Cross-conditioning control equipment fault detection system based on causal correction and full space modeling

Abstract

The application discloses a cross-working condition control equipment fault detection system based on causal correction and full-space modeling. The system comprises a control equipment detection instrument, a control equipment fault database, a data processing module, a full-space fault modeling module, a causal correction module and a control equipment fault display and control module. The full-space modeling module and the causal correction module respectively solve the problems of data sparsity and sample selection bias existing in traditional methods, and are intended to realize unbiased estimation of fault probability under multiple working conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fault detection and causal inference algorithm, and particularly relates to a cross-condition control equipment fault detection system based on causal correction and full-space modeling. BACKGROUND

[0002] As an important and challenging task, the accurate estimation of cross-condition fault occurrence rate is the software foundation of weapon equipment in modern war and a hot issue of research by domestic and foreign scholars. However, the true value of fault occurrence of non-target condition samples is difficult to observe, which brings difficulties to the training of fault occurrence rate estimator. The common method for solving this problem is to train the fault occurrence rate only using target condition data, but this method has been proved to face the problem of sample selection bias due to the existence of sample selection bias; this method also faces the problem of data sparsity due to the small amount of target condition data.

[0003] The present application proposes a full-space fault modeling module on the basis of multi-task learning, effectively solving the problem of data sparsity; based on advanced causal inference technology, the counterfactual distribution is constructed by sample weighting to eliminate the bias of cross-domain estimation of fault occurrence rate, which is expected to further improve the online effect and has broad application prospects. SUMMARY

[0004] In order to overcome the sample selection bias problem of existing fault detection technology under multiple conditions, the present application provides an intelligent detection system which has simple implementation method, high recognition accuracy and good robustness, and can correct sample selection bias.

[0005] The technical scheme provided by the present application is: a cross-condition control equipment fault detection system based on causal correction and full-space modeling, which is composed of six modules of control equipment detection instrument, control equipment fault database, data processing module, full-space fault modeling module, causal correction module and control equipment fault display and control module.

[0006] 1) Control equipment monitoring instrument: collects control equipment bearing temperature, transmission case oil temperature and oil level, water pressure, motor current and temperature information, and stores control equipment information into the database;

[0007] 2) Control equipment fault database: contains historical control equipment data and corresponding fault type labels, and the fault type labels mainly include bearing fault, auxiliary pump fault, hydraulic motor fault, motor overload and cooling system fault;

[0008] 3) Data processing module: processes fault data into a representation that can be directly input into the model;

[0009] 4) Full-space modeling module: based on multi-task learning method, model the condition tendency score and failure rate of each working condition;

[0010] 5) Causal correction module: based on causal inference method, correct the failure rate estimate;

[0011] 6) Control equipment failure display and control module: visualize the control equipment failure diagnosis performance through confusion matrix, precision-recall curve and operator characteristic curve.

[0012] 1. Data processing module: process offline fault logs into representations that can be directly input into the model, the specific steps are as follows:

[0013] 1) Feature bucketing: continuous features such as engine oil level, water pressure, motor current, etc. are converted into discrete features through bucketing, where the bucketing strategy includes but is not limited to equidistant bucketing, equifrequency bucketing and adaptive bucketing based on XGBoost;

[0014] 2) Frequency filtering: for discrete features, some values may only be hit a few times. Discretizing these values into one-hot vectors cannot improve the generalization performance of the model, and will also cause additional storage overhead. Therefore, count the number of occurrences of each value of all discrete features in the training data, and remove samples with occurrence less than the threshold;

[0015] 3) One-hot: considering that most discrete features are unordered features, therefore for each discrete feature, each integer value is mapped to a unique binary vector, and the index corresponding to the integer value is marked as 1, and the rest are zero values;

[0016] 4) Embedding: project the high-dimensional sparse one-hot vector into a low-dimensional dense embedding space. For each sample:

[0017] E i = ψ i (X i ),i=1,2,…,n

[0018] where X i is the one-hot vector corresponding to the i-th feature; is the embedding vector corresponding to the i-th feature, d is the dimension of the embedding vector; ψ i is the embedding query operator corresponding to the i-th feature, with the one-hot vector X i as the index, output the corresponding embedding vector E i in the embedding query table. Finally, the obtained embedding representation is concatenated into a vector

[0019] E=concat(E i ),i=1,2,…,n

[0020] 5) Data split: split the dataset into training set, test set, validation set in a reasonable time sequence.

[0021] 2) Full space modeling module: model the working condition tendency score and the failure occurrence rate, and the specific steps are as follows:

[0022] 1) Construct a network static computation graph to obtain the estimation of the working condition tendency score and the estimation of the failure occurrence rate

[0023] 2) Calculate the loss function based on and

[0024] 3) For the network static computation graph described above, the application further provides an implementation method based on a double-tower model. First, the spliced embedding vector E is input into a shared neural network to extract shallow features:

[0025] Z share =φ share (E;θ share )

[0026] Wherein, φ share is a mapping function corresponding to the shared network, and the parameter is θ share ; Z share is a shared feature output by the shared network modeling. Further, two neural networks are used to model the working condition tendency score and the failure occurrence rate, respectively:

[0027]

[0028]

[0029] Wherein, φ mode is a mapping function corresponding to the working condition modeling network, containing parameters θ mode , outputting the working condition tendency score estimation φ err is a mapping function corresponding to the failure occurrence rate modeling network, containing parameters θ err , outputting the failure occurrence rate estimation

[0030] For the network static computation graph described above, the application further provides an implementation method based on a double-tower model. First, the spliced embedding vector E is input into a shared neural network to extract shallow features:

[0031] Z share =φ share (E;θ share )

[0032] ​where φ share is a shared neural network with parameters θ share ; Z share is the shared feature output by the shared network. Further, the shared feature is input into K expert sub-networks φ k :

[0033] Z k = k (Z share ; k ), k = 1, 2, …, K.

[0034] where Z k is the representation output by each expert sub-network, and θ k is the corresponding parameter of each expert sub-network. Then the weight of each expert sub-network on the working condition tendency score prediction task and the failure rate prediction task is calculated:

[0035]

[0036]

[0037] where g k is the gating function corresponding to the kth expert sub-network, containing the parameter ψ k ; is the weight of the kth expert sub-network on the working condition tendency score prediction task, is the weight of the kth expert sub-network on the failure rate prediction task. Further, Z k is weighted and summed by Z and Z to obtain the representation Z mode of the working condition tendency score prediction task and the representation Z err of the failure rate prediction task:

[0038]

[0039]

[0040] Modeling working condition tendency score and failure rate:

[0041]

[0042]

[0043] where φ mode is the mapping function corresponding to the working condition tendency score modeling network, containing the parameter θ mode , outputting the working condition tendency score estimate φ err is the mapping function corresponding to the failure rate modeling network, with parameters θerr , output failure rate prediction

[0044] For the loss function calculation, the application further provides a calculation method based on binary cross entropy (BCE) loss: first, based on the chain rule, the network estimates the working condition-failure rate

[0045]

[0046] Further based on binary cross entropy, the working condition tendency score and the working condition-failure rate loss are calculated:

[0047]

[0048]

[0049] Wherein, BCE(x, y) = x*log(y) + (1-x)*log(1-y), y mode is the label of the working condition tendency score, which is 1 if E belongs to the target working condition, otherwise 0; y merr is the label of the working condition-failure rate, which is 1 if E belongs to the target working condition and a failure occurs, otherwise 0.

[0050] 3. Causal correction module: the estimated value of the failure rate is corrected causally, and the specific steps are as follows:

[0051] 1) Construct propensity score: based on the working condition tendency score estimated value of the target model, calculate the practical propensity index score p;

[0052] 2) Construct unbiased loss: based on the propensity index score, construct the counterfactual unbiased estimate L err of the failure rate loss;

[0053] 3) Parameter update: use the back propagation algorithm to calculate the derivative of the loss function to the network parameters, and perform incremental update;

[0054] 4) Real-time evaluation and online deployment: after each parameter update, the performance measurement parameters of the model are calculated on the validation set to obtain the final classifier and deploy online.

[0055] For the propensity score construction, the application further provides an implementation method based on the truncated propensity index. First, use the stop_gradient operation to truncate the gradient of other parameters, which is to avoid the wrong gradient propagation when calculating L err :

[0056]

[0057] Further, the truncated value is recorded as the practical propensity score p:

[0058]

[0059] This operation can prove to reduce the variance of subsequent unbiased loss and enhance training stability.

[0060] For the construction of the unbiased loss, the application further gives an implementation method based on inverse propensity score. Since the label y err Only the target working condition sample is observable, so the current fault detection method mostly uses the working condition-fault occurrence rate label of the target working condition to train the fault occurrence rate estimator when training the fault occurrence rate estimator:

[0061]

[0062] Where δ is the loss function corresponding to the method; for the target working condition sample o=1, for the non-target working condition sample o=0. It is easy to prove that the loss is biased for the non-target working condition sample. To solve this problem, the constructed propensity score is used for compensation to obtain the practical fault occurrence rate loss:

[0063]

[0064] It is easy to prove that the compensated loss function L err is unbiased, which can guide to converge uniformly to y err .

[0065] For the construction of the unbiased loss, the application further gives an implementation method based on double robust estimation. For the target working condition sample, the fault occurrence rate error can be calculated by the following formula:

[0066]

[0067] Where δ is the loss function corresponding to the method. It can be proved that this loss is biased. To solve this problem, the constructed propensity score is used for compensation to obtain the practical fault occurrence rate loss:

[0068]

[0069] Where, err is the mapping function corresponding to the fault occurrence rate modeling network. The double robustness of the loss lies in that as long as one of the estimators and p is unbiased, L err is unbiased for the expectation of all samples. In the above formula,​ is an estimated value of the modeling error of the failure occurrence rate, i.e., an estimated error attribution value, and is modeled using a neural network φ dis Modeling:

[0070]

[0071] where φ dis contains parameters θ err ; share is a shared feature modeled by the shared network in the full-space modeling module. Considering that the true value of the failure occurrence rate modeling error is known (i.e., δ) for the target working condition sample, φ dis is trained using the following formula to accurately model the "estimated error attribution value" and improve the estimation quality of L err

[0072]

[0073] Compared with the prior art, the technical concept and benefits of the present application are:

[0074] 1) In view of the performance degradation problem of the existing fault diagnosis model in the multi-working condition scene and the single-working condition data sparsity problem, the present application proposes a full-space modeling module, which utilizes the working condition tendency score and the working condition-failure occurrence rate estimation two auxiliary tasks for multi-task modeling, effectively utilizes a large amount of non-target working condition data, alleviates the data sparsity problem of the failure occurrence rate estimation task, and improves the generalization ability of the model;

[0075] 2) The present application further utilizes the causal inference method to weight the samples in the causal correction module based on the full-space multi-task modeling, thereby calibrating the sample selection bias of the failure occurrence rate estimation;

[0076] 3) The causal correction module introduced by the present application does not need to introduce additional data and labels, and has no high requirements for specific application scenarios, and can provide extensive support for downstream applications. BRIEF DESCRIPTION OF DRAWINGS

[0077] Figure 1 A flowchart of a cross-working condition control equipment fault detection system implementation based on causal correction and full-space modeling;

[0078] Figure 2 A network structure diagram of a cross-working condition control equipment fault detection system based on causal correction and full-space modeling. DETAILED DESCRIPTION

[0079] The present application is further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present application is not limited thereby. ​

[0080] As Figure 1 , 2 shown, a cross-condition control equipment fault detection system based on causal correction and full space modeling is composed of six modules, namely a control equipment detection instrument 1, a control equipment fault database 2, a data processing module 3, a full space fault modeling module 4, a causal correction module 5, and a control equipment fault display and control module 6.

[0081] 1) Control equipment monitoring instrument 1: collects control equipment bearing temperature, transmission case oil temperature and oil level, water pressure, motor current and temperature information, and stores control equipment information into the database;

[0082] 2) Control equipment fault database 2: contains historical control equipment data and corresponding fault type labels, and the fault type labels mainly include bearing fault, auxiliary pump fault, hydraulic motor fault, motor overload, and cooling system fault;

[0083] 3) Data processing module 3: processes fault data into a representation that can be directly input into a model;

[0084] 4) Full space modeling module 4: based on a multi-task learning method, models each condition tendency score and fault occurrence rate;

[0085] 5) Causal correction module 5: based on a causal inference method, corrects the fault occurrence rate estimate;

[0086] 6) Control equipment fault display and control module 6: visualizes the control equipment fault diagnosis performance through a confusion matrix, precision-recall curve, and operator characteristic curve.

[0087] Data processing module 3: processes offline fault logs into a representation that can be directly input into a model, and the specific steps are as follows:

[0088] 1) Feature bucketing: continuous features such as engine oil level, water pressure, and motor current are converted into discrete features through bucketing, and the bucketing strategies include but are not limited to equal interval bucketing, equal frequency bucketing, and adaptive bucketing based on XGBoost;

[0089] 2) Frequency filtering: for discrete features, some values may only be hit a few times. Discretizing these values into one-hot vectors cannot improve the generalization performance of the model, and will cause additional storage overhead. Therefore, the occurrence frequency of all discrete feature values in the training data is counted, and samples with an occurrence frequency less than a threshold are removed;

[0090] 3) One-hot: considering that most discrete features are unordered features, each integer value of each discrete feature is mapped to a unique binary vector, and the index corresponding to the integer value is marked as 1, and the rest are zero values;

[0091] 4) Embedding: Project the high-dimensional sparse one-hot vector into a low-dimensional dense embedding space. For each sample:

[0092] E i i=1,2,…,n i i

[0093] where X i is the one-hot vector corresponding to the i-th feature; is the embedding vector corresponding to the i-th feature, and d is the dimension of the embedding vector; ψ i is the embedding query operator corresponding to the i-th feature, indexed by the one-hot vector X i , and outputs the corresponding embedding vector E i in the embedding query table. Finally, the obtained embedding representation is concatenated into a vector

[0094] E=concat(E i ),i=1,2,…,n

[0095] 5) Data splitting: Reasonably split the dataset into training set, test set and validation set in chronological order.

[0096] Full space modeling module 4: Model the condition tendency score and the failure rate, and the specific steps are as follows:

[0097] 1) Construct a network static computation graph to obtain the prediction of the condition tendency score and the prediction of the failure rate

[0098] 2) Calculate the loss function based on and

[0099] For the construction of the network static computation graph, the application further provides an implementation method based on a double tower model. First, the concatenated embedding vector E is input into a shared neural network to extract shallow features:

[0100] Z share =φ share (E;θ share )

[0101] where φ share is the mapping function corresponding to the shared network, and the parameter is θ share ; Z share is the shared feature output by the shared network modeling. Further, two neural networks are used to model the condition tendency score and the failure rate, respectively:

[0102] ​​​

[0103]

[0104] Where, φ mode It is the mapping function corresponding to the working condition modeling network, which includes the parameter θ. mode Output working condition tendency score prediction φ err It is the mapping function corresponding to the failure rate modeling network, containing the parameter θ. err Output failure rate prediction

[0105] Regarding the construction of the static computation graph of the network, this invention further provides an implementation method based on a dual-tower model. First, the concatenated embedding vector E is input into a shared neural network to extract shallow features:

[0106] Z share =φ share (E;θ share )

[0107] Where, φ share It is a shared neural network with parameters θ share Z share These are the shared features output from the shared network model. Furthermore, the shared features are input into K expert subnetworks φ. k :

[0108] Z k = k (Z share ; k ),=1,2,…,

[0109] Z k It is the representation of the output of each expert subnetwork, θ k These are the parameters corresponding to each expert subnetwork. Next, the weights of each expert subnetwork for the operating condition propensity score prediction task and the failure rate prediction task are calculated:

[0110]

[0111]

[0112] Among them, g k It is the gating function corresponding to the k-th expert subnetwork, containing the parameter ψ. k ; It represents the weight of the k-th expert subnetwork for the task of predicting the work condition propensity score. Z is the weight of the k-th expert subnetwork for the failure rate prediction task. Furthermore, Z... k Use respectively and a weighted sum to obtain a representation Z of the working condition tendency score prediction task mode and the failure rate prediction task err :

[0113]

[0114]

[0115] Modeling working condition tendency scores and failure rates:

[0116]

[0117]

[0118] where φ mode is a mapping function corresponding to the working condition tendency score modeling network, containing parameters θ mode , and outputs the working condition tendency score prediction φ err is a mapping function corresponding to the failure rate modeling network, with parameters θ err , and outputs the failure rate prediction

[0119] For the calculation of the loss function, the application further provides a calculation method based on binary cross entropy (BCE) loss: first, based on the chain rule, the network estimates the working condition-failure rate

[0120]

[0121] Further based on binary cross entropy, the loss of working condition tendency score and working condition-failure rate is calculated:

[0122]

[0123]

[0124] where BCE(x, y) = x*log(y) + (1-x)*log(1-y), y mode is the label of the working condition tendency score, which is 1 if E belongs to the target working condition, and 0 otherwise; y merr is the label of the working condition-failure rate, which is 1 if E belongs to the target working condition and a failure occurs, and 0 otherwise.

[0125] Causal correction module 5: performing causal correction on the failure rate prediction value, the specific steps are as follows:

[0126] 1) Construct propensity score: based on the propensity score of the target model, calculate the practical propensity index score p;

[0127] 2) Construct unbiased loss: based on the propensity index score, construct the counterfactual unbiased estimate L of the failure rate loss err ;

[0128] 3) Parameter update: use the back propagation algorithm to calculate the derivative of the loss function to the network parameters, and perform incremental update;

[0129] 4) Real-time evaluation and online deployment: after each parameter update, calculate the performance measurement parameters of the model on the validation set to obtain the final classifier and deploy it online.

[0130] For the construction of the propensity score, the application further gives an implementation method based on the truncated propensity index. First, use the stop_gradient operation to truncate the gradient of other parameters, which is to avoid the wrong gradient propagation when calculating L err :

[0131]

[0132] Further, truncate and record the truncated value as the practical propensity index score p:

[0133]

[0134] This operation has been proved to reduce the variance of the subsequent unbiased loss and enhance the training stability.

[0135] For the construction of the unbiased loss, the application further gives an implementation method based on the inverse propensity index. Since the label y err of the failure rate (under the target working condition) is only observable for the target working condition samples, most of the current fault detection methods use the working condition-failure rate label of the target working condition to train the failure rate estimator:

[0136]

[0137] Where δ is the loss function corresponding to this method; for target working condition samples o=1, for non-target working condition samples o=0. It is easy to prove that the loss is biased for non-target working condition samples. To solve this problem, the constructed propensity score is used for compensation to obtain the practical failure rate loss:

[0138]

[0139] It is easy to prove that the compensated loss function L erris unbiased, can guide converges to y err .

[0140] For the constructed unbiased loss, the application further gives an implementation method based on double robust estimation. For the target working condition sample, the fault occurrence rate error can be calculated by the following formula:

[0141]

[0142] Where δ is the loss function corresponding to the method. It can be proved that this loss is biased. To solve this problem, the constructed propensity score is used for compensation to obtain the practical fault occurrence rate loss:

[0143]

[0144] Where, err is the mapping function corresponding to the fault occurrence rate modeling network. The double robustness of the loss lies in that as long as one of the estimators and p is unbiased, L err is unbiased for all samples. In the above formula, is the estimated value of the fault occurrence rate modeling error, that is, the estimated error attribution value, which is modeled by using the neural network φ dis :

[0145]

[0146] Where, φ dis contains parameters θ err ; share is the shared feature modeled by the shared network in the full-space modeling module. Considering that for the target working condition sample, the true value of the fault occurrence rate modeling error is known (that is, δ). Therefore, in order to make φ dis accurately model the estimated error attribution value, so as to improve the estimation quality of L err , φ

[0147]

Claims

1. A cross-condition control equipment fault detection system based on causal correction and full space modeling for unbiased estimation of fault probability under multiple conditions; the system comprises a control equipment detection instrument, a control equipment fault database, a data processing module, a full space modeling module, a causal correction module, a control equipment fault display and control module; wherein, The control equipment monitoring instrument: collects the information of control equipment bearing temperature, transmission case oil temperature and oil level, water pressure, motor current and temperature, and stores the control equipment information into the database; The control equipment fault database: contains historical control equipment data and its corresponding fault type label, and the fault type label mainly includes: bearing fault, auxiliary pump fault, hydraulic motor fault, motor overload, cooling system fault; The data processing module: processes the fault data into a representation that can be directly input into the model; The full space modeling module: based on the multi-task learning method, models the condition tendency score and the fault occurrence rate; The causal correction module: based on the causal inference method, corrects the estimated value of the fault occurrence rate; The control equipment fault display and control module: visualizes the control equipment fault diagnosis performance through the confusion matrix, precision-recall curve and operator characteristic curve.

2. The cross-condition control equipment fault detection system based on causal correction and full-space modeling of claim 1, wherein: The data processing module processes the offline fault log into a representation that can be directly input into the model by the following steps: (1.1) Feature bucketing: continuous features including engine oil level, water pressure, motor current are converted into discrete features through bucketing, and the bucketing strategies include but are not limited to equidistant bucketing, equal frequency bucketing and adaptive bucketing based on XGBoost; (1.2) Frequency filtering: for discrete features, some values may only be hit a few times, discretizing these values into one-hot vectors cannot improve the generalization performance of the model, and will cause additional storage overhead; therefore, the occurrence frequency of all discrete feature values in the training data is counted, and samples with occurrence frequency less than the threshold are removed; (1.3) One-hot: considering that most discrete features are unordered features, therefore for each discrete feature, each integer value is mapped to a unique binary vector, and the index corresponding to the integer value is marked as 1, and the rest are zero; (1.4) Embedding: project the high-dimensional sparse one-hot vector into a low-dimensional dense embedding space; for each sample: E i = ψ i (x i ), i = 1, 2,..., n where X i is the one-hot vector corresponding to the i-th feature; is the embedding vector corresponding to the i-th feature, d is the dimension of the embedding vector; ψ i is the embedding query operator corresponding to the i-th feature, indexed by the one-hot vector X i , outputs the corresponding embedding vector E i in the embedding query table; finally, the obtained embedding representations are concatenated into a vector : E = concat(E i ), i = 1, 2,..., n (1.5) Data splitting: reasonably split the dataset into training set, test set and validation set in time sequence.

3. The cross-condition control equipment fault detection system based on causal correction and full-space modeling of claim 1, wherein: The full space modeling module models the condition tendency score and the fault occurrence rate by the following steps: (2.1) Construct a network static computation graph to obtain the prediction of the working condition tendency score and the prediction of the failure occurrence rate (2.2) based on and computing a loss function.

4. The cross-condition control equipment fault detection system based on causal correction and full-space modeling of claim 3, wherein, The step (2.1) is specifically: first input the concatenated embedding vector E into a shared neural network to extract shallow features: Z share = φ share (E; θ share ) wherein φ share is a mapping function corresponding to the sharing network, with parameter θ share ; Z share is a shared feature output by the sharing network; Two neural networks are used to model the condition tendency score and the fault occurrence rate respectively: wherein φ mode is a mapping function corresponding to the operating condition modeling network, containing parameters θ mode ; is the operating condition tendency score prediction value; φ err is a mapping function corresponding to the failure occurrence rate modeling network, containing parameters θ err ; is the output failure occurrence rate prediction value.

5. The cross-condition control equipment fault detection system based on causal correction and full-space modeling of claim 3, wherein, The step (2.1) is specifically: first input the concatenated embedding vector E into a shared neural network to extract shallow features: Z share = φ share (E; θ share ) where φ share is a shared neural network, parameterized by θ share ; Z share is shared features modeled by the shared network; inputting the shared features into K expert subnetworks φ k : Z k = φ k (Z share ; θ k ), k = 1, 2,..., K where Z k is the output of each expert subnetwork, θ k is the corresponding parameter of each expert subnetwork; then the weight of each expert subnetwork to the working condition tendency score prediction task and the failure occurrence rate prediction task is calculated: wherein g k is the gating function corresponding to the kth expert subnetwork, containing the parameters ψ k ; is the weight of the kth expert subnetwork for the working condition tendency score prediction task, is the weight of the kth expert subnetwork for the failure occurrence rate prediction task; Z k are respectively calculated by weighted summation, to obtain the representation Z mode of the working condition tendency score prediction task err :​ Model the condition tendency score and the fault occurrence rate: wherein φ mode is a mapping function corresponding to the working condition tendency score modeling network, containing parameters θ mode , and outputting working condition tendency score prediction φ err is a mapping function corresponding to the failure occurrence rate modeling network, with parameters θ err , and outputting failure occurrence rate prediction 6. The cross-condition control equipment fault detection system based on causal correction and full-space modeling of claim 3, wherein, The step (2.2) is specifically: firstly, calculating the network's estimation of the working condition-fault occurrence rate based on the chain rule : Based on binary cross entropy, calculate the loss of condition tendency score and condition-fault occurrence rate: where BCE(x, y) = x*log(y) + (1 - x)*log(l - y), y mode is the label of working condition tendency score, taking 1 if E belongs to the target working condition, otherwise taking 0; y merr is the label of working condition-failure occurrence rate, taking 1 if E belongs to the target working condition and failure occurs, otherwise taking 0.

7. The cross-condition control equipment fault detection system based on causal correction and full-space modeling of claim 1, wherein: The causal correction module corrects the estimated value of the fault occurrence rate by the following steps: (3.1) Constructing tendency score: based on the target model, the working condition tendency score is calculated to estimate the practical tendency index score p; (3.2) Constructing an unbiased loss: Based on the propensity index score, construct a counterfactual unbiased estimate L of the failure rate loss err ; (3.3) Parameter updating: using the back propagation algorithm to calculate the derivative of the loss function to the network parameters, and performing incremental updating; (3.4) Real-time evaluation and online deployment: after each parameter update, the performance measurement parameters of the model are calculated on the validation set to obtain the final classifier and deploy it online.

8. The cross-condition control equipment fault detection system based on causal correction and full-space modeling of claim 7, wherein: In the step (3.1), the output working condition tendency score prediction is first truncated using a stop_gradient operation Gradient on other parameters: On numerical truncation, the truncated value is denoted as the practical inclination index score p:

9. The cross-condition control equipment fault detection system based on causal correction and full-space modeling of claim 7, wherein: In the step (3.2), the constructed tendency score is used for compensation to obtain the practical failure rate loss: The loss function L after the easy evidence compensation err is unbiased and can guide the failure rate estimation converges to the failure rate y err .

10. The cross-condition control equipment fault detection system based on causal correction and full space modeling of claim 7, wherein: In the step (3.2), for the target working condition sample, the failure rate error is calculated by the following formula: where δ is the loss function corresponding to the method, y mode is the working condition tendency score, y merr is the working condition-failure rate label, represents the output failure rate estimate with parameters θ err ; then the constructed tendency score is used for compensation to obtain the practical failure rate loss: In the above formula, is an estimate of the "modeling error" of the failure occurrence rate, i.e., an "estimated error attribution value", using a neural network φ dis Modeling: where φ dis contains the parameter θ err ; Z share is a shared feature modeled in the full space modeling module from the shared network modeling output; training :

Images