Method for optimizing service life of automobile thermal manager based on deep learning

Abstract

The invention discloses an automobile thermal manager life optimization method based on deep learning, and the method comprises the following steps: collecting data, recording a timestamp, and forming a multi-source time sequence; executing preprocessing and resampling, and constructing a running state vector sequence; constructing an input fragment sequence and an operation state prediction target sequence of the predictive coding network; setting an operation state prediction path and a life state prediction path, and generating an operation state prediction sequence and a life state prediction sequence; calculating an operation prediction error sequence, and calculating a life prediction error sequence; executing error-driven updating and updating parameters to obtain a predictive coding network subjected to life constraint optimization; calculating a life cost index and a thermal management performance index to jointly construct a control target; and updating a heat management control strategy, and generating and applying a control instruction sequence. The predictive coding network and deep learning are adopted, the service life constraint is introduced into the automobile thermal management control process, and service life evolution perception and active optimization are achieved.

Description

Technical Field

[0001] This invention relates to the field of thermal manager lifetime optimization, and more particularly to a deep learning-based method for optimizing the lifetime of automotive thermal managers. Background Technology

[0002] With the continuous improvement of vehicle electrification and intelligence, the automotive thermal manager, as a key component ensuring the stable operation of the powertrain, electric drive system, and passenger compartment thermal environment, has evolved from traditional control strategies based on empirical rules or static thresholds to intelligent control methods centered on model predictive control, energy consumption optimization control, and data-driven approaches. Existing technologies typically use target temperature deviation, system energy consumption, or instantaneous operating performance as primary control objectives, adjusting operating parameters such as temperature, pressure, and flow rate to achieve stable operation of the thermal manager under complex conditions.

[0003] However, existing thermal management control methods generally treat the thermal manager's lifetime as a post-event assessment or offline analysis object. Lifetime prediction results are mostly used for maintenance decisions or health assessments, failing to continuously perceive the cumulative impact of control behavior on lifetime degradation during normal system operation. In particular, they cannot model and constrain the irreversible evolution process of lifetime depletion. Furthermore, existing deep learning-based or optimization-based control methods mostly focus on operational state prediction or performance optimization, lacking technical means to directly incorporate lifetime prediction errors into model updates and control decisions. This results in thermal manager lifetime degradation often only being detected near the failure stage, making it difficult to proactively suppress and optimize lifetime depletion through control strategies. Summary of the Invention

[0004] One objective of this invention is to propose a deep learning-based method for optimizing the lifespan of automotive thermal management systems. This invention employs predictive coding networks and deep learning to introduce lifespan constraints into the automotive thermal management control process, thereby achieving lifespan evolution perception and proactive optimization.

[0005] A method for optimizing the lifespan of an automotive thermal manager based on deep learning, according to an embodiment of the present invention, includes the following steps: Collect data and record timestamps to form multi-source time series; Preprocessing and resampling are performed on the multi-source time series to generate a sequence of synchronous data frames. Each synchronous data frame is then concatenated according to a preset feature order to obtain a sequence of running state vectors. The input segment sequence of the predictive coding network is constructed based on the sequence of running state vectors, and the target sequence of running state prediction is formed. A predictive coding network is constructed based on the running state vector sequence. Running state prediction path and lifetime state prediction path are set up. The input segment sequence is input into the predictive coding network to generate running state prediction sequence and lifetime state prediction sequence. Calculate the operation prediction error sequence between the operation status prediction sequence and the operation status prediction target sequence, and calculate the lifetime prediction error sequence; Based on the running prediction error sequence and lifetime prediction error sequence, error-driven updates are performed on the internal state of the predictive coding network and the parameters are updated to obtain a lifetime-constrained optimized predictive coding network. Based on the lifetime-constrained optimized predictive coding network and the updated internal state output lifetime state quantity and lifetime consumption rate, and calculate the lifetime cost index, the control objective is constructed together with the thermal management performance index. The thermal management control strategy is updated under the control objective constraints, and a sequence of control instructions is generated and applied to the execution object of the thermal manager.

[0006] Optionally, the formation of the multi-source time series includes: Collect key temperature data and form a key temperature sequence in chronological order; Collect key pressure data and key flow data, and form key pressure sequences and key flow sequences in chronological order, respectively; Collect key current data and key voltage data, and form key current sequences and key voltage sequences in chronological order, respectively; Collect valve opening data, pump speed data, compressor speed data, and heating power data, and generate valve opening sequence, pump speed sequence, compressor speed sequence, and heating power sequence in chronological order, respectively. Collect vehicle operating condition data and form a vehicle operating condition sequence in chronological order; When collecting key temperature sequences, key pressure sequences, key flow sequences, key current sequences, key voltage sequences, valve opening sequences, pump speed sequences, compressor speed sequences, heating power sequences, and vehicle operating condition sequences, corresponding timestamps are added to each sequence to form multi-source time series.

[0007] Optionally, the generation of the running state vector sequence includes: The timestamps of each running data sequence in the multi-source time series are uniformly mapped, and the running data sequences are aligned to the same time axis to obtain the time-aligned multi-source time series. The time-aligned multi-source time series is processed for missing data. Interpolation is performed on the data at the adjacent time indexes at the locations where missing data exists to obtain the missing-completed multi-source time series. Outlier removal is performed on the multi-source time series after missing data completion. By comparing the data change amplitude at adjacent time indices, outlier data points are identified and removed to obtain the multi-source time series after outlier removal. Noise filtering is performed on the multi-source time series after outlier removal. By smoothing the data at adjacent time indices, high-frequency fluctuations are suppressed to obtain the filtered multi-source time series. The amplitude normalization process is performed on the filtered multi-source time series to obtain the normalized multi-source time series; Based on a unified sampling period, the normalized multi-source time series is resampled, and the data corresponding to each running data sequence is extracted at the same sampling time to generate a synchronous data frame sequence. Each synchronization data frame in the synchronization data frame sequence is spliced ​​together according to a preset feature order to obtain a running state vector sequence.

[0008] Optionally, the composition of the target sequence for predicting the running state includes: Read the running status vector sequence and determine the preset window length according to the time index order; In the sequence of running state vectors, according to the time index order, a continuous running state vector with a preset window length is extracted starting from each time index position to form an input segment corresponding to that time index; The input segments formed at each time index position are arranged in time index order to form an input segment sequence; For each input segment in the input segment sequence, the running state vector corresponding to the next time point after the time index of that input segment is selected from the running state vector sequence and used as the running state prediction target for that input segment. The running state prediction targets corresponding to each input segment are arranged in time index order to form a running state prediction target sequence.

[0009] Optionally, the generation of the operating state prediction sequence and the lifetime state prediction sequence includes: Read the running state vector sequence, determine the input feature dimension of the predictive coding network based on the running state vector sequence, and initialize the current internal state of the predictive coding network; Construct a runtime state prediction path and a lifetime state prediction path for a predictive coding network, and set up network structures in the runtime state prediction path and the lifetime state prediction path respectively. At each time index, the input segment sequence is fed into the predictive coding network, which updates its current internal state under the drive of the input segment sequence. After the current internal state is updated, the operating state prediction sequence corresponding to the time index is output based on the operating state prediction path, and the lifetime state prediction sequence corresponding to the time index is output based on the lifetime state prediction path.

[0010] Optionally, the step of calculating the operational state prediction error sequence between the operational state prediction sequence and the operational state prediction target sequence, and calculating the lifetime prediction error sequence corresponding to the lifetime state prediction sequence under lifetime constraints, includes: Under each time index, obtain the running state prediction sequence and lifetime state prediction sequence output by the predictive coding network, and obtain the running state prediction target sequence corresponding to the time index. For the predicted running state sequence and the target running state sequence, the running state vectors at corresponding positions in the two are aligned at each time index to obtain the running prediction error value corresponding to the time index. Arrange the runtime prediction error values ​​corresponding to each time index in order of time index to form a runtime prediction error sequence; For the lifetime state quantity sequence in the lifetime state prediction sequence, the lifetime state quantity corresponding to the time index is compared with the lifetime state quantity corresponding to the previous time index under each time index to obtain the lifetime prediction error value corresponding to the time index. For the lifetime consumption rate sequence in the lifetime state prediction sequence, at each time index, it is determined whether the lifetime consumption rate corresponding to that time index is negative. When the lifetime consumption rate is negative, the absolute value of the lifetime consumption rate is added to the lifetime prediction error value corresponding to that time index. When the lifetime consumption rate is non-negative, the lifetime prediction error value corresponding to that time index is kept unchanged. The lifetime prediction error values ​​corresponding to each time index are arranged in order of time index to form a lifetime prediction error sequence.

[0011] Optionally, the generation of the lifetime-constrained optimized predictive coding network includes: At each time index, read the running prediction error value corresponding to the time index in the running prediction error sequence and the lifetime prediction error value corresponding to the time index in the lifetime prediction error sequence, and read the internal state of the predictive coding network at the time index as the current internal state. Under the time index, the error-driven information is composed of the running prediction error value and the lifetime prediction error value. The adjustment direction of the current internal state is determined based on the error-driven information. According to the preset state update step size, perform an update operation on the current internal state along the adjustment direction to obtain the updated internal state; Under the time index, the corresponding runtime prediction error value and lifetime prediction error value are recalculated based on the updated internal state, and it is determined whether the runtime prediction error value and lifetime prediction error value have decreased compared with the previous version, so as to obtain the updated internal state. Based on the updated internal state, the parameters of the predictive coding network are updated to obtain a lifetime-constrained optimized predictive coding network.

[0012] Optionally, the construction of the control target includes: Within each control cycle, a lifetime-constrained predictive coding network is used, and based on the updated internal state, the lifetime state quantity and lifetime consumption rate corresponding to the control cycle are output. Within the control cycle, the lifetime cost index is calculated based on the lifetime state quantity and lifetime consumption rate, and the target temperature deviation index and energy consumption index are obtained. Within the control period, the life cost index, the target temperature deviation index, and the energy consumption index are used together to construct the control target.

[0013] Optionally, the step of updating the thermal management control strategy under control target constraints, generating a sequence of control instructions, and applying them to the execution object of the thermal manager includes: Within each control cycle, the control objective corresponding to the control cycle is read, and the execution object is determined. Under the constraints of the control objective, a sequence of candidate control instructions is generated around the execution object. For each candidate control command sequence, the corresponding target temperature deviation index, energy consumption index and lifetime cost index are calculated within the control cycle. In candidate control command sequences where the lifetime cost index does not exceed the preset lifetime cost threshold, the target temperature deviation index and energy consumption index are substituted into the control target to obtain the control target value corresponding to the candidate control command sequence. Within the control cycle, the control target values ​​corresponding to each candidate control instruction sequence are compared, and the candidate control instruction sequence with the smallest control target value is selected as the control instruction sequence, which is then applied to the execution object.

[0014] The beneficial effects of this invention are: This invention introduces predictive coding networks into the automotive thermal management control process, achieving synchronous modeling and joint optimization control of the thermal manager's operating state and lifespan evolution state. This effectively overcomes the problem in existing technologies where lifespan is only used as a post-event evaluation indicator and cannot participate in control decisions. Compared to traditional control methods that only focus on target temperature deviation or system energy consumption, this invention constructs operating state prediction pathways and lifespan state prediction pathways driven by multi-source operating data. This allows lifespan state variables and lifespan consumption rates to be continuously predicted and updated during normal system operation, thereby enabling the control system to have real-time perception capabilities of lifespan degradation trends. By explicitly introducing lifespan constraints in the prediction stage, prediction results that violate the physical laws of lifespan are transformed into lifespan prediction errors. This makes the lifespan degradation process no longer a hidden variable but a quantifiable and feedback-based optimization object, significantly improving the effectiveness and usability of lifespan prediction results in the control system.

[0015] Furthermore, this invention drives the internal state and parameter updates of the predictive coding network by jointly running the prediction error sequence and the lifetime prediction error sequence. This allows the model to continuously correct its understanding of the operating and lifetime states during online operation, thereby improving prediction accuracy and system adaptability. At the control decision level, this invention constructs the control objective by combining the lifetime cost index with the target temperature deviation index and energy consumption index. During the control strategy update process, candidate control command sequences are constrained and screened, ensuring that the generated control commands meet thermal management performance requirements while actively suppressing the growth of the lifetime consumption rate. Thus, this invention realizes the transformation of thermal management control from a single performance orientation to a lifetime constraint orientation. It can effectively reduce the irreversible lifetime loss of the thermal manager without sacrificing thermal management performance, extend its service life, and improve the reliability and economy of the entire vehicle operation, demonstrating significant engineering application value. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0017] Figure 1 This is an overall flowchart of a deep learning-based method for optimizing the lifespan of an automotive thermal manager proposed in this invention. Figure 2 This is a schematic diagram illustrating the construction of life prediction error in a deep learning-based automotive thermal manager life optimization method proposed in this invention. Figure 3 This is a schematic diagram illustrating the generation and execution of control commands in a deep learning-based method for optimizing the lifespan of an automotive thermal manager proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A deep learning-based method for optimizing the lifespan of automotive thermal managers includes the following steps: Data from the vehicle's thermal manager during vehicle operation is collected, and timestamps are recorded for each data point to form a multi-source time series. Preprocessing and resampling are performed on the multi-source time series to generate a sequence of synchronous data frames. Each synchronous data frame is then concatenated according to a preset feature order to obtain a sequence of running state vectors. The input segment sequence of the predictive coding network is constructed based on the sequence of running state vectors, and the target sequence of running state prediction is formed. A predictive coding network is constructed based on the running state vector sequence. Running state prediction path and lifetime state prediction path are set up. At each time index, the input segment sequence is input into the predictive coding network to generate running state prediction sequence and lifetime state prediction sequence. Calculate the operation prediction error sequence between the operation state prediction sequence and the operation state prediction target sequence, and calculate the lifetime prediction error sequence corresponding to the lifetime state prediction sequence under lifetime constraints; Based on the running prediction error sequence and lifetime prediction error sequence, error-driven updates are performed on the internal state of the predictive coding network and parameter updates are performed to obtain a lifetime-constrained optimized predictive coding network. Within each control cycle, a lifetime-constrained predictive coding network is used to output the lifetime state quantity and lifetime consumption rate corresponding to the current control cycle based on the updated internal state, and the lifetime cost index is calculated. The lifetime cost index and the thermal management performance index are used together to construct the control objective. The thermal management control strategy is updated under the control objective constraints, and a sequence of control instructions is generated and applied to the execution object of the thermal manager.

[0020] In this embodiment, the formation of the multi-source time series includes: Collect key temperature data to characterize the operating status of the vehicle's thermal manager and form a key temperature sequence in chronological order; Key pressure and key flow data were collected to characterize the operating status of the automotive thermal manager, and key pressure and key flow sequences were formed in chronological order, respectively. Key current and key voltage data were collected to characterize the operating status of the automotive thermal manager, and key current and key voltage sequences were formed in chronological order, respectively. Valve opening data, pump speed data, compressor speed data, and heating power data are collected to characterize the operating status of the automotive thermal manager, and then the valve opening sequence, pump speed sequence, compressor speed sequence, and heating power sequence are formed in chronological order. Collect vehicle operating condition data to characterize the vehicle's operating status and form a vehicle operating condition sequence in chronological order. When collecting key temperature sequences, key pressure sequences, key flow sequences, key current sequences, key voltage sequences, valve opening sequences, pump speed sequences, compressor speed sequences, heating power sequences, and vehicle operating condition sequences, corresponding timestamps are added to each sequence to form multi-source time series.

[0021] In this embodiment, the generation of the running state vector sequence includes: The timestamps of each running data sequence in the multi-source time series are uniformly mapped, and the running data sequences are aligned to the same time axis to obtain the time-aligned multi-source time series. The time-aligned multi-source time series is processed for missing data. Interpolation is performed on the data at the adjacent time indexes at the locations where missing data exists to obtain the missing-completed multi-source time series. Outlier removal is performed on the multi-source time series after missing data completion. By comparing the data change amplitude at adjacent time indices, outlier data points are identified and removed to obtain the multi-source time series after outlier removal. Noise filtering is performed on the multi-source time series after outlier removal. By smoothing the data at adjacent time indices, high-frequency fluctuations are suppressed to obtain the filtered multi-source time series. Amplitude normalization is performed on the filtered multi-source time series to map the data amplitudes of each running data series to a uniform numerical range, resulting in a normalized multi-source time series. Based on a unified sampling period, the normalized multi-source time series is resampled, and the data corresponding to each running data sequence is extracted at the same sampling time to generate a synchronous data frame sequence. Each synchronization data frame in the synchronization data frame sequence is spliced ​​together according to a preset feature order to obtain a running state vector sequence.

[0022] In this embodiment, the composition of the target sequence for predicting the running state includes: Read the sequence of running status vectors and determine the preset window length for constructing the input segment sequence according to the time index order; In the sequence of running state vectors, according to the time index order, a continuous running state vector with a preset window length is extracted starting from each time index position to form an input segment corresponding to that time index; The input segments formed at each time index position are arranged in time index order to form an input segment sequence; For each input segment in the input segment sequence, the running state vector corresponding to the next time point after the time index of that input segment is selected from the running state vector sequence and used as the running state prediction target for that input segment. The running state prediction targets corresponding to each input segment are arranged in time index order to form a running state prediction target sequence.

[0023] In this embodiment, the generation of the operating state prediction sequence and the lifetime state prediction sequence includes: Read the running state vector sequence, determine the input feature dimension of the predictive coding network based on the running state vector sequence, and initialize the current internal state of the predictive coding network; Construct a runtime state prediction path for a predictive coding network, and set up a network structure in the runtime state prediction path to perform state mapping on the current internal state, so that the runtime state prediction path generates a runtime state prediction sequence given the current internal state. Construct a lifetime state prediction path for a predictive coding network, and set up a network structure in the lifetime state prediction path to perform state mapping on the current internal state, so that the lifetime state prediction path generates a lifetime state prediction sequence given the current internal state. At each time index, the input segment sequence is fed into the predictive coding network, which updates its current internal state under the drive of the input segment sequence. After the current internal state is updated, the operating state prediction sequence corresponding to the time index is output based on the operating state prediction path, and the lifetime state prediction sequence corresponding to the time index is output based on the lifetime state prediction path.

[0024] In this embodiment, the step of calculating the operation prediction error sequence between the operation state prediction sequence and the operation state prediction target sequence, and calculating the lifetime prediction error sequence corresponding to the lifetime state prediction sequence under lifetime constraints, includes: Under each time index, obtain the running state prediction sequence and lifetime state prediction sequence output by the predictive coding network, and obtain the running state prediction target sequence corresponding to the time index. For the predicted sequence of running states and the target sequence of running states, the running state vectors at corresponding positions in the two are aligned at each time index. The difference between the predicted value and the target value is calculated for each feature component in the corresponding running state vector. The difference is then squared and accumulated to obtain the running prediction error value corresponding to the time index. Arrange the runtime prediction error values ​​corresponding to each time index in order of time index to form a runtime prediction error sequence; For the lifetime state quantity sequence in the lifetime state prediction sequence, at each time index, the lifetime state quantity corresponding to the time index is compared with the lifetime state quantity corresponding to the previous time index. When the lifetime state quantity corresponding to the time index is greater than the lifetime state quantity corresponding to the previous time index, the difference between the two is included in the lifetime prediction error value corresponding to the time index. When the lifetime state quantity corresponding to the time index is not greater than the lifetime state quantity corresponding to the previous time index, the lifetime prediction error value corresponding to the time index is set to zero. For the lifetime consumption rate sequence in the lifetime state prediction sequence, at each time index, it is determined whether the lifetime consumption rate corresponding to that time index is negative. When the lifetime consumption rate is negative, the absolute value of the lifetime consumption rate is added to the lifetime prediction error value corresponding to that time index. When the lifetime consumption rate is non-negative, the lifetime prediction error value corresponding to that time index is kept unchanged. The lifetime prediction error values ​​corresponding to each time index are arranged in order of time index to form a lifetime prediction error sequence.

[0025] In this embodiment, the generation of the lifetime-constrained optimized predictive coding network includes: At each time index, read the running prediction error value corresponding to the time index in the running prediction error sequence and the lifetime prediction error value corresponding to the time index in the lifetime prediction error sequence, and read the internal state of the predictive coding network at the time index as the current internal state. Under the time index, the running prediction error value and the lifetime prediction error value together constitute the error driving information. Based on the error driving information, the adjustment direction of the current internal state is determined, and the adjustment direction is to simultaneously reduce the running prediction error value and the lifetime prediction error value. According to the preset state update step size, perform an update operation on the current internal state along the adjustment direction to obtain the updated internal state; Under the time index, the corresponding runtime prediction error value and lifetime prediction error value are recalculated based on the updated internal state. It is then determined whether the runtime prediction error value and lifetime prediction error value have decreased compared to before the update. If the runtime prediction error value and lifetime prediction error value have not decreased simultaneously, the internal state update operation is repeated based on the recalculated runtime prediction error value and lifetime prediction error value until the runtime prediction error value and lifetime prediction error value no longer increase before and after two adjacent internal state updates, thus obtaining the updated internal state. The parameters of the predictive coding network are updated based on the updated internal state, so that the parameters of the predictive coding network are adjusted synchronously with the updated internal state, resulting in a lifetime-constrained optimized predictive coding network.

[0026] In this embodiment, the construction of the control target includes: Within each control cycle, a lifetime-constrained predictive coding network is used, and based on the updated internal state, the lifetime state quantity and lifetime consumption rate corresponding to the control cycle are output. Within the control period, a lifetime cost index is calculated based on the lifetime state quantity and the lifetime consumption rate. The lifetime cost index is obtained by accumulating the lifetime consumption rate along the time index order within the control period and weighting it together with the absolute value of the difference between the lifetime state quantity and the start and end times of the control period. Within the control cycle, the target temperature deviation index and energy consumption index are obtained. The target temperature deviation index is obtained by accumulating the difference between the actual temperature and the target temperature within the control cycle along the time index. The energy consumption index is obtained by accumulating the energy consumption of the pump, valve, compressor and heater within the control cycle along the time index. Within the control cycle, the life cost index, the target temperature deviation index, and the energy consumption index are used together to construct the control target corresponding to the control cycle.

[0027] In this embodiment, updating the thermal management control strategy under control target constraints, generating a sequence of control instructions, and applying them to the execution object of the thermal manager includes: Within each control cycle, the control objective corresponding to the control cycle is read, and the execution object for executing the control objective is determined. The execution object includes pumps, valves, compressors, and heaters. Under the constraint of the control target, a candidate control instruction sequence is generated around the execution object. The candidate control instruction sequence is generated by applying incremental changes within a preset amplitude limit to the pump speed, valve opening, compressor speed and heating power based on the control instruction sequence of the previous control cycle, so that the candidate control instruction sequence maintains continuity between adjacent control cycles. For each candidate control command sequence, the corresponding target temperature deviation index, energy consumption index and lifetime cost index are calculated within the control cycle. In candidate control command sequences where the lifetime cost index does not exceed the preset lifetime cost threshold, the target temperature deviation index and energy consumption index are substituted into the control target to obtain the control target value corresponding to the candidate control command sequence. Within the control cycle, the control target values ​​corresponding to each candidate control command sequence are compared, and the candidate control command sequence with the smallest control target value is selected as the control command sequence. The control command sequence is applied to the execution object so that the execution object can meet the target temperature deviation index and energy consumption index constraints while suppressing the growth of the life cost index.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to the operation scenario of the vehicle thermal management system of a new energy passenger vehicle. This vehicle is equipped with an electric drive system, a power battery system, and a passenger compartment thermal management system. The thermal manager needs to operate continuously under various conditions during vehicle operation, including frequent start-stop cycles, prolonged high-speed driving, rapid temperature rise in low-temperature environments, and continuous heat dissipation in high-temperature environments. Because the thermal manager is in a state of dynamic adjustment for extended periods, its internal key components experience irreversible lifespan degradation due to repeated exposure to heat, pressure, and flow rate changes. Traditional thermal management control strategies primarily focus on temperature control accuracy and energy consumption levels, making it difficult to identify and suppress the continuous impact of control behavior on lifespan during normal system operation. Consequently, lifespan degradation issues are often only discovered during later maintenance or when the vehicle is nearing failure, exhibiting a significant passive nature.

[0029] In this application scenario, the vehicle operates continuously during daily use, and the onboard control system collects real-time operational data from the thermal manager, attaching a unified time reference to various data types. This data covers temperature, pressure, flow, current, and voltage changes of the thermal manager at different operating stages, as well as the working status of the actuators. It also includes operating condition information reflecting the overall operating status of the vehicle. All collected data is organized into a multi-source time series within the system, and then time alignment, anomaly processing, noise suppression, and resampling are automatically performed within the control system, enabling synchronous analysis of data from different sensor sources on a unified time axis. The preprocessed data is further concatenated into an operational state vector sequence to comprehensively characterize the operational features of the thermal manager over continuous time.

[0030] Building upon this foundation, the system constructs a predictive coding network to model the sequence of operating state vectors. Structurally, the predictive coding network possesses both operating state prediction and lifetime state prediction capabilities, enabling joint prediction of the future operating state and lifetime evolution trend of the heat manager under the same internal state representation. During actual operation, the system uses continuous segments of operating states as input, causing the predictive coding network to continuously update its internal state. The deviation between the network's output operating state prediction and the actual operating state is used to form the operating prediction error. Simultaneously, the network's output lifetime state prediction is transformed into a lifetime prediction error under lifetime constraints. In this way, lifetime-related information no longer exists merely as an independent prediction result, but directly participates in the network's internal state update process in the form of error.

[0031] During continuous vehicle operation, the predictive coding network is not trained once and used statically, but is continuously updated based on error-driven changes. The system continuously adjusts the network's internal state and updates its parameters synchronously based on changes in operational and lifetime prediction errors, enabling the model to gradually adapt to the constantly changing environmental conditions and operating modes of the vehicle in real-world use. Through this online update mechanism, the predictive coding network can more accurately reflect the dynamic characteristics of the thermal manager under real-world operating conditions, thereby improving the stability and consistency of predictions regarding operational and lifetime states.

[0032] At the control application level, within each control cycle, the system utilizes a lifetime-constrained optimized predictive coding network to output lifetime state variables and lifetime consumption rates, and constructs a lifetime cost index accordingly. This lifetime cost index, along with target temperature deviation and energy consumption information reflecting thermal management performance, constitutes the control objective, enabling the control strategy to simultaneously consider performance requirements and lifetime constraints during the decision-making process. Under this control objective constraint, the system generates candidate control instruction sequences around the thermal manager's execution components, and through a comprehensive evaluation of lifetime cost and thermal management performance, selects the control instruction that best meets the control objective and applies it to the execution object. In this way, the control strategy can maintain smooth changes during continuous operation, avoiding frequent large-scale adjustments, thereby reducing the long-term impact on critical components of the thermal manager.

[0033] During long-term vehicle operation, continuous recording and analysis of operational data reveals that the thermal management system employing the method of this invention exhibits a more stable lifespan evolution trend. The system can adjust its control strategy in advance by predicting changes in the lifespan consumption rate before abnormal changes occur, effectively suppressing accelerated lifespan degradation. Simultaneously, the thermal management performance remains stable throughout the entire operation, without significant control lag or performance degradation due to the introduction of lifespan constraints, indicating that this invention achieves a good balance between lifespan optimization and performance assurance.

[0034] By comprehensively analyzing the vehicle's operation over a long period, this invention further verifies that the proposed method transforms thermal management control from a passive, performance-oriented approach to an active, optimized control method with continuous awareness of lifespan evolution. The system can identify potential lifespan risks during normal operation and intervene by dynamically adjusting control behavior, preventing lifespan issues from surfacing in later stages. Therefore, this invention not only improves the reliability and stability of the thermal manager in actual operation but also provides more effective technical support for the operation management of the vehicle throughout its entire lifecycle, fully demonstrating its feasibility and significant beneficial effects in practical applications.

[0035] Table 1. Comparison of experimental results between the method of the present invention and the traditional method.

[0036] As can be seen from Table 1, although traditional thermal management control methods can maintain basic temperature control requirements, they show significant deficiencies in life-related indicators. The average life consumption rate and fluctuation range are both at a high level, indicating that control behavior has a continuous and uncontrollable impact on the life of the thermal manager.

[0037] The performance-optimized deep learning method improves upon traditional methods in terms of target temperature deviation and energy consumption. The average lifetime consumption rate decreases to 0.86, which is about 14% lower than that of traditional methods, indicating that it reduces lifetime loss caused by high-frequency adjustment to some extent. However, this method still does not take lifetime as a core constraint variable, and the fluctuation range of lifetime consumption remains at 0.31, indicating that the lifetime degradation trend is still obvious in long-term operation.

[0038] The method of this invention exhibits superior overall performance across various metrics. The average lifetime degradation rate is further reduced to 0.68, approximately 32% lower than traditional methods and approximately 21% lower than existing deep learning methods. Simultaneously, the lifetime degradation fluctuation is significantly reduced to 0.19, indicating a smoother lifetime degradation process. Combined with the prediction stabilization convergence time and long-term continuous operation stability score, it can be seen that this invention, through a lifetime prediction error-driven update mechanism, enables the model to adapt to changes in operating conditions more quickly and actively suppresses lifetime degradation in control decisions. This significantly improves the long-term reliability of the system while ensuring thermal management performance.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for optimizing the lifetime of an automotive thermal manager, characterized in that, The method comprises the following steps: Collecting data and recording time stamps to form a multi-source time series; Performing preprocessing and resampling on the multi-source time series to generate a synchronous data frame sequence, and splicing each synchronous data frame according to a preset feature sequence to obtain a running state vector sequence; Based on the running state vector sequence, an input segment sequence of a prediction encoding network is constructed, and a running state prediction target sequence is formed; Based on the running state vector sequence, a prediction encoding network is constructed, a running state prediction path and a life state prediction path are set, the input segment sequence is input into the prediction encoding network, and a running state prediction sequence and a life state prediction sequence are generated; A running prediction error sequence between the running state prediction sequence and the running state prediction target sequence is calculated, and a life prediction error sequence is calculated; Based on the running prediction error sequence and the life prediction error sequence, error-driven updating is performed on the internal state of the prediction encoding network and the parameters are updated to obtain a prediction encoding network optimized by life constraint; Based on the prediction encoding network optimized by life constraint and the updated internal state, a life state quantity and a life consumption rate are output, and a life cost index is calculated, and a control target is constructed together with a thermal management performance index; Under the constraint of the control target, a thermal management control strategy is updated, a control instruction sequence is generated and is applied to an execution object of a thermal manager.

2. The method of claim 1, wherein, The formation of the multi-source time series comprises: Collecting key temperature data and forming a key temperature sequence in time sequence; Collecting key pressure data and key flow data, and forming a key pressure sequence and a key flow sequence in time sequence respectively; Collecting key current data and key voltage data, and forming a key current sequence and a key voltage sequence in time sequence respectively; Collecting valve opening data, pump speed data, compressor speed data and heating power data, and forming a valve opening sequence, a pump speed sequence, a compressor speed sequence and a heating power sequence in time sequence respectively; Collecting vehicle operating condition data and forming a vehicle operating condition sequence in time sequence; When collecting the key temperature sequence, the key pressure sequence, the key flow sequence, the key current sequence, the key voltage sequence, the valve opening sequence, the pump speed sequence, the compressor speed sequence, the heating power sequence and the vehicle operating condition sequence, corresponding time stamps are attached to each sequence to form the multi-source time series.

3. The method of claim 1, wherein, The generation of the running state vector sequence comprises: Uniformly mapping the time stamps of each running data sequence in the multi-source time series to align each running data sequence to the same time axis to obtain a time-aligned multi-source time series; Performing missing data completion processing on the time-aligned multi-source time series, and performing interpolation completion based on the data at the adjacent time index at the position where the missing data exists to obtain a missing data completed multi-source time series; Performing outlier removal processing on the missing data completed multi-source time series, determining abnormal data points by comparing the data change amplitudes at adjacent time indexes, and removing the abnormal data points to obtain an outlier removed multi-source time series; The noise filtering processing is performed on the multi-source time series after the abnormal values are removed, high-frequency fluctuations are suppressed by smoothing the data at adjacent time indexes, and a filtered multi-source time series is obtained; The amplitude normalization processing is performed on the filtered multi-source time series, and a normalized multi-source time series is obtained; The resampling is performed on the normalized multi-source time series based on a unified sampling period, the data corresponding to each operation data sequence is extracted at the same sampling time, and a synchronous data frame sequence is generated; Each synchronous data frame in the synchronous data frame sequence is spliced according to a preset characteristic sequence, and an operation state vector sequence is obtained.

4. The method of claim 1, wherein, The composition of the operation state prediction target sequence includes: The operation state vector sequence is read, and a preset window length is determined according to the time index sequence; In the operation state vector sequence, the operation state vector with a continuous length of the preset window length is cut off from each time index position according to the time index sequence, and an input segment corresponding to the time index is formed; The input segments formed at each time index position are arranged in the time index sequence to form an input segment sequence; For each input segment in the input segment sequence, the operation state vector corresponding to the next time of the time index of the input segment is selected from the operation state vector sequence as the operation state prediction target corresponding to the input segment; The operation state prediction target sequences are arranged in the time index sequence to form the operation state prediction target sequence.

5. The method of claim 1, wherein, The generation of the operation state prediction sequence and the life state prediction sequence includes: The operation state vector sequence is read, the input feature dimension of the prediction coding network is determined based on the operation state vector sequence, and the current internal state of the prediction coding network is initialized; The operation state prediction path and the life state prediction path of the prediction coding network are constructed, and the network structure is set in the operation state prediction path and the life state prediction path; At each time index, the input segment sequence is input into the prediction coding network, and the prediction coding network updates the current internal state under the driving of the input segment sequence; After the current internal state is updated, the operation state prediction sequence corresponding to the time index is output based on the operation state prediction path, and the life state prediction sequence corresponding to the time index is output based on the life state prediction path.

6. The method of claim 1, wherein, The calculation of the operation prediction error sequence between the operation state prediction sequence and the operation state prediction target sequence, and the calculation of the life prediction error sequence corresponding to the life state prediction sequence under the life constraint condition include: At each time index, the operation state prediction sequence and the life state prediction sequence output by the prediction coding network are obtained, and the operation state prediction target sequence corresponding to the time index is obtained; For the operation state prediction sequence and the operation state prediction target sequence, the operation state vectors at the corresponding positions in the two sequences are aligned at each time index to obtain the operation prediction error value corresponding to the time index; The operation prediction error values corresponding to each time index are arranged in the time index sequence to form the operation prediction error sequence. The life state quantity sequence in the life state prediction sequence is compared at each time index, and the life state quantity at the time index is compared with the life state quantity at the previous time index to obtain a life prediction error value corresponding to the time index; The life consumption rate sequence in the life state prediction sequence is judged at each time index, and whether the life consumption rate corresponding to the time index is negative is determined. When the life consumption rate is negative, the absolute value of the life consumption rate is accumulated with the life prediction error value corresponding to the time index. When the life consumption rate is non-negative, the life prediction error value corresponding to the time index is maintained unchanged; The life prediction error values corresponding to the time indexes are arranged in time index order to form a life prediction error sequence.

7. The method of claim 1, wherein, The generation of the life-constrained optimization prediction coding network includes: At each time index, the running prediction error value corresponding to the time index in the running prediction error sequence and the life prediction error value corresponding to the time index in the life prediction error sequence are read, and the internal state of the prediction coding network at the time index is read as the current internal state; At the time index, the running prediction error value and the life prediction error value jointly constitute error driving information, and the adjustment direction of the current internal state is determined based on the error driving information; According to a preset state update step, the current internal state is updated along the adjustment direction once to obtain an updated internal state; At the time index, the running prediction error value and the life prediction error value corresponding to the updated internal state are recalculated based on the updated internal state, and it is determined whether the running prediction error value and the life prediction error value are smaller than before the update to obtain an updated internal state; Based on the updated internal state, the parameter of the prediction coding network is updated to obtain the life-constrained optimization prediction coding network.

8. The deep learning based method for lifetime optimization of an automotive thermal manager according to claim 1, wherein, The construction of the control target includes: In each control period, the life-constrained optimization prediction coding network is used, and based on the updated internal state, the life state quantity and the life consumption rate corresponding to the control period are output; In the control period, the life cost index is calculated based on the life state quantity and the life consumption rate, and the target temperature deviation index and the energy consumption index are obtained; In the control period, the life cost index, the target temperature deviation index and the energy consumption index are jointly constructed to form the control target.

9. The deep learning based method for lifetime optimization of an automotive thermal manager according to claim 1, wherein, The updating of the thermal management control strategy under the control target constraint, the generation of the control instruction sequence and the application to the execution object of the thermal manager include: In each control period, the control target corresponding to the control period is read, and the execution object is determined. Under the control target constraint, the candidate control instruction sequence is generated around the execution object; For each candidate control instruction sequence, the target temperature deviation index, the energy consumption index and the life cost index corresponding to the control instruction sequence are calculated in the control period, and the target temperature deviation index and the energy consumption index are jointly substituted into the control target to obtain the control target value corresponding to the candidate control instruction sequence in the candidate control instruction sequence whose life cost index does not exceed the preset life cost threshold. In the control period, the control target values corresponding to the candidate control instruction sequences are compared, and the candidate control instruction sequence with the minimum control target value is selected as the control instruction sequence, and the control instruction sequence is applied to the execution object.

Images