current power battery fault handling efficiency is low and the emergency response mechanism is not sound enough. Therefore, the fault control of new energy vehicle power batteries still has deficiencies in diagnostic accuracy and efficiency. Summary of the invention

[0004] In view of the above shortcomings of the prior art, the present application discloses a power battery fault management method and system, which are used to solve the technical problems of low accuracy and efficiency of power battery fault diagnosis in the prior art.

[0005] In a first aspect, the present application provides a method for power battery fault management, the method comprising: obtaining a fault data set and a knowledge graph, the fault data set comprising various signal time-frequency diagrams of the power battery and fault descriptions corresponding to each of the signal time-frequency diagrams, the fault descriptions comprising fault phenomena and fault keywords, the knowledge graph comprising the relationship between fault phenomena, fault keywords, fault causes and management decisions; constructing a multi-task large model, and performing multi-task reasoning training on the multi-task large model according to the fault data set to obtain an initial fault management model, the multi-task comprising identification of the fault keywords, analysis of the fault causes and formulation of the management decisions; using the relationships in the knowledge graph as reasoning constraints, performing multi-task reasoning optimization training on the initial fault management model to obtain a target fault management model; and using the target fault management model to perform fault management on the power battery.

[0006] In one embodiment of the present application, the method for constructing the multi-task large model includes: constructing a basic large model, the basic large model includes at least one sub-model, the sub-model integrates a visual transformer structure and a bidirectional encoder representation transformation structure, the visual transformer structure is used to process the signal time-frequency diagram, and the bidirectional encoder representation transformation structure is used to process the fault description; fine-graining the forward feedback network in the visual transformer structure and the bidirectional encoder representation transformation structure to obtain multiple fine-grained experts, and setting at least one shared expert among the multiple fine-grained experts, the fine-grained expert is used to process the unit features in the signal time-frequency diagram or the fault description, and the shared expert is used to process the shared features of each of the unit features; initializing the gating values ​​of each of the fine-grained experts and each of the shared experts, and constructing a cross-expert multi-task loss function to obtain a multi-task large model, the multi-task loss function is composed of a task loss function and an expert interaction loss function, and the multi-task large model is a cross-expert multi-task large model including at least one cross-expert multi-task sub-model.

[0007] In one embodiment of the present application, multi-task reasoning training is performed on the multi-task large model according to the fault data set to obtain an initial fault control model, including: labeling each group of signal time-frequency diagrams and corresponding fault descriptions in the fault data set with actual labels, and the actual labels include actual fault keywords, actual fault causes and actual control decisions; using the labeled fault data set to perform multi-task reasoning training on the multi-task large model, iteratively updating the visual transformer structure in the cross-expert multi-task sub-model and the number of shared experts in the bidirectional encoder representation transformation structure and the learning parameters of the path network in the cross-expert multi-task sub-model, and the learning parameters include an expert allocation weight matrix and a noise adjustment weight matrix; until the loss value of the multi-task loss function converges to a preset loss value, the initial fault control model is obtained.

[0008] In one embodiment of the present application, the method for constructing the multi-task loss function includes: constructing a subtask loss function based on the prediction error of each subtask in the multi-task, and determining the task loss function based on the preset task weights of each subtask and each subtask loss function; for each cross-expert multi-task sub-model, constructing the interaction information between the multi-task and the fine-grained expert and the shared expert, and determining the expert interaction loss function based on the interaction information corresponding to each cross-expert multi-task sub-model and the preset information sharing influence weight; constructing the multi-task loss function based on the task loss function and the expert interaction loss function.

[0009] In one embodiment of the present application, the construction of the interactive information between the multi-tasks and the fine-grained experts and the shared experts includes: constructing a joint probability that any subtask is assigned to any expert, constructing a first marginal probability that any subtask is not assigned to any expert, and constructing a second marginal probability that any expert is not assigned any subtask; constructing a joint entropy between the multi-tasks and the fine-grained experts and the shared experts based on the joint probability, constructing a first marginal entropy of the multi-tasks based on the first marginal probability, and constructing a second marginal entropy between the fine-grained experts and the shared experts based on the second marginal probability; constructing the interactive information based on the joint entropy, the first marginal entropy and the second marginal entropy.

[0010] In one embodiment of the present application, the mutual relationships in the knowledge graph are used as reasoning constraints, and multi-task reasoning optimization training is performed on the initial fault control model to obtain a target fault control model, including: inputting the fault data set into the initial fault control model, and extracting paths from the knowledge graph based on the fault phenomena and the fault keywords corresponding to each of the signal time-frequency graphs to determine an initial path set, where the paths are from fault keywords to fault causes and then to control decisions; calculating the occurrence probability of each path in the initial path set, and performing search and reasoning on each path in the initial path set based on the occurrence probability to determine a target path set, where the occurrence probability of each path in the target path set is higher than a preset probability threshold; performing multi-task reasoning optimization training on the initial fault control model based on each path in the initial path set and each path in the target path set to obtain the target fault control model.

[0011] In one embodiment of the present application, the initial fault control model is subjected to multi-task reasoning optimization training according to each path in the initial path set and each path in the target path set to obtain the target fault control model, including: constructing a composite loss function for multi-task reasoning optimization training, the composite loss function being composed of a first probability of generating a reliable path in path extraction and a second probability of generating a correct path in retrieval reasoning, the first probability being determined by evaluating the reliability of each path in the initial path set, and the second probability being determined by evaluating the correctness of each path in the target path set; with the goal of maximizing the first probability and the second probability, the weight parameters in the initial fault control model are iteratively updated to obtain the target fault control model.

[0012] In one embodiment of the present application, there are multiple cross-expert multi-task sub-models; before obtaining the multi-task large model, it also includes: setting a regularization layer between two of the cross-expert multi-task sub-models, and performing residual connection on the multiple cross-expert multi-task sub-models; setting a learnable embedding vector in the hidden layer of each cross-expert multi-task sub-model, and the embedding vector is used to realize the perception of the multiple tasks by each of the fine-grained experts and each of the shared experts.

[0013] In one embodiment of the present application, the use of the target fault management and control model to perform fault management on the power battery includes: obtaining a target signal time-frequency diagram and a corresponding target fault description of the power battery to be tested; inputting the target signal time-frequency diagram and the target fault description into the target fault management and control model to obtain target fault keywords, target fault causes and target management and control decisions corresponding to the power battery to be tested.

[0014] In the second aspect, the present application provides a power battery fault management and control system, the system comprising: an acquisition module, used to acquire a fault data set and a knowledge graph, the fault data set comprising various signal time-frequency diagrams of the power battery and fault descriptions corresponding to each of the signal time-frequency diagrams, the fault descriptions comprising fault phenomena and fault keywords, the knowledge graph comprising the relationship between fault phenomena, fault keywords, fault causes and management and control decisions; a pre-training module, used to construct a multi-task large model, and perform multi-task reasoning training on the multi-task large model according to the fault data set to obtain an initial fault management and control model, the multi-task comprising the identification of the fault keywords, the analysis of the fault causes and the formulation of the management and control decisions; an optimization training module, used to use the relationship in the knowledge graph as a reasoning constraint, perform multi-task reasoning optimization training on the initial fault management and control model to obtain a target fault management and control model; a management and control module, used to perform fault management and control on the power battery using the target fault management and control model.

[0015] As described above, a power battery fault control method and system provided in the embodiments of the present application have the following beneficial effects:

[0016] First, the fault data set and knowledge graph are obtained. The fault data set includes various signal time-frequency diagrams of the power battery and the fault descriptions corresponding to each signal time-frequency diagram. The fault description includes fault phenomena and fault keywords. The knowledge graph includes the relationship between fault phenomena, fault keywords, fault causes and management decisions. Then, a multi-task large model is constructed, and multi-task reasoning training is performed on the multi-task large model according to the fault data set to obtain an initial fault management model. The multi-task includes the identification of fault keywords, the analysis of fault causes and the formulation of management decisions. Then, the relationships in the knowledge graph are used as reasoning constraints, and the initial fault management model is optimized through multi-task reasoning training to obtain the target Fault control model, and finally use the target fault control model to control the power battery fault. The target fault control model is obtained by combining the multi-task large model and the knowledge graph. Using this target fault control model, the fault keywords and fault causes of the power battery can be accurately predicted and the corresponding control decisions can be obtained. Moreover, through the automated process of fault identification, cause analysis and measure formulation, fault diagnosis can be completed quickly and corresponding measures can be obtained, which reduces manual intervention and decision-making time, improves fault decision-making efficiency and the intelligent level of fault diagnosis, ensures the fault identification and early warning capabilities of the power battery, and ensures the accuracy and efficiency of power battery fault diagnosis.

[0017] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings herein are incorporated into the specification and constitute a part of the specification, showing embodiments consistent with the present application, and together with the specification, are used to explain the principles of the present application. Obviously, the drawings described below are only some embodiments of the present application, and for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0019] Figure 1 is a schematic diagram of an implementation environment of a power battery fault management and control system shown in an exemplary embodiment of the present application;

[0020] Figure 2 is a flow chart of a power battery fault control method shown in an exemplary embodiment of the present application;

[0021] Figure 3 is a structural diagram of a cross-expert multi-task large model shown in an exemplary embodiment of the present application;

[0022] Figure 4 is a structural diagram of a transformer architecture shown in an exemplary embodiment of the present application;

[0023] Figure 5 is a schematic diagram of the evolution of a sub-model in a large cross-expert multi-task model shown by an exemplary embodiment of the present application;

[0024] Figure 6 is a structural diagram of a target fault management model shown in an exemplary embodiment of the present application;

[0025] Figure 7 It is a block diagram of a power battery fault management system shown in an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0026] The following will describe the implementation methods of the present application with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. The present application can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be understood that the preferred embodiments are only for illustrating the present application, not for limiting the scope of protection of the present application.

[0027] It should be noted that the illustrations provided in the following embodiments are only used to illustrate the basic concept of the present application in a schematic manner, and therefore the illustrations only show components related to the present application rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the form, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0028] In the following description, a large number of details are discussed to provide a more thorough explanation of the embodiments of the present application. However, it is obvious to those skilled in the art that the embodiments of the present application can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present application difficult to understand.

[0029] At present, there are various types of power battery failures in new energy vehicles, including but not limited to excessive battery pressure difference, incomplete battery charge, battery life decline, vehicle power loss, etc. These failures not only affect the normal use of the vehicle, but may also pose a potential threat to driving safety. Although the BMS of new energy vehicles is constantly developing, its fault identification and early warning capabilities still need to be improved. The inventors of this application have found that existing fault diagnosis often relies on fixed threshold judgments, which is difficult to adapt to complex and changeable actual operating conditions, resulting in untimely fault warnings or frequent false alarms. In addition, the monitoring and diagnosis capabilities of the internal microscopic state of the battery are relatively weak, making it difficult to detect potential safety hazards in advance. Once a power battery fails, it is crucial to quickly and accurately locate the fault point and take corresponding treatment measures.

[0030] Therefore, see Figure 1 , Figure 1 FIG. 1 is a schematic diagram of an implementation environment of a power battery fault management system shown in an exemplary embodiment of the present application. Figure 1 As shown, the implementation environment includes a vehicle 110 and a power battery fault management and control system 120, wherein the power battery fault management and control system 120 is embedded in the vehicle 110, and is used to realize fault management of the power battery in the vehicle 110. The power battery fault management and control system 120 includes but is not limited to a vehicle system, an on-board computer, etc., and obtains a target fault management and control model by combining a multi-task large model with a knowledge graph. By using the target fault management and control model, the fault keywords and fault causes of the power battery can be accurately predicted and corresponding management and control decisions can be obtained. Moreover, through the automated process of fault identification, cause analysis and measure formulation, fault diagnosis can be quickly completed and corresponding measures can be obtained, thereby reducing manual intervention and decision-making time, improving fault decision-making efficiency and the intelligent level of fault diagnosis, and ensuring the fault identification and early warning capabilities of the power battery, so that the accuracy and efficiency of power battery fault diagnosis are guaranteed.

[0031] See also Figure 2 , Figure 2 is a flowchart of a power battery fault control method shown in an exemplary embodiment of the present application. The method can be applied to Figure 1 The implementation environment shown is a schematic diagram. It should be understood that the method may also be applicable to other exemplary implementation environments, and this embodiment does not limit the implementation environment to which the method is applicable.

[0032] like Figure 2 As shown, in an exemplary embodiment, the power battery fault management method includes at least steps S210 to S240, which are described in detail as follows:

[0033] Step S210, obtaining a fault data set and a knowledge graph, wherein the fault data set includes various signal time-frequency diagrams of the power battery and the fault descriptions corresponding to each signal time-frequency diagram, the fault descriptions include fault phenomena and fault keywords, and the knowledge graph includes the relationship between fault phenomena, fault keywords, fault causes and management decisions.

[0034] In this embodiment, the various signal time-frequency diagrams include time-frequency diagrams of signals such as current, voltage and temperature obtained by wavelet transform.

[0035] In a possible embodiment, the knowledge graph includes fault keywords, fault phenomena, fault causes and the mutual relationship between management and control decisions, specifically including causal correspondence, response strategy correspondence, fault cause compound relationship, fault concurrency relationship and fault sequence relationship, etc. Among them, the causal correspondence, such as the relationship between the fault cause and the fault keyword, indicates that a specific cause leads to the occurrence of a certain fault, such as the relationship between the fault keyword and the fault phenomenon, indicating that a certain fault will show a specific phenomenon; the response strategy correspondence, such as the relationship between the fault keyword and the management and control decision, indicates the recommended management and control measures for a certain fault, such as the relationship between the fault cause and the management and control decision, indicates the recommended management and control measures for a certain fault cause; the fault cause compound relationship, such as the relationship between the fault cause and the fault cause, indicates that multiple causes work together to cause a fault; the fault concurrency relationship, such as the relationship between the fault keyword and the fault keyword, indicates that two or more faults occur at the same time; the fault sequence relationship, such as the relationship between the fault keyword and the fault keyword, indicates that one fault may cause another fault to occur.

[0036] Step S220, construct a multi-task large model, and perform multi-task reasoning training on the multi-task large model according to the fault data set to obtain an initial fault management and control model. The multi-tasks include recognition of fault keywords, analysis of fault causes, and formulation of management and control decisions.

[0037] Among them, the fault keyword is the fault type or fault name, the fault cause is the cause of a certain fault, and the management and control decision is the operation and maintenance measures corresponding to the specific fault.

[0038] In this embodiment, various signal time-frequency diagrams of the power battery and the fault descriptions corresponding to each signal time-frequency diagram are input into the multi-task big model as training samples, and the multi-task big model is trained with multi-task reasoning for fault keyword recognition, fault cause analysis, and control decision-making. In this way, an automated process of fault identification, cause analysis, and measure formulation can be implemented based on the big model, thereby improving the efficiency of power battery fault control and the intelligence level of fault diagnosis.

[0039] In one embodiment, a method for constructing a multi-task large model includes: constructing a basic large model, the basic large model includes at least one sub-model, the sub-model integrates a visual transformer structure and a bidirectional encoder representation transformation structure, the visual transformer structure is used to process a signal time-frequency diagram, and the bidirectional encoder representation transformation structure is used to process a fault description; fine-graining the forward feedback network in the visual transformer structure and the bidirectional encoder representation transformation structure to obtain a plurality of fine-grained experts, and setting at least one shared expert among the plurality of fine-grained experts, the fine-grained experts are used to process unit features in a signal time-frequency diagram or a fault description, and the shared experts are used to process shared features of each unit feature; initializing the gating values ​​of each fine-grained expert and each shared expert, and constructing a cross-expert multi-task loss function to obtain a multi-task large model, the multi-task loss function is composed of a task loss function and an expert interaction loss function, and the multi-task large model is a cross-expert multi-task large model including at least one cross-expert multi-task sub-model.

[0040] See also Figure 3 , Figure 3 is a structural diagram of a cross-expert multi-task large model shown in an exemplary embodiment of the present application. Figure 3As shown, the input of the cross-expert multi-task large model is the various signal time-frequency diagrams of the power battery and the corresponding fault descriptions, and the output is three fault control subtasks: fault keywords, fault causes, and control decisions. The cross-expert multi-task large model is composed of several multi-task large model sub-modules (i.e., cross-expert multi-task sub-models) including VIT (Vision Transformer) and BERT (Bidirectional Encoder Representations from Transformers) structures. The VIT structure is used to capture the high-dimensional features of the signal time-frequency diagram, and the BERT structure is used to capture the text features in the fault description. The number of cross-expert multi-task sub-models can be specified according to the training data to meet the needs of deep feature mining of different data. In each cross-expert multi-task sub-model, the VIT structure and the BERT structure both contain multiple transformer architectures. In the FFN (Feed-Forward Network) layer of the transformer architecture, a fine-grained expert strategy and a designated shared expert strategy are used to introduce multiple experts. That is, in the transformer architecture, the FFN is first replaced by N FFNs with smaller dimensions. Each small FFN represents an expert, and multiple fine-grained experts are obtained to keep the expert parameters and the model calculation amount consistent, and at least one shared expert is designated. Among them, the transformer architecture includes multiple FFNs, all FFNs can be fine-grained, or some FFNs can be fine-grained or fine-grained with different dimensions reduced according to needs to meet the needs of mining different data features. In addition, in each cross-expert multi-task sub-model, there is a gating network to set the gating value of each expert, which is used to realize the allocation between different unit feature tokens and different experts.

[0041] See also Figure 4 , Figure 4 is a structural diagram of a transformer architecture shown in an exemplary embodiment of the present application. Figure 4As shown in the figure, the left side is the Encoder part and the right side is the Decoder part. In the Encoder part, the data is first input into the Input Embedding layer to convert the input word into a vector representation, and then added to the Positional Encoding to retain the position information of the word in the input sequence. Next, there are multiple encoder layers (N layers), each of which contains a multi-head attention mechanism and a forward feedback network. In each encoder layer, the multi-head attention mechanism is used to capture the dependencies between words in the input sequence, and the forward feedback network is used for nonlinear transformation; in the Decoder part, the output word is first converted into a vector representation through the Output Embedding layer, and then added to the Positional Encoding. Encoding (position embedding) is added to retain the position information of the word in the output sequence, followed by multiple decoder layers (N layers), each decoder layer contains a masked multi-head attention mechanism, a multi-head attention mechanism and a forward feedback network. In each decoder layer, the masked multi-head attention mechanism is used to capture the dependencies between words in the output sequence while preventing the decoder from accessing future information. The multi-head attention mechanism is used to capture the dependencies between the input sequence and the output sequence, and the forward feedback network is used for nonlinear transformation. Finally, the output of the decoder passes through a linear layer and a Softmax layer to generate the final output probability distribution.

[0042] In the transformer architecture, it mainly includes the attention mechanism and FFN. The attention mechanism can be described as:

[0043]

[0044] Where T represents the number of tokens; represents the output sequence of all tokens in the lth layer; Att represents the attention mechanism; Represents the hidden state of all tokens in the previous layer (l-1 layer); Indicates that the attention mechanism is applied to the hidden states of all tokens in the previous layer (l-1 layer); It means that the output sequence of all tokens in layer l is the sum of the original hidden state and the result adjusted by the attention mechanism. show is a real matrix of size T×d.

[0045] The un-fine-grained improved FFN can be described as:

[0046]

[0047] in, It represents the hidden state of the tth token in the lth layer. FFN represents the feed-forward network. Represents the result of the lth layer after the tth token is adjusted based on the attention mechanism of the previous layer output; Indicates that Conduct a feed-forward network; It means that the final hidden state of the lth layer at the tth token is the direct addition of the output of the forward feedback network and the input. show is a d-dimensional real vector.

[0048] In this embodiment, when the number of experts is limited, each expert will learn a lot of different knowledge, which will lead to excessive model calculation and ineffective use of knowledge. However, if each token is assigned to multiple experts, the diverse knowledge will be decomposed and easy to learn.

[0049] Therefore, in this embodiment, experts are divided into finer granularities to enhance the flexibility and adaptability of expert collaboration and ensure that expert parameters and model calculations are consistent. Thus, mN experts with smaller granularity are obtained, which improves the model's ability to learn diverse knowledge. After adopting the fine-grained expert strategy, the process of FFN mining features in each cross-expert multi-task sub-model can be described as:

[0050]

[0051] in, The hidden state finally generated by the t-th token in the l-th layer; N represents the number of experts before fine-graining; mN represents the number of experts after fine-graining; i represents the i-th expert; It represents the result of the adjustment of the t-th token in the l-th layer based on the attention mechanism of the output of the previous layer; Indicates that the i-th expert will For the forward feedback network, g i,t Represents the gating value of the ith expert for the tth token.

[0052] The calculation formula of the gate value is:

[0053]

[0054] Among them, g i,t represents the gate value of the ith expert for the tth token; s i,t represents the correlation score between the ith expert and the tth token; mN represents the number of fine-grained experts; Topk({s j,t|1≤j≤mN},mK) represents the set of relevance scores of the top K experts most relevant to the t-th token selected from all experts associated with the t-th token; j is an index variable used to traverse all experts (from 1 to mN), and for each expert j, calculate its relevance score for the t-th token.

[0055] The formula for calculating the relevance score is:

[0056]

[0057] Among them, s i,t represents the correlation score between the i-th expert and the t-th token; Indicates that Perform transposition; represents the i-th expert in the l-th layer; Softmax i (·) indicates normalization processing.

[0058] In addition, considering that when each token is assigned to multiple experts, experts need to share certain knowledge to ensure faster model convergence, but this will also cause redundancy in model parameters. Therefore, designated experts can be used to capture the knowledge shared by different tokens, thereby effectively alleviating redundancy.

[0059] In this embodiment, shared experts are designated from the fine-grained experts to share knowledge, thereby reducing the amount of model calculation and alleviating the problem of parameter redundancy. In VIT, fine-grained experts are used to process unit features in the signal time-frequency diagram, and shared experts are used for the shared features of each unit feature in the signal time-frequency diagram; in BERT, fine-grained experts are used to process unit features in fault descriptions, and shared experts are used for the shared features of each unit feature in the fault description. After adopting the strategy of designating shared experts, the process of FFN mining features in each cross-expert multi-task sub-model can be described as:

[0060]

[0061] Where mN represents the number of all experts; K s represents the number of shared experts; mN-K s represents the number of fine-grained experts; All shared expert processed results; represents the result of all fine-grained expert processing; g i,t represents the gating value of the i-th fine-grained expert for the t-th token.

[0062] If the number of all experts mN = 4, the number of shared experts K s =1, then the number of fine-grained experts mN-K s =3.

[0063] See also Figure 5 , Figure 5 FIG. 1 is a schematic diagram of the evolution of a sub-model in a large cross-expert multi-task model, shown in an exemplary embodiment of the present application. Figure 5 As shown in Figure 2, before FFN is fine-grained, there are a total of N experts; then the dimension of each FFN is reduced to the original Then the number of all experts becomes 2N. If one shared expert is designated, that is, the first expert is the shared expert, then the remaining 2N-1 experts are fine-grained experts.

[0064] In this way, the cross-expert multi-task sub-model replaces all FFNs in the transformer architecture with FFNs of smaller dimensions. Each small FFN represents an expert, and adopts fine-grained expert strategies and designated shared expert strategies, which can improve the specificity of experts in the model, reduce the amount of model calculation, and alleviate the problem of parameter redundancy.

[0065] In addition, in this embodiment, considering the problem that the multi-task gradient direction and convergence speed are inconsistent in the training of the multi-task large model, resulting in a large difference in the magnitude of the task loss value, a multi-task loss function across experts is constructed to alleviate the problem that the multi-task gradient direction and convergence speed are inconsistent in the training of the multi-task large model. Considering that the degree of information interaction between experts will affect the accuracy of the model, the process of expert collaborative processing of tasks is considered, and the multi-task loss function includes not only the task loss function part of each subtask in the multi-task, but also adds the expert interaction loss function part.

[0066] In this way, the task loss function and the expert interaction loss function are combined to form a multi-task loss function to achieve model convergence training, alleviate the gradient contradiction between multiple tasks, and ensure the effective interaction between different experts when processing multiple tasks, thereby improving the performance and accuracy of the model.

[0067] As a possible embodiment, in the power battery fault management problem, there are three tasks T1, T2, T3 and a set of fault descriptions and time-frequency signal graphs of the power battery of a new energy vehicle in the Q segment I1, I2, ..., I Q , the task T can be defined as a raw data I consisting of fault description and signal frequency diagram q Mapped to task T(I q ), then the data set D can be defined as the training pair (I q ; T i (I q)). In the power battery fault management problem, each task contains only one training pair, but the raw data consisting of a fault description and a signal frequency graph corresponds to the three subtasks of fault keyword recognition, fault cause analysis, and control decision making. The input of the cross-expert multi-task sub-model is the raw data consisting of a fault description and a signal frequency graph. O , it passes through a series of experts E1, E2, ..., E N The path network G captures the deep features of the input and outputs the weighted sum of the fault features mined by each expert. The output of the cross-expert multi-task sub-model can be simplified as:

[0068]

[0069] Where y represents the output feature vector; N represents the number of all experts, including fine-grained experts and shared experts; x represents the input of each expert; E k (x) represents the fault feature mined by the kth expert, G k (x) represents the weight value corresponding to the kth expert. That is, the fault features mined by each expert are weighted and summed to obtain the final output feature vector.

[0070] For a path network G, it is a network with parameters W g and W noise The formula for calculating the expert weight of the Top-K network is:

[0071] G(x)=TopK(Softmax(xW g +N(0,1)Softplus(xW noise ))) Formula (10)

[0072] Where G(x) represents the expert’s weight value; x represents the expert’s input; Wg represents the expert’s weight matrix; W noise represents the noise adjustment weight matrix; Softplus represents the smoothed estimate of the Relu function; Softmax represents the Softmax function; N(0, 1) represents the standard normal distribution; TopK(·) means setting all elements of the vector except the largest K values ​​to 0.

[0073] The smoothing estimate of the Relu function is expressed as:

[0074] Softplus(x)=log(1+exp(x)) Formula (11)

[0075] In one embodiment, there are multiple cross-expert multi-task sub-models; before obtaining the multi-task large model, it also includes: setting a regularization layer between two cross-expert multi-task sub-models, and performing residual connection on multiple cross-expert multi-task sub-models; setting a learnable embedding vector in the hidden layer of each cross-expert multi-task sub-model, and the embedding vector is used to realize the perception of multiple tasks by each fine-grained expert and each shared expert.

[0076] Among them, the regularization layer is used to prevent overfitting in the neural network architecture; the residual connection is used to solve the gradient disappearance or explosion problem encountered in deep network training; the hidden layer can be the input hidden layer.

[0077] In this embodiment, a regularization layer is set between two cross-expert multi-task sub-models to prevent overfitting, improve model generalization, and perform residual connections on multiple cross-expert multi-task sub-models, thereby alleviating the gradient vanishing and gradient exploding problems in model training. In addition, in order to make each expert aware of the multi-task of fault control, a learnable embedding vector is also added to the hidden layer in the cross-expert multi-task sub-model. In this way, the performance and accuracy of the model multi-task reasoning are further improved.

[0078] Please continue to see Figure 3 , a regularization layer is set between the two cross-expert multi-task sub-models, and a learnable embedding vector is set in the input hidden layer of each cross-expert multi-task sub-model.

[0079] In this embodiment, the output of the cross-expert multi-task sub-model after adding the embedding vector can be expressed as:

[0080]

[0081] Where y represents the output feature vector; N represents the number of all experts, including fine-grained experts and shared experts; x represents the input of each expert; i represents the task id; e i represents the embedding vector of task i; E k (x+e i ) represents the kth expert combined with e i The fault features mined in the case of Represents the weight value of the k-th expert corresponding to the i-th task.

[0082] As a possible embodiment, in the multi-task of power battery fault management and control, regarding the embedding vector, for example, the embedding vector corresponding to the fault keyword recognition task can be (0,0,1), the embedding vector corresponding to the fault cause analysis task can be (0,1,0), and the embedding vector corresponding to the control decision-making task can be (1,0,0).

[0083] In one embodiment, multi-task reasoning training is performed on a multi-task large model according to a fault data set to obtain an initial fault control model, including: annotating each set of signal time-frequency diagrams and corresponding fault descriptions in the fault data set with actual labels, wherein the actual labels include actual fault keywords, actual fault causes, and actual control decisions; using the labeled fault data set to perform multi-task reasoning training on the multi-task large model, iteratively updating the number of shared experts in the visual transformer structure and the bidirectional encoder representation transformation structure in the cross-expert multi-task sub-model and the learning parameters of the path network in the cross-expert multi-task sub-model, wherein the learning parameters include an expert allocation weight matrix and a noise adjustment weight matrix; until the loss value of the multi-task loss function converges to a preset loss value, the initial fault control model is obtained.

[0084] In this embodiment, the multi-task large model has powerful natural language processing and data analysis capabilities. By performing multi-task reasoning training on a large amount of data with detailed labels, the large model is used to deeply mine and analyze the operating data of the power battery. With the convergence of the multi-task loss function as the goal, an accurate fault control model is established, so that the model can more accurately identify faults, analyze the causes of faults, and provide corresponding control decisions, thereby improving the robustness of the model.

[0085] In one embodiment, a method for constructing a multi-task loss function includes: constructing a subtask loss function based on the prediction error of each subtask in the multi-task, and determining the task loss function based on the preset task weights of each subtask and the loss function of each subtask; for each cross-expert multi-task sub-model, constructing the interaction information between the multi-task and the fine-grained experts and shared experts, and determining the expert interaction loss function based on the interaction information corresponding to each cross-expert multi-task sub-model and the preset information sharing influence weight; constructing a multi-task loss function based on the task loss function and the expert interaction loss function.

[0086] In this embodiment, the multi-task loss function is:

[0087]

[0088] Among them, L1 represents the multi-task loss function; M represents the number of tasks; T i represents the i-th task; represents the task weight of the i-th task; represents the subtask loss function of the i-th task; Indicates multiple tasks T1, T2...T M The loss is weighted accumulation; w MI represents the influence weight of total interactive information on total loss (i.e., the influence weight of preset information sharing); T represents the multi-task set {T1, T2, ..., T M}; EY represents the set of all fine-grained experts and shared experts in the cross-expert multi-task sub-model {E1, E2, ..., T K};I(T;E Y ) represents the task T and the expert E in the cross-expert multi-task sub-model Y interactive information; Represents the task T and expert E in multiple cross-expert multi-task sub-models Y The mutual information summation.

[0089] In formula (13), The part represents the task loss function, The part represents the expert interaction loss function.

[0090] In this embodiment, the multi-task loss function composed of the task loss function and the expert interaction loss function can not only alleviate the gradient contradiction problem between multiple tasks, but also ensure the effective interaction between different experts when processing multiple tasks, thereby ensuring the performance and accuracy of multi-task reasoning of cross-expert multi-task large model.

[0091] In a possible embodiment, a subtask loss function is constructed according to the prediction error of each subtask in the multi-task, that is, according to the error between the predicted fault keywords and the actual fault keywords, the error between the predicted fault causes and the actual fault causes, and the error between the predicted management and control decisions and the actual management and control decisions. The loss function of each subtask can be a cross-entropy loss function.

[0092] In one embodiment, constructing interactive information between multi-tasks and fine-grained experts and shared experts includes: constructing a joint probability that any subtask is assigned to any expert, constructing a first marginal probability that any subtask is not assigned to any expert, and constructing a second marginal probability that any expert is not assigned any subtask; constructing a joint entropy between multi-tasks and fine-grained experts and shared experts based on the joint probability, constructing a first marginal entropy of multi-tasks based on the first marginal probability, and constructing a second marginal entropy of fine-grained experts and shared experts based on the second marginal probability; constructing interactive information based on the joint entropy, the first marginal entropy and the second marginal entropy.

[0093] In this embodiment, in order to simulate the process of experts collaborating to handle tasks, a probability model is defined on the task set T and the expert set E. When the model is initialized, a certain probability distribution is randomly assigned to the task T.

[0094] For each cross-expert multi-task sub-model, the task T is transformed into j Assigned to Expert E j The probability (joint probability) can be expressed as:

[0095]

[0096] Among them, E j represents the jth expert; T i represents the i-th task; P(E j |T i ) indicates that task T i Assigned to Expert E j The joint probability of Represents task T i The number of corresponding fault descriptions; Represents task T i By Expert E j Processing from fault description I k The eigenvector x k probability.

[0097] According to the joint probability calculation formula P(E, T) = P(E|T)P(T), we can get P(E) = ∑ T P(E, T). Considering that experts should have certain expertise in a certain task, we use interactive information to reflect it. Interactive information can be expressed as:

[0098]

[0099] Where T represents a multi-task set {T1, T2, ..., T M}; E represents the set of all fine-grained experts and shared experts {E1, E2, ..., T K}; I(T; E) represents the interactive information between multi-task and fine-grained experts and shared experts; M represents the number of tasks; N represents the number of experts; T i represents the i-th task; E j represents the jth expert; P(T i , E j ) represents task T i Assigned to Expert E j The joint probability of i ) represents task T i Not assigned to Expert E j The marginal probability of j ) indicates expert E j Unassigned tasks T i The marginal probability of .

[0100] The joint probability P(T i , E j ) refers to task T i With expert E j The probability of simultaneous occurrence; marginal probability P(T i ) represents task T iThe probability of occurrence, regardless of any particular expert; the marginal probability P(E j ) indicates expert E j The probability of being selected, without regard to any specific task.

[0101] According to formula (15), if experts assign the same frequency to all tasks, the mutual information I(T; E) is 0. If each expert assigns one task, the mutual communication between experts will be captured to the maximum extent, and the mutual information I(T; E) will reach the maximum value of 1. In order to better understand the mutual information, it can be further expressed as:

[0102]

[0103] Where T represents a multi-task set {T1, T2, ..., T M}; E represents the set of all fine-grained experts and shared experts {E1, E2, ..., T K}; I(T; E) represents the interactive information between multi-task and fine-grained experts and shared experts; M represents the number of tasks; N represents the number of experts; T i represents the i-th task; E j represents the jth expert; P(T i , E j ) represents task T i Assigned to Expert E j The joint probability of logP(T i , E j ) represents P(T i , E j )’s natural logarithm; represents the joint entropy between task T and expert E; P(T i ) represents task T i The marginal probability of not being assigned to expert Ej; logP(T i ) represents P(T i )’s natural logarithm; represents the marginal entropy of task T not assigned to expert E; P(E j ) indicates expert E j Unassigned tasks T i The marginal probability of logP(E j ) represents P(E j )’s natural logarithm; represents the marginal entropy of expert E not being assigned task T.

[0104] In this embodiment, the relationship between different subtasks and experts is quantified by defining joint probability, marginal probability and entropy based on these probabilities, and interactive information between multi-tasks and fine-grained experts and shared experts is constructed. In this way, the synergy between different types of tasks can be enhanced, so that the model can perform better when dealing with complex multi-task problems.

[0105] In addition, in formula (16), the first term represents P(T i , E j )=P(E j |T i )P(T i ) of the negative entropy, maximizing this favorable stimulus condition distribution P(E j |T i ) is sharp, because P(T i ) is a constant determined by the data distribution and is not affected by the model’s weight parameters. The second term represents the i ) is a constant determined by , which can be ignored. The third term represents P(E j ) determines the entropy. Maximizing this allows experts to use the data set evenly. Y ) and w MI Construct a cross-expert multi-task loss function to ensure effective interaction between different experts when processing multiple tasks, alleviate the gradient contradiction between multiple tasks, and effectively improve the performance of the model.

[0106] Step S230, using the mutual relationships in the knowledge graph as reasoning constraints, performing multi-task reasoning optimization training on the initial fault control model to obtain the target fault control model.

[0107] In this embodiment, considering that the large model itself is a generative model, it may have hallucination problems when outputting answers, that is, it generates problematic and inaccurate answers. Therefore, after the multi-task large model pre-training is completed to obtain the initial fault management model, the graph reasoning method is used to mine the knowledge in the knowledge graph to enhance the reasoning ability of the large model, thereby alleviating the errors caused by the hallucination problem of the large model and improving the accuracy and efficiency of intelligent operation and maintenance decisions of power battery faults.

[0108] In this way, the cross-expert multi-task large model has powerful natural language processing and data analysis capabilities, and can deeply understand the failure phenomena, causes and impacts of power batteries. Combined with the knowledge graph, it can systematize and structure the failure knowledge of power batteries to form a complete failure knowledge base, making the fault identification and diagnosis process more accurate, and can quickly locate the fault point and reduce misdiagnosis and missed diagnosis.

[0109] In one embodiment, the mutual relationships in the knowledge graph are used as reasoning constraints, and the initial fault control model is trained for multi-task reasoning optimization to obtain a target fault control model, including: inputting the fault data set into the initial fault control model, extracting paths from the knowledge graph based on the fault phenomena and fault keywords corresponding to each signal time-frequency graph, and determining an initial path set, where the paths are from the fault keywords to the fault causes and then to the control decisions; calculating the occurrence probability of each path in the initial path set, and performing search and reasoning on each path in the initial path set according to the occurrence probability to determine a target path set, where the occurrence probability of each path in the target path set is higher than a preset probability threshold; performing multi-task reasoning optimization training on the initial fault control model according to each path in the initial path set and each path in the target path set to obtain a target fault control model.

[0110] Since the cross-expert multi-task large model does not have prior knowledge of the relationship between fault phenomena, fault keywords, fault causes and control decisions in the knowledge graph, the large model cannot directly generate reasoning paths based on the knowledge graph. In addition, the large model cannot correctly understand the reasoning path and cannot perform appropriate reasoning based on the path to obtain the correct answer to the problem.

[0111] Therefore, in this embodiment, the multi-task big model is first pre-trained using the signal time-frequency diagram and fault description of the power battery, and then the pre-trained cross-expert multi-task big model is combined with the knowledge graph for re-training using the graph reasoning method, so as to utilize the relationship between fault phenomena, fault keywords, fault causes and management decisions in the knowledge graph to effectively solve the hallucination problem of the big model, and based on the path extraction and retrieval reasoning method, knowledge is extracted from the knowledge graph into the cross-expert multi-task big model to generate a reliable relationship path, and the cross-expert multi-task big model is driven to perform reasoning training based on the extracted path. First, the knowledge graph is used to generate paths related to multiple tasks, and then the paths are screened, and the answers to the multiple tasks are inferred according to the screened paths. In this way, the accuracy and efficiency of the target fault control model regarding power battery fault control can be effectively improved.

[0112] In addition, the cross-expert multi-task large model combined with the knowledge graph ensures the accuracy of the model's prediction results. It can not only further obtain highly targeted and reliable fault management strategies, including fault prevention and control, such as replacing aging batteries in advance, adjusting charging strategies, etc., thereby reducing the occurrence rate of faults and extending battery life, but also automatically recommend better solutions based on the type of fault and its severity, thereby improving decision-making efficiency and intelligence. At the same time, as a structured knowledge representation method, the knowledge graph facilitates knowledge sharing and collaborative work between different departments and teams. In the new energy vehicle power battery fault management, by building a unified knowledge graph, cross-departmental and cross-domain knowledge sharing and communication can be achieved, further promoting the effectiveness of power battery fault management.

[0113] In this embodiment, each edge in the knowledge graph that represents the relationship between fault phenomena, fault keywords, fault causes and management decisions is provided with an edge weight, wherein the edge weight represents the probability of the existence of a relationship between two entities, and the edge weight can be learned from the historical fault data of the power battery. In the process of multi-task reasoning optimization training of the initial fault control model, firstly, based on the fault phenomena and fault keywords corresponding to each signal time-frequency graph, all possible paths are determined from the knowledge graph to form an initial path set, and then the occurrence probability of each path in the initial path set is calculated according to the edge weight, and the paths with an occurrence probability higher than the probability threshold are integrated into the target path set, and the initial fault control model is constrained to ensure the credibility of the paths in the initial path set and the target path set, so as to realize multi-task reasoning optimization training, so that the prediction performance of the target fault control model can be further guaranteed.

[0114] See also Figure 6 , Figure 6 FIG. 1 is a structural diagram of a target fault control model shown in an exemplary embodiment of the present application. Figure 6 As shown in the figure, the target fault control model includes a cross-expert multi-task large model and a graph reasoning structure. First, the multi-task large model is pre-trained using the signal time-frequency diagram of the power battery and the corresponding fault description to obtain a cross-expert multi-task large model. Then, the pre-trained cross-expert multi-task large model is combined with the knowledge graph and trained again using the graph reasoning structure, thereby effectively solving the hallucination problem of the cross-expert multi-task large model using the knowledge graph. In the graph reasoning structure, it includes a path extraction module and a retrieval reasoning module. The path extraction module mainly uses the knowledge graph to generate paths related to the problem. The retrieval reasoning module mainly screens the paths extracted by the path extraction module and infers the answer to the problem based on the screened paths.

[0115] In this embodiment, the graph reasoning method can be expressed as:

[0116] P θ(a|q, G) = ∑ z∈Z P θ (a|q,z,G)P θ (z|q) Formula (17)

[0117] Among them, θ represents the weight parameter of the cross-expert multi-task large model; P θ (a|q, G) represents the probability of the cross-expert multi-task model generating the answer a given question q and knowledge graph G; z represents the path generated by the cross-expert multi-task model; Z represents the set of possible paths; P θ (z|q) represents the prior probability of generating path z when given question q, which can be calculated by the path extraction module; P θ (a|q, z, G) represents the probability of the cross-expert multi-task large model generating answer a given question q, path z and knowledge graph G, which can be calculated through the retrieval reasoning module. In addition, given question q, the fault cause and control decision are inferred based on the fault phenomenon and fault keywords; answer a is the path, which includes from fault keywords to fault causes and then to control decisions; path z is the reasoning route from question q to potential answer a; path z is essentially a series of intermediate steps or reasoning processes, and answer a is the final result, which is the specific answer obtained through the guidance of path z and combined with the learning and reasoning of the model.

[0118] In one embodiment, multi-task reasoning optimization training is performed on the initial fault control model according to each path in the initial path set and each path in the target path set to obtain a target fault control model, including: constructing a composite loss function for multi-task reasoning optimization training, the composite loss function being composed of a first probability of generating a reliable path in path extraction and a second probability of generating a correct path in retrieval reasoning, the first probability being determined by evaluating the reliability of each path in the initial path set, and the second probability being determined by evaluating the correctness of each path in the target path set; with the goal of maximizing the first probability and the second probability, iteratively updating the weight parameters in the initial fault control model to obtain the target fault control model.

[0119] In this embodiment, by maximizing the probability of generating a reliable path in path extraction and the probability of generating a correct path in retrieval reasoning, that is, by jointly optimizing path extraction and retrieval reasoning, the overall performance of the target fault management model can be significantly improved, making the model more efficient and reliable when performing power battery fault management.

[0120] In this embodiment, the variational lower bound in formula (17) is optimized to obtain:

[0121] logP(a|q,G)≥E z～Q(z) [logP θ (a|q,z,G)]-DKL (Q(z)||P θ (z|q)) Formula (18)

[0122] Where P(a|q,G) represents the probability of the cross-expert multi-task model generating the answer a given the question q and the knowledge graph G; z represents the path generated by the cross-expert multi-task model; Q(z) represents the posterior distribution of the path z generated based on the knowledge graph G, that is, the probability distribution of the path z generated after observing the question q and the knowledge graph G; P θ (a|q, z, G) represents the probability of the cross-expert multi-task large model generating answer a given question q, path z, and knowledge graph G; P θ (z|q) represents the prior probability of generating path z when given problem q; D KL (Q(z)||P θ (z|q) represents the prior probability P θ KL divergence (Kullback-Leibler Divergence, relative entropy) between (z|q) and the posterior distribution Q(z); E z～Q(z) [logP θ (a|q, z, G)] represents the expectation of generating reliable answers based on the path and knowledge graph.

[0123] In formula (17), it is necessary to minimize the KL divergence and maximize the expectation to promote the generation of reliable paths by the cross-expert multi-task large model. By maximizing this variational lower bound, the approximate optimal path distribution Q(z) is found, making the probability of the generated answer a as high as possible while maintaining the rationality of the path distribution.

[0124] In path extraction, knowledge is extracted from the knowledge graph to generate paths in the large model, and the reliability of the inferred fault path is improved by minimizing the KL divergence. For example, given a question q and an answer a, the connection e can be obtained through the knowledge graph. q and e a The fault path w z (e q , e a )=e q , the corresponding relationship set z = {r1, r2, ..., r l}, the posterior distribution Q(z) can be estimated as:

[0125]

[0126] Assume that all paths conform to uniform distribution, where Q(z) represents the probability distribution of path z after observing question q and knowledge graph G; Q(z|a, q, G) represents the probability of the cross-expert multi-task large model generating answer a given question q, path z and knowledge graph G; Z represents the number of paths; Indicates that in the knowledge graph G, there is a path between question q and answer a.

[0127] According to formula (19), the objective function of path extraction is:

[0128]

[0129] Among them, L plan represents the objective function of path extraction; D KL (Q(z)||P θ (z|q)) and the transformed D KL (Q(z|a,q,G)||P θ (z|q)) represents the prior probability P θ The KL divergence between (z|q) and the posterior distribution Q(z); z represents the path generated by the cross-expert multi-task large model; Z * represents the set of reliable paths between question q and answer a; Represents the reliable path set Z * The probability of each reliable path in .

[0130] By optimizing formula (20), the probability of the cross-expert multi-task large model extracting knowledge from the knowledge graph and obtaining reliable paths can be maximized, and the reliable paths whose probabilities exceed a certain threshold can be integrated into the initial path set.

[0131] In retrieval reasoning, the cross-expert multi-task large model is used to select appropriate paths from the initial path set obtained by path extraction and infer the answer to the question. The probability of performing retrieval reasoning on multiple reliable paths can be defined as:

[0132]

[0133] in, It means to perform a product operation on each path in the initial path set Z.

[0134] Using K path solutions to approximate the expected value, the objective function of retrieval reasoning can be expressed as:

[0135]

[0136] Among them, L reason represents the objective function of retrieval reasoning; represents the initial path set obtained by path extraction; Represents maximizing the expectation of generating the correct answer based on the path and knowledge graph.

[0137] Through formula (22), the probability of the cross-expert multi-task large model extracting knowledge from the knowledge graph and obtaining the correct path can be maximized, and the correct paths with probabilities higher than a certain threshold can be integrated into the target path set.

[0138] Therefore, in multi-task reasoning optimization training, the composite loss function of the cross-expert multi-task large model can be expressed as:

[0139]

[0140] Among them, L2 represents the composite loss function; represents the objective function of path extraction; Represents the objective function of retrieval reasoning.

[0141] In a possible embodiment, path extraction mainly uses knowledge graphs to infer and generate fault paths, and designs the following prompt words and path formats:

[0142] <prompt>Please generate a path based on the problem reasoning

[0143] <question>

[0144] <response> <path>r1<SEP>r2<SEP>...<SEP>r l < / PATH>

[0145] in, <path>, <SEP>, < / PATH> represent the start point, separator, and end point of the path respectively. For example, they are the fault keyword, fault cause, and control decision in sequence.

[0146] In a possible embodiment, the optimization objective in path extraction can be expressed as:

[0147]

[0148] where, is the decomposition form of logP θ (z|q); the path z is regarded as a series of relationships r1, r2,... r |z| , r i represents the i-th relationship in the path z, and r<i represents the sequence of the first i - 1 relationships.

[0149] As a possible embodiment, it means finding the weight parameter θ of the model that maximizes the objective function.

[0150] In a possible embodiment, the breadth-first algorithm is used in retrieval and reasoning to screen the fault path w z , and this screening process can be expressed as:

[0151] W z = {w z (e q , e * )|w z (e q , e * ) = e q →e1→...→e a* , w z (e q , e * ) ∈ G} Formula (25)

[0152] However, screening the fault path w according to the breadth-first algorithm z may contain some noise. Therefore, the set of screened fault paths W z and the question q are imported into the cross-expert multi-task large model together, so as to reduce the error caused by noise, improve the accuracy of question answering, and obtain the correct answer a. This process can be expressed as:

[0153]

[0154] where, is 's decomposition form; represents the product of the per-word probabilities of generating the answer a; t i represents the i-th token in generating the answer a; t<i represents the first i - 1 tokens; P θ (t i |t < i, q; w z represents the probability of generating t z when given t < i, question q, and path w i .

[0155] As a possible embodiment, it is indicated that the weight parameter θ of the model that needs to maximize the objective function is to be found.

[0156] In a possible embodiment, the iterative update of the weight parameters in the initial fault control model includes the number of shared experts in the cross-expert multi-task large model and the learning parameters of the path network, and the learning parameters include the expert assignment weight matrix and the noise adjustment weight matrix.

[0157] Step S240, using the target fault control model to perform fault control on the power battery.

[0158] Among them, performing fault control on the power battery includes obtaining the fault keywords, fault causes, and control decisions of the power battery.

[0159] In an embodiment, using the target fault control model to perform fault control on the power battery includes: obtaining the target signal time-frequency diagram and the corresponding target fault description of the power battery to be tested; inputting the target signal time-frequency diagram and the target fault description into the target fault control model to obtain the target fault keywords, target fault causes, and target control decisions corresponding to the power battery to be tested.

[0160] In this embodiment, the target fault control model is obtained by combining the cross-expert multi-task large model and the knowledge graph. Using the target fault control model to perform fault control on the power battery to be tested improves the efficiency and accuracy of the fault control of the power battery to be tested.

[0161] In a possible embodiment, using the target fault control model to perform fault control on the power battery can obtain the fault keywords, various fault causes corresponding to the fault keywords, and the control decisions corresponding to each fault cause.

[0162] The above-mentioned power battery fault control method first obtains a fault data set and a knowledge graph. The fault data set includes various signal time-frequency diagrams of the power battery and the fault descriptions corresponding to each signal time-frequency diagram. The fault description includes fault phenomena and fault keywords. The knowledge graph includes the relationship between fault phenomena, fault keywords, fault causes and control decisions. Then, a multi-task large model is constructed, and multi-task reasoning training is performed on the multi-task large model according to the fault data set to obtain an initial fault control model. The multi-task includes the recognition of fault keywords, the analysis of fault causes and the formulation of control decisions. Then, the relationships in the knowledge graph are used as reasoning constraints to perform multi-task reasoning optimization on the initial fault control model. Training is performed to obtain the target fault control model, and finally the target fault control model is used to control the power battery fault. The target fault control model is obtained by combining the multi-task large model and the knowledge graph. The target fault control model can accurately predict the fault keywords and fault causes of the power battery and obtain the corresponding control decisions. Moreover, through the automated process of fault identification, cause analysis and measure formulation, fault diagnosis can be completed quickly and corresponding measures can be obtained, which reduces manual intervention and decision-making time, improves fault decision-making efficiency and the intelligent level of fault diagnosis, ensures the fault identification and early warning capabilities of the power battery, and ensures the accuracy and efficiency of power battery fault diagnosis.

[0163] See also Figure 7 , Figure 7 is a block diagram of a power battery fault management system shown in an exemplary embodiment of the present application. The system can be applied to Figure 1 The implementation environment shown is a schematic diagram. It should be understood that the system may also be applicable to other exemplary implementation environments, and this embodiment does not limit the implementation environment to which the system is applicable.

[0164] like Figure 7 As shown, in an exemplary embodiment, the power battery fault management and control system 700 includes at least an acquisition module 710, a pre-training module 720, an optimization training module 730 and a management and control module 770, which are described in detail as follows:

[0165] An acquisition module 710 is used to acquire a fault data set and a knowledge graph, wherein the fault data set includes various signal time-frequency graphs of the power battery and the fault descriptions corresponding to the signal time-frequency graphs, wherein the fault descriptions include fault phenomena and fault keywords, and the knowledge graph includes the relationships among the fault phenomena, fault keywords, fault causes and management decisions;

[0166] The pre-training module 720 is used to build a multi-task large model and perform multi-task reasoning training on the multi-task large model according to the fault data set to obtain an initial fault control model. The multi-task includes the recognition of fault keywords, the analysis of fault causes, and the formulation of control decisions.

[0167] The optimization training module 730 is used to use the mutual relationships in the knowledge graph as reasoning constraints to perform multi-task reasoning optimization training on the initial fault control model to obtain the target fault control model;

[0168] The control module 740 is used to perform fault control on the power battery using the target fault control model.

[0169] It should be noted that the power battery fault management system provided in the above embodiment and the power battery fault management method provided in the above embodiment belong to the same concept, wherein the contents of the operations performed by each module have been described in detail in the method embodiment and will not be repeated here.

[0170] The flowchart and block diagram in the accompanying drawings illustrate the possible architecture, functions and operations of the system, method and computer program product according to various embodiments of the present application. Wherein, each box in the flowchart or block diagram can represent a module, a program segment, or a part of the code, and the above-mentioned module, program segment, or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0171] The units involved in the embodiments described in this application may be implemented by software or hardware, and the units described may also be set in a processor. The names of these units do not, in some cases, constitute limitations on the units themselves.

[0172] The above embodiments are merely illustrative of the principles and effects of the present application, and are not intended to limit the present application. Anyone familiar with the technology may modify or change the above embodiments without violating the spirit and scope of the present application. Therefore, all equivalent modifications or changes made by a person of ordinary skill in the art without departing from the spirit and technical ideas disclosed in the present application shall still be covered by the claims of the present application.< / path> < / path> < / response> < / question> < / prompt>

Claims

1. A power battery fault control method, characterized in that: The method comprises: Acquire a fault data set and a knowledge graph, wherein the fault data set includes various signal time-frequency graphs of the power battery and fault descriptions corresponding to the signal time-frequency graphs, wherein the fault descriptions include fault phenomena and fault keywords, and the knowledge graph includes the relationship between the fault phenomena, fault keywords, fault causes and management decisions; Constructing a multi-task large model, and performing multi-task reasoning training on the multi-task large model according to the fault data set to obtain an initial fault control model, wherein the multi-task includes identification of the fault keywords, analysis of the fault causes, and formulation of the control decision; Using the mutual relationships in the knowledge graph as reasoning constraints, performing multi-task reasoning optimization training on the initial fault control model to obtain a target fault control model; The target fault management model is used to perform fault management on the power battery.

2. The power battery fault control method according to claim 1, characterized in that: The method for constructing the multi-task large model includes: Constructing a basic large model, the basic large model includes at least one sub-model, the sub-model integrates a visual transformer structure and a bidirectional encoder representation transformation structure, the visual transformer structure is used to process the signal time-frequency diagram, and the bidirectional encoder representation transformation structure is used to process the fault description; Fine-graining the feedback-forward network in the visual transformer structure and the bidirectional encoder representation transformation structure to obtain a plurality of fine-grained experts, and setting at least one shared expert among the plurality of fine-grained experts, wherein the fine-grained expert is used to process unit features in the signal time-frequency diagram or the fault description, and the shared expert is used to process shared features of each of the unit features; Initialize the gating values ​​of each of the fine-grained experts and each of the shared experts, and construct a cross-expert multi-task loss function to obtain the multi-task large model, wherein the multi-task loss function is composed of a task loss function and an expert interaction loss function, and the multi-task large model is a cross-expert multi-task large model including at least one cross-expert multi-task sub-model.

3. The power battery fault control method according to claim 2, characterized in that: The performing multi-task reasoning training on the multi-task large model according to the fault data set to obtain an initial fault management and control model includes: Annotate each set of signal time-frequency diagrams and corresponding fault descriptions in the fault data set with actual labels, wherein the actual labels include actual fault keywords, actual fault causes, and actual control decisions; Performing multi-task reasoning training on the multi-task large model using the annotated fault data set, iteratively updating the number of shared experts in the visual transformer structure and the bidirectional encoder representation transformation structure in the cross-expert multi-task sub-model and the learning parameters of the path network in the cross-expert multi-task sub-model, wherein the learning parameters include an expert allocation weight matrix and a noise adjustment weight matrix; The initial fault management model is obtained under the condition that the loss value of the multi-task loss function converges to a preset loss value.

4. The power battery fault control method according to claim 2, characterized in that: The multi-task loss function is constructed in the following way: Constructing a subtask loss function according to the prediction error of each subtask in the multitask, and determining the task loss function according to the preset task weights of each subtask and each subtask loss function; For each of the cross-expert multi-task sub-models, construct the interaction information between the multi-task and the fine-grained expert and the shared expert, and determine the expert interaction loss function according to the interaction information corresponding to each of the cross-expert multi-task sub-models and the influence weight of the preset information sharing; The multi-task loss function is constructed according to the task loss function and the expert interaction loss function.

5. The power battery fault control method according to claim 4, characterized in that: The constructing of interactive information between the multi-tasks and the fine-grained experts and the shared experts includes: constructing a joint probability that any subtask is assigned to any expert, constructing a first marginal probability that any subtask is not assigned to any expert, and constructing a second marginal probability that any expert is not assigned any subtask; constructing a joint entropy between the multiple tasks and the fine-grained expert and the shared expert according to the joint probability, constructing a first marginal entropy of the multiple tasks according to the first marginal probability, and constructing a second marginal entropy of the fine-grained expert and the shared expert according to the second marginal probability; The interactive information is constructed according to the joint entropy, the first marginal entropy and the second marginal entropy.

6. The power battery fault control method according to claim 1, characterized in that: The method of using the mutual relationships in the knowledge graph as reasoning constraints, performing multi-task reasoning optimization training on the initial fault control model, and obtaining a target fault control model includes: The fault data set is input into the initial fault control model, and based on the fault phenomena and the fault keywords corresponding to each of the signal time-frequency graphs, a path is extracted from the knowledge graph to determine an initial path set, where the path is from the fault keyword to the fault cause and then to the control decision; Calculating the occurrence probability of each path in the initial path set, and performing search and reasoning on each path in the initial path set according to the occurrence probability to determine a target path set, wherein the occurrence probability of each path in the target path set is higher than a preset probability threshold; According to each path in the initial path set and each path in the target path set, multi-task reasoning optimization training is performed on the initial fault management and control model to obtain the target fault management and control model.

7. The power battery fault management method according to claim 6, characterized in that: The performing multi-task reasoning optimization training on the initial fault control model according to each path in the initial path set and each path in the target path set to obtain the target fault control model includes: Constructing a composite loss function for multi-task reasoning optimization training, the composite loss function is composed of a first probability of generating a reliable path in path extraction and a second probability of generating a correct path in retrieval reasoning, the first probability is determined by evaluating the reliability of each path in the initial path set, and the second probability is determined by evaluating the correctness of each path in the target path set; With the goal of maximizing the first probability and the second probability, weight parameters in the initial fault management and control model are iteratively updated to obtain the target fault management and control model.

8. The power battery fault management method according to any one of claims 2 to 5, characterized in that: There are multiple cross-expert multi-task sub-models; Before obtaining the multi-task large model, the method further includes: Setting a regularization layer between two of the cross-expert multi-task sub-models, and performing residual connections on a plurality of the cross-expert multi-task sub-models; A learnable embedding vector is set in the hidden layer of each of the cross-expert multi-task sub-models, and the embedding vector is used to realize the perception of the multi-task by each of the fine-grained experts and each of the shared experts.

9. The power battery fault management method according to any one of claims 1 to 7, characterized in that: The method of using the target fault control model to control the power battery fault includes: Obtain the target signal time-frequency diagram and the corresponding target fault description of the power battery to be tested; The target signal time-frequency diagram and the target fault description are input into the target fault management and control model to obtain target fault keywords, target fault causes and target management and control decisions corresponding to the power battery to be tested.

10. A power battery fault management and control system, characterized in that: The system comprises: An acquisition module, used to acquire a fault data set and a knowledge graph, wherein the fault data set includes various signal time-frequency graphs of the power battery and fault descriptions corresponding to the signal time-frequency graphs, wherein the fault descriptions include fault phenomena and fault keywords, and the knowledge graph includes the relationship between the fault phenomena, fault keywords, fault causes and management decisions; A pre-training module, which is used to build a multi-task large model and perform multi-task reasoning training on the multi-task large model according to the fault data set to obtain an initial fault management and control model, wherein the multi-task includes the recognition of the fault keywords, the analysis of the fault causes and the formulation of the management and control decisions; An optimization training module, used to use the mutual relationships in the knowledge graph as reasoning constraints, perform multi-task reasoning optimization training on the initial fault control model, and obtain a target fault control model; The control module is used to perform fault control on the power battery using the target fault control model.

Images