A method for generating fault samples of engineering machinery based on diffusion probability model

Abstract

This invention relates to the field of sample construction technology, and more particularly to a method for generating engineering machinery fault samples based on a diffusion probability model. The method includes: extracting multiple feature vectors and constructing equipment operation feature vectors; extracting operating condition encoding vectors; inputting the operating condition encoding vectors into a diffusion probability model and a condition adjustment model respectively to obtain a current-step noise estimation vector and an operating condition adjustment coefficient; calculating a current intermediate feature state; re-inputting the current intermediate feature state, the operating condition encoding vector, and the current-step noise estimation vector into the diffusion probability model to update the current intermediate feature state; iteratively updating the current intermediate feature state along a preset diffusion step sequence until candidate fault samples are generated; the generator outputs a correction vector consistent with the candidate fault samples; optimizing the candidate fault samples based on the correction vector; establishing constraints to perform secondary optimization on the optimized candidate fault samples; and generating a comprehensively optimized candidate fault sample.

Description

Technical Field

[0001] This invention relates to the field of sample construction technology, and more particularly to a method for generating engineering machinery fault samples based on a diffusion probability model. Background Technology

[0002] In the field of modern construction machinery, fault diagnosis technology has become an important technical means to ensure reliable equipment operation, reduce maintenance costs, and minimize downtime losses. With the increasing workload of excavators, loaders, cranes, and large transmission and hydraulic equipment, key components are subjected to long-term high loads, strong impacts, frequent start-stop cycles, and alternating complex environmental conditions. Equipment faults exhibit significant operational condition correlations and state coupling. Existing fault diagnosis methods for construction machinery typically rely on monitoring signals such as vibration, temperature, and current, selecting features through manual experience, or directly training recognition models using existing samples.

[0003] These methods can achieve certain results when the number of samples is sufficient and the operating conditions are relatively simple. However, in real engineering scenarios, they are often constrained by two aspects: First, the samples of failures of key components are scarce, especially the samples of serious failures, early failures, and failures under specific load conditions are difficult to obtain, making it difficult for the training data to cover the real failure state space. Second, the manifestation of the same failure under different operating conditions is significantly different. Vibration, temperature rise, and current load are not isolated changes, but rather form a coupled shift with the operating conditions, which makes the existing samples, even if limited in number, often show uneven distribution and insufficient coverage of operating conditions.

[0004] Most existing data augmentation methods expand the sample size by simple perturbation, slicing and recombination, or general generation methods. While these methods can increase the number of samples, they usually lack explicit expression of the operating conditions and multi-source state coupling relationships of engineering machinery. The generated results are prone to deviating from the actual fault distribution, and it is particularly difficult to balance fault structure preservation, operating condition consistency, and usability in subsequent diagnostic training. Therefore, how to construct a unified fault state expression and operating condition expression for multi-source monitoring signals based on the actual operating scenarios of engineering machinery, and then generate fault samples within this expression space that are consistent with the target operating conditions and match the actual fault distribution, and further effectively use them for diagnostic model training, has become a key problem that urgently needs to be solved in current fault sample generation and intelligent diagnostic technologies. Summary of the Invention

[0005] To address the above problems, this invention provides a method for generating engineering machinery fault samples based on a diffusion probability model.

[0006] To achieve the above objectives, this invention proposes a method for generating engineering machinery fault samples based on a diffusion probability model, comprising:

[0007] Multiple sets of transportation data sequences are divided into a unified time window. Multiple feature vectors are extracted from the time window and linearly weighted and fused to construct equipment operation feature vectors. Operating condition coding vectors are extracted from the equipment operation feature vectors.

[0008] The operating condition encoding vector is input into the diffusion probability model and the condition adjustment model respectively to obtain the current step noise estimation vector and the operating condition adjustment coefficient. The current intermediate feature state is calculated based on the operating condition adjustment coefficient and the equipment operation feature vector. The current intermediate feature state, the operating condition encoding vector and the current step noise estimation vector are re-input into the diffusion probability model to update the current intermediate feature state. The current intermediate feature state is iteratively updated along the preset diffusion step sequence until candidate fault samples are generated.

[0009] The generator receives the candidate fault samples and the operating condition coding vector, and outputs a correction vector consistent with the candidate fault samples. Based on the correction vector, the generator optimizes the candidate fault samples, establishes structural proximity constraints and operating condition readback constraints to perform secondary optimization on the optimized candidate fault samples, and generates a comprehensively optimized candidate fault sample.

[0010] The optimized candidate fault samples are combined with the real equipment operation feature vectors to form a joint training sample set. The fault diagnosis model is trained using the joint training sample set. The fault diagnosis model receives the equipment operation feature vectors and outputs the predicted probability of each fault category.

[0011] In some embodiments, the fault diagnosis model incorporates a loss function during training. The parameters of the loss function include sample weights, cross-entropy terms, true fault category labels, fault category prediction probabilities, adjustment coefficients, fault category prediction probabilities for candidate fault samples, and fault category prediction probabilities for the actual equipment operating feature vectors in the same operating condition pair.

[0012] In some embodiments, after optimizing the candidate fault samples based on the correction vector, the method further includes:

[0013] The discriminator is used to judge the optimized candidate fault samples and output the probability that the candidate fault samples belong to the true distribution.

[0014] In some embodiments, the diffusion probability model is a denoising network structure, which includes an input mapping layer, a fully connected residual block, and an output mapping layer.

[0015] In some embodiments, the condition adjustment model includes a fully connected layer, which is used to map the operating condition encoding vector to an intermediate dimension and output the operating condition adjustment coefficient.

[0016] In some embodiments, the multiple sets of transport data sequences include the original vibration sequence, the original temperature sequence, and the original current sequence.

[0017] In some embodiments, extracting the operating condition coding vector from the device operating feature vector specifically includes:

[0018] Extract the relevant components of the operating conditions from the temperature feature vector and current feature vector within the same time window and combine them to form the operating condition coding vector.

[0019] In some embodiments, the diffusion probability model includes a forward diffusion function, which is used to calculate the current intermediate feature state. The parameters of the forward diffusion function include the fidelity coefficient of the diffusion step, the operating condition adjustment coefficient, the equipment operating feature vector, and the disturbance vector.

[0020] In some embodiments, the diffusion probability model further includes a reverse update function, which is used to update the current intermediate feature state. The parameters of the reverse update function include a noise estimation vector, a fidelity coefficient, and a working condition adjustment coefficient.

[0021] In some embodiments, the noise estimation vector includes the current intermediate feature state, the current diffusion step, and the operating condition coding vector.

[0022] The beneficial effects of this invention are as follows:

[0023] This application first addresses the differences in sampling methods, change rates, and state sensitivity of multi-source monitoring signals such as vibration, temperature, and current, and constructs a unified representation of equipment operation characteristics. It then extracts operating condition codes related to load and thermal state from these codes, enabling fault states and operating conditions to be expressed correspondingly within the same time window. Subsequently, a diffusion generation process constrained by operating conditions is introduced into this unified feature space, allowing the expansion direction and magnitude of fault samples to be adaptively adjusted according to changes in the target operating conditions, thereby obtaining candidate fault samples that match the actual operating conditions.

[0024] Based on this, the adversarial optimization mechanism under conditional constraints is further utilized to adjust the local structural differences between the preliminary samples and the real samples, and the matching degree of the generated samples in the real fault distribution is improved by the working condition consistency constraint. Finally, the optimized fault samples and the real fault samples are used together for fault diagnosis model training, so that the model can not only learn the fault center state corresponding to the real samples, but also cover the boundary state and transition state caused by the scarcity of samples under complex working conditions.

[0025] Through the above technical approach, this invention no longer focuses on simply increasing the number of samples, but organizes fault state expression, working condition constraint generation, sample quality optimization, and diagnostic model training into a continuous technical whole, thereby improving the degree of fit of the generated samples to real engineering machinery scenarios and enhancing the stability and recognition ability of the fault diagnosis model under complex working conditions and small sample conditions. Attached Figure Description

[0026] Figure 1 This is a flowchart of a method for generating engineering machinery fault samples based on a diffusion probability model in a specific embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] refer to Figure 1 As shown, this application proposes a method for generating engineering machinery fault samples based on a diffusion probability model, including:

[0029] S1: Divide multiple sets of transportation data sequences into a unified time window, extract multiple feature vectors from the time window, and perform linear weighted fusion to construct an equipment operation feature vector. Extract the operating condition coding vector from the equipment operation feature vector, specifically including:

[0030] Key components of construction machinery are typically equipped with condition monitoring systems during long-term operation. For example, vibration sensors are placed in locations such as slewing bearings, main pumps, motor drive units, or gearboxes; temperature sensors are placed on the outer ring of bearings, on the housing surface, or near oil passages; and current acquisition modules are configured in the drive circuit. These three types of sensors continuously output time-series data, including the original vibration sequence. The original temperature sequence reflects mechanical shock, wear, and meshing abnormalities. The original current sequence reflects the state of frictional heating and heat dissipation. This reflects changes in drive load and operating conditions. To ensure these three types of signals can be used uniformly in subsequent sample generation, the three sequences are first segmented according to a unified time window. In practice, continuously acquired vibration, current, and temperature data can be divided into multiple window segments with the same time boundary, each window corresponding to a short-term operating state of the equipment. For example, in a certain loading cycle, a fixed window covers the complete segment of "start-load-unload," and then synchronously extracts data from within that window. , , Vibration signals, due to their more dense sampling, directly retain the complete sequence within the window and extract amplitude distribution and frequency band energy characteristics; temperature signals are extracted for variation trends and fluctuation characteristics within the window; and current signals are extracted for load variation amplitude and trends within the window. The vibration characteristics mentioned here... Temperature characteristics Current characteristics These are not original values, but rather feature vectors formed after statistical analysis and spectral analysis within their respective windows. The extraction of vibration frequency band energy comes from classical spectral analysis methods, which first decompose the vibration sequence within the window and then statistically analyze the energy proportion within a preset frequency range. Temperature change trends are derived from discrete-time series difference analysis, which uses the amplitude of changes in adjacent sampled values ​​to describe the rate of change in thermal state. Current load characteristics are derived from current fluctuation statistics commonly used in motor operation monitoring, which uses the amplitude and rate of change of the current sequence within the window to characterize the current load state. Since the three types of features originally come from different physical signals and their numerical ranges are inconsistent, they are first dimensionlessly processed based on the equipment's historical operating data before fusion. , , The relative degree of change is expressed on the same scale. After this processing, all three types of features can be regarded as dimensionless vectors. When performing linear combination, the dimensions on both sides remain consistent, and the calculation results also conform to the common sense in engineering that "the same state quantity is jointly represented by multiple source features".

[0031] After obtaining the dimensionless eigenvectors, linear weighted fusion is used to construct the device operation feature vectors. This approach originates from the classic weighted average concept in multi-sensor information fusion. Its basic principle is that when multiple observations jointly describe the state of the same object, different contribution ratios can be assigned according to the sensitivity of each observation to the target state, thus forming a unified state representation. In the scenario described in this application, directly using the classic weighted average is insufficient because the ability of various signals to characterize faults differs under different operating conditions of engineering machinery. Therefore, a contribution coefficient obtained from historical operating samples is introduced on top of the classic weighted fusion approach. , , The following fusion expression is obtained, corresponding to the degree of influence of vibration, temperature, and current on the expression of fault state:

[0032] ;

[0033] in, This represents the device's operational feature vector within the current time window. Indicates the original vibration sequence Extract and complete the dimensionless vibration feature vector within this time window; Represents the original temperature sequence Extract and complete the dimensionless temperature feature vector within this time window; Represents the original current sequence Extract and complete the dimensionless current feature vector within this time window; , , These represent the contribution coefficients of three types of characteristics: vibration, temperature, and current. The values ​​of these three coefficients are non-negative and their sum is 1. Thus Maintaining a uniform relative scale. The derivation logic of this formula is to first consider the equipment state as being jointly determined by "mechanical anomaly characteristics," "thermal state characteristics," and "load condition characteristics," and then combine the three types of characteristics using contribution coefficients. Because , , It has been transformed into a dimensionless vector of the same dimension and scale, so the terms on the right can be directly added together, and the terms on the left are obtained as follows: It remains a dimensionless state vector. In practical implementation, coefficients can be set based on the statistical results of historical operating samples. For example, for gearbox components, vibration characteristics are more sensitive to faults, so a higher coefficient can be chosen. For equipment operating under continuous heavy loads, current and temperature are more sensitive to changes in operating conditions, so the current and temperature should be appropriately increased. and The fusion process is illustrated using a three-dimensional feature example. Assume that after feature extraction and dimensionless transformation, a certain time window yields... , , Based on the statistical results of historical samples of this component, , , After substituting, we get This result indicates that the device status within the current window simultaneously contains significant vibration anomaly information and moderate load change information, which can serve as direct input for subsequent diffusion generation steps.

[0034] In formation Subsequently, temperature characteristics within the same time window and current characteristics Extract the relevant operating condition components and combine them to form the operating condition coding vector. The specific method is to... The component reflecting heat accumulation and the trend of heat change, and The components reflecting load level and load fluctuation are spliced ​​together in a fixed order to obtain... For example, if the thermal change component of the temperature characteristic in a certain window is... The load component of the current characteristic is Then it can form This construction method ensures and Originating from the same time window, both describe the same runtime segment. Ultimately, this step outputs the device runtime feature vector. and working condition encoding vector .

[0035] S2: Input the operating condition encoding vector into the diffusion probability model and the condition adjustment model respectively to obtain the current step noise estimation vector and the operating condition adjustment coefficient. Calculate the current intermediate feature state based on the operating condition adjustment coefficient and the equipment operating feature vector. Re-input the current intermediate feature state, the operating condition encoding vector, and the current step noise estimation vector into the diffusion probability model to update the current intermediate feature state. Iterate and update the current intermediate feature state along a preset diffusion step sequence until candidate fault samples are generated. Specifically, this includes:

[0036] S2 takes the output of S1 and converts the device operation feature vector. and working condition encoding vector As input, candidate fault samples are generated within the unified fault feature space already established in S1. The distribution of engineering machinery fault samples depends not only on the fault itself, but also on the load level and thermal state at the time of the fault. For example, slewing bearing wear mainly manifests as abnormal vibration under light load conditions, while under continuous heavy load conditions, it will simultaneously lead to temperature rise and drive current increase. Therefore, the generation process needs to be synchronized between the fault structure and the operating condition structure. Based on this scenario characteristic, this step adopts the forward denoising-backward denoising idea of ​​the diffusion probability model as the starting point of the algorithm. Its initial source belongs to the classic Gaussian diffusion process in the probability generation model. The basic principle is to gradually add random perturbations to the samples in the feature space, so that the original sample distribution gradually transitions to a distribution close to Gaussian noise, and then trains the denoising network to learn to gradually recover the original samples from the noise state. Combining the characteristic of state drift of engineering machinery fault samples under different operating conditions, this step introduces an operating condition adjustment term on the basis of classic Gaussian diffusion, so that the diffusion intensity can be adaptively adjusted with changes in operating conditions. In specific implementation, a diffusion model is first established. , A one-dimensional vector denoising network structure is adopted, and the inputs are the current diffusion state, the sequence code of the current diffusion step, and the working condition code vector. The network consists of an input mapping layer, two fully connected residual blocks, and an output mapping layer. The input mapping layer concatenates the current diffusion state with the working condition code and projects it into the hidden space. Each residual block consists of a linear layer, a nonlinear activation layer, and a linear layer, and retains the state of the previous layer through a shorting structure. The output mapping layer provides the noise estimation result for the current diffusion step. Simultaneously, the working condition code vector... Entering the conditional adjustment model, this module consists of two fully connected layers. The first layer will... Mapped to the intermediate dimension, the second layer outputs the operating condition adjustment coefficient. And activated by Sigmoid The value falls between zero and one. This is how it is obtained. This can reflect the degree of influence of the current operating conditions on the diffusion amplitude; for example, heavy load and continuous temperature rise conditions will result in higher values. Light load stable conditions will result in lower .

[0037] The forward diffusion function is based on the classical Gaussian diffusion formula. In the classical form, the intermediate state is obtained by weighting the original sample and random noise with a fidelity coefficient. This step adds the operating condition adjustment coefficient to this formula. Multiplying by a noise term allows changes in operating conditions to directly alter the search range of the diffusion process, thereby covering the actual fault state offset region of the construction machinery under different operating conditions. The expression for the forward diffusion function is:

[0038] ;

[0039] in, Indicates the first The current intermediate feature state under each diffusion step; For the first The fidelity coefficient for each diffusion step is given by a preset diffusion schedule and gradually decreases as the number of diffusion steps increases; For the working condition encoded vector The operating condition adjustment coefficient is calculated by the condition adjustment module; For the first The random perturbation vector obtained by sampling in each diffusion step has a dimension of ... Same, generated by a Gaussian sampler. Because... It is already in the dimensionless feature space. It is also a random vector of the same dimension and scale. and Since they are all pure numbers, both sides of the formula maintain the same scale. The logical relationship of the formula is: first, from S1, we obtain... and Then by Calculated by the condition adjustment module Finally , and the Step random sampling obtained Substituting into the above equation forms a diffusion intermediate state. .

[0040] After forward diffusion is complete, the reverse recovery process follows the stepwise denoising principle of the diffusion model. Specifically, starting from the maximum diffusion step, the current intermediate state is... Diffusion step sequence number encoding and the same working condition encoding vector Reinput Diffusion Model ,Depend on Output the noise estimation result of the current step, and then use it to reconstruct the vector that is closer to the true fault state in the previous step. The expression of the reverse update function is written as:

[0041] ;

[0042] in, Indicates the first The recovered state vector of each diffusion step, i.e. the current intermediate feature state after reverse update; This indicates that the diffusion model takes the current state as input. Current diffusion step and working condition encoding vector The output noise estimation vector. This formula corresponds one-to-one with the forward diffusion formula: the random perturbation term added in forward diffusion is... In reverse recovery, the noise estimation result output by the diffusion model is used instead. And deduct from the current state, then divide by Complete the state backtracking. Because... and They all reside in the same dimensionless feature space. and Since it is a pure number, all terms in the formula maintain a consistent scale.

[0043] Based on the example in S1, a complete calculation implementation can be given. Let the device operation feature vector output by S1 be... The working condition coding vector is .Will After inputting the condition adjustment module, the operating condition adjustment coefficient is obtained. Let the current diffusion step be... The fidelity coefficient given in this step for the preset diffusion schedule is... Then there is , Let the random perturbation vector generated by the Gaussian sampler in this step be... ,but The noise term is calculated first as Multiply by get ; Sample retention items are from Calculated as Adding the two parts together, we get the first part. intermediate state vector under each diffusion step Then Diffusion step sequence number encoding and working condition encoding vector Input into the diffusion model together For example, if the current step The estimated noise vector is Substituting this into the reverse recovery formula, we have , , First calculate the denoising term. Then by After subtracting the denoising term, we get Finally, divide by ,get .

[0044] By continuing the iterative recovery from large to small along the preset diffusion step sequence, candidate fault samples can eventually be obtained. Through this generation process, this step expands the single real state representation given in S1 into candidate fault samples consistent with the target operating condition.

[0045] S3: The generator receives the candidate fault samples and the operating condition encoding vector, and outputs a correction vector consistent with the candidate fault samples. Based on the correction vector, it optimizes the candidate fault samples, establishes structural proximity constraints and operating condition readback constraints to perform secondary optimization on the optimized candidate fault samples, and generates a comprehensively optimized candidate fault sample, specifically including:

[0046] S3 continues along the generation chain already established by S2, processing the candidate fault sample pairs output by S2. .in, This represents candidate fault samples obtained from the diffusion process in a unified fault feature space. This represents the operating condition code corresponding one-to-one with the sample. S2 has already enabled the samples to have consistent operating conditions; this step further addresses the question of whether the samples sufficiently approximate the fault distribution of real equipment. In engineering machinery scenarios, the same type of fault often exhibits a clear coupling relationship in real equipment. For example, slewing bearing wear samples under heavy load conditions often simultaneously show increased vibration components, shifted current components, and increased thermal components; under intermittent light load conditions, the vibration components often change first, followed by the thermal components. S2 generates... The system has already fallen near the target operating condition, but the local coupling relationship still needs further correction. Therefore, this step uses a conditional generative adversarial network. Perform structural refinement. In specific implementation, the generator... Received samples Operating condition coding Output a value that is the same as... Dimensionally consistent correction vector; discriminator The generator receives the result of concatenating the sample vector with the working condition code and outputs the probability that the sample belongs to the true distribution. The generator uses a residual correction structure, first... and After concatenation, the input is a linear mapping layer, then passes through two fully connected residual blocks and an output mapping layer, and outputs a correction value. The discriminator uses two fully connected layers and one sigmoid output layer. The input is the result of concatenating the sample vector and the working condition code, and the output is a probability between zero and one.

[0047] Furthermore, in order to explicitly incorporate the constraint that "fault states should correspond to correct operating conditions" from the engineering machinery samples into the training process, a working condition readback matrix is ​​also constructed. .matrix The set of historical real samples formed from S1 specifically consists of a set of sample pairs composed of multiple sets of real equipment operation feature vectors and their corresponding operating condition encoding vectors. The matrix is ​​obtained through least squares fitting. It can map the state vector in the unified fault feature space back to the operating condition coding space, thereby checking whether any sample is consistent with the target operating condition in terms of operating condition representation.

[0048] The generator's output does not directly replace the result obtained from S2. Instead, the optimized sample is obtained by using residual correction. This expression originates from the residual learning concept in deep learning, its original form being the decomposition of the learning target into "original input + correction value". In the scenario of this invention, S2 has already provided samples with the correct working condition orientation. Therefore, the generator only needs to learn "how to adjust the local structure to resemble the real fault state while maintaining consistent operating conditions," and the corresponding optimization expression is:

[0049] ;

[0050] in, This represents the fault samples optimized using a conditional generative adversarial network. This indicates that the generator is based on the current sample. and corresponding working condition codes The calculated residual correction amount. Because... and Both are located in the unified fault feature space established by S1, therefore they maintain the same scale and can be obtained by direct addition. and The physical meaning is consistent. After the above formula gives the sample generation method, it is also necessary to define the training objective. The training objective is based on the minimax game form of classical conditional generative adversarial networks. Its basic idea is to make the discriminator distinguish between real samples and generated samples as much as possible, while making the generator deceive the discriminator as much as possible. On this classical form, this step superimposes two types of constraints specifically for engineering machinery fault samples: the first type is structural proximity constraints, which make the optimization result revolve around the value given by S2. The first type involves local adjustments; the second type is operational condition readback constraints, which cause the optimized candidate fault samples to be processed by a matrix. The code is still coded with the target operating condition after readback. To maintain consistency, including the comprehensively optimized candidate fault samples The expression is:

[0051] ;

[0052] in, This indicates that the discriminator correctly identifies a set of real sample pairs. The judgment result, S1 represents a real device operation feature vector from the historical real sample set. This is the corresponding operating condition encoding vector; This indicates that the discriminator optimizes sample pairs. The judgment result; This represents the balance coefficient. The logical relationship of this formula is: first, the optimized sample is obtained from the previous formula. , and then and The data is fed into a discriminator to determine the true / false distribution, and simultaneously... calculate Check whether the working condition expression corresponding to this sample is still close to the target working condition. The two squared distances in the formula occur in the same feature space or the same working condition coding space, thus maintaining a consistent scale.

[0053] A complete set of calculation examples can be given based on the numerical results of S2. Let the candidate fault samples output by S2 be... The corresponding working condition code is .Will and After concatenation, the input is given to the generator. If the residual correction value output by the generator is... Then the optimized sample is obtained from the previous equation. Then, from the real sample library formed by S1, take a real equipment operation feature vector corresponding to this working condition, denoted as... Its corresponding working condition code is denoted as After inputting into the discriminator, the result is set to... At the same time, optimize the sample pairs Input discriminator obtains Let's assume the working condition readback matrix is ​​obtained by fitting the S1 sample pair. Acting on Later obtained Take the balance coefficient At this point, the constraint terms in the objective function can be calculated one by one: first calculate the structural proximity terms, The sum of squares is Next, calculate the operating condition readback item. The sum of squares is The sum of the two items is: multiplied by Later obtained The opposing terms can be calculated using the natural logarithm. , The total of the opposing parts is approximately Therefore, the total target value corresponding to this batch of samples is approximately After multiple iterations, a final set of optimized samples was obtained. The sample distribution will be centered around S2, and will be closer to the real sample in the sense of the discriminator and closer to the target working condition code in the sense of working condition readback.

[0054] S4: Combine the optimized candidate fault samples with the real equipment operation feature vectors to form a joint training sample set. Train the fault diagnosis model using the joint training sample set. The fault diagnosis model receives the equipment operation feature vectors and outputs the predicted probability of each fault category, specifically including:

[0055] S4 starts with the optimized fault samples output from S3, combining them with the real equipment operating feature vectors that have already been aligned with the operating conditions during the generation process to form a joint training sample set, and uses this set to train the fault diagnosis model. The processing object here is still the state vector in the unified fault feature space established by S1. Therefore, the model input is no longer the original vibration sequence, temperature sequence, and current sequence, but rather the feature vector after the fault state has been explicitly expressed. This design forms a continuous relationship with the previous steps: S1 compresses the multi-source monitoring signals into state expressions with fault and operating condition significance; S2 extends the state expressions along the given operating conditions to the generative region; S3 then refines the generated samples to a position closer to the real fault distribution; and S4 truly transforms these high-quality samples into fault diagnosis capabilities.

[0056] In practical implementation, the fault diagnosis model A one-dimensional vector classification network structure is adopted. The input layer receives the fault state vector. The first hidden layer is a fully connected mapping layer, which is used to project the input state onto the diagnostic feature space. The second hidden layer is a residual layer with a shorting structure, which is used to maintain the relative relationship between fault-sensitive components. The output layer is a linear mapping layer, which outputs the predicted probability of each fault category.

[0057] The fault diagnosis model parameters are updated through batch training, with each batch containing feature vectors from real equipment operation. And optimized fault samples aligned with its operating conditions The labels for the real samples come from equipment maintenance records, disassembly and inspection records, or historical diagnostic conclusions, thus optimizing the fault samples. The label is inherited from the real sample corresponding to the starting point of its generation chain. The labels, namely S2 and S3, are expanded and adjusted around the same real fault state under the same time window and the same working condition coding constraints, so the labels remain consistent.

[0058] The training objective is based on cross-entropy loss in classic classification learning, the original idea of ​​which is to measure the difference between the model's output probability and the true class. This step, combined with the scenario of engineering machinery fault diagnosis, makes two targeted modifications to this classic objective. The first modification is to introduce sample source weights, so that the contributions of real samples and generated samples in parameter updates are different; the second modification is to introduce a working condition pairing consistency term, so that the predicted distributions of real samples and generated samples in the model output layer remain similar under the same working condition, thus continuing to pass the working condition consistency established in S3 into the diagnostic model. The expression for the loss function is:

[0059] ;

[0060] in, This represents the total loss during the training phase; Indicates the first The source weight of each sample is set, with higher values ​​for real samples and slightly lower values ​​for optimized fault samples. Indicates the first The cross-entropy term of each sample, This is the label for the actual fault category corresponding to the sample. For fault diagnosis model The output is a vector of predicted class probabilities. The adjustment coefficient represents the condition consistency term; Indicates the first Optimize fault samples in operating condition pairing The predicted probability vector; Represents the actual equipment operating feature vector in the same working condition pair. The predicted probability vector. The first part of the formula comes from the classic cross-entropy classification objective, used to ensure that the model converges to the correct class for each sample; the second part comes from the Euclidean distance constraint idea, which is used in this step to constrain the output consistency of samples under the same working conditions, so that while learning to correctly classify real samples, the model also learns that generated samples and real samples under the same working conditions should give similar judgments. The logical relationship inside the formula is: first, sort all samples in the batch by index Calculate the weighted classification items, and then pair the real samples and optimized samples that have already established a correspondence with the same working conditions according to the index. Calculate and output consistent terms.

[0061] A set of numerical examples that are sequentially connected with the aforementioned steps can be provided. Let the first... The actual equipment operating feature vector in the working condition pairing is The corresponding optimized fault sample is Both originate from the same time window and the same operating condition code, and the actual fault category is Class I fault in both cases. Input fault diagnosis model Then, output the predicted probability vector. ;Will After inputting the same model, the output is a predicted probability vector. The corresponding real label vector is Let the weights of the real samples be... Optimize sample weights The adjustment coefficient for the consistent operating conditions term is taken as follows: For real samples, the cross-entropy term is: Even after weighting, it is still the same. For the optimized sample, its cross-entropy term is: After weighting, it becomes Then calculate the consistent operating conditions. The sum of squares is multiplied by Later obtained Therefore, the contribution of this pair of samples to the total loss is During training, if the model continues to iterate and obtains new outputs... Then the cross-entropy of the optimized sample will decrease to After weighting, it is approximately Meanwhile, the new output difference is The sum of squares is The operating condition consistency item was reduced to This significantly reduces the total loss, indicating that the model has simultaneously improved its ability to classify optimized samples and made its output more closely resemble the diagnostic results of real samples under similar operating conditions. The resulting fault diagnosis model, trained using this process, yields... In actual deployment, the received feature vector is still the same as that of S1, and the output is the probability result of each fault category. Since the training set contains both the true state center and the boundary states expanded and adjusted by S2 and S3, the model can provide more stable fault identification results when facing complex working conditions, few-sample faults, and engineering machinery scenarios with significant state drift.

[0062] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for generating engineering machinery fault samples based on a diffusion probability model, characterized in that, include: Multiple sets of transportation data sequences are divided into a unified time window. Multiple feature vectors are extracted from the time window and linearly weighted and fused to construct equipment operation feature vectors. Operating condition coding vectors are extracted from the equipment operation feature vectors. The operating condition encoding vector is input into the diffusion probability model and the condition adjustment model respectively to obtain the current step noise estimation vector and the operating condition adjustment coefficient. The current intermediate feature state is calculated based on the operating condition adjustment coefficient and the equipment operation feature vector. The current intermediate feature state, the operating condition encoding vector and the current step noise estimation vector are re-input into the diffusion probability model to update the current intermediate feature state. The current intermediate feature state is iteratively updated along the preset diffusion step sequence until candidate fault samples are generated. The generator receives the candidate fault samples and the operating condition coding vector, and outputs a correction vector consistent with the candidate fault samples. Based on the correction vector, the generator optimizes the candidate fault samples, establishes structural proximity constraints and operating condition readback constraints to perform secondary optimization on the optimized candidate fault samples, and generates a comprehensively optimized candidate fault sample. The optimized candidate fault samples are combined with the real equipment operation feature vectors to form a joint training sample set. The fault diagnosis model is trained using the joint training sample set. The fault diagnosis model receives the equipment operation feature vectors and outputs the predicted probability of each fault category.

2. The method for generating engineering machinery fault samples based on a diffusion probability model according to claim 1, characterized in that, The fault diagnosis model incorporates a loss function during training. The parameters of the loss function include sample weights, cross-entropy terms, true fault category labels, fault category prediction probabilities, adjustment coefficients, fault category prediction probabilities for candidate fault samples, and fault category prediction probabilities for the actual equipment operating feature vectors in the same operating condition pair.

3. The method for generating engineering machinery fault samples based on a diffusion probability model according to claim 1, characterized in that, After optimizing the candidate fault samples based on the correction vector, the process further includes: The discriminator is used to judge the optimized candidate fault samples and output the probability that the candidate fault samples belong to the true distribution.

4. The method for generating engineering machinery fault samples based on a diffusion probability model according to claim 1, characterized in that, The diffusion probability model is a denoising network structure, which includes an input mapping layer, a fully connected residual block, and an output mapping layer.

5. The method for generating engineering machinery fault samples based on a diffusion probability model according to claim 1, characterized in that, The condition adjustment model includes a fully connected layer, which is used to map the operating condition encoding vector into an intermediate dimension and output the operating condition adjustment coefficient.

6. The method for generating engineering machinery fault samples based on a diffusion probability model according to claim 1, characterized in that, The multiple sets of transport data sequences include the original vibration sequence, the original temperature sequence, and the original current sequence.

7. The method for generating engineering machinery fault samples based on a diffusion probability model according to claim 1, characterized in that, The step of extracting the operating condition coding vector from the equipment operating feature vector specifically includes: Extract the relevant components of the operating conditions from the temperature feature vector and current feature vector within the same time window and combine them to form the operating condition coding vector.

8. The method for generating engineering machinery fault samples based on a diffusion probability model according to claim 1, characterized in that, The diffusion probability model includes a forward diffusion function, which is used to calculate the current intermediate feature state. The parameters of the forward diffusion function include the fidelity coefficient of the diffusion step, the operating condition adjustment coefficient, the equipment operating feature vector, and the disturbance vector.

9. The method for generating engineering machinery fault samples based on a diffusion probability model according to claim 1, characterized in that, The diffusion probability model also includes a reverse update function, which is used to update the current intermediate feature state. The parameters of the reverse update function include a noise estimation vector, a fidelity coefficient, and a working condition adjustment coefficient.

10. The method for generating engineering machinery fault samples based on a diffusion probability model according to claim 8, characterized in that, The noise estimation vector includes the current intermediate feature state, the current diffusion step, and the operating condition coding vector.

Images