Suspension control device failure diagnosis method and system based on inspection log

Abstract

The application discloses a suspension control equipment fault diagnosis method and system based on inspection logs, and belongs to the technical field of automobile suspension control equipment diagnosis, and comprises the following steps: S1, inspection log preprocessing and vectorization; S2, density matrix representation of the inspection log; S3, LSTM time sequence coding; and S4, fault diagnosis. The application represents each inspection log as an enhanced density matrix, and the semantic correlation between words is explicitly coded through the non-diagonal elements of the density matrix. The representation method breaks through the limitations of the traditional bag-of-words model and the word vector average method, and for the first time, the density matrix is applied to the fault diagnosis field of industrial equipment text logs.

Description

Technical Field

[0001] This invention relates to the field of automotive suspension control equipment diagnostic technology, specifically to a method and system for diagnosing suspension control equipment faults based on inspection logs. Background Technology

[0002] In the intelligent operation and maintenance of suspension systems in new energy vehicles, fault diagnosis technology is crucial for ensuring vehicle ride comfort and handling. Currently, fault diagnosis of suspension control equipment (including controllers, sensors, actuators, and communication modules) mainly relies on the following technical approaches.

[0003] Threshold-based monitoring (physical signal level): This method uses acceleration sensors, displacement sensors, or current sensors placed on the suspension system to collect physical signals such as vibration, displacement, or current during vehicle operation in real time. The system compares the collected signal amplitude with a preset fixed threshold, and triggers an alarm if the threshold is exceeded.

[0004] Diagnostics based on traditional deep learning (text vector sequence level): To utilize the text logs output by on-board diagnostic (OBD) systems, existing technologies typically employ Natural Language Processing (NLP) techniques. Specific solutions often use Recurrent Neural Networks (RNNs) or Long Short-Term Memory Networks (LSTMs) to process the log text. These methods usually segment the text into word sequences, mapping each word to a low-dimensional vector (Word Embedding), and then inputting these vectors into LSTM units step-by-step, attempting to capture temporal dependencies through hidden states.

[0005] Attention-based models (feature weighting level): To address the problem of long sequence dependencies, existing improvements have introduced attention mechanisms. This method attempts to capture the relationship between key failure words (such as "Error" and "Fail") and the context by assigning different weights to different words in the log text, thereby generating failure-specific sentence representations.

[0006] Existing text log diagnostic methods based on natural language processing (NLP) have weak semantic representation capabilities. Traditional LSTM or RNN methods typically represent words as one-dimensional vectors, which struggle to fully capture the multidimensional semantic relationships and potential logical entanglements of words in complex contexts (such as contradictory semantics like "the system is normal but occasionally reports errors"). Furthermore, existing methods tend to ignore subspace relationships between words, and current technologies mostly operate in real space, failing to effectively model the complex dependencies between "fault phenomena" and "fault codes" in log text. They also exhibit poor noise resistance; inspection logs often contain a large amount of irrelevant characters or interfering information (noise), making attention-based methods susceptible to this noise and leading to misjudgments of fault polarity (e.g., misjudging "no serious fault found" as a serious fault). Therefore, this paper proposes a fault diagnosis method and system for suspension control equipment based on inspection logs. Summary of the Invention

[0007] The technical problem to be solved by this invention is: how to solve it, and a method for diagnosing suspension control equipment faults based on inspection logs is provided.

[0008] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps:

[0009] S1: Inspection Log Preprocessing and Vectorization

[0010] The inspection log text of the suspension control equipment is collected and preprocessed, word embedding vectorization and state enhancement processing are performed in sequence to obtain the word enhancement state corresponding to each word.

[0011] S2: Density matrix representation of the inspection log

[0012] Based on the vocabulary enhancement state of all words in a single inspection log text, a density matrix of the single inspection log text is constructed.

[0013] S3: LSTM timing coding

[0014] The density matrix is ​​temporally encoded using the LSTM-M mechanism to obtain the enhanced fault feature vector.

[0015] S4: Fault Diagnosis

[0016] Classification calculations are performed based on the enhanced fault feature vectors, and fault diagnosis results are output.

[0017] Furthermore, in step S1, the specific processes of preprocessing and word embedding vectorization are as follows:

[0018] S11: Perform Chinese word segmentation, stop word removal, and fault keyword extraction on the inspection log text in sequence. Among them, fault keywords are words in the inspection log text that can directly indicate or associate the abnormal state, fault mode, or performance degradation of the suspension control equipment.

[0019] S12: Match the word embedding vectors corresponding to each word from the word embedding lookup table, and perform L2 normalization on the word embedding vectors to obtain the word state vector, wherein the word is the fault keyword.

[0020] Furthermore, in step S2, the specific process of state enhancement processing is as follows:

[0021] S13: Regarding the vocabulary state vector Introducing complex-valued probability amplitude ,in, Indicates amplitude. Representing phase, and thus constructing lexical enhancement states. for:

[0022] ;

[0023] in, , It is a complex space.

[0024] Furthermore, in step S2, the formula for constructing the density matrix is:

[0025] ;

[0026] in, It is a state of vocabulary enhancement. The outer product of itself and its transpose yields a Matrix; It is the first The learnable weights of each word satisfy , ; It is the density matrix of a single inspection log text, and has the properties of being symmetric, positive semi-definite, and having a trace of 1.

[0027] Furthermore, in the density matrix In the middle, diagonal elements Represents the j-th feature, i.e., the abstract semantic attribute corresponding to the j-th dimension in the word embedding space; off-diagonal elements This represents the correlation between the j-th and k-th dimensions of the feature, i.e., the semantic association strength between words in the inspection log text.

[0028] Furthermore, in step S3, in the LSTM-M mechanism, let the current time be... The input density matrix is The hidden state at the previous moment was The cell state at the previous moment was ;

[0029] The input gate is used to process the input density matrix, and the processing formula is as follows:

[0030] ;

[0031] in, Let be the activation vector of the input gate. It is the Sigmoid activation function. It is a trainable weight matrix. This means flattening the density matrix into dimensional vector, Corresponding to the hidden state in the previous time step The trainable weight matrix, It is the bias vector;

[0032] The forget gate is used to access the cell state at the previous time step. The element-wise weighted summation is performed using the following formula:

[0033] ;

[0034] in, It is the activation vector of the forget gate. For a trainable weight matrix, It is the bias vector;

[0035] The above formula is used in conjunction with the forget gate output to calculate the candidate cell state. Update to get the current cell state:

[0036] ;

[0037] in, It is the cell state updated at the current moment. Represents element-wise product. It represents the cell state at the previous moment;

[0038] Output gates are used to output hidden states. As the enhanced fault feature vector, the processing formula is as follows:

[0039] ;

[0040] ;

[0041] in, It is the activation vector of the output gate. It is the hyperbolic tangent activation function. For a trainable weight matrix, It is the bias vector.

[0042] Furthermore, in step S4, the fault diagnosis result includes the fault type probability distribution and the fault probability corresponding to the current inspection log. The fault feature vector is input into the fully connected layer for classification calculation, and the fault type probability distribution of the suspension control device is output. The fault types include normal, valve body jamming, spring crack, sensor offset, and foreign object intrusion.

[0043] This invention also provides a suspension control equipment fault diagnosis system based on inspection logs, applied to the above-mentioned method, including:

[0044] The inspection log preprocessing and vectorization module is used to collect the inspection log text of the suspension control equipment, and perform preprocessing, word embedding vectorization and state enhancement processing in sequence to obtain the word enhancement state corresponding to each word.

[0045] The density matrix representation module of the inspection log is used to construct the density matrix of a single inspection log text based on the word enhancement state of all words in the single inspection log text.

[0046] The LSTM timing coding module is used to perform timing coding on the density matrix using the LSTM-M mechanism to obtain the enhanced fault feature vector.

[0047] The fault diagnosis module is used to perform classification calculations based on the enhanced fault feature vectors and output fault diagnosis results.

[0048] The present invention has the following advantages over the prior art:

[0049] 1. Each inspection log (or the daily inspection log set) is represented as an enhanced density matrix, and the semantic relevance between words is explicitly encoded through the off-diagonal elements of the density matrix. This representation method breaks through the limitations of the traditional bag-of-words model and word vector averaging method, and is the first to apply density matrix to the field of fault diagnosis of industrial equipment text logs.

[0050] 2. Traditional LSTMs can only process vector inputs and cannot directly process density matrices. The LSTM-M mechanism designed in this invention achieves efficient processing of matrix-form inputs by flattening the density matrix, performing a linear transformation, and then inputting it into the gating unit of the LSTM. This mechanism preserves the correlation information encoded in the density matrix and integrates it into the temporal modeling process. Attached Figure Description

[0051] Figure 1is a schematic structural diagram of a fault diagnosis and prediction system for new energy vehicle suspension intelligent control equipment based on inspection log text recognition processing in the embodiment of the present invention;

[0052] Figure 2 is a schematic flowchart of a fault diagnosis method for suspension control equipment based on inspection logs in the embodiment of the present invention. DETAILED DESCRIPTION

[0053] The embodiments of the present invention will be described in detail below. The present embodiment is implemented on the premise of the technical solution of the present invention, and provides detailed implementation modes and specific operation processes, but the protection scope of the present invention is not limited to the following embodiments.

[0054] Embodiment 1

[0055] The present embodiment provides a technical solution: a fault diagnosis and prediction system for new energy vehicle suspension intelligent control equipment based on inspection log text recognition processing. The core idea is to introduce a density matrix that can enhance the correlation features between texts in inspection logs into natural language processing tasks, perform enhanced state and enhanced matrix characterization on inspection log texts, and combine a long short-term memory network (LSTM) to realize fault type diagnosis and fault trend prediction of the new energy vehicle suspension intelligent control equipment. The present invention is mainly composed of four core modules, and its schematic structural diagram is shown in Figure 1 as shown in the figure.

[0056] The specific modules are introduced as follows:

[0057] 1. Inspection Log Preprocessing and Vectorization Module

[0058] Text data such as daily inspection logs of new energy vehicle suspension control equipment are collected. The text data is preprocessed as follows: a BERT-based tokenizer is used for Chinese word segmentation; stop words, such as meaningless words like "de", "le", "shi" are removed; fault keywords are extracted, which refer to core words that can directly indicate or strongly correlate with the abnormal state, fault mode or performance degradation of the suspension control equipment in texts such as inspection logs, such as "abnormal noise", "stuck", "offset", "no response", etc.

[0059] From the word embedding lookup table look up the set of word embedding vectors for each word in the sentence in the preprocessed text data such as inspection logs (fault keyword) , where represents the size of the vocabulary, represents the dimension of the word embedding vector, and each is the word embedding vector of the word in sentence 𝑠 , , .

[0060] Each word Word embedding vectors After normalization, the word state vector representation of the segmented words in the inspection log is obtained as follows:

[0061] ;

[0062] in, , Indicates L2 normal form;

[0063] Lexical state vector of inspection log word segmentation Introducing complex-valued probability amplitude ,in Indicates amplitude. Representing the phase, the lexical augmentation state of the inspection log is constructed as follows:

[0064] ;

[0065] in, , It is a complex space.

[0066] 2. Density matrix representation module of inspection log

[0067] The density matrix of this inspection log is constructed as follows:

[0068] ;

[0069] in: It is an enhanced state of the inspection log vocabulary. The outer product of itself and its transpose yields a Matrix; It is the first The learnable weights of each word satisfy , ; It is the density matrix of the inspection log, which has the properties of being symmetric, positive semi-definite, and having a trace of 1.

[0070] Physical meaning of the density matrix:

[0071] diagonal elements Let d represent the j-th feature, which is the abstract semantic attribute corresponding to the j-th dimension in the word embedding space. Since the word embedding vector (with dimension d) is a distributed representation learned in a high-dimensional semantic space through unsupervised pre-training, its single dimension does not directly correspond to a specific, nameable physical quantity or Chinese word. Instead, it represents a certain potential semantic component of the word in a specific context (e.g., semantic intensity related to "motion state", semantic field related to "mechanical structure"). Its magnitude represents the discriminative importance of the feature dimension in the current inspection log semantic expression. Semantic intensity, semantic field, etc., are automatically acquired through end-to-end learning during the network model training process.

[0072] off-diagonal elements This represents the correlation between the j-th and k-th dimensions of the features, i.e., the semantic association strength between words (fault keywords) in the inspection log, which is automatically acquired through end-to-end learning during the network model training process.

[0073] 3. LSTM timing coding module

[0074] Traditional LSTMs accept vector inputs, while the LSTM-M mechanism of this invention accepts matrix inputs (density matrices). Therefore, this invention designs an LSTM variant suitable for matrix inputs:

[0075] Let the current time be The input density matrix is The hidden state at the previous moment was Cell state is .

[0076] (1) Input gate (processing matrix input):

[0077]

[0078] in, is the activation vector of the input gate, used to control how much of the input information at the current time step is written into the cell state, and its value ranges from (0,1). It is the Sigmoid activation function. It is a trainable weight matrix. This means flattening the density matrix into dimensional vector, Corresponding to the hidden state in the previous time step The trainable weight matrix, This is the bias vector.

[0079] (2) Gate of Oblivion:

[0080]

[0081] in, It is the activation vector of the forget gate, used to control the cell state at the previous time step. The amount of information retained is within the range of (0,1); For a trainable weight matrix, It is the bias vector.

[0082] The candidate cell status is as follows:

[0083]

[0084] Cell status update:

[0085]

[0086] in, It is the candidate cell state, which is new information generated by the current input and the hidden state of the previous time step; It represents the cell state updated at the current moment, and is the element-wise product (Hadamard product). This indicates that the gating vector and the state vector are multiplied at corresponding positions. Let be the activation vector of the input gate. It represents the cell state at the previous moment.

[0087] (3) Output gate:

[0088]

[0089]

[0090] in, This is the activation vector of the output gate, used to control how many of the cell states at the current time step are output as hidden states, and its value ranges from (0,1). It is the hyperbolic tangent activation function. For a trainable weight matrix, Corresponding to the hidden state in the previous time step The trainable weight matrix, It is a bias vector. It is the hidden state at the current moment. It is the cell state updated at the current moment. It is the hidden state from the previous moment.

[0091] Enhanced fault characteristics: Hiding the output state of LSTM-M The enhanced fault feature vector is used for subsequent classification and prediction.

[0092] 4. Fault Diagnosis and Prediction Output Module

[0093] Fault classification: Input a fully connected layer and output a probability distribution of fault types (normal / valve body jamming / spring cracking / sensor misalignment / foreign object intrusion, etc.).

[0094] Fault probability: Outputs the probability of fault occurrence corresponding to the current inspection log. .

[0095] Trend prediction: For a series of inspection logs spanning multiple days, a daily density matrix is ​​used to predict the trend. As input, LSTM-M outputs a fault probability change curve to achieve fault trend prediction.

[0096] like Figure 2 As shown, this embodiment also provides a method for diagnosing suspension control equipment faults based on inspection logs. Based on the above system implementation, it specifically includes the following steps:

[0097] Step 1: Inspection Log Preprocessing and Vectorization

[0098] The inspection log text of the suspension control equipment is collected and preprocessed, word embedding vectorization and state enhancement processing are performed in sequence to obtain the word enhancement state corresponding to each word.

[0099] Step 2: Density Matrix Representation of Inspection Logs

[0100] Based on the vocabulary enhancement state of all words in a single inspection log text, a density matrix of the single inspection log text is constructed.

[0101] Step 3: LSTM timing coding

[0102] The density matrix is ​​temporally encoded using the LSTM-M mechanism to obtain the enhanced fault feature vector.

[0103] Step 4: Fault Diagnosis

[0104] Classification calculations are performed based on the enhanced fault feature vectors, outputting fault diagnosis results and supporting trend prediction.

[0105] Example 2

[0106] This invention effectively addresses semantic ambiguity. In inspection logs, the same word may correspond to different faults in different contexts. For example, "jerking" might refer to "response delay" when describing a solenoid valve, but "abnormal stiffness" when describing a spring. This invention enhances the lexical states and density matrix of the inspection logs, enabling the model to simultaneously represent multiple possible semantic interpretations, with subsequent LSTM selecting the optimal interpretation based on the context.

[0107] This invention introduces a density matrix into the fault detection modeling task of inspection logs. Experimental results on the TREC-QA and WIKI-QA datasets are shown in Table 1. The TREC-QA dataset is the standard benchmark dataset used in the question-answering track of the Text Retrieval Conference (TREC). The Wiki-QA dataset is an open-domain question-answering dataset released by Microsoft Research. We use it for fault detection, assuming that each question has at least one correct answer, corresponding to whether a fault exists.

[0108] Table 1 Experimental Results

[0109]

[0110] We use Mean Precision (MAP) and Mean Reciprocal Rank (MRR) as evaluation metrics, which are frequently used for the same tasks on the same dataset. The experiments above demonstrate that introducing a density matrix that enhances the inter-lexical association features in the inspection logs can significantly improve the performance of fault diagnosis modeling for intelligent control equipment of new energy vehicle suspension based on inspection logs.

[0111] In summary, the above-described method and system for fault diagnosis of suspension control equipment based on inspection logs proposes a fault diagnosis and prediction system for intelligent suspension control equipment of new energy vehicles based on inspection log text recognition processing, significantly enhancing the semantic relevance modeling capability in inspection logs. Traditional text processing methods treat words as independent features, failing to capture the semantic associations between "abnormal noise" and "valve body," or "jamming" and "action delay." This invention explicitly encodes the relevance between words through the off-diagonal elements of the density matrix, enabling the understanding of the deep semantic structure of fault descriptions.

[0112] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for fault diagnosis of suspension control equipment based on inspection logs, characterized in that, Includes the following steps: S1: Inspection Log Preprocessing and Vectorization The inspection log text of the suspension control equipment is collected and preprocessed, word embedding vectorization and state enhancement processing are performed in sequence to obtain the word enhancement state corresponding to each word. S2: Density matrix representation of the inspection log Based on the vocabulary enhancement state of all words in a single inspection log text, a density matrix of the single inspection log text is constructed. S3: LSTM timing coding The density matrix is ​​temporally encoded using the LSTM-M mechanism to obtain the enhanced fault feature vector. S4: Fault Diagnosis Classification calculations are performed based on the enhanced fault feature vectors, and fault diagnosis results are output.

2. The method for fault diagnosis of suspension control equipment based on inspection logs according to claim 1, characterized in that, In step S1, the specific processes of preprocessing and word embedding vectorization are as follows: S11: Perform Chinese word segmentation, stop word removal, and fault keyword extraction on the inspection log text in sequence. Among them, fault keywords are words in the inspection log text that can directly indicate or associate the abnormal state, fault mode, or performance degradation of the suspension control equipment. S12: Match the word embedding vectors corresponding to each word from the word embedding lookup table, and perform L2 normalization on the word embedding vectors to obtain the word state vector, wherein the word is the fault keyword.

3. The suspension control equipment fault diagnosis method based on inspection logs according to claim 2, characterized in that, In step S2, the specific process of state enhancement processing is as follows: S13: Regarding the vocabulary state vector Introducing complex-valued probability amplitude ,in, Indicates amplitude, Representing phase, and thus constructing lexical enhancement states. for: ； in, , It is a complex space.

4. The suspension control equipment fault diagnosis method based on inspection logs according to claim 3, characterized in that, In step S2, the formula for constructing the density matrix is: ； in, It is a state of vocabulary enhancement. The outer product of itself and its transpose yields a Matrix; It is the first The learnable weights of each word satisfy , ; It is the density matrix of a single inspection log text, and has the properties of being symmetric, positive semi-definite, and having a trace of 1.

5. The suspension control equipment fault diagnosis method based on inspection logs according to claim 4, characterized in that, In the density matrix In the middle, diagonal elements Represents the j-th feature, i.e., the abstract semantic attribute corresponding to the j-th dimension in the word embedding space; off-diagonal elements This represents the correlation between the j-th and k-th dimensions of the features, i.e., the semantic association strength between words in the inspection log text.

6. The suspension control equipment fault diagnosis method based on inspection logs according to claim 4, characterized in that, In step S3, in the LSTM-M mechanism, let the current time be... The input density matrix is The hidden state at the previous moment was The cell state at the previous moment was ; The input gate is used to process the input density matrix, and the processing formula is as follows: ； in, Let be the activation vector of the input gate. It is the Sigmoid activation function. It is a trainable weight matrix. This means flattening the density matrix into dimensional vector, Corresponding to the hidden state in the previous time step The trainable weight matrix, It is the bias vector; The forget gate is used to access the cell state at the previous time step. The element-wise weighted summation is performed using the following formula: ； in, It is the activation vector of the forget gate. For a trainable weight matrix, It is the bias vector; The above formula is used in conjunction with the forget gate output to calculate the candidate cell state. Update to get the current cell state: ； in, It is the cell state updated at the current moment. Represents element-wise product. It represents the cell state at the previous moment; Output gates are used to output hidden states. As the enhanced fault feature vector, the processing formula is as follows: ； ； in, It is the activation vector of the output gate. It is the hyperbolic tangent activation function. For a trainable weight matrix, It is the bias vector.

7. The method for fault diagnosis of suspension control equipment based on inspection logs according to claim 6, characterized in that, In step S4, the fault diagnosis results include the fault type probability distribution and the fault probability corresponding to the current inspection log. The fault feature vector is input into the fully connected layer for classification calculation, and the fault type probability distribution of the suspension control device is output. The fault types include normal, valve body jamming, spring crack, sensor offset, and foreign object intrusion.

8. A suspension control equipment fault diagnosis system based on inspection logs, characterized in that, The method applied to any one of claims 1 to 7 includes: The inspection log preprocessing and vectorization module is used to collect the inspection log text of the suspension control equipment, and perform preprocessing, word embedding vectorization and state enhancement processing in sequence to obtain the word enhancement state corresponding to each word. The density matrix representation module of the inspection log is used to construct the density matrix of a single inspection log text based on the word enhancement state of all words in the single inspection log text. The LSTM timing coding module is used to perform timing coding on the density matrix using the LSTM-M mechanism to obtain the enhanced fault feature vector. The fault diagnosis module is used to perform classification calculations based on the enhanced fault feature vectors and output fault diagnosis results.

Images