A method for adjusting nitrogen and oxygen concentrations of a cold chain controlled atmosphere machine based on a deep neural network

Abstract

The application discloses a kind of nitrogen oxygen concentration regulation method of cold chain controlled atmosphere machine based on deep neural network, comprising the following steps: step one: constructing latent feature representation;Step two: latent feature representation is input into multi-objective variational inversion network, and inference obtains oxygen potential disturbance field;Step three: oxygen content prediction curve is generated based on oxygen potential disturbance field in combination with historical regulation behavior response relationship;Step four: decision calculation module generates regulation strategy sequence;Step five: regulation strategy sequence is synchronized to each equipment;Step six: after each regulation period ends, feedback sample is written into incremental training cache;Step seven: when incremental training cache reaches the set capacity threshold, the weight of multi-objective variational inversion network is updated using adaptive learning rate mechanism;Step eight: when identifying that oxygen potential disturbance field disturbance reasoning anomaly occurs, start field map tracking recovery mechanism.The application combines multi-objective variational inversion and graph tracking recovery mechanism to intelligently generate cold chain oxygen control strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cold chain logistics gas regulation, and in particular to a nitrogen-oxygen concentration regulation method for a cold chain gas regulation machine based on a deep neural network. BACKGROUND

[0002] With the continuous expansion of cross-regional circulation of fresh agricultural products and the increasing demand for preservation quality of goods in cold chain transportation, cold chain gas regulation technology based on gas composition regulation has become an important means to ensure the quality of goods storage and transportation. Existing cold chain gas regulation machines mostly use fixed control thresholds or simplified feedback regulation algorithms to control nitrogen injection and fan circulation through oxygen concentration monitoring values to maintain a set low-oxygen environment, but there are generally the following problems in complex actual applications:

[0003] The collected oxygen concentration, temperature and humidity, and pressure difference sensor data have problems of sparse spatial distribution and inconsistent dynamic response, which makes it difficult to fully reflect the real-time and space evolution process of gas exchange inside the cold chain cabinet, leading to insufficient information for the control strategy and response lag; traditional control methods mostly use linear regression, proportional-integral control algorithms or empirical logic rule modeling, which lack deep modeling capabilities for goods metabolism behavior, nitrogen disturbance effects and changes in system airtightness, and cannot accurately predict the comprehensive effects of regulation actions in actual operation, which is prone to under-regulation or over-regulation; the regulation system lacks effective memory and use of causal relationships between historical behavior and disturbance results, making it difficult for the regulation strategy to adaptively update with changes in goods varieties, loading capacity and seasonality, and the model performance is prone to degradation in long-term operation, reducing control accuracy; when sensor drift, model misalignment or execution unit abnormalities occur, existing devices mostly run with fixed parameters, lack of recovery mechanisms based on historical similar scene reasoning, and cannot maintain optimal control capability in abnormal state, resulting in regulation blind spots and preservation risks.

[0004] Therefore, how to provide a nitrogen-oxygen concentration regulation method for a cold chain gas regulation machine based on a deep neural network is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a nitrogen-oxygen concentration regulation method for a cold chain gas regulation machine based on a deep neural network. The present application combines a multi-objective variational inversion network and an oxygen content prediction network to intelligently construct an oxygen potential disturbance field and generate a cold chain gas regulation control strategy, solving the problems of inaccurate oxygen concentration prediction, lagging regulation strategy response, and unstable control under complex disturbances, and having the advantages of accurate regulation, strong adaptability, stable operation, and high abnormal recovery capability, suitable for intelligent gas regulation control requirements under various goods and environmental conditions.

[0006] According to the nitrogen-oxygen concentration regulation method for a cold chain gas regulation machine based on a deep neural network, the following steps are included:

[0007] Step one: Collect multi-dimensional historical operation data and perform cross-dimensional embedding coding to construct latent feature representation;

[0008] Step two: Input the latent feature representation into a multi-objective variational inference network to generate a current metabolic state hidden variable and infer an oxygen potential disturbance field;

[0009] Step three: Based on the oxygen potential disturbance field, in combination with the historical adjustment behavior response relationship, generate an oxygen content prediction curve within a future target time window;

[0010] Step four: The decision calculation module generates an adjustment strategy sequence based on the oxygen content prediction curve and the upper and lower threshold values of the corresponding oxygen concentration of the goods, in combination with the oxygen potential disturbance field, the adjustment strategy sequence including nitrogen production set value, fan frequency, air valve opening and oxygen reduction trigger value;

[0011] Step five: The adjustment strategy sequence is synchronized to the nitrogen generating device, fan and air valve through the execution control module to realize dynamic cooperation of nitrogen injection, oxygen replacement, gas circulation and differential pressure control;

[0012] Step six: After each adjustment cycle, the real-time sampling sequence, adjustment strategy sequence and final oxygen content are packaged as feedback samples and written into an incremental training cache;

[0013] Step seven: When the incremental training cache reaches a set capacity threshold, use an adaptive learning rate mechanism to update the weights of the multi-objective variational inference network;

[0014] Step eight: When a disturbance inference anomaly is identified in the oxygen potential disturbance field, start the field map tracking recovery mechanism to match the most similar oxygen potential disturbance field state from the historical operation database, quickly generate emergency recovery control parameters and immediately issue them for execution.

[0015] Optionally, the multi-dimensional historical operation data includes goods variety, loading capacity, cold chain equipment environment temperature, humidity, air pressure, membrane separation nitrogen generator power gear, nitrogen flow, valve opening, fan frequency, oxygen concentration in the cabinet, nitrogen purity and differential pressure.

[0016] Optionally, the cross-dimensional embedding coding to construct latent feature representation is specifically:

[0017] Extract the features of the multi-dimensional historical operation data, and after normalization, input them into the corresponding feature embedding subnetwork;

[0018] The feature embedding subnetwork is an encoding module containing a fully connected layer and a ReLU activation function, used to map the original features of each dimension to an embedding vector of a set length;

[0019] input the embedding vector into a cross-attention mechanism module, the cross-attention mechanism module is composed of a query network, a key-value network and a weighted fusion module:

[0020] The query network takes the embedding vector of the cargo variety and the loading capacity as input to generate an attention query matrix;

[0021] The key-value network takes the remaining dimension feature embedding vector as input to generate an attention key matrix and an attention value matrix;

[0022] The remaining dimension features are cold chain equipment environmental temperature, humidity, air pressure, membrane separation nitrogen making machine power gear, nitrogen flow, valve opening, fan frequency, oxygen concentration in the cabinet, nitrogen purity and differential pressure characteristics;

[0023] The weighted fusion module calculates the attention weight according to the dot product similarity between the attention query matrix and the attention key matrix, and applies the attention weight to the attention value matrix to obtain a context-aware representation;

[0024] The context-aware representation is spliced with the original embedding vector to form a latent feature representation as input to a multi-objective variational inversion network.

[0025] Optionally, the step two is specifically:

[0026] input the latent feature representation into a multi-objective variational inversion network, the objective variational inversion network includes an encoder subnetwork, a variational sampling module, a decoder subnetwork and a perturbation mapping layer;

[0027] The encoder subnetwork includes a plurality of sequentially connected fully connected layers and ReLU activation functions, for encoding the latent feature representation into a mean vector and a log variance vector of a latent distribution;

[0028] The variational sampling module generates a current metabolic state latent variable according to the mean vector and the log variance vector through a reparameterization method;

[0029] The decoder subnetwork includes an oxygen metabolic rate estimation branch, a nitrogen replacement intensity inversion branch and a differential pressure adaptability branch;

[0030] The outputs of the three branches are combined and input into a perturbation mapping layer, the perturbation mapping layer includes a spatial position encoding unit, a time recursion generation unit and a tensor combination module;

[0031] The spatial position encoding unit is used to receive the spatial arrangement parameters of the goods in the cold chain cabinet, the air duct distribution map and the sensor layout map, discretize and grid encode the spatial structure information to generate a spatial position matrix;

[0032] The time recursion generation unit is a set of shared weight gated recurrent unit, receives the outputs of three branches, and combines the historical adjustment behavior sequence to perform recursive iteration calculation, and outputs a disturbance state vector with time sequence characteristics;

[0033] The tensor combination module receives the spatial position matrix and the disturbance state vector, constructs an oxygen potential disturbance field with a three-dimensional structure through point-by-point mapping and channel fusion, and the channel dimension of each position point represents the oxygen concentration disturbance trend of the position point at future time steps.

[0034] Optionally, the oxygen metabolism rate estimation branch includes a time convolution layer and a temperature modulation unit, the time convolution layer is used to extract the time characteristics of the current metabolic state hidden variable, and the temperature modulation unit adjusts the amplitude of the convolution result according to the ambient temperature, and is used to generate the oxygen metabolism consumption per unit time under the current state;

[0035] The nitrogen replacement intensity inversion branch includes two serially connected fully connected layers and an attention weighting mechanism, the attention weighting mechanism allocates feature weights according to the coupling relationship between the nitrogen injection rate and the oxygen concentration change rate, and is used to generate an oxygen concentration change intensity factor caused by nitrogen injection under the current state;

[0036] The differential pressure adaptability branch includes two residual convolution units and an airtightness adjustment factor, the residual convolution unit is used to extract multi-scale flow structure features, and the airtightness adjustment factor corrects the differential pressure estimation value according to the equipment sealing level and the cabinet structure parameter, and is used to output a differential pressure compensation value.

[0037] Optionally, the step three is specifically:

[0038] The oxygen potential disturbance field is input into an oxygen content prediction network, the oxygen content prediction network includes a disturbance feature extraction module, a historical response memory module and a sequence curve generation module;

[0039] The disturbance feature extraction module includes a plurality of stacked two-dimensional convolution layers and a residual connection structure, receives the oxygen potential disturbance field, extracts the spatial features of the oxygen potential disturbance field, and outputs a spatial disturbance feature map;

[0040] The historical response memory module includes a gated recurrent unit network, receives a historical adjustment behavior parameter sequence and a corresponding historical oxygen concentration measured sequence, extracts the time sequence response characteristics between the historical adjustment behavior and the oxygen concentration, and outputs a historical response feature vector;

[0041] The sequence curve generation module comprises an attention fusion network and a recurrent decoding network, the attention fusion network is used for fusing the spatial disturbance feature map and the historical response feature vector to form a disturbance response feature, and the recurrent decoding network is used for decoding the disturbance response feature step by step to output an oxygen concentration prediction value of each time step in a future target time window.

[0042] The oxygen concentration prediction value of each time step is sequentially combined into a complete sequence to generate an oxygen content prediction curve in the future target time window.

[0043] Optionally, the fourth step specifically comprises:

[0044] The decision calculation module identifies all time sections in which the predicted oxygen concentration exceeds the upper and lower threshold values of the corresponding oxygen concentration of the goods in the target time window according to the oxygen content prediction curve, and records the maximum deviation value of each over-limit section and the over-limit duration .

[0045] According to the oxygen concentration change amount caused by the input of each adjustment behavior unit in the oxygen potential disturbance field, a disturbance response coefficient is defined: .

[0046] wherein, represents the disturbance response coefficient of the i th adjustment behavior unit, represents the oxygen concentration change caused by the adjustment amount of the i th adjustment behavior unit . represents the adjustment amount of the i th adjustment behavior unit.

[0047] In each oxygen concentration over-limit section, an adjustment input satisfying the following optimization objective is calculated: .

[0048] If multiple satisfy the optimization objective, the one with the minimum energy consumption is selected.

[0049] When the maximum deviation value is less than a preset early intervention deviation threshold value, the adjustment behavior with the maximum disturbance response coefficient is selected to perform a low-amplitude compensation operation to suppress the rising trend of the oxygen concentration; that is, when it is predicted that the oxygen concentration is slightly deviated but not serious, the control means with the highest adjustment efficiency is selected, and only a mild intervention is performed to prevent the problem from being enlarged and to avoid excessive control.

[0050] If there is a continuous over-limit duration in the oxygen content prediction curve, which is greater than a preset over-limit time judgment threshold value or the maximum deviation value If the maximum deviation is greater than the preset upper target limit, a reduction oxygen trigger state is set, the reduction oxygen trigger determination value is set to 1, the nitrogen injection rate and the fan frequency corresponding to the time period are set to the maximum adjustment value, and the air supply valve is closed; when the reduction oxygen trigger determination value is set to 0, the reduction oxygen trigger state is not started.

[0051] Finally, the control parameter set is combined in time sequence into an adjustment strategy sequence, which includes a nitrogen production set value, a fan frequency, an air supply valve opening degree, and a reduction oxygen trigger determination value.

[0052] Optionally, the air supplied by the air supply valve is air, which is used to increase the oxygen concentration.

[0053] Optionally, the step seven is specifically:

[0054] All feedback samples in the incremental training cache are combined into several batches of training data in the order of the collection time, each feedback sample including a data sequence sampled in a complete adjustment cycle, a corresponding actual execution adjustment strategy sequence, and an ultimately obtained oxygen concentration measurement result;

[0055] The existing weight parameters in the current multi-objective variational inversion network are used as initial parameters, the training data batches are input into the multi-objective variational inversion network one by one, and the error between the calculated output and the actual measurement result is determined to determine the size of the current loss function;

[0056] In the training process, an adaptive learning rate mechanism is used, after each training batch is completed, the change trend of the loss function compared with the loss function of the last batch is evaluated, if the loss function continuously appears a trend of a descending speed less than a set speed threshold or even rising, the current learning rate is reduced to avoid fluctuations or model performance degradation in the parameter updating process;

[0057] If the loss function continuously appears a descending speed greater than the set speed threshold, the current learning rate is increased to speed up the convergence speed of the weight parameters;

[0058] If the descending speed of the loss function is equal to the set speed threshold, the current learning rate is maintained unchanged to maintain stable training effect;

[0059] The training process sets an early stop condition, when the multi-objective variational inversion network model is continuously trained for multiple batches, the change amplitude of the loss function is less than a set change lower threshold, indicating that the performance improvement is no longer obvious, the current fine-tuning process is terminated;

[0060] Replace the original weight parameters with the latest multi-objective variational inversion network model weight parameters obtained after fine-tuning, empty the incremental training cache, and wait for the accumulation of new feedback samples again to maintain the dynamic adaptability of the multi-objective variational inversion network model to the cold chain system environment and equipment state.

[0061] Optionally, the field map tracking recovery mechanism specifically comprises:

[0062] In a continuous plurality of sampling periods, when the deviation between the oxygen potential disturbance field prediction output and the oxygen concentration measured value exceeds the dynamic tolerance threshold, or the oxygen potential disturbance field output appears value vacancy or decoding failure, the disturbance reasoning abnormality is identified;

[0063] After identifying the disturbance reasoning abnormality, the data samples containing the oxygen potential disturbance field state label in the historical operation database are retrieved, the joint similarity measurement function is constructed based on the spatial distribution structure of the oxygen potential disturbance field and the time sequence characteristics of the corresponding adjustment behavior, the matching calculation of the current abnormal state and the historical oxygen potential disturbance field state is performed, and a plurality of history oxygen potential disturbance field state samples with optimal similarity are determined as a candidate set;

[0064] The corresponding control parameters in the candidate set are weighted and combined to generate emergency recovery control parameters for the current control period, and the emergency recovery control parameters include nitrogen production set value, fan frequency and air supplement valve opening degree;

[0065] The emergency recovery control parameters are issued to each corresponding control object to replace the original adjustment strategy sequence for execution;

[0066] During the execution process, the disturbance reasoning state and the actual response data are collected, when the deviation between the oxygen potential disturbance field prediction output and the oxygen concentration measured value is less than or equal to the dynamic tolerance threshold, and the oxygen potential disturbance field output does not appear value vacancy and decoding failure, the original control path is restored, and the field map tracking recovery process is ended.

[0067] The beneficial effects of the present application are:

[0068] The present application proposes a regulation strategy generation method jointly driven by cross-dimensional latent feature expression and disturbance response inversion, which solves the problems of large disturbance uncertainty, complex regulation response mechanism and poor model adaptability in oxygen concentration regulation of cold chain controlled atmosphere system. The embedded context representation based on goods variety, loading capacity and environmental parameters is used to realize nonlinear interaction and fusion of multi-dimensional features through attention mechanism, and metabolic hidden variable representation suitable for various operating states is obtained. In the inversion stage, oxygen metabolism, nitrogen disturbance and pressure difference compensation are modeled as independent branches, and their dynamic causal relationships are extracted and fused into the oxygen potential disturbance field with spatiotemporal structure, which further drives the generation of oxygen content prediction curve in the target time period. The disturbance response coefficient is introduced to optimize the input amplitude at the control level, and the minimum control rate path is matched according to the target oxygen concentration deviation, which realizes the precise regulation of oxygen content. In terms of model self-updating, a feedback sample buffer area with target labels is constructed, and a dynamic loss trend adjustment learning rate mechanism is used to realize online fine-tuning and updating of the disturbance inversion network, improving the model's adaptability to seasonal changes and goods characteristics. In terms of abnormal state processing, the field map tracking recovery mechanism of oxygen potential disturbance field is first introduced. When reasoning fails or abnormally fluctuates, similar field states are quickly retrieved from historical samples to construct recovery control parameters, ensuring the stability and safety of the system in complex situations. Finally, the stability of oxygen content prediction in cold chain controlled atmosphere cabinet, low-energy consumption control and abnormal recovery ability are comprehensively improved, which is significantly better than the regulation effect of traditional feedback control and linear modeling methods. BRIEF DESCRIPTION OF DRAWINGS

[0069] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application and are used to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0070] Figure 1 The overall flowchart of a nitrogen and oxygen concentration regulation method for cold chain controlled atmosphere machine based on deep neural network proposed by the present application;

[0071] Figure 2 The multi-objective variational inversion network structure diagram of a nitrogen and oxygen concentration regulation method for cold chain controlled atmosphere machine based on deep neural network proposed by the present application;

[0072] Figure 3 The matching and emergency regulation flowchart of the field map tracking recovery mechanism of a nitrogen and oxygen concentration regulation method for cold chain controlled atmosphere machine based on deep neural network proposed by the present application. DETAILED DESCRIPTION

[0073] The application will be described in further detail below with reference to the drawings. These drawings are simplified schematic diagrams and only show the basic structure of the application in a schematic manner, and thus only show the components relevant to the application.

[0074] Reference Figures 1-3 A nitrogen-oxygen concentration adjustment method for a cold-chain CA machine based on a deep neural network, comprising the following steps:

[0075] Step one: collect multi-dimensional historical operation data and perform cross-dimensional embedding coding to construct a latent feature representation;

[0076] Step two: input the latent feature representation into a multi-objective variational inversion network to generate a current metabolic state hidden variable and infer an oxygen potential disturbance field;

[0077] Step three: based on the oxygen potential disturbance field, in combination with a historical adjustment behavior response relationship, generate an oxygen content prediction curve within a future target time window;

[0078] Step four: a decision calculation module generates an adjustment strategy sequence based on the oxygen content prediction curve and the upper and lower threshold values of the corresponding oxygen concentration of the goods, in combination with the oxygen potential disturbance field, the adjustment strategy sequence including a nitrogen production set value, a fan frequency, a makeup air valve opening degree, and a deoxygenation trigger determination value;

[0079] Step five: the adjustment strategy sequence is synchronously issued to the nitrogen generating device, the fan, and the makeup air valve through an execution control module to realize dynamic cooperation of nitrogen injection, oxygen replacement, gas circulation, and pressure difference control;

[0080] Step six: after each adjustment period ends, the real-time sampling sequence, the adjustment strategy sequence, and the final oxygen content are packaged as feedback samples and written into an incremental training buffer;

[0081] Step seven: when the incremental training buffer reaches a set capacity threshold, the multi-objective variational inversion network is updated in weight using an adaptive learning rate mechanism;

[0082] Step eight: when a disturbance inference anomaly of the oxygen potential disturbance field is identified, a field map tracking recovery mechanism is started to match an oxygen potential disturbance field state with the best similarity from a historical operation database, to quickly generate emergency recovery control parameters and immediately issue and execute them.

[0083] In this embodiment, the multi-dimensional historical operation data includes the type of goods, the loading capacity, the cold chain equipment environment temperature, the humidity, the air pressure, the power gear of the membrane separation nitrogen generator, the nitrogen flow, the valve opening degree, the fan frequency, the oxygen concentration in the cabinet, the nitrogen purity, and the pressure difference.

[0084] The generated nitrogen is preferentially supplied to the gas regulating box. When the nitrogen concentration in the gas regulating box reaches the set value, the nitrogen generating system continues to run, and the excess nitrogen is converted into liquid nitrogen by the condensing system for storage. The liquid nitrogen is used for subsequent emergency or peak period rapid replenishment to ensure the continuity and stability of the system operation.

[0085] In this embodiment, the cross-dimensional embedding encoding is performed to construct a latent feature representation, specifically:

[0086] The features of the multi-dimensional historical operation data are extracted and input to the corresponding feature embedding sub-network after normalization processing;

[0087] The feature embedding sub-network is an encoding module containing a fully connected layer and a ReLU activation function, which is used to map the original features of each dimension to embedding vectors of a set length;

[0088] The embedding vectors are input to a cross-attention mechanism module, which is composed of a query network, a key-value network, and a weighted fusion module:

[0089] The query network takes the embedding vectors of the cargo variety and loading capacity as input to generate an attention query matrix;

[0090] The key-value network takes the embedding vectors of the remaining dimension features as input to generate an attention key matrix and an attention value matrix;

[0091] The remaining dimension features are cold chain equipment environmental temperature, humidity, air pressure, membrane separation nitrogen generator power gear, nitrogen flow, valve opening, fan frequency, oxygen concentration in the cabinet, nitrogen purity, and differential pressure features;

[0092] The weighted fusion module calculates the attention weight according to the dot product similarity between the attention query matrix and the attention key matrix, and applies the attention weight to the attention value matrix to obtain a context-aware representation;

[0093] The context-aware representation and the original embedding vector are spliced to form a latent feature representation, which is used as input to the multi-objective variational inversion network.

[0094] In this embodiment, step two is specifically:

[0095] The latent feature representation is input to a multi-objective variational inversion network, which includes an encoder sub-network, a variational sampling module, a decoder sub-network, and a perturbation mapping layer;

[0096] The encoder sub-network includes a plurality of sequentially connected fully connected layers and ReLU activation functions, which are used to encode the latent feature representation into a mean vector and a log variance vector of a latent distribution;

[0097] The variational sampling module generates a current metabolic state hidden variable through a reparameterization method according to the mean vector and the logarithmic variance vector;

[0098] The decoder subnetwork includes an oxygen metabolic rate estimation branch, a nitrogen replacement intensity inversion branch, and a pressure difference adaptability branch;

[0099] The outputs of the three branches are combined and input into a disturbance mapping layer, which includes a spatial position encoding unit, a time recursive generation unit, and a tensor combination module;

[0100] The spatial position encoding unit is used to receive spatial arrangement parameters of goods in the cold chain cabinet, an air duct distribution map, and a sensor layout map, discretely grid encode the spatial structure information, and generate a spatial position matrix;

[0101] The time recursive generation unit is a set of gated recurrent units sharing weights, receives the outputs of the three branches, and combines the historical adjustment behavior sequence to perform recursive iterative calculation, and outputs a disturbance state vector with time sequence characteristics;

[0102] The tensor combination module receives the spatial position matrix and the disturbance state vector, and constructs an oxygen potential disturbance field with a three-dimensional structure through point-by-point mapping and channel fusion. The channel dimension of each position point represents the oxygen concentration disturbance trend of the position point at future time steps.

[0103] In this embodiment, the oxygen metabolic rate estimation branch includes a time convolution layer and a temperature modulation unit. The time convolution layer is used to extract the time characteristics of the current metabolic state hidden variable, and the temperature modulation unit adjusts the amplitude of the convolution result according to the ambient temperature, and is used to generate the oxygen metabolic consumption per unit time in the current state;

[0104] The nitrogen replacement intensity inversion branch includes two fully connected layers connected in series and an attention weighting mechanism. The attention weighting mechanism allocates feature weights according to the coupling relationship between the nitrogen injection rate and the oxygen concentration change rate, and is used to generate an oxygen concentration change intensity factor caused by nitrogen injection in the current state;

[0105] The pressure difference adaptability branch includes two residual convolution units and an air tightness adjustment factor. The residual convolution unit is used to extract multi-scale flow structure features, and the air tightness adjustment factor corrects the pressure difference estimation value according to the equipment sealing level and the cabinet structure parameters, and is used to output a pressure difference compensation value.

[0106] In this embodiment, step three is specifically:

[0107] inputting the oxygen potential disturbance field into an oxygen content prediction network, the oxygen content prediction network comprising a disturbance feature extraction module, a historical response memory module, and a sequence curve generation module;

[0108] The disturbance feature extraction module comprises a plurality of stacked two-dimensional convolution layers and a residual connection structure, receives the oxygen potential disturbance field, extracts spatial features of the oxygen potential disturbance field, and outputs a spatial disturbance feature map;

[0109] The historical response memory module comprises a gated recurrent unit network, receives a historical adjustment behavior parameter sequence and a corresponding historical oxygen concentration measured sequence, extracts time sequence response features between the historical adjustment behavior and the oxygen concentration, and outputs a historical response feature vector;

[0110] The sequence curve generation module comprises an attention fusion network and a recurrent decoding network, the attention fusion network is used to fuse the spatial disturbance feature map and the historical response feature vector to form a disturbance response feature, and the recurrent decoding network decodes the disturbance response feature step by step to output an oxygen concentration prediction value at each time step in a future target time window;

[0111] The oxygen concentration prediction value at each time step is sequentially combined into a complete sequence to generate an oxygen content prediction curve in the future target time window.

[0112] In the embodiment, the step four is specifically:

[0113] The decision calculation module identifies, according to the oxygen content prediction curve, all time sections in which the predicted oxygen concentration exceeds the upper and lower limit thresholds of the corresponding oxygen concentration of the goods in the target time window, records the maximum deviation value of each over-limit section and the over-limit duration ;

[0114] According to the oxygen concentration change caused by each adjustment behavior unit in the oxygen potential disturbance field, a disturbance response coefficient is defined: ;

[0115] wherein, represents the disturbance response coefficient of the i-th type of adjustment behavior unit, represents the oxygen concentration change caused by the adjustment amount of the i-th type of adjustment behavior unit ; represents the adjustment amount of the i-th type of adjustment behavior unit;

[0116] In each oxygen concentration over-limit section, an adjustment input satisfying the following optimization objective is calculated: ;

[0117] If a plurality of satisfy the optimization objective, the one with the minimum energy consumption is selected.

[0118] When the maximum deviation value is less than the preset early intervention deviation threshold value, the disturbance response coefficient with the largest adjustment behavior is selected, a low-amplitude compensation operation is performed to suppress the rising trend of oxygen concentration; that is, when it is predicted that the oxygen concentration has deviated slightly but is not serious, the control means with the highest adjustment efficiency is selected, and only a slight intervention is performed to prevent the problem from being enlarged and to avoid excessive control;

[0119] If there is a continuous over-limit duration in the oxygen content prediction curve greater than the preset over-limit time determination threshold value or the maximum deviation value greater than the maximum deviation value of the preset upper limit, the oxygen reduction trigger state is set, the oxygen reduction trigger determination value is set to 1, the nitrogen injection rate and the fan frequency corresponding to the time period are specified as the maximum adjustment value, and the air supply valve is closed; when the oxygen reduction trigger determination value is 0, the oxygen reduction trigger state is not started.

[0120] Finally, the control parameter set is combined in time sequence to form an adjustment strategy sequence, which includes the nitrogen production set value, the fan frequency, the air supply valve opening degree, and the oxygen reduction trigger determination value.

[0121] In this embodiment, the air supplied by the air supply valve is air, which is used to increase the oxygen concentration.

[0122] In this embodiment, step seven is specifically:

[0123] All feedback samples in the incremental training cache are combined in time sequence according to the collection time to form several batches of training data, each feedback sample including a data sequence sampled in a complete adjustment period, a corresponding actual adjustment strategy sequence, and an oxygen concentration measurement result finally obtained;

[0124] The existing weight parameters in the current multi-objective variational inversion network are used as initial parameters, the training data batches are input into the multi-objective variational inversion network one by one, and the error between the calculated output and the actual measurement result is calculated to determine the size of the current loss function;

[0125] The loss function of the multi-objective variational inversion network is composed of multiple sub-targets, aiming to simultaneously optimize the prediction accuracy of the oxygen metabolism rate estimation, the nitrogen replacement intensity inversion, and the pressure difference compensation task, and to ensure the expression stability of the latent space. The loss function mainly consists of two parts:

[0126] One is the reconstruction error term, which is used to measure the deviation between the model output and the real sampling data, and to ensure that the prediction result of each output branch can accurately reflect the actual system state;

[0127] The second is a variational regularization term, which is used to constrain the difference between the latent variable distribution and the standard normal distribution, and improve the generalization ability and stability of the model.

[0128] In the training process, an adaptive learning rate mechanism is adopted. After each training batch, the trend of the loss function compared to the loss function of the previous batch is evaluated. If the loss function continuously appears to decrease at a speed less than a set speed threshold or even increases, the current learning rate is reduced to avoid fluctuations or performance degradation in the parameter update process.

[0129] If the loss function continuously appears to decrease at a speed greater than the set speed threshold, the current learning rate is increased to speed up the convergence of the weight parameters.

[0130] If the loss function decreases at a speed equal to the set speed threshold, the current learning rate is maintained to maintain stable training results.

[0131] The training process sets an early stopping condition. When the multi-objective variational inversion network model is continuously trained for multiple batches, the change in the loss function is less than a set lower threshold, indicating that the performance improvement is no longer significant, the fine-tuning process is terminated.

[0132] The latest multi-objective variational inversion network model weight parameters obtained after fine-tuning are replaced with the original weight parameters, and the incremental training cache is emptied to accumulate new feedback samples again to maintain the dynamic adaptability of the multi-objective variational inversion network model to the cold chain system environment and equipment state.

[0133] In this embodiment, the field map tracking recovery mechanism is specifically:

[0134] When the deviation between the oxygen potential disturbance field prediction output and the oxygen concentration measured value exceeds the dynamic tolerance threshold, or the oxygen potential disturbance field output has value gaps or decoding fails within a plurality of consecutive sampling periods, the disturbance reasoning anomaly is identified.

[0135] After identifying the disturbance reasoning anomaly, data samples containing oxygen potential disturbance field state labels in the historical operation database are retrieved, a joint similarity measure function is constructed based on the spatial distribution structure of the oxygen potential disturbance field and the time sequence characteristics of the corresponding adjustment behavior, the current abnormal state and the historical oxygen potential disturbance field state are matched and calculated, and a plurality of historical oxygen potential disturbance field state samples with optimal similarity are determined as a candidate set.

[0136] Let the oxygen potential disturbance field of the current abnormal state be a three-dimensional tensor , where X and Y are spatial dimensions representing the horizontal and vertical grid coordinates of the cold chain cabinet, and T is the time dimension representing the future prediction time step.

[0137] Let the oxygen potential disturbance field of the kth candidate sample in the historical operation database be The corresponding adjustment behavior sequence is wherein represents the control parameter vector of the kth candidate sample at time step T;

[0138] The current adjustment behavior sequence is wherein represents the control parameter vector of the current adjustment behavior sequence at time step t;

[0139] The joint similarity measure function is defined as follows: ;

[0140] wherein, represents the weight coefficient, represents the difference measure of the spatial disturbance field: ;

[0141] wherein, represents the Frobenius norm;

[0142] represents the difference measure of the adjustment behavior sequence: ;

[0143] wherein, represents the inner product operation, represents the two-norm;

[0144] The final candidate set is: ;

[0145] wherein, TopK represents that K candidate samples with the smallest similarity scores are selected from N historical samples, and the candidate set is composed of the K candidate samples.

[0146] The corresponding control parameters in the candidate set are combined by weighting to generate emergency recovery control parameters for the current control period, and the emergency recovery control parameters include a nitrogen production set value, a fan frequency, and a supplementary air valve opening degree.

[0147] The emergency recovery control parameters are issued to each corresponding control object to replace the original adjustment strategy sequence for execution.

[0148] During the execution process, the disturbance inference state and the actual response data are collected, and when the deviation between the oxygen potential disturbance field prediction output and the oxygen concentration measured value is less than or equal to the dynamic tolerance threshold, and the oxygen potential disturbance field output does not appear value vacancy and decoding failure, the original control path is restored, and the field diagram tracking recovery process is ended. Embodiment

[0149] ​In order to verify the feasibility of the application in implementation, the application is applied to the cold chain logistics environment control system of a certain food storage enterprise. Facing the problem of oxygen concentration fluctuation during the transportation of high-value perishable goods (such as fresh fruits and seafood), the conventional PID regulation strategy responds slowly in a complex dynamic environment, cannot realize the real-time coupling of goods respiration metabolism and replacement regulation, and is easy to cause the oxygen concentration to be too high or to decrease too fast, affecting the preservation effect. In order to verify the effectiveness of the nitrogen and oxygen concentration regulation method based on deep neural network proposed by the application in the cold chain scene, a batch of high-activity tropical fruits are selected as the target goods, and a continuous regulation experiment is carried out. The multi-dimensional operating state is sampled in real time during the experiment, and the method of the application is used for online regulation decision and execution control.

[0150] In actual deployment, first, based on the multi-dimensional historical operating data of goods varieties, loading capacity, cold chain equipment environment temperature, humidity, air pressure, membrane separation nitrogen making machine power gear, nitrogen flow, valve opening, fan frequency, oxygen concentration in the cabinet, nitrogen purity and pressure difference, potential feature representation is constructed and input to the multi-objective variational inversion network to generate the oxygen potential disturbance field, and then combined with the historical regulation behavior response relationship, the oxygen concentration change trend in the next 30 minutes is predicted. After the prediction result shows that the oxygen concentration will exceed the target upper limit in the next 5 minutes, the current optimal regulation behavior unit is calculated according to the disturbance response coefficient, and the regulation strategy sequence is intelligently generated and executed.

[0151] The system executes 11 continuous control cycles (each cycle lasts about 2 minutes), and Table 1 shows the cold chain regulation experiment data table, which records the oxygen concentration, regulation parameters and disturbance state in each cycle.

[0152] Table 1 Cold chain regulation experiment data table

[0153] Control period Initial oxygen concentration (%) Nitrogen production setpoint (L / min) Blower frequency (Hz) Makeup valve opening (%) Oxygen concentration forecast (%) Oxygen concentration measured (%) Pressure differential (Pa) T1 20.9 25 40 0 18.2 18.3 120 T2 18.2 28 45 0 15.6 15.5 125 T3 15.6 30 50 0 13.4 13.5 130 T4 13.4 32 55 0 11.2 11.1 135 T5 11.2 35 60 0 10.1 10.0 138 T6 10.1 30 55 5 10.4 10.3 135 T7 10.4 28 50 10 10.2 10.2 132 T8 10.2 27 48 12 10.3 10.2 130 T9 10.3 26 46 14 10.1 10.2 128 T10 10.1 25 45 15 10.2 10.1 126 T11 10.2 25 45 15 10.2 10.2 125

[0154] From the data in Table 1 above, it can be seen that the initial oxygen concentration is 20.9%, and in the next 3 cycles, the oxygen concentration is rapidly reduced to about 10.1% by increasing the nitrogen production and fan frequency, and then the strategy is automatically switched to a stable state, and the oxygen concentration is maintained in the target interval by small air supply adjustment and frequency slow adjustment. The overall pressure difference is stable between 125~135Pa. The system realizes smooth transition from high oxygen environment to target interval without obvious overshoot or oscillation.

[0155] In order to test the recovery ability of the application under abnormal state recognition, after control cycle T7, some oxygen concentration sensors are intentionally disconnected, the system recognizes that the disturbance field prediction is abnormal, and triggers the field diagram tracking recovery mechanism. In the recovery mechanism, the system retrieves samples containing complete disturbance state labels from the historical operation database, and matches the four most similar field states through a similarity measurement function.

[0156] Table 2 Field map tracking recovery mechanism candidate sample data table

[0157] Sample number Disturbance space difference score Behavior sequence difference score Joint similarity score Sample weight coefficient Control parameters (nitrogen / blower / valve) H001 0.183 0.209 0.196 0.30 85 / 42 / 25 H004 0.166 0.241 0.202 0.26 83 / 40 / 28 H007 0.191 0.214 0.202 0.24 86 / 43 / 27 H011 0.176 0.232 0.204 0.20 84 / 41 / 29

[0158] As shown in Table 2, after identifying the disturbance reasoning anomaly, the system successfully triggers the field map tracking recovery mechanism, retrieves and matches four historical disturbance field state samples with the optimal joint similarity score from the historical operation database, namely H001, H004, H007 and H011. The numerical values of each sample in the disturbance space difference score and the behavior sequence difference score are small, and the overall joint similarity score is not more than 0.205, indicating that it has high similarity in spatial disturbance structure and control behavior characteristics with the current abnormal state.

[0159] Among them, the sample H001 has the lowest joint similarity score (0.196), indicating that it has the strongest coupling degree with the current state, so it is given the highest sample weight (0.30); followed by H004, H007 and H011, which are given weight coefficients of 0.26, 0.24 and 0.20 respectively. The system performs weighted fusion on the control parameters of the four candidate samples according to the weight proportion, and the finally generated emergency recovery control parameters are nitrogen production of 84.52 L / min, fan frequency of 41.52 Hz, and air valve opening of 27.06%. The fan frequency and the air valve opening are within the acceptable range; the nitrogen production is expressed in total flow rate, which is relatively large, but since it is triggered in an emergency recovery under abnormal conditions, it is allowed to use conservative large flow rate for a short time to ensure the stability of oxygen concentration. After the emergency recovery control parameters are immediately issued and put into execution, the oxygen concentration in the subsequent period is successfully maintained within the target range, and the pressure difference changes stably without fluctuation anomaly or system jitter, fully verifying the response efficiency and regulation accuracy of the field map tracking recovery mechanism in actual operation.

[0160] Through this round of emergency response experiment, it can be seen that the abnormal identification mechanism of the present application can quickly perceive the disturbance reasoning anomaly, the field map tracking recovery mechanism has high-precision historical sample matching capability, and can automatically generate control parameters with high adaptability in a very short time, ensuring the continuity and stability of oxygen concentration regulation of the cold chain gas regulation system under abnormal conditions. The overall system shows excellent robustness and adaptability, which is significantly better than the traditional rule-based preset recovery strategy, and has wide practical value.

[0161] This embodiment demonstrates the real-time analysis capability of the present application for multi-target nonlinear disturbance in complex cold chain environment and the adjustment robustness of online response to abnormal state, effectively solves the pain points of traditional methods in prediction inaccuracy, unstable control and emergency failure, and is especially suitable for fresh food transportation, vaccine storage and transportation and precision equipment cold chain transportation scenes with high precision requirements.

[0162] The above merely provides the preferred embodiment of the present application, and the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical scheme and the inventive concept of the present application, can make equivalent replacements or changes within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for adjusting nitrogen and oxygen concentration in a cold chain controlled atmosphere machine based on a deep neural network, characterized in that, Includes the following steps: Step 1: Collect multidimensional historical operational data and perform cross-dimensional embedding encoding to construct latent feature representations; Step 2: Input the latent feature representation into a multi-objective variational inversion network to generate latent variables of the current metabolic state and infer the oxygen potential perturbation field; Step 3: Based on the oxygen potential perturbation field and combined with the historical regulation behavior response relationship, generate the oxygen content prediction curve within the future target time window; Step 4: The decision calculation module generates an adjustment strategy sequence based on the oxygen content prediction curve and the upper and lower limits of the oxygen concentration corresponding to the cargo, combined with the oxygen potential disturbance field. The adjustment strategy sequence includes nitrogen production setpoint, fan frequency, gas replenishment valve opening degree and oxygen reduction trigger judgment value. Step 5: The control module synchronously sends the adjustment strategy sequence to the nitrogen generator, blower, and gas supply valve; Step 6: After each adjustment cycle, the real-time sampling sequence, the adjustment strategy sequence, and the final oxygen content are packaged into a feedback sample and written into the incremental training cache; Step 7: When the incremental training cache reaches the set capacity threshold, the weights of the multi-objective variational inversion network are updated using an adaptive learning rate mechanism; Step 8: When an abnormality in the inference of the oxygen potential disturbance field is identified, the field map tracking and recovery mechanism is activated. The oxygen potential disturbance field state with the best similarity is matched from the historical operation database, and emergency recovery control parameters are quickly generated and immediately issued for execution. Step three specifically involves: The oxygen potential perturbation field is input into the oxygen content prediction network, which includes a perturbation feature extraction module, a historical response memory module, and a sequence curve generation module. The perturbation feature extraction module includes multiple stacked two-dimensional convolutional layers and residual connection structures, receives the oxygen potential perturbation field, extracts the spatial features of the oxygen potential perturbation field, and outputs a spatial perturbation feature map. The historical response memory module includes a gated recurrent unit network, which receives the historical regulation behavior parameter sequence and the corresponding historical oxygen concentration measured sequence, extracts the time-series response features between historical regulation behavior and oxygen concentration, and outputs the historical response feature vector. The sequence curve generation module includes an attention fusion network and a recurrent decoding network. The attention fusion network is used to fuse the spatial perturbation feature map and the historical response feature vector to form a perturbation response feature. The recurrent decoding network decodes the perturbation response feature step by step and outputs the oxygen concentration prediction value for each time step within the future target time window. The oxygen concentration prediction values ​​at each time step are sequentially combined into a complete sequence to generate an oxygen content prediction curve within the future target time window. Step four specifically involves: Based on the oxygen content prediction curve, the decision calculation module identifies all time segments within the target time window where the predicted oxygen concentration exceeds the upper and lower limits of the corresponding oxygen concentration for the cargo, and records the maximum deviation value for each exceeding segment. Duration of exceeding the limit ; The disturbance response coefficient is defined based on the change in oxygen concentration caused by each unit input of the regulating behavior in the oxygen potential disturbance field: ; in, This represents the disturbance response coefficient of the i-th type of regulatory behavior unit. This represents the unit adjustment amount for the i-th type of regulatory behavior. The resulting change in oxygen concentration, This represents the unit adjustment amount for the i-th type of regulatory behavior; Within each oxygen concentration exceeding the limit, calculate the adjustment input that satisfies the following optimization objective: ; If multiple If the optimization objective is met, then the one with the lowest energy consumption per unit input is selected. When the maximum deviation value When the deviation is less than the preset early intervention threshold, select the disturbance response coefficient. The most significant regulatory action is to perform low-amplitude compensation to suppress the upward trend of oxygen concentration. If the oxygen content prediction curve shows a continuous duration of exceeding the limit The time limit exceeds the preset threshold or the maximum deviation value. If the maximum deviation exceeds the preset target limit, the oxygen reduction trigger state is set, the oxygen reduction trigger judgment value is set to 1, the nitrogen injection rate and fan frequency for the corresponding time period are set to the maximum adjustment value, and the gas supply valve is closed; when the oxygen reduction trigger judgment value is set to 0, the oxygen reduction trigger state is not activated. Finally, the set of control parameters is combined into an adjustment strategy sequence according to time sequence. The adjustment strategy sequence includes nitrogen production setpoint, fan frequency, gas supply valve opening degree and oxygen reduction trigger judgment value.

2. The method for adjusting nitrogen and oxygen concentration in a cold chain controlled atmosphere machine based on a deep neural network according to claim 1, characterized in that, The multidimensional historical operating data includes cargo type, loading volume, ambient temperature, humidity, air pressure of cold chain equipment, power level of membrane separation nitrogen generator, nitrogen flow rate, valve opening, fan frequency, oxygen concentration in cabinet, nitrogen purity and pressure difference.

3. The method for adjusting nitrogen and oxygen concentration in a cold chain controlled atmosphere machine based on a deep neural network according to claim 1, characterized in that, The process of performing cross-dimensional embedding encoding to construct latent feature representations specifically involves: The features of the multidimensional historical operation data are extracted, normalized, and then input into the corresponding feature embedding sub-networks. The feature embedding subnetwork is an encoding module containing fully connected layers and ReLU activation functions, used to map the original features of each dimension into an embedding vector of a set length; The embedding vector is input into the cross-attention mechanism module, which consists of a query network, a key-value network, and a weighted fusion module. The query network takes the embedding vectors of cargo type and loading quantity as input to generate an attention query matrix; The key-value network takes the feature embedding vectors of other dimensions as input to generate an attention key matrix and an attention value matrix; The weighted fusion module calculates attention weights based on the dot product similarity between the attention query matrix and the attention key matrix, and applies the attention weights to the attention value matrix to obtain a context-aware representation. The context-aware representation is concatenated with the original embedding vector to form a latent feature representation, which is then used as the input to the multi-objective variational inversion network.

4. The method for adjusting nitrogen and oxygen concentration in a cold chain controlled atmosphere machine based on a deep neural network according to claim 1, characterized in that, Step two specifically involves: The latent feature representation is input into a multi-objective variational inversion network, which includes an encoder subnetwork, a variational sampling module, a decoder subnetwork, and a perturbation mapping layer. The encoder subnetwork includes multiple sequentially connected fully connected layers and a ReLU activation function, used to encode the latent feature representation as the mean vector and log-variance vector of the latent distribution; The variational sampling module generates latent variables of the current metabolic state based on the mean vector and the log-variance vector through a reparameterization method. The decoder subnetwork includes an oxygen metabolism rate estimation branch, a nitrogen replacement intensity inversion branch, and a pressure difference adaptation branch; The outputs of the three branches are merged and input into the perturbation mapping layer, which includes a spatial location encoding unit, a temporal recursive generation unit, and a tensor combination module. The spatial location encoding unit is used to receive the spatial arrangement parameters of goods in the cold chain cabinet, the air duct distribution map and the sensor layout map, and to perform discrete grid encoding of the spatial structure information to generate a spatial location matrix. The time recursive generation unit is a set of gated loop units with shared weights. It receives the output of three branches and performs recursive iterative calculations in combination with the historical regulation behavior sequence to output a perturbation state vector with time series characteristics. The tensor combination module receives the spatial position matrix and the perturbation state vector, and constructs an oxygen potential perturbation field with a three-dimensional structure by point-by-point mapping and channel fusion. The channel dimension of each position point represents the oxygen concentration perturbation trend of the position point in the future several time steps.

5. The method for adjusting nitrogen and oxygen concentration in a cold chain controlled atmosphere machine based on a deep neural network according to claim 4, characterized in that, The oxygen metabolism rate estimation branch includes a temporal convolutional layer and a temperature modulation unit. The temporal convolutional layer is used to extract the temporal features of the latent variables of the current metabolic state, and the temperature modulation unit adjusts the amplitude of the convolution result according to the ambient temperature to generate the oxygen metabolism consumption per unit time under the current state. The nitrogen replacement intensity inversion branch includes two tandem fully connected layers and an attention weighting mechanism. The attention weighting mechanism assigns feature weights based on the coupling relationship between the nitrogen injection rate and the oxygen concentration change rate, which is used to generate the oxygen concentration change intensity factor caused by nitrogen injection in the current state. The differential pressure adaptive branch includes two residual convolutional units and an airtightness adjustment factor. The residual convolutional units are used to extract multi-scale flow structure features, and the airtightness adjustment factor corrects the differential pressure estimate based on the equipment sealing level and internal structural parameters of the cabinet, and is used to output the differential pressure compensation value.

6. The method for adjusting nitrogen and oxygen concentration in a cold chain controlled atmosphere machine based on a deep neural network according to claim 1, characterized in that, The gas supplied by the gas replenishment valve is air, which is used to increase the oxygen concentration.

7. The method for adjusting nitrogen and oxygen concentration in a cold chain controlled atmosphere machine based on a deep neural network according to claim 1, characterized in that, Step seven specifically includes: All feedback samples in the incremental training cache are grouped into several batches of training data according to the order of their collection time. Each feedback sample contains a real-time sampled data sequence within a complete adjustment cycle, the corresponding actual execution adjustment strategy sequence, and the final oxygen concentration measurement result. Using the existing weight parameters in the current multi-objective variational inversion network as initial parameters, the training data is input into the multi-objective variational inversion network in batches, and the current loss function size is determined by calculating the error between the output and the actual measurement results. An adaptive learning rate mechanism is used during training. After each training batch, the trend of the loss function compared with the previous batch is evaluated. If the loss function shows a continuous decreasing rate that is less than the set decreasing rate threshold or even an increasing trend, the current learning rate is reduced. If the loss function continues to decrease at a rate greater than the set rate of decrease, then the current learning rate is increased; If the rate of decrease of the loss function is equal to the set rate of decrease threshold, then the current learning rate is maintained. The training process is set with an early stopping condition. When the change in the loss function of the multi-objective variational inversion network model is less than the set lower limit threshold after multiple consecutive batches of training, indicating that the performance improvement is no longer significant, the current fine-tuning process is terminated. Replace the original weight parameters with the latest multi-objective variational inversion network model weight parameters obtained after fine-tuning, and clear the incremental training cache to wait for the accumulation of new feedback samples.

8. The method for adjusting nitrogen and oxygen concentration in a cold chain controlled atmosphere machine based on a deep neural network according to claim 1, characterized in that, The field map tracking and recovery mechanism is as follows: If the deviation between the predicted output of the oxygen potential perturbation field and the measured value of oxygen concentration exceeds the dynamic tolerance threshold within multiple consecutive sampling periods, or if the output of the oxygen potential perturbation field has a missing value or fails to decode, it is identified as an abnormal perturbation inference. After identifying the perturbation inference anomaly, data samples containing oxygen potential perturbation field state labels are retrieved from the historical operation database. A joint similarity measurement function is constructed based on the spatial distribution structure of the oxygen potential perturbation field and the temporal characteristics of the corresponding regulation behavior. The current abnormal state and the historical oxygen potential perturbation field state are matched and calculated to determine multiple historical oxygen potential perturbation field state samples with the best similarity as a candidate set. The control parameters corresponding to the candidate set are weighted and combined to generate emergency recovery control parameters for the current control cycle. The emergency recovery control parameters include nitrogen production set value, fan frequency and gas supply valve opening. The emergency recovery control parameters are sent to the corresponding control objects to replace the original adjustment strategy sequence for execution. During execution, the perturbation inference state and actual response data are collected. When the deviation between the predicted output of the oxygen potential perturbation field and the measured value of oxygen concentration is less than or equal to the dynamic tolerance threshold, and there are no missing values ​​or decoding failures in the output of the oxygen potential perturbation field, the original control path is restored and the field map tracking recovery process ends.

Images